European voters say they support higher defense spending. But when asked whether they would accept higher taxes or cuts to welfare to fund it, approval falls from 40% to 28% in France and from 37% to 24% in Germany. That gap is Europe’s real defense problem.

The political ceiling on defense spending has collapsed. Germany is committing hundreds of millions to startups, procurement laws are being rewritten, and factories are expanding. Yet rearmament is entering its harder phase.

I attended the Munich Security Conference this year with that tension in mind: welfare vs. warfare.

The tone shifted, but expectations haven’t

Last year in Munich, US Vice President JD Vance delivered a blunt message that Europe’s vulnerabilities were internal as well as external, and that American support would not be unconditional. The speech was widely read as a warning that political alignment and defense spending were now intertwined.

This year, Secretary of State Marco Rubio struck a more conventional tone, emphasizing alliance durability and shared strategic interests. While the rhetoric softened, the expectation did not: Europe must assume greater responsibility for its own security.

But will voters accept its fiscal implications?

Subscribe now

Germany is serious

Through 2025 and into early 2026, Berlin accelerated approvals and contracting, with defense spending approaching €80-90 billion and a rising share directed toward equipment. Advance payments and multi-year commitments have enabled firms such as Rheinmetall to expand ammunition and air defense capacity. Germany’s new procurement acceleration law - the Bundeswehrbeschaffungsbeschleunigungsgesetz - seeks to compress timelines, even if its name suggests bureaucracy dies hard.

The fiscal architecture has shifted as well. The €100B special fund for the Bundeswehr and greater borrowing flexibility have created near-term space for rearmament without immediate cuts elsewhere. But that window is finite: once the fund is exhausted, elevated spending must be embedded in the core budget and in Germany’s fiscal culture.

The more consequential shift is industrial. When the defense ministry awards €536M in strike drone contracts to Helsing and Stark, with potential follow-on tranches pushing the total toward €4.32B, it is not simply buying hardware. It is conferring market validation. That validation attracts private capital, often in multiples of the original contract. Government becomes buyer of first resort. Investors finance expansion. Supply chains localize. Talent concentrates.

This is how industrial gravity forms. Energetics production, advanced components, robotics integration, and secure software increasingly need to sit domestically for resilience. The dynamic resembles the US AI buildout, where large-scale investment in compute triggered reshoring of semiconductor fabrication and energy infrastructure. Defense, like AI, generates similar downstream pull.

Meanwhile, the rest of Europe risks underestimating how quickly that gravitational effect can consolidate Germany’s advantage.

If you are not tested in Ukraine, you are not serious

This was one of the clearer takeaways from Munich. Operational credibility now determines status. Systems deployed and iterated in Ukraine command attention while those that remain untested struggle for relevance. Survivability under electronic warfare, speed of iteration, and demonstrated impact increasingly define reputation.

Indeed, Ukraine has become Europe’s sorting mechanism. It has also exposed the physics of modern war: ammunition, interceptors, drones, armored vehicles, and replacement systems are recurring expenditures consumed at a tempo measured in weeks and months. And this is where rearmament could slow.

Defense companies will not invest in new production lines unless they believe demand will persist beyond the immediate crisis. Governments, however, remain largely in emergency mode, purchasing finished hardware in large batches for delivery, stockpiling it, or transferring it to Ukraine.

The question is what follows the first wave. If a country acquires 100,000 strike drones but does not deploy them, the timing and scale of the next contract become uncertain. In the interim, production lines slow or require subsidy. If the next conflict demands 500,000 rather than 100,000, Europe must decide whether it prefers warehouses of depreciating inventory or factories capable of sustained surge output.

Rearmament built on episodic hardware purchases will struggle to scale unless procurement evolves.

Subscribe now

From buying equipment to buying capacity

The initial procurement surge was necessary. The harder transition is from emergency buying to structural capacity.

A more durable model may resemble cloud infrastructure rather than traditional arms purchasing. Instead of procuring only finished inventory, governments could contract for guaranteed production capacity, paying to maintain throughput and activating full-rate output when required. In cloud computing, customers distinguish between spot usage and reserved capacity. Defense procurement may require a similar distinction between stockpiled hardware and maintained surge capability.

Paying for capacity rather than only inventory aligns incentives with permanence. It enables firms to invest in workforce and supply chains without relying on irregular mega-orders and reduces the risk that expanded lines contract once urgency fades.

If Europe intends rearmament to be structural rather than episodic, procurement models must reflect that intent.

Welfare versus warfare

The fiscal constraint remains central and the United Kingdom offers a cautionary example. Commitments to raise defense spending toward 2.5% of GDP have been prominent, yet debates about deployable mass and readiness persist:

The contrast with the Europeans is becoming embarrassing. The [British] army will have 148 Challenger 3 battle tanks by 2030 but currently has more operational command headquarters than it does artillery pieces, having given 19 howitzers to Ukraine and replaced them with just 14 guns. In contrast, Poland will soon have 980 tanks and 685 self-propelled guns. Finland can mobilise 300,000 troops. Britain’s regular and reserve army totals only 90,000.

Across Europe, polling suggests that voters increasingly believe the world is becoming more dangerous:

As a result, a large proportion of citizens in Canada, the UK, Germany, France and the US support higher defense spending in principle. But when that support is framed in terms of higher taxes, increased borrowing, or reductions in social spending, voter support falls. For example, in Germany defense spending is one of the least popular uses of government funds, topped only by overseas aid. In the last year, voter approval for defense spending subject to these tradeoffs drop from 40% to 28% in France and 37% to 24% in Germany.

That gap defines the central structural challenge. Rearmament at 3-5% of GDP is not incremental: it represents structural reprioritization within economies built around expansive welfare states. Governments can announce multi-year defense plans and approve emergency packages. Sustaining elevated baselines requires durable consent across electoral cycles and economic downturns.

Without that consent, industrial expansion rests on fragile foundations.

Subscribe now

Allies, exports, and domestic priority

Another tension receives less attention: balancing support for allies with domestic resilience. In peacetime, exports reinforce alliances and sustain scale. In wartime, priorities shift. Following the October 7 attacks, Israel redirected production toward domestic requirements, and debates over US munitions supply underscored how quickly allied dependence can become politically sensitive.

When conflict escalates, self-defense takes precedence. If multiple European states were drawn into high-intensity conflict simultaneously, their industrial bases would face similar allocation pressures. Fiscal durability is one constraint. Production allocation under stress is another.

Sharing the upside?

If governments are committing multi-year contracts that de-risk entire sectors, they may also reconsider how value is distributed.

Taking minority equity stakes in companies receiving substantial public contracts would align incentives and allow taxpayers to participate in long-term upside when early demand is state-driven. When the state acts as customer of first resort and absorbs initial risk, it is operating as a strategic investor. Sharing in long-term returns reflects that reality.

Cultural legitimacy

One striking shift in Munich, reinforced in conversations with engineers and AI researchers in Zurich, concerned talent sentiment. A year ago, many technical candidates were hesitant to work on defense. This year, defense work is increasingly viewed as necessary and technically serious, particularly in autonomy, AI, robotics, and advanced manufacturing.

Cultural normalization is a precondition for scale. Europe cannot expand its defense production base without attracting the software and systems talent that previously defaulted to consumer or enterprise sectors.

Europe’s fiscal test

Munich 2025 marked the end of complacency. Munich 2026 clarified the next phase.

Ukraine defines operational credibility. Germany is reshaping Europe’s industrial center of gravity. Procurement models remain misaligned with permanence. Public support weakens once tradeoffs become explicit.

Europe possesses the resources, technology, and industrial base required to rearm. The decisive question is whether it can reconcile warfare with welfare not for a single budget cycle, but for a generation.

Rearmament can survive crisis. Whether it survives normal politics will determine Europe’s strategic trajectory.

Subscribe now

Europe’s security assumptions changed decisively in 2025. The Munich Security Conference marked the moment when American security guarantees were no longer treated as automatic, and Ukraine ceased to be an exception to Europe’s defense model, which had by this time expired.

Since then, defense spending across European NATO members rose sharply, with total outlays moving toward €380-400B and procurement spending rising far faster than budgets overall. More than twenty countries increased defense spending, many by double digits. What had once been treated as a ceiling became a floor.

What has not yet sufficiently changed, however, is how Europe actually builds and buys weapons, and the distance between announced intent and delivered capability remains wide.

The political constraint has lifted

In spending terms, 2025 marked a break with the past. A growing bloc of European states - led by Poland and several Baltic and Nordic countries - began openly backing defense spending levels closer to 5% of GDP over the medium term, pushing the long-standing 2% benchmark from target to baseline. At the EU level, new procurement instruments such as the €150B Security Action for Europe (SAFE) facility saw Brussels motivate direct industrial enablement.

The political argument over whether Europe should spend largely collapsed, even if delivery remained uneven across countries. The harder question became what that spending could actually buy, and how quickly it could be turned into usable military capability. Germany illustrates the constraint. Between 2025 and early 2026, the six-month moving average of domestic defense orders rose by roughly 2x, while domestic sales increased by about 25%. Industrial production, by contrast, edged up only marginally over the same period, highlighting how quickly demand is now outpacing output.

A defense base built to manage decline

But, Europe ran into the production gap almost immediately. The war in Ukraine’s high-intensity fighting consumes immense amounts of artillery shells per day, along with drones, interceptors, and spare parts, at rates that invalidate peacetime assumptions. By late 2025, the EU and its member states had together provided more than €60B in cumulative military assistance to Ukraine since 2022, much of it drawn directly from European stockpiles. But fresh military allocations in 2025 were far smaller - on the order of only a few billion euros - underscoring the gap between cumulative support and the pace of new production. Europe’s defense industry was not built for this environment: it was built to manage decline post-World War 2 and the Cold War. That legacy reflects decades of unpredictable demand, stop-start procurement, and capital discipline that rewarded efficiency and predictability over stockpile and surge capacity.

The UK offers a clear case study. Detailed analysis of British defense procurement shows a system optimised for procedural compliance rather than delivery, with shifting requirements, program churn and weak accountability for delay. Large programs arrive late and compromised, while smaller suppliers struggle to navigate acquisition pathways designed around legacy primes.

Germany, again, exhibits a different dynamic. In 2025 it accelerated approvals and contracting by passing the Bundeswehrbeschaffungsbeschleunigungsgesetz (literal translation Federal Armed Forces Procurement Acceleration Act) - a procurement‑acceleration law whose ambition to simplify process was clearer in intent than in nomenclature. Defense spending rose to nearly €80-90B, with a disproportionate share flowing into equipment. Contractors such as Rheinmetall report advance payments, fast‑tracked parliamentary approvals and flexible contracting to support rapid capacity expansion, particularly in ammunition and air defense. Budgeted procurement and maintenance spending is now set to rise from roughly €32B in 2024 to around €100B by 2029, while large orders requiring parliamentary approval more than quadrupled from about €20B in 2020 to roughly €80B by 2025.

Capital allocation reinforces these dynamics. For two decades, Europe’s major defense contractors optimised for stable margins, predictable returns and low political risk. They returned capital, avoided aggressive acquisitions and treated excess capacity as waste. And now, even as governments speak openly about urgency, European primes continued to prioritise dividends and buybacks to the tune of $5B in 2025, a problem we wrote about on Air Street Press two years ago. At a time when we need to boost R&D and turbocharge the innovation economy to fight a rapidly evolving war, this is behavior we cannot collectively afford to incentivize.

To make matters worse, ESG-driven investment frameworks has treated defense exposure as reputational risk rather than strategic necessity. The contradiction has become explicit. In Norway, parliamentarians have criticised rules barring the sovereign wealth fund from investing in defense contractors such as Lockheed Martin even as the state buys 52 F-35 fighter jets from the same supplier. Similar tensions now run through Europe’s financial system as banks and asset managers struggle to align stale ESG policies with governments’ rearmament priorities, reinforcing a bias toward stability and incrementalism at odds with the need for scale and sustained production.

What delivery now means

Europe does not need additional strategies. It needs evidence of output.

That means factories running at capacity, missile and interceptor lines sized for replenishment rather than scarcity, contracts long enough to justify expansion, and procurement systems that tolerate speed and accept risk. It also means forces that can be sustained in high-intensity operations, not merely displayed for deterrence.

In 2025, Europe announced seriousness, and in 2026, that seriousness has to show up in production. The continent is capable of doing so because it doesn’t lack the money, talent or motivation. It lacks time. And it cannot complain that its defense industry lacks dynamism while rewarding it for behaving like a bond - safe, predictable and slow.

See you at the Munich Security Conference next month. 🫡

Bonus: come join the Air Street Munich AI meetup on Tuesday 17th Feb 2026 and the Air Street Zurich AI meetup on Thursday 19th Feb 2026!

Subscribe now

Key takeaways

• This essay assesses the state of Europe’s defense industry in 2025 and the execution risks heading into 2026.

• Europe has crossed the political threshold on defense spending, but not the industrial one.

• Defense demand is growing several times faster than production capacity.

• Germany’s acceleration in 2025 relied on exceptional measures rather than systemic reform.

• Capital allocation and procurement incentives still favor stability over surge.

• 2026 will test whether Europe can translate spending into sustained output.

A call for urgency and innovation

At Air Street Capital, we were invited to make a submission to the SDR process last year. Our message was simple: the UK has no time to waste:

1. We urged the MOD to treat artificial intelligence not as a future curiosity but as an organising principle of modern warfare.

2. We argued that venture-backed dual-use companies must be brought into the fold on equal terms with traditional suppliers.

3. We advocated for systemic procurement reform to allow the UK to field capability at software speed.

Strategic signals of progress

To its credit, the SDR takes major steps in this direction. Most notably, it creates UK Defence Innovation (UKDI), a new centre of gravity for sourcing and deploying novel technologies, backed by a ring-fenced budget of at least £400 million and a mandate that 10% of the equipment budget be spent on emerging capabilities. The MOD also commits to establishing a Defence Investors’ Advisory Group, echoing our recommendation that venture capital not be lumped in with generic SME engagement. These reforms signal a new willingness to treat the innovation economy not as a sideshow but as a strategic partner.

The review is equally clear-eyed on force design. Long-range, attritable autonomous systems—such as land-based Recce-Strike swarms and hybrid carrier drone wings—are no longer edge-cases, they’re core architecture. By 2027, the MOD aims to field a digital targeting web backed by a Secret-level cloud network and a new Digital Warfighter cadre. These shifts demand exactly the kind of layered autonomy stack that companies like Delian Alliance Industries are building: GPS-denied navigation, passive sensor fusion, real-time jamming, and machine-speed strike systems designed for contested environments. The MOD will also establish a CyberEM Command, fusing cyber operations and electromagnetic spectrum dominance into a coherent domain, and bring export controls back inside its procurement portfolios.

The industrial base, too, is set for rearmament. The Secretary of State’s foreword promises always-on munitions manufacturing, six new energetics plants, and 7,000 long-range missiles. This "arsenal pipeline" aligns closely with our submission's call for permanent surge capacity—one that doesn't require geopolitical shocks to get moving.

The review also validates several of our specific proposals. It adopts our segmentation model of procurement timelines, including a commitment to let contracts for novel capabilities in as little as three months. It even gestures toward our recommendation that the MOD hold security clearances on behalf of start-ups, committing by April 2026 to "reform regulations, Intellectual Property handling, and security clearance requirements" as part of a broader package to reduce barriers for industrial partners. And it acknowledges, in language echoing our submission, that defence procurement is bogged down in Cold War-era cycles that are no longer fit for purpose.

Subscribe now

But critical gaps remain

Yet for all this progress, the SDR still falls short in critical ways.

First, the MOD stops shy of creating a truly arms-length procurement agency with the freedom to run rapid-fire competitions and iterative fly-offs. Instead, power remains concentrated in the traditional hierarchy, now coordinated by a new National Armaments Director. While the role is centralised, it is not independent, nor is it led by external commercial talent. Without outside leadership, institutional gravity may pull good intentions back into old habits that we’ve documented in detail in our work on European Dynamism in late 2023.

Second, start-ups still face massive onboarding friction. While the clearance proposal was positively signaled in passing, implementation details remain vague. There is no firm delivery date, no operational framework, and no broader reform of Single-Source Contracting or Defence Standards. The UK’s most capable dual-use founders will continue to ask themselves: do I build for the MOD, or ship globally?

Third, the review ducks the capital question. While a “Defence AI Investment Fund” is trailed for 2026, there is no sovereign growth vehicle capable of taking patient equity risk. Nor is there any resolution to the question of GPU supply for defence AI workloads. In a world where the US has created a secure DoD-dedicated cloud with Microsoft and Palantir, and where countries like the UAE are buying compute at sovereign scale, Britain’s silence on this point is not good enough.

Fourth, the role of defence primes remains largely untouched. Despite the SDR echoing our concerns that many primes prioritise dividends over R&D and struggle to attract AI talent. there is no meaningful restructuring of their role. The Defence and Security Accelerator remains in place, still partially managed by a prime contractor. Nor are there new restrictions to prevent primes from operating innovation units, despite the clear conflict this poses. The review's calls for "streamlining" innovation hubs do not go far enough to ensure that smaller, more agile firms have a fair chance.

Fifth, the review leans heavily on the promise of dual-use technologies to drive both defence capability and economic growth. While that framing is politically convenient, it sidesteps a harder truth: some of the most strategically important technologies have no viable civilian market. In such cases, dual-use thinking delays delivery to the frontline. Our position remains that dual-use is not enough. A track must exist for purely defence-native capabilities.

Finally, the SDR embraces the need for speed in principle but doesn’t always back it up in practice. Despite acknowledging that many "next-generation" capabilities are already commercially available, it still categorises them as future-looking. The timelines for major implementation milestones—2026 and beyond—are out of sync with the geopolitical urgency the UK itself identifies.

Subscribe now

A mandate for follow-through

In sum, the SDR embraces many of the right principles at we at Air Street also endorse: speed, autonomy, modularity, digital-first command structures, and a more open relationship with external innovators. But it tempers nearly every radical proposal with institutional caution. The result is a defence strategy that reflects the world as it is, but not yet the speed or scale of change required to meet it.

The next phase will matter even more. The forthcoming Defence Industrial Strategy and Defence Investment Plan are where execution must meet intention. They must go further: enshrining the 10% innovation budget for non-prime startups in statute, creating sufficient domestic compute reserves, deploying capital at large scale, and delivering rapid procurement reform not just in spirit, but in practice.

From blueprint to battlefield: what great would look like in 12 Months

By mid-2026, progress should be visible across five fronts:

• UK Defence Innovation operating independently with authority to run competitive trials.

• MOD-held security clearance system live and accessible to startups.

• Defence AI Investment Fund deployed with venture-scale cheque sizes.

• Sovereign compute access provisioned for defence use.

• 10% novel-tech procurement floor formalised with transparent reporting.

These are not moonshots. They’re the minimum conditions for credibility.

The UK’s Strategic Defence Review shows that it understands the new grammar of deterrence. It embraces autonomy, digital warfare, and distributed production as pillars of credible 21st century military power. But it hesitates where it matters most: the enabling conditions that allow new suppliers to deliver at speed and scale. The review diagnoses the future with clarity. What remains to be seen is whether Whitehall can build the muscles to meet it.

Read the official review

Breathless headlines around the geopolitical implications of AI dominate our newsfeeds: “AI Arms Race”, “AI War Heats Up”, “AI Doomers Versus AI Accelerationists Locked In Battle For Future Of Humanity”. But where’s the European Union in all of this?

We’ll give it to you straight: it’s quite likely Europe's ambitious regulatory initiatives will be obsolete before they're even fully enacted.

The European Union appears to be operating on an entirely different timeline, one where AI's risks warrant methodical, unhurried deliberation rather than focused enablement.

The world moves on while the EU (waits to) AI Act

The groundwork for the EU AI Act was laid over four years ago when the European Commission published a proposal for "harmonised rules on artificial intelligence" in April 2021. It was then made law in August 2024.

In the technological equivalent of several geological eras, OpenAI has since shipped 14 distinct versions of ChatGPT. GPT-4, which thrust generative AI into the mainstream, wasn't even released until March 2023; almost two years after the proposal for legislation was published.

During this regulatory gestation period, AI startups have raised over $300 billion globally, while a basket of five prominent AI companies have added $4.5 trillion in enterprise value.

Meanwhile, Western narratives around AI safety have undergone a spectacular pivot. Once dominated by philosophers and futurists warning of superintelligent machines that might view humanity as dispensable (think Skynet), today's conversation has, almost overnight, become simultaneously accelerationist and defanged.

The EU is crafting regulation at a snail’s pace, ostensibly oblivious to the technological renaissance erupting around them. While the capabilities of models and applications are evolving exponentially, bureaucrats are seemingly focused on trying to grasp an ever fleeting ‘just past’.

Put simply, the Act was born in a world before GPT-4, and will take effect in a world shaped by GPT-7. Let’s consider a live example. Matthias Fuchs is the founder of Tiger Cow Studios, a Berlin-based AI automation agency working with clients across German industry. He describes how regulatory uncertainty is blocking new development:

"Roughly 60% of the AI prototypes we’ve built for German Mittelstand firms have never left the lab. Potential clients err on the side of caution and completely ban LLMs like ChatGPT, due to uncertainties around EU AI Act compliance. When rules and regulations stop experiments in their infancy, firms miss out on learning loops and see their competitive edges wither."

The fundamental question has become unavoidable: is the EU AI Act a masterclass in regulatory obsolescence—carefully mapping a territory that will have transformed beyond recognition before the cartography is complete?

Let’s dive in.

Subscribe now

The EU AI Act: A beautiful architecture but is it buildable?

The Act introduces a sophisticated, four-tier risk classification system:

1. Prohibited AI Practices: Systems deemed unacceptably harmful are banned outright, including social scoring systems and most real-time biometric identification in public spaces

2. High-Risk AI Systems: Applications in sensitive domains like healthcare face extensive requirements

3. Limited Risk AI Systems: Systems with moderate transparency concerns must meet disclosure requirements

4. Minimal Risk AI Systems: All other AI applications remain largely unregulated

This meticulous classification is overseen by a multi-level governance Frankenstein that would make Mary Shelley proud. At the European supranational level, the AI Office serves as the central coordinator while the European AI Board is meant to ensure harmonized application.

A scientific panel of independent experts hovers nearby, tasked with providing technical opinions on model risks and capabilities; presumably when they're not engaged in academic turf wars.

Completing this triumvirate is an advisory forum where industry representatives, civil society organizations, and standards bodies can voice concerns that will likely be meticulously documented and then elegantly filed away.

Complicated enough? We feel you. So here’s a chart that illustrates the same:

Meanwhile, each member state (read: EU country) must establish its own implementation machinery, including market surveillance authorities and regulatory sandboxes. Here’s how this is supposed to work:

National competent authorities oversee market surveillance while notifying authorities evaluate conformity assessment bodies that, in turn, certify high-risk systems for compliance and user safety. These are complemented by a constellation of AI regulatory sandboxes where innovation supposedly thrives under controlled conditions.

Perhaps the most revealing aspect of the Act is the implementation timeline. While the EU AI Act formally entered into force in August 2024, full compliance for public authorities' high-risk systems isn't required until August 2030 – six years later, or almost ten years after the initial proposal. In technological terms, this is equivalent to regulating today's smartphones based on BlackBerry design principles.

Dr. Neo Christopher Chung, a Warsaw-based AI researcher and the founder of Obz AI by Alethia, reveals there is much to be read between the lines of the Act:

"Think tanks and research institutes that do AI security, safety, and governance are writing 'code of practice' documents based on the EU AI Act. This is the devil in the details of European policy making. The EU creates the high-level framework and then outsources its implementation to advisory organizations (more here, think European twists on Brookings) who shape how the guidelines are applied in practice. For now, most people are waiting and dragging their feet because no one has told them what to do and how to comply."

Almost one year after the Act’s ratification, four years since the proposal for legislation was first published, and probably less than 5% of what will ultimately be implemented exists, if that.

At the same time, the Commission has established ambitious benchmarks: 75% of EU enterprises adopting AI by 2030, doubling AI startups to 12,000, and reducing high-risk AI incidents by 65%. Yet, one wonders if these targets measure bureaucratic self-perpetuation rather than meaningful innovation.

Maybe, though, we’re being too tough and the EU’s approach is just a reflection of what is simply realistic. Let’s look at other systems in an attempt to put Europe’s regulatory framework in context.

Share

Competing Approaches: America and China are both playing to win, albeit with quite different systems

Under Trump's second administration, the regulatory focus is squarely on competitiveness with China, rather than safety guardrails like in Europe. Federal initiatives favor voluntary guidelines over comprehensive legislation, while states like Colorado and California have established their own "regulatory laboratories"^, in an attempt to conceive of a future national standard.

The US lacks a centralized governance structure ‘built for purpose’, operating instead through existing authorities like the Federal Trade Commission (FTC). This deliberate move reflects America's competitive ethos but could risk dangerous gaps in oversight.^,

For now, it’s probably prudent to temper any rigorous analysis of the American regulatory landscape. In July of this year, “AI Czar” David Sacks and his team are expected to publish an ‘Artificial Intelligence Action Plan’ that will set out the regulatory objectives of the second Trump presidency.

China, meanwhile, has been implementing AI regulation since 2021, with rapid developments in the past 18 months. Their regulatory system assigns differentiated obligations across the AI ecosystem – from service providers to users – creating a shared responsibility model. Recent standards provide concrete technical requirements for security incident response, content filtering, and bias prevention.

Most notably, China's approach integrates ethics reviews at the development stage rather than focusing solely on deployed technologies. The Cyberspace Administration coordinates with six other ministries, creating a comprehensive regulatory ecosystem that balances innovation with state interests.

Subscribe now

Case Study: EuroHPC's Bureaucratic Bottleneck

No example better illustrates Europe's regulatory folly than the European High Performance Computing Joint Undertaking (EuroHPC), established to deploy the continent’s supercomputing infrastructure.

Despite being impressive technology, accessing it requires navigating an administrative labyrinth. Researchers must provide extensive technical documentation, undergo multiple evaluation stages, and wait up to six months for approval (translate that to 3.5 years in ‘normal’ sectors). The process includes admin checks, technical assessments, peer review, committee evaluations, and final approval by the governing board – all before any actual computation begins.

Once approved, users face substantial reporting requirements documenting everything from energy consumption to resource utilization compared to plan. While China and the US race ahead with AI development, Europe's researchers are crushed by the weight of paperwork.

For time-sensitive AI research, such delays are tantamount to killing the golden goose before it’s hatched. During the waiting period, technologies evolve, research questions become outdated, and competitors (especially in other jurisdictions) accelerate. This administrative overhead diverts valuable researcher time away from actual innovation, undermining Europe's competitive position before the first training runs even begin.

Share

A Common Sense Approach for Europe

Europe too often overthinks. If it wants to be the preferred destination for leading AI researchers and developers, it should simply allow top talent to work and focus on maintaining the high quality of life that attracts them in the first place. As the chronically online tech folks on Twitter/X say, “let the cracked AI engineers cook.”

Europe’s admirable regulatory intentions require urgent recalibration to ensure global competitiveness:

• Streamline Administrative Processes: Significantly reduce bureaucratic burdens and shorten approval timelines to keep up with pace of AI innovation.

• Centralized Authority: Empower an independent supranational body capable of swift and decisive regulatory implementation to prevent conflicting national approaches.

• Incentive-Based Regulations: Shift regulatory strategies from restrictive compliance mandates to incentive-based mechanisms, fostering innovation and compliance through rewards rather than penalties.

• Market-Driven Resource Management: Transition resource-intensive projects such as EuroHPC to private entities or strategic cloud partnerships, leveraging market-driven efficiency to optimize resource allocation.

Conclusion

Europe has the potential to lead in AI. Not through micromanagement, but by offering an environment where innovation flourishes alongside its genuinely commendable social system. Europe must focus on incentives (enablement) rather than restrictions (control). Europe can establish a regulatory approach that respects its values while ensuring it remains competitive with China and the US.

The EU AI Act, for all its architectural ambition, epitomizes Europe’s idealistic challenge. It attempts the holy grail of balancing an admirable commitment to ethical principles with the practical realities of a competitive technological world order. But if Europe fails to synchronize its regulatory tempo with the cadence of development, it risks becoming an elaborate bureaucratic archive. Carefully cataloging yesterday's innovations while the tools of tomorrow are built in Silicon Valley and Shenzhen, far from the land of the Enlightenment, is fit for a natural history museum, not the modern-day arena.

Sources

Article 5, EU AI Act. See also European Commission, "Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act)," issued February 4, 2025.

Article 5, EU AI Act. See also European Commission, "Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act)," issued February 4, 2025.

Articles 6-15, EU AI Act. For expanded analysis, see European Commission, "Impact Assessment Report on the AI Act," SWD(2021) 84 final, 21.4.2021.

Commission Decision (EU) 2024/1459 of 24 January 2024 establishing the European Artificial Intelligence Office, OJ L 42, 24.1.2024, pp. 32-38.

Articles 65-66, EU AI Act. See also European Commission, "Communication on the establishment of the European AI Board," COM(2025) 12 final, 15.1.2025.

Articles 57-59, EU AI Act. For implementation guidelines, see Commission Implementing Regulation (EU) 2025/XXX establishing detailed arrangements for AI regulatory sandboxes.

European Commission, “Third Draft of the General-Purpose AI Code of Practice published, written by independent experts”, March 11, 2025

European Commission, "Monitoring and Evaluation Framework for the AI Act," SWD(2024) 238 final, published alongside the final text of the AI Act, 13.6.2024

"Removing Barriers to American Leadership in Artificial Intelligence," The White House, January 23, 2025

"AI legislation in the US: A 2025 overview," Software Improvement Group, 2025, p. 8-9, 10-11

"AI Watch: Global regulatory tracker - United States," White & Case LLP, March 31, 2025, p. 7-8, 12

"Removing Barriers to American Leadership in Artificial Intelligence," The White House, January 23, 2025

“Tracing the Roots of China's AI Regulations”, Carnegie Endowment, February 2024

"China's New AI Regulations," Latham & Watkins, August 16, 2023, Number 3110, p. 4-5.

"Basic Requirements for the Security of Generative Artificial Intelligence Services," referenced in "AI Ethics: Overview (China)," China Law Vision, January 20, 2025.

"AI Watch: Global regulatory tracker - China," White & Case LLP, 31 March 2025.

EuroHPC JU, "Access Policy for the allocation of the Union's share of access time of the EuroHPC Joint Undertaking Supercomputers and AI Factories," Version 2025 04 09, p. 6-7.

EuroHPC JU, "Extreme Scale Access - Project Scope and Plan," 2025.

EuroHPC JU, "Extreme Scale Access - Full Call Details," 2025, p. 4-5.

EuroHPC JU, "Extreme Scale Access - Terms of Reference," 2025, p. 12.

EuroHPC JU, "Extreme Scale Access - Progress Report," 2025; "Extreme Scale Access - Final Report," 2025.

Reacting to the UK’s announcement of a rise in defence spending to 2.5% of GDP by 2027, Sky News declared: “Starmer has put Britain on a war footing”.

If only that were the case. The massive gulf between the Brave1 defence tech conference in Kyiv last week, and an AI in defence conference in London this week, should worry anyone thinking about the UK’s position in a world undergoing huge geopolitical and technological change. Note: Brave1 is a Government of Ukraine platform that brings together innovative companies with products that can be used in the defence of Ukraine.

Of course, neither conference can be wholly representative of each country’s embrace of defence tech - but equally it would be a mistake to complacently ignore the differences.

Brave1 showed what it means to actually be “on a war footing”. And this was not just because of the occasional air raid siren, or the groups of people huddled over phones showing recent action on the frontline.

Walking into Brave1 you were instantly struck by the intensity of the event. It was packed. Everyone was in a rush. We faced a similar barrage of questions from everyone we spoke to: What are you building? How cheap is it? Where’s it deployed? When can we test it? How many can you make?

Cost. Scalability. Availability. Urgency.

That’s all anyone cared about.

They also peppered conversations with references to how various systems were performing at the front. Sharing important product insights based on ever evolving countermeasures and tactics as each side iterates to get an edge.

When new draftees are sent to the front after just weeks of training, and face a similar life expectancy once there, nobody has time for decks, proposals, or waiting 6 months for the next “industry briefing day” that results in another 6-12 months of drafting requirements.

Back to the conference room

Contrast this with the London military AI conference. Some of the basic scenery was the same - bad coffee, people adding each other on LinkedIn, roll up banners with questionable aesthetic choices - but whereas Brave1 was marked by action, London can only be characterised by its opposite.

In London there was a keen sense that this time next year you could have the same people discussing the same points on the same panels. The “Synergy breakout room" would still be there. Still full of “disruptors”. There would be no clarity on how AI would magically infuse UK defence (despite the Prime Minister calling AI our “defining opportunity”), but maybe nobody would even mind, or notice. A sense that we can keep going at a glacial pace because surely, if things do kick off, it will be alright on the night.

There was also a wide gap in how AI is being applied. At Brave1, the emphasis was on AI’s potential to transform low cost, highly scalable platforms into effective weapon systems. An example of this was a computer vision enabled strike drone that was literally held together by duct tape.

In London, AI’s role in conflict was almost an after thought. It was instead more of a back office efficiency tool, or simply a convenient trend for consultancies to now sell to MoD. AI for streamlining large organisations has its place, but it’s not super helpful for allies defending their countries from invading autocrats.

From a startup perspective, Brave1 shows what it means to have a genuine industry - military partnership centred on adopting new technology as quickly as possible. Short procurement cycles. Access to rigorous testing with a clear pathway to adoption if successful. Contracts. Rapid end user feedback.

This is not to say Ukraine is perfect. We are not naive about corruption, or what can happen to your IP after you give your product to a ‘partner’, but it is lightyears ahead of where we are today in the UK.

Putting innovation theatre to bed

The UK remains wedded to the traditional model of hoping small companies can somehow invest in R&D while bouncing from innovation grant to innovation grant. With a startup’s prospects determined more by its proposal writing ability than the quality of its product.

None of this matters if you believe the UK does not face a dangerous world. Or if you think the defence Primes are best placed to harness emerging technologies and build low cost, intelligent, unmanned systems.

But if, like us, you believe differently, we must learn lessons from Brave1 and Ukraine. Our first step should be a shift in mindset.

We don’t have the luxury of hoping that a small rise in spending in any way reflects our preparedness for the wars of the future.

- Matthew Wright, VP Operations, Delian Alliance Industries

Learn more about Delian Alliance Industries at www.delian.ai

For more on our defence writing, we invite you to read the following pieces:

Introduction

A few weeks ago, the French Minister of the Armed Forces Sébastien Lecornu confirmed to French parliamentarians that the ‘Eurodrone’ project was set to be delayed by another year. The Eurodrone - the European Medium Altitude Long Endurance Remotely Piloted Aircraft System (MALE RPAS) - is an EU project to create an unmanned combat aerial vehicle that’s been running for over a decade. 

Conceived in 2014, following the Russian seizure of Crimea and mounting European concern about the dominance of American, Chinese, and Israeli firms in the drone market - the project has been a fiasco from start to … not finish. It’s been hit by cost overruns, contractual fights, spats between governments, and the final vehicle - if it’s ever delivered - is likely never to see service. In fact, Lecornu has acknowledged that they are only continuing with the programme, because cancellation would be costly. Textbook economic sunk cost fallacy right there. 

So far, business as usual in the world of European defense contracting. 

But we believe that the Eurodrone project can also teach us some useful lessons, which are increasingly relevant as the nature of war shifts and European governments begin to adjust their defense posture out of last resort.

The Eurodrone: a brief history

In the spring of 2014, Dassault, Airbus, and Alenia submitted a proposal to the French, German, and Italian governments for a study to shape the requirements for the Eurodrone. If actioned speedily, they believed (perhaps optimistically) that a working prototype could be ready for testing in 2020. Given the history of failed European drone projects, the companies wanted the governments to pay for the study.

By the end of 2015, the three governments hadn't given the study the go ahead. This was because they were unable to agree on any basic details among themselves - including the speed, mission profiles, whether it would be armed, how many engines it should have, or who would manufacture it. It would take until September 2016 to finally authorize it.

Why did it take so long to agree on the basics?

While many European governments use similar terminology, when you scratch under the surface, it tends to mean different things. This is particularly the case in anything relating to defense, resilience, or sovereignty. 

Back in 2013, the EU began to use the term ‘strategic autonomy’ for the first time in a defense and security setting. In its earliest versions, it acted as a floating signifier - its meaning would depend on which European capital you happened to be in.

For France, ‘strategic autonomy’ was a call-back to the era of Charles De Gaulle, when the country attempted to formulate a defense and security policy that, to some extent, diverged from that of the US. In Berlin, it was read as a statement about the need to maintain wider industrial competitiveness. Italy saw it as building EU capacity to act more in the Mediterranean for security and migration management. Meanwhile, Poland and the Baltic nations, acutely aware of the importance of the US security umbrella, were inclined to view it as posturing. 

Unmanned aerial vehicles (UAVs) were relatively new for European governments. While the US had been using Predator drones to strike targets since 2001, except for the UK, European countries well into the 2010s shied away from buying them at any scale. 

After being forced to rely on US Predator drones while running their counterinsurgency operation in Mali, the French were keen for a drone that could be armed.

Germany, and the other Eurodrone partner nations, were much more focused on border surveillance. While Germany had begun to dabble in armed UAVs, it was unenthusiastic. 

In 2014, after their lease on unarmed Israeli Heron drones expired, the German government purchased a small handful of armed vehicles - only after extensive lobbying by the military and against a political backlash. Even then, the German Armed Forces didn’t want the armed drones to be too lethal. In the words of one insider: “Ideally, we could have a UAV carrying a number of very small bombs – and only a few of them would be fitted with explosives. The other would have no explosive material but would produce only a loud bang or smoke without causing harm.”

It was against this backdrop that the Eurodrone was conceived. Countries agreed there should be a drone, but were unwilling to concede that the drone you use for domestic surveillance is unlikely to be the same one used for fighting extremists in West Africa. 

This need to accommodate different countries’ different requests and to divide up the manufacturing across different local suppliers led to years of squabbling and negotiations. This produced a costly and complex design that could accommodate everyone’s different requirements. 

For example, to fly in non-segregated civilian airspace, UAVs need to adhere to higher safety and redundancy standards. Regulators need to be confident that in the event of engine failure, the UAV won’t plummet into a population center. Meanwhile, to support strike capabilities, UAVs need higher payload capacity and structural reinforcement.

To tick these boxes and use European components, the final design employed two Catalyst engines, designed by General Electric’s Italian subsidiary, Avio Aero. 

This is unusual in the world of UAVs. Military vehicles like the US Predator or Israeli Heron do not usually use two engines, because it’s complicated, expensive, fuel inefficient, and … heavy.

The design and definition were finally agreed in 2019 - a year before the manufacturers had first hoped to start testing prototypes. And it pleased no one. French parliamentarians criticized the “obesity” caused by “German specifications”, with a French Senate report claiming that “with two motors and a weight of 10 tons, this drone will be too heavy, too expensive and therefore, too difficult to export”. The Germans complained about spiraling costs caused by immature technologies and began to recommend including American components.

The drone, if it’s ever delivered, will now arrive after 2030 and at least over 40% over budget. Naturally, contractors in different European countries have fallen out. The preliminary design review was completed in May this year, 10 years after Dassault and Airbus submitted their original proposal.

European governments will have spent approximately €8B for a drone that’s … not really optimized for anything. Given the small handful of orders - this will work out at €120M per drone.

In the 16 years between the conception and earliest completion date of the project, a lot will have changed: the decline of big platforms, the rise of attritable warfare, the use of AI and autonomy to improve accuracy, the ballooning complexity of electronic warfare. None of this will be reflected in the final design.

What can we learn?

We aren’t retelling the story of the Eurodrone so we can laugh at the participants. Instead, it’s a source of some positive lessons on how to procure technology. There are a few that we can identify. 

If something works, just buy it

The Eurodrone joins the long list of attempts by national governments to replicate capabilities that others have delivered competently. This is motivated by two differing errors in reasoning. 

Firstly, an incoherent vision of sovereignty. While it’s possible to understand the theoretical appeal of truly sovereign technology for defense, it’s rarely realizable. 

If the US decided that it was going to cut off support for every piece of military equipment it had sold to European governments, a single ‘autonomous’ drone programme would not be enough to rescue us from the ensuing disaster. It’s the equivalent of semiconductor sovereignty efforts that derisk Chinese disruption to the supply chain by heavily subsidizing the onshore construction of one or two fabs. Expensive? Yes. An actual guarantor of sovereignty? No.

Secondly, this belief that individual countries’ needs are special. Again, it’s true that different militaries’ exact uses for equipment vary. But once you attempt to redesign technology that already works by committee, it pretty quickly stops working. The archetypal example of this is Watchkeeper - the botched British attempt to replicate the highly capable Israeli Hermes 450 drone. 265 user requirements, 1,910 additional system requirement, and £1.4B later - none of the drones are on operational deployments.

Big defense contractors aren’t your friend

Time is money for start-ups. It isn’t for a defense prime … they literally bill for it. Early-stage companies have lean teams, need to ship quickly to build a track record, and need to be paid. Defense primes with billions of dollars sat on their balance sheets … do not.

Increases in complexity, project delays, last-minute changes in requirements - these are all a nightmare for a start-up. But they’re great news for primes. They can suck them up in the short-term and bill handsomely for them later. Anduril delivered its prototype autonomous sub to the Australian Navy after three years - one year ahead of schedule. Contrast that to the mystifying timelines in this story.

A start-up will tell you that it will be impossible to build your all-purpose uber-drone to the timelines and specifications you’re demanding. They will suggest something that actually works and push back on extra requirements. Defense primes will happily take your money in the knowledge that it’s a bad (yet lucrative) idea.

You can just do stuff

We’ve written in the past about how there’s an urgency divide across European countries. In our 2023 report on defense procurement, we touched on how Poland had built a significant tank fleet by purchasing off the shelf platforms from South Korea. 

There are, of course, theoretical advantages to integrated European defense programmes. But these advantages don’t mean anything if the equipment is obsolete by the time it’s delivered, or doesn’t ship at all. An unintegrated, real system beats a perfectly harmonized, undelivered one. 

This is particularly important in the current technological context. UAV suppliers operating on the frontline are having to upgrade their equipment every few months in the face of evolving electronic war capabilities. It becomes essentially impossible if every adjustment involves securing sign-off across an elaborate chain of defense ministries and subcontractors.

The era of the exquisite platform is over

The Eurodrone is a project from a different era. Whether it’s the conflict in Ukraine, Iran’s strikes on Israel, clashes with the Houthis over shipping - it’s becoming clear that war has changed. As Michael Horowitz, the former U.S. Deputy Assistant Secretary of Defense for Force Development and Emerging Capabilities, argued in a recent essay, we are now entering the age of ‘precise mass’:

Militaries find themselves in a new era in which more and more actors can muster uncrewed systems and missiles and gain access to inexpensive satellites and cutting-edge commercially available technology. With these tools, they can more easily conduct surveillance and stage accurate and devastating attacks. Its imperatives already shape warfare in Ukraine and the Middle East, influence dynamics in the Taiwan Strait, and inform planning and procurement in the Pentagon.

In the era of precise mass, war will be defined in large part by the deployment of huge numbers of uncrewed systems, whether fully autonomous and powered by artificial intelligence or remote-controlled, from outer space to under the sea. 

A program that produces tens of drones at a cost of billions of Euros has no role in this future. It’s why we’re so bullish on companies combining commercial off-the-shelf hardware with advanced AI, like our friends at Delian Alliance Industries.

Learn when to cut your losses

If a program isn’t working - just kill it already. 

When government ministers say it’s as expensive to scrap something as it is to finish it - they’re almost always wrong. This calculation does not account for the cost of maintaining and supporting the system once it’s delivered. Or the time and effort spent managing the process. Remember - time is (yours and our taxpayer) money! This could all be dedicated elsewhere. To be innovative, you need to internalize the Econ 101 lesson about sunk costs.

Closing thoughts

We’ve written about European defense procurement repeatedly under the loose heading of European Dynamism. The reason we talk about ‘dynamism’ is because we aren’t doomers - we believe that all of these problems are fundamentally fixable. If you’re working on procurement reform or are building in European defense - email us. It’s never too early.

We believe that European countries have the capital, talent, and technology required to defend themselves. Instead, they lack the political willpower and institutions to mobilize these resources. In our response to the SDR, we outline the lack of ambition that has characterized the UK’s approach to technology in defense, as well as a potential roadmap for change.

You can read our submission in full below.

Introduction

Air Street Capital is a London-based venture capital (VC) firm investing a second institutional fund of $121 million in AI-first companies across Europe and the USA. We are one of a small number of European VC firms that are willing and able to invest in pure defence start-ups. 

While we are strong believers in investors’ responsibility to act, much of the market remains unconvinced. 

Despite the UK being the third largest venture capital market in the world, there is no meaningful ecosystem of investors in defence start-ups. Early-stage defence investment in the UK remains a grab-bag of individuals, small funds, and government vehicles. These funds collectively have little capital to deploy and are focused on dual-use technology.

This stems from a perception that defence is ‘uninvestable’ for VC. Domestic and international investors simply believe that the government i) is not serious about adopting new technology; ii) does not understand that time is money for start-ups; iii) will, in the final analysis, direct orders to the same oligopoly of incumbents. 

This state of affairs needs to be urgently rectified. As we see every day on the frontline in Ukraine, the shape of conflict has fundamentally changed. The teams that have adapted best to this change have been agile specialists, incentivized to build cheaply and effectively, not the manufacturers of traditional platforms. The Ukrainian military, motivated by a sense of urgency, battle-tested challengers’ equipment and gave them the opportunity to demonstrate their impact at scale. We believe that the UK should be trying to replicate this model as much as possible, rather than maintaining the pretence that it is possible to achieve start-up outcomes by supporting incumbents. 

In this submission, we lay out our analysis of where government thinking about technology and early-stage companies need to change, lay out some potential policy measures, and traps to avoid.

Urgent acceleration in ambition

Despite the contours of future conflict being obvious to even non-specialist observers, Defence appears to have focused on ‘AI-washing’ over adoption. The MOD has produced overlapping plans and frameworks for AI, and run many demo days, but evidence of actual change is thin on the ground. 

Many of these plans and frameworks amount to aspirational lists of capabilities, occasionally with an approximate price tag, but with no clear plans for acquisition and implementation. A good example of this is the department’s Defence Drone Strategy, a 12-page document that amounts to a statement of objectives, padded out with images and case studies.

There appears to be little appreciation of just how behind the curve Defence is. For example, the Department’s Defence Artificial Intelligence Strategy and accompanying playbook, divides capabilities into ‘AI Now’ (mature technology for immediate adoption), ‘AI Next’ (‘next-generation’ or ‘generation after next’ capabilities), and ‘AI Future’ capabilities.

Many of the technologies listed in the latter two categories are both available commercially off-the-shelf and being used in live conflicts around the world. Indeed, no R&D is required to develop better internal document discovery. 

This lack of ambition stems from a number of institutional failings.

Firstly, a lack of education. Senior leadership seem unaware of the current state of the art in AI research and what reasonable performance they should expect from suppliers. This means routine or underwhelming work is hailed as ‘world-leading’. 

This is exacerbated by a second problem, namely, dependence on a small number of defence primes. Despite repeated pledges to diversify the MOD’s supplier base, the department’s own statistics show that the ten largest defence holding companies receive a greater proportion of MOD spend than they did a decade ago. These companies do not attract or incentivize top AI talent and have prioritized share buybacks and dividends over investment into their own R&D. They also have no incentive to improve the education levels of their customers.

Finally, a cultural tendency towards reinventing the wheel. As the Watchkeeper fiasco demonstrates, the bias towards creating bespoke British versions of equipment that already works is frequently counterproductive. It is also a recipe for being ripped off by primes, who can increase the bill with every change in specification. 

Recommendations:

• The proposed timeline in the Defence Artificial Intelligence Strategy needs to be radically revised and accelerated to reflect both technological and geopolitical reality. No capabilities that are available on the market and deployed in current conflicts should be marked as ‘AI next’.

• Defence’s default position should be to buy technology off-the-shelf or to make only light adaptations to off-the-shelf solutions, unless there is a compelling reason to do otherwise.

Subscribe now

From SMEs to start-ups to winners

Defence’s attempts to diversify its supplier base are hampered by a category error. The umbrella term ‘SME’ lumps together companies with wildly different goals, needs, and growth potential. A local building materials supplier and a venture-backed AI start-up should not be given the same label, either for the purposes of statistics or policy-making. 

This is not just a matter of semantics. Current policy, as laid out in the Defence Small and Medium-sized Enterprise Action Plan, is focused on integrating small businesses into the defence supply chain. It does not envision these businesses ever scaling and playing more than a supporting role to the oligopoly. 

This signals a lack of understanding in government about the economics of venture capital. VC firms do not invest in “SMEs”, they invest in companies that begin life as start-ups but have the potential to scale into becoming multi billion-pound public companies.

Investing early in private companies is an inherently high-risk endeavor. Most investments will not succeed. To justify taking a risk with their investors’ money, VC firms need to believe that there is a realistic chance of a successful exit with significant upside. Currently, only a small handful of firms believe that this opportunity exists in UK defence.

The ‘SME mindset’ is exemplified in the Defence and Security Accelerator (DASA). DASA awards small grants that might be valuable to individual research groups or regional businesses, but it has next-to-no track record of helping early-stage innovators scale to the next stage of their journey. With much of its day-to-day operations managed by a defense prime (which happens to be the biggest single recipient of DASA grants), it is not a serious vehicle for proto-winners. 

With the government-sanctioned gateways for startups largely a dead-end, their only realistic alternatives are either to pursue single-source procurement or to subcontract from the primes.

Single-source procurement is theoretically a workable avenue for start-ups, but is contingent on access to a strong network of insiders and granular familiarity with MOD budgeting. It leads to seed capital that should be spent on developing better products being used to pay lobbyists and advisers. 

Subcontracting from the primes can be a quicker entry route for start-ups, but it has crucial limitations. Not only is intellectual property under threat, but primes can restrict start-ups ability to compete with them in the future either via contractual terms or by gatekeeping customer relationships. By itself, this model will not drive the investment required to create a thriving ecosystem. 

Recommendations:

• DASA and any other vehicles for awarding innovation grants should be barred from awarding any contracts to companies that are at Series B or later.

• Build inhouse capacity and remove prime contractors from the operation of accelerators or innovation units. 

• Unbundle hardware and software components of major contracts by default, allowing companies to bid for one or both.

• Make interoperability a requirement instead of guidance, without exceptions.

A radical reset

For the next generation of capabilities, the priority should be to move away from the system of paper-based requirements as quickly as possible. 

While Defence has the best understanding of the challenges that it faces, neither civil servants nor the forces are expert in designing technical solutions. Additionally, the system of paper-based requirements is one of core pillars of the primes’ advantage over new entrants, due to both their ability to shape them and to respond to repeated changes.

For capabilities like affordable mass, autonomy on the edge, and real-time situational awareness, the government should invite potential suppliers into a real competition. Different bidders’ capabilities would be tested against each other in a series of practical trials, with one bidder eliminated per round. Between rounds, entrants would have the opportunity to iterate on their product. 

As well as providing a real test of capabilities, the process would also reflect how modern technology is actually developed, from MVP through cycles of iteration. 

For this process to succeed, it could not be run as another version of DASA. It needs to be a competition for real, long-term, multi-million pound contracts - not innovation grants. In his 2009 review of defense acquisition, Sir Bernard Gray suggested that defence procurement could be run by a separate organization, under private sector leadership. While this was rejected at the time for being too politically contentious, it could be revived in narrower form to focus on the question of advanced technology. This could be a means of both avoiding prime capture and bringing in outside expertise.

Recommendations:

• Create an advanced procurement agency, with control over a significant proportion of MOD procurement spend, freed from civil service business and pay scales to operate a version of the competitive system outlined above.

• Grant this organization ARIA-style legislative safeguards to prevent it falling victim to changes in political direction.

Dual use isn’t enough

In our interactions with government stakeholders, a frequent refrain is that investors could avoid these acquisition problems by focusing on start-ups that build dual-use technologies. The rationale being that by acquiring private-sector customers early, this removes the risk of depending on long procurement cycles. This is theoretically appealing, but practically unwise.

Firstly, many of the crucial AI-enabled capabilities that defence needs are unlikely to experience meaningful demand in the private sector. For example, there are likely no real commercial use cases for drone swarming technologies.

Secondly, it delays getting crucial capabilities to the frontline. Start-ups working on dual use will inevitably focus on prioritizing the most responsive customer and will focus on establishing product-market fit in those industries. By the time Defence buys from them and the technology is adapted to fit its needs, it will likely be out-of-date.

Thirdly, revealed buyer preferences. Defence does not have a strong track record of purchasing dual-use technology, due to exactly the biases we described around bespoke technology and overspecification. This last point is not a UK-specific problem, it is endemic in many defence establishments, and has worsened as the market power of the primes has increased. The notorious 1993 ‘Last Supper’ in US defence was a significant accelerant, even though the greater use of dual use technology became official Pentagon policy in 1995.

Interim steps

The competition reform would require a radical shift in how the UK procures technology and would naturally take time to implement. There are some smaller steps that Defence could take in the meantime. However, it is important to stress that by themselves these are not an adequate substitute for more radical reform. While they will make life for existing founders easier, they will not catalyze greater investment by themselves. 

Recommendations:

• Overhauling current framework contracts to remove the arbitrary entry windows. These were designed for bureaucratic convenience and can potentially leave start-ups waiting for years before being able to bid for work.

• The current system of clearances creates a chicken and egg scenario, where start-ups are unable to win work without clearances, but unable to obtain clearances without work. This often results in either costly subcontracting arrangements or having to spend additional money hiring staff with clearances from previous roles. The MOD should hold clearances on behalf of start-ups at a cost, as they do for contractors and their own employees.

• The requirements for holding classified information, laid out in the Facility Security Clearance Policy and Guidance, are costly for start-ups and MOD-provided facilities frequently in remote locations. The MOD could create a compliant facility in central London and sublet space to start-ups.

  Subscribe now

As believers in openness as a driver of progress, we share our unvarnished views publicly, not just behind closed doors. So in that spirit, we’re sharing our submission in full. 

Share

Matt, 

Thank you for the opportunity to engage as part of this review. As promised, we’re sharing our follow up from the No 10 roundtable below.

We emphasize the need for a whole-of-government approach to technology policy. That means, using the different levers distributed across government departments, rather than treating AI as a single sector in need of a new funding settlement from DSIT.

While we’ve broken out the topics separately - they are, of course, all connected. For example, a better functioning spinout ecosystem will benefit from a greater number of specialist solo GPs.

The recommendations below are by no means exhaustive. We’ve left out areas where we believe others will make a more detailed contribution (e.g. building regulatory capacity) or where we believe plans are already underway (e.g. a national data library). We believe these to be crucial areas of work, but we’re keen to avoid duplicating recommendations.

Recommendations

Government as a customer, not an investor

Government’s usual lever of choice for early-stage company support is subsidy (either via VCs indirectly or via direct investment). Government is a poor direct investor and if a deep tech company at Series B needs a government vehicle as a direct shareholder, it has likely failed to find product-market fit. It should be allowed to either fail or be acquired, rather than acting as a sink for talent and capital.

• Government should cease any direct investment in start-ups beyond Series A. Pre-A investments should only be in companies with a clear tie to national security.

Instead of acting as an investor, the government should focus on procurement reform. Its spending accounts for approximately 45% of the UK’s GDP, but access to this for start-ups is conditional on deep pockets, insider networks, and luck. This is a rare combination and not a scalable formula for success.

Procurement has traditionally been challenging for start-ups for several reasons:

• Prohibitively long timelines combined with a highly labour-intensive process. Start-ups are unable to devote significant leadership, engineering, and business development resources to processes that may not arrive at an outcome for well over a year. Recent moves, such as the Ministry of Defence’s decision to lower the limit for ministerial approval for spending from £2M to £50k are retrograde steps.

• Fragmentation, forcing them to build relationships with a byzantine network of regional stakeholders, many of whom individually have limited spending power. 

• Overemphasis on paper requirements, rather than developing the best answer to a specific user need. While government bodies will know the problems they face better than any outsider, they are less likely to be well-placed to know the exact shape of the solution. 

• The process becoming an end in itself. If a company tries to hire for a role, but doesn’t find a suitable applicant,  it will usually pause the search, rethink, and try again. This philosophy is rare in procurement and often leads to subpar or inappropriate solutions produced by incumbents being chosen out of desperation. 

• Little promise of pay-off at the end, with start-ups often being limited to pitching for small grants and contracts from innovation units and accelerators that usually lack the potential to develop into larger-scale work. This means revenue is highly unpredictable, which is challenging for both hiring and fundraising. 

We have laid out more specific recommendations around defence previously, but believe that many of the same principles could be applied to other areas of public procurement. For some capabilities, this could involve:

• Testing MVPs against each other round-by-round, eliminating them one at a time and providing feedback to the survivors. 

• Compensating start-ups for participating, based on how many rounds they make it through. Companies eliminated in the first round or that have pitched an obviously unsuitable solution receive nothing.

• Ensuring a real, multi-year contract lies at the end of the process, rather than a pilot or forced partnership with a larger provider.

Harnessing innovation in our universities

The previous government’s Independent review of university spin-out companies was a valuable first step in correcting the UK’s historic poor performance in commercializing the technological breakthroughs in its universities. It confirmed the slow deal times, unfair equity takes, and lack of transparency that we exposed via Spinout.fyi and outlined a welcome set of recommendations around best practice. 

However, the final recommendations were voluntary and too generous towards universities. The review explicitly adopted the equity splits recommended by TenU - the technology transfer offices’ lobbying arm - giving universities scope to claim up to 10% equity in software start-ups and 25% equity in life sciences spinouts. These terms still risk rendering promising companies essentially uninvestable.

UK universities’ role in the spinout process should be more akin to that of their US counterparts - in-house TTOs whose role is confined to quickly turning around legal documents. The UK model, where universities act as a cross between a lawyer, a VC, and a venture-builder, results in universities that provide an expensive service badly. Universities should focus on building alumni ecosystems that make successful graduates want to give back to their alma mater, rather than trying to use founders as a substitute for an endowment. 

• TTOs should uniformly adopt a Simple Agreement to Spinout (SAS) that sets a transparent standard specifically for spinouts. The SAS gives TTOs the option to select either i) 1-5% common equity or ii) in the case of clearly identified IP over a drug molecule or medical device, 1% royalty on net sales (direct or sub-licensing based), or iii) 1% of the exit consideration upon M&A or IPO, in addition to a one-off £20,000 per patent that the spinout in-licenses. If the spinout does not use licensed IP for 24 months, it must return it to the University for recycling. No other fees are levied.

• Compliance should be mandatory, with universities that don’t adhere being subject to meaningful funding allocation penalties. They should also consider penalizing universities that restrict access about academic work to their chosen partner investors, giving them an unfair advantage.

• The government should reject any proposals from universities for additional subsidy for spinouts, beyond the proof of concept funds outlined in the review to help spinouts make a go/no go decision. Given the volumes of early-stage capital on offer in the UK, anything beyond this is unnecessary. Founders that are serious about creating global winners normally prefer to raise money from outside firms with a track record, not universities. 

Target support at emerging managers, not incumbents

Government financing vehicles played a role in creating the UK’s VC ecosystem, most recently under the banner of the British Business Bank (BBB). But many of these schemes’ origins date back over a decade and a half to a time when the financial crash was hitting and the UK’s VC scene was an international backwater. This is no longer the case. The UK does not have a funding gap and the BBB should not be promoting this narrative.

Once a manager has established a track record over one or two funds, they should not require government support. Instead, they should be able to raise from university endowments, non-profit institutions, entrepreneurs’ family offices, funds of funds, sovereign wealth, and corporates. 

Instead, government should focus on the real bottlenecks in the ecosystem: 

• Focus on emerging funds, or solo general partners and start-up operators turned investors. These are precisely the specialist investors a reformed spinout ecosystem needs. At the moment, solo GPs are not eligible for BBB funding. An emerging manager programme should have looser requirements on the team, while moving significantly faster. It should take no more than six months to secure 20% of a fund.

• Create a follow-on fund to support managers who can’t afford to exercise their pro-rata rights. The government could create a follow-on fund which invests in UK companies that had their A or B round led by a Tier 1 US VC firm. The government would invest via a special purpose vehicle, with the manager selling their pro-rata rights in the round in exchange for carry. The VC fund would manage the SPV and the government wouldn’t talk to the company or run additional due diligence.

• Remove unnecessary regulatory requirements. Consider adopting an SEC-style exemption for VC, to remove reporting and compliance requirements that were often designed to protect retail investors investing in public markets. For smaller funds, the £100k a year on compliance (vs $30k in the US) they need to spend on compliance consumes a substantial portion of their management fee. An FCA sandbox for emerging managers, which maintained AML and KYC requirements while losing other unnecessary overheads, could act as a draw for talented managers from other European countries that also maintain these requirements.

Infrastructure fit for the 21st century

The government’s Independent Review of The Future of Compute laid bare the poor state of the UK’s compute infrastructure, including our lack of supercomputing capabilities and the low number of GPUs available to researchers.

• The government’s decision to scrap investment in a national supercomputing facility was short-sighted and wrong. It should be reversed.

Assuming the government does decide to build out capacity after all, the next stage will be in getting the access model right. Making this worthwhile will require a degree of both risk-taking and speed, while avoiding turning into subsidy for incumbents. 

Any future access model should consider:

• Allowing access to industry (pre-Series B only) as well as academia;

• As short an application process as possible. One of the main gaps in the market is ‘bursts’ (e.g. testing if scaling laws in a domain work before committing to a larger training run) - a months-long application process would defeat the objective;

• Using external yardsticks of quality (e.g. publishing track record of the team, submissions to major conferences etc.) to judge applicants, rather than the government attempting to run its assessment process.

• The government should also consider whether it is best placed to operate the cluster day-to-day, or whether an external firm (e.g. CoreWeave or Crusoe) could provide a better standard of service more quickly. 

Subscribe now

Prefer narration? Listen to the audio version of this post.

Introduction

It’s been close to six months since we published our report covering the challenges early-stage defense companies face in Europe. Since then, we’ve seen a rush of defense-related announcements from European governments, speculation about the future of US support for the Ukrainian war effort, and the Russian military continuing to step up its offensive.

Given the pace of developments, we’re re-examining our initial work and assessing what’s changed, what hasn’t, some points we didn’t evaluate last time, and where we go from here.

The innovation arms race

The war in Ukraine underscores the importance of both acting urgently and focusing on the right challenge. Our previous report correctly touched on Ukrainian innovation, which has been widely documented in media coverage. Unsurprisingly, however, innovation in this conflict has not been one-sided. 

While initially caught off-guard, the Russian military has quickly sought to learn from Ukrainian tactics and weaponise them. Both sides have demonstrated considerable agility in copying from each other throughout this conflict.

Fighting takes place under a drone-infested sky, with 90% of injuries in some areas stemming from FPV drones and drone-dropped munitions. These range from more expensive Shahed drones, imported from Iran or assembled domestically, or the less sophisticated Orlan-10, which is used for reconnaissance. The Orlan, despite being critiqued for its often crude assembly, has played a vital role in helping find targets, collect intelligence, and support Russian electronic warfare. Crucially, it is also cheaper and easier to produce than it is to shoot down. It’s also helpfully powered by a series of western components:

Meanwhile, the vast majority of drone-dropped munitions that don’t rely on improvised volunteer drones use off-the-shelf DJI drones from China.

These efforts have been supplemented by an array of volunteer efforts that are reminiscent of some of those seen on the Ukrainian side.

It is difficult to verify the prevalence and effectiveness of many of these volunteer efforts, as much of the source material comes from pro-Russian Telegram accounts. However, we do know that they are receiving government support and there is no shortage of publicly available video of their drones apparently being deployed to the frontline.

A Russian studio has also created a pirated version of a powerful FPV simulation game, which had originally been created by a Finnish company to support the Ukrainian war effort. A contact of ours, close to UK drone efforts, has warned us that is eerily realistic and competitive with Ukrainian equivalents. While available openly via Steam, we wouldn’t recommend downloading, as we don’t know what spyware may be lurking within. 

While we definitely saw a mini-exodus of Russian tech talent in 2022, obituaries for the industry now seem premature. There are reports that large numbers of people with valuable tech skills have in fact returned to the country, lured by the promise of generous payouts from the government. It’s also clear that in any future conflict, it is unlikely that we will be able to rely on a sanctions regime to cut off access to advanced technology. As well as European components semi-regularly turning up inside Russian drones, Russian middlemen aren’t struggling to get hold of Starlink terminals.

The other concerning development has been the failure of many western innovators to make a significant difference on the front-line. The clearest case of this are the challenges US drone manufacturers have run into when faced with Russian electronic warfare. So far, this has proven able to adapt quicker than equipment manufacturers can roll out updates. This technology has also been used to inflict chaos on Estonian airspace and even grounded the UK defense secretary’s plane, which was missing crucial defenses due to budget cuts.

Peacetime versus wartime mindset

Given the clear importance of both a strong industrial base, flexibility, and affordable mass, we should be expecting to see a gear-shift in Europe.

In our original report, we highlighted the ‘peacetime’ vs ‘wartime’ split in the European psyche, predominantly along geographical lines. This meant that Eastern European countries have largely been increasing defense spending rapidly, rearming, and pushing for more substantial aid to Ukraine. Meanwhile, further west, things have been proceeding at a comparatively lethargic pace.

Readiness in the Eastern European and Baltic countries remains a priority, even during challenging economic conditions. Not only are these countries continuing to increase defense spending as a proportion of GDP, we’ve seen:

• Estonia announce a €50M defense technology fund, with the potential to grow larger, which crucially, will not focus on dual use;

• Lithuania cut the time required for defense manufacturers to set up operations in the country from 2 years to 6 months by streamlining approvals  and has introduced generous tax incentives;

• Poland increased its purchase of off-the-shelf UAVs, while the country is gearing up to launch its first military satellites next year.

Moving further west, the picture is more mixed. While investment levels are definitely increasing, there are lingering questions about strategy, readiness, and procurement methods.

France has certainly shifted tonally, with President Macron’s surprising suggestion that French troops could be deployed to Ukraine. We’ve also seen France increase spending on military R&D, pledging €300M a year on an agency dedicated to military applications of AI. Government-back technology investment in France, however, is not famed for its speed, effectiveness, or flexibility. It’s crucial that any body doesn’t simply subsidize long-standing incumbents and that there’s a clear roadmap for bringing the product of this R&D to the frontline at scale.

We see some of these typical patterns in French policy reasserting themselves in their approach to defense M&A exits. The government appears to be restricting these to acquisition by a very small pool of French corporations, making it difficult to see how the sector will be venture-backable. This is concerning, given how uncertain the exit market already looks for defense technology.

In Western Europe the lack of readiness extends beyond procurement. There have been a string of embarrassing incidents, ranging from China hacking the UK Ministry of Defence through to missile misfires and revelations about the woeful state of German military communications, which suggest that readiness remains in a poor place. Meanwhile, the drip feed of revelations about former Wirecard COO and Russian agent Jan Marsalek’s thorough penetration of the Austrian government should concern us all.

Subscribe now

The UK: revisiting our central case study

Our report took the UK as its primary case study, finding a combination of hazy strategy, underfunding, and botched execution. 

Perhaps the most significant single move was the government’s announcement that defense spending would increase from 2.2% of GDP to 2.5% by 2030. This was extravagantly branded as a ‘turning point’ for European security. While additional funding is always welcome, reaction in the UK was mixed. The government announced its intention to find the additional money through unspecified civil service job cuts. Similarly, the claim that this would total £75B in extra funding was quickly shown to be a 3x exaggeration, produced by disingenuous benchmarking. 

The other significant bright spot has been the news that the AUKUS nations are preparing to exempt each other from their defense export control regimes. Not only does this open up major new export markets for innovators, it makes partnership and the sharing of classified information significantly easier. It also makes the UK the most logical country in Europe by far to incorporate a defense start-up. Our friends at Lambda Automata building a significant London presence being a case in point.

Beyond that, there have been a few interesting acquisition developments.

The new Challenger 3 tank is entering final production under an £800M contract with Rheinmetall BAE Systems Land. 148 of these will become the Army’s Main Battle Tank, remaining in service until 2040. This suggests it’s business as usual at the Ministry of Defence.

While the initial delivery of the Challenger 2 was announced with great fanfare, it’s been far from a game-changer in Ukraine. Traditional platforms that take years to build and which are specified down to the last bolt are difficult to adapt in the face of field conditions. That’s why only half of those sent to Ukraine are in operation, due to reliability and maintenance issues, while it takes months for spare parts to arrive. Meanwhile their weight causes them to become regularly stuck in the mud and require towing. Concerningly, the Challenger 3 will be even heavier.

Beyond questionable choices around traditional hardware, the government also published the Defence Drone Strategy in February. The 12 page document (8 of which are the covers, introduction, photos, glossary, and past case studies) lists a £4.6B resource commitment over the next 10 years, but is light on detail. It largely repeats long-standing aspirations around R&D, cooperating closely with industry, simplifying acquisition, and embracing open architectures. While these may be good things, they are not firm, measurable commitments or a plan.

It remains unclear if the UK government will be able to end its dependence on expensive, bespoke platforms in this arena. When it comes to UAVs, this approach has already been a disaster. The UK’s Watchkeeper programme resulted in a £1.5B outlay over several years to deliver 54 drones years late, none of which are on operational deployment. Originally meant to be a copy of the successful Israeli Hermes 450, in classic MOD style, 2,000 systems requirements were added to the programme, making the drones too heavy. It also transpired that they could not handle bad weather. Meanwhile, the RAF’s specialist drone squadron has not conducted a single trial since 2020, due to budget shortfalls. This stands in contrast to the delivery of the Ghost Shark autonomous submarine project by Anduril, which came in on budget and early.

As with France’s R&D agency, it’s worth questioning the extent to which European governments necessarily need to engage in highly capital-intensive R&D projects or multi-year partnerships with industry when so much technology is already available off-the-shelf. 

Whether it’s the Defence Drone Strategy, the R&D boosts built into the defense spending uplift, admittedly cool lasers, this expensive domestic innovation drive risks becoming a distraction. When components like drones, optical and thermal cameras, edge computing hardware, and guidance and navigation devices are essentially commoditized, it seems bizarre to approach procuring them like a traditional manned platform. Especially given the clear value that cheap drones, treated as ammunition rather than aircraft, have demonstrated. Fresh R&D spending needs to avoid becoming another round of innovation theater. This will divert attention away from fixing a fundamentally broken acquisition system. Scale and efficiency are much more likely to win out than elegance. 

Closing thoughts

While much has happened, it feels that the same reluctance to tackle fundamental challenges persists in many European countries. Symbolically throwing good money after bad may be easy, but it’s wasteful and only creates a smokescreen of reassurance. As we said in our original report, we don’t believe change is likely to be driven by the small group of incumbents whose business model depends on multi-year contracts for bespoke platforms. It’ll be innovators, whether it’s Lambda Automata, the people who recently spent their weekend at a London defense hackathon, and other early-stage innovators. 

It’s our belief that good institutional wiring and agility are usually more important than raw investment. The solution here is either root-and-branch reform or hiving procurement out into a separate, well-funded specialist agency exempt from the usual rules. It does not lie in a constellation of new innovation hubs, state-led R&D programmes, or government subsidies for the VC sector. The UK is gearing up to launch a new Defence Innovation Agency in 2025 - this is the perfect opportunity to be radical. Based on current trends, however, we are on course for the missed opportunity of a lifetime.

Share

As the government’s last major fiscal event before the upcoming general election, this was essentially their final opportunity to swing big for the UK’s tech sector. Did they rise to the occasion? 

Missing the mark

In our estimation, he unveiled a package devoid of ambition or resources, seemingly designed to avoid tackling any of the real problems faced by people trying to build AI-first companies in the UK.

In the past, we’ve written about government technology investment that’s simultaneously expensive and unimpactful.

Increasing funding for the Alan Turing Institute, the UK’s national data science center, to £100M, is a classic example of this. Bluntly, very few people who are not directly employed or funded by the Institute take it seriously and it wholly failed to prepare the government for the GenAI explosion of the past few years. This is why the Tony Blair Institute has recommended all but shuttering it. Government would have received a better (and entirely free) briefing on advances in NLP and the power of the transformer architecture from reading the State of AI Report. 

The Budget also included measures designed to boost the UK’s capital markets. In addition to last year’s agreement aimed at increasing pension funds’ investment in venture capital, the Budget will force some pension funds to publicly disclose their split of UK and international investments. As far as we can tell, the objective is to pressure funds to invest more of their assets into UK equities, correcting the “valuation discount” versus the US, and creating a stronger exit market for companies. 

We’re old-fashioned and believe that asset managers should make investments based on what they believe is in the best interests of their clients, not because they’ve been accused of being unpatriotic by the Treasury. Far less attention is being given to the question of why UK-listed shares don’t attract more interest - possibly because the government may not like the answer...

But the most galling part of the Budget for us was not the tech-related announcements. 

Buried in the small print, the Treasury acknowledged that not only would UK defense not receive additional funding, elements of its budget would face actual cuts. Its capital departmental expenditure limit will be slashed from £20.3 billion in 2022-23 to £18.9 billion in 2024-25. This follows on from the news that the UK’s drone squadron has not conducted a single test since being formed in 2020, due to funding shortfalls. In the context of poor UK defense readiness, an ongoing European conflict, and our action in Yemen, this is mystifying.

Share

The purpose of a system is what it does

So where does this gap between the UK’s science and technology rhetoric and action come from?

Team Air Street have attended three events in London in recent months that each give us a piece of the puzzle.

Starting with the most recent, Alex attended a conference focused on the application of AI in a national security context. An investor with insight into UK policy thinking argued that there was a strong cultural bias against taking advances in AI seriously. Especially in contrast to the Bay Area, it’s ‘cool’ to be sober about technology in the UK, rather than excited about its transformative potential. This leads to complacency in policy-making.

Alex was also at a ministerial roundtable focused on helping start-ups to access compute in the UK, potentially through a national cluster. While there was agreement in the room about the importance of this work and the potentially positive role, the discussion did not address the elephant in the room: the absence of a national compute cluster or a serious plan to build one. 

While the Budget and other official documents routinely reference a £1.5 billion investment in future UK compute capabilities, the government has never published an itemized list of these commitments. We believe that ‘future compute’ has likely been defined rather broadly. As far as we can tell, the majority of this money will go into one facility in Edinburgh, which will not, alongside any of its other investments, be open to entrepreneurs. 

Last year, Nathan was in front of the House of Lords Communications and Digital Select Committee to discuss the LLM opportunity in the UK. The theme of his evidence was that the UK needed to be more ambitious, in terms of both investment and fundamental reform. This received pushback from a senior member of the committee on the grounds that many other European countries faced the same challenges. 

What do these three events tell us? 

In short, the UK can’t decide how serious it is about AI leadership. Against a dire economic backdrop, the government is scrabbling around for a science and technology future that simultaneously requires i) little upfront investment ii) no institutional change iii) no risk. 

A serious vision of technology leadership would involve:

• A vast upscale in our compute ambitions, while opening up access to this infrastructure to innovative start-ups

• Wholesale procurement reform, so start-ups aren’t locked out of working with the country’s biggest customer

• Opening up access to government data to innovative start-ups so they can build systems that can tackle crucial challenges for the public sector

• Reform to the early-stage VC ecosystem to reduce the burden on new funds, encouraging specialists or former operators to enter the sector

• An end to British Business Bank subsidy for VC incumbents or meandering startups, and the reallocation of funds to support strategically important winners

• Copy France is creating a competitive offer for overseas founders to build their business in the UK, including tax breaks and discounts on hardware

• A line-by-line examination of current recipients of technology-focused funding from the government to determine if they’re builders or merely lobbyists.

We’re not denying that this would involve significant upheaval. Not only does the UK have all the raw ingredients to do this, the upside would be tremendous. However, while those in authority radiate unseriousness, the system will behave unseriously. 

Technology leadership will turn into subsidizing an ineffective data science institute and browbeating pension funds into buying more Rio Tinto shares. A commitment to spend 3% of GDP on defense will turn into reductions in capex. And we’ll all have a friendly conversation about the best way to allocate resources from a compute cluster that doesn’t yet exist. 

Share

The first phase of our work on European Dynamism and defense examined how the sector was stuck in a doom loop. Thanks to a broken procurement system, new entrants rarely scale, investors see next to no chance of any attractive return on the capital they risk, and founders opt for a surer playbook in the safer lands of SaaS.

We’ll have more to say on procurement reform in the coming weeks, but today, we’re looking at the other side of the market: exits.

There’s been a lively discussion in recent weeks among investors about the ‘exit problem’ in European defense. With lower margins than sectors like enterprise software, potential returns are already structurally more constrained. Add to that the low likelihood of many challengers making it to IPO. The defense primes often don’t pursue tech acquisitions aggressively, while historic consolidation of the buyer universe has narrowed the pool of potential acquirers. 

Following World War II and then the Cold War, western governments actively encouraged this consolidation to counter falling demand. This was most famous in the US with the so-called ‘last supper’:

However, the same trend is visible in Europe. For example, BAE Systems is a product of decades of mergers and acquisitions, fusing together at least 12 different companies.

In this piece, we’ll offer a perspective on the barriers to exit, including how the defense primes could move from being the part of the problem to part of the solution.

We aren’t going to spill much ink here on the well-trodden subject of public markets reform, as many others have done this already. Instead, we feel there’s a valuable case study in an apparently different industry, which has some surprising areas of overlap. One where the industry is dominated by a small number of old-school, publicly traded companies with large amounts of cash on their balance sheets that are also subject to strict government regulations when creating and selling new products. We are, of course, talking about big pharma.

Share

Strategic partnerships and an M&A boom

To be clear, we don’t believe big pharma companies are the star pupils of the public company world. They have faced many of the same criticisms as the defense primes, whether it’s anti-competitive commercial practices or focusing too much on dividends and buybacks. When compared to the defense primes, however, they are model corporate citizens. 

When big pharma spotted an emerging wave of new start-ups that significantly outclassed its own R&D, it saw an opportunity. The precedent was set in the 1980s by Genentech and Roche. Genentech pioneered and commercialized the field of recombinant DNA technology and is widely seen as the first ever biotechnology company. Back then, while biotechnology was confined to academic research labs, Genentech was out successfully raising venture funding and bringing products to market.

Roche leaned into the opportunity to partner with a business that was beating the rest of the field on R&D. First, it bought the rights to Interferon, a drug for the treatment of hepatitis C and hairy cell leukemia. It similarly bought the marketing rights to Activase, a drug used for treating heart attacks and strokes. As Genentech ran into financial difficulties, Roche purchased a 60% stake in 1990 for $2.1 billion.

While this purchase, and a later complete acquisition, forged closer research ties, Roche was content to allow Genentech to operate as an independent unit with far-reaching autonomy. This allowed it to preserve its culture, ability to attract talent, and flexibility, while benefiting from Roche’s commercial reach and firepower. Genentech is all the stronger for it, poaching academic luminaries to build out a 400-person computational sciences team, while striking a strategic partnership with NVIDIA.

This set a precedent in the industry that other firms soon began to replicate. The entire biopharma industry is predicated on big pharma outsourcing early and very risky R&D to a lively ecosystem of biotech startups that are funded with venture dollars. 

Once biotechs demonstrate their ability to discover, design and successfully advance promising medicines to a certain stage of maturity, pharma is more than happy to step in with large acquisitions in the billions - either of companies or individual product lines. This passing of the baton enables promising medicines to be adequately capitalized for large clinical trials and regulatory compliance procedures until they obtain clearance from the FDA such that they can be administered to patients.  

In 1993, we saw the number of strategic partnerships between biotech companies and big pharma hit 69, before soaring to 502 by 2004 and continuing to climb. Strategic partnerships then tipped into an M&A boom in the sector. This was combined with healthy R&D investment, averaging out at 20% of annual revenues.

Defense primes underinvest and taxpayers pick up the bill

By contrast, while acquisitions happen in defense, volumes are smaller. Life sciences M&A in 2023 alone hit $191 billion, while the $38 billion for defense was regarded as a good year. Meanwhile, the largest defense firms spend precious little on R&D, especially compared to pharma. We struggled to find a defense contractor that spent over 5% of net revenues, with most sitting at around 2-3%. For example, AbbVie, which is in the middle of the pharma R&D league table, spends more than the three largest defense contractors combined.

Source: Company Form 10-Ks (2022)

While it’s true that pharma companies operate on significantly larger margins, we analyzed the balance sheets of a number of defense primes, and found no shortage of spare cash. Instead, we found those companies were simply more interested in paying out dividends and buying back shares than they were in investing in R&D. 

It’s also important not to be taken in by headline R&D investment figures, as defense primes usually pass the cost of R&D back onto the customer. In essence, the taxpayer provides subsidized R&D for multinationals, the product of which is then sold on to foreign customers.

For example, QinetiQ, who we charted as the biggest winner of DASA grants, recorded £328 million of R&D expenditure in 2023, but £313.8 million of this was funded by the customer. QinetiQ naturally paid its shareholders a 2% dividend. Last year, Lockheed Martin returned £9.1 billion to shareholders through dividends and buybacks. Company-funded R&D totalled £1.7 billion in 2022. Similarly, BAE Systems unveiled a £1.5 billion share buyback programme in 2022, while spending £287 million of its own money on R&D. 

In 2023, the US Department of Defense conducted a study of defense financing arrangements. Industry representatives claimed that the profitability from government contracts simply wasn’t enough to finance further investment. 

The study flatly disagreed finding that: “Despite increased profit and cash flow, defense contractors chose to reduce the overall share of revenue spent on IR&RD and Capital Expenditures (CapEx), while significantly increasing the overall share of revenue paid to shareholders in cash dividends and share by buybacks by 73%”.

Source: Department of Defense: Contract Finance Study Report 2023 

A question of incentives

So why has pharma upped its game, while defense hasn’t? There is likely a lack of imagination on the part of the primes, but it’s not the main factor. 

In short, there’s been no commercial reason for them to do so. If a pharma business manages to secure a patent on a crucial new drug, this can lock in billions of dollars of revenue for decades, which it would otherwise not receive. Its merits are a discussion for another time, but pharma patents create a winner-takes-all dynamic. This just does not exist in defense contracting.

Due to a desire to preserve jobs and maintain capacity in the industrial base, programmes of record are often divided up among the primes. This makes the competition among the primes a distinctly lower stakes endeavor, while shutting out challengers. There are almost no circumstances in which governments would allow these businesses to fail - no matter how often they fail to deliver. 

At the risk of sounding like a broken record, the incentive to engage with start-ups at scale - whether that’s through investment, licensing, or acquisition, will only come when the customer rewards innovation. 

As we established in our previous report, efforts to increase the proportion of defense spending directed to smaller businesses benefits plumbers and electricians who happen to operate near military installations, but it does little to modernize capabilities. However we cut the problem, whether it’s entry or exit, we end up back in the same place. Fundamental reform is unavoidable and anything short of this is wishful thinking. 

Multi-billion dollar companies aren’t going to spontaneously cut their buyback programs to focus on R&D out of patriotism. Start-ups are going to struggle to disrupt them when the front door remains sealed shut.

We will continue to speak out on this issue and invite anyone who agrees with our mission or is working on an idea in this space to get in touch.

Subscribe now

These include:

• Resolving structural barriers (e.g. visa rules for talent);

• Building the right infrastructure (e.g. building out a proper national compute cluster);

• Acting as a customer (e.g. fixing procurement). 

Anyone who’s spent time on the UK technology, life science, or venture scenes will be aware of the ‘funding gap’ discourse. The argument essentially runs as follows: the UK is a great place to attract early-stage investment, but when you reach Series B, you will hit up against a lack of domestic capital.

As a result, promising companies either fail to get through their growth stage or have to turn to deep-pocketed US investors. Sam Gyimah, UK science minister turned venture partner at Lakestar, went as far as saying that the UK faced a $1.8 trillion funding gap. As a result, the government needs to step in and ‘catalyze investment’ (read: ‘prop businesses up with taxpayers’ money’).  

We frequently heard this argument when we were campaigning for spinout reform. The failure of more than 95% of university spinouts to raise more than £25 million was not because universities had sabotaged them through a flawed process. Instead, it was because of a funding gap in [insert incredibly well-capitalized sector of your choice].

This always confused us, because we’d read about these funding gaps in the same newspapers that would run stories about record volumes of VC ‘dry powder’. 

Our friends at Dealroom have helpfully compiled data showing that UK VCs are currently sitting on $18 billion of unspent money - twice as much as five years ago. 

Source: Dealroom and Air Street Capital analysis

The notion that an industry with this much capital at its disposal is unable to lead B or C rounds without the government stepping in isn’t credible. So what’s going on?

Share

A stop gap excuse 

Firstly, we think ‘funding gaps’ offer a comforting defeatism. By underpricing the potential impact of difficult institutional change, whether it’s capital markets reform or improvements to procurement, we can convince ourselves that the problem is out of our hands. It’s easy to lament that we can’t all be as rich as the US. Getting out the cheque book is easier than rewiring the operating system, no matter how ineffective it usually is.

Secondly, it’s straightforward lobbying for grant money. While the UK government rightly valorizes its science and technology sector for the quality of its output, the companies are motivated by the same incentives as other parts of the economy. We saw first-hand just how aggressively universities fight for their short-term commercial interests. If lobbying the government gets you free money, you’ll do it. 

Finally, a refusal to accept the basic reality of free markets. The majority of venture-backed companies will not succeed and we would expect many of them to fail at the ‘funding gap’ stage. Good technology can fail to achieve product-market fit. A company’s storytelling may fail to persuade customers or investors.  

Someone else can come along and build something better or more quickly or get to customers faster…or all three! 

While regrettable for everyone involved, it’s not a market failure the government needs to correct. In fact, it’s a sign that the market is functioning properly. 

Throwing good taxpayer money after bad VC money is not only wasteful, it’s actively harmful to the ecosystem. It distorts investment by creating a false demand signal and locks talent into businesses with weak long-term prospects. 

By contrast, Silicon Valley accepts entrepreneurship as a high-risk path. There are many successful founders who initially experienced failure, before going on to build something better. If the government had stepped in to catalyze investment and keep their first business afloat, even after it had failed commercially, the later success would never have been born. This comes with a staggering opportunity cost in human potential.

It often feels like governments are acting on an outdated model of the VC ecosystem. 10-15 years ago, the venture industry in the UK was genuinely very small and the government played an important role in bringing it to life. 

However, times have changed. Instead of backstopping investors that can afford to stand on their own two feet, governments should focus their attention on the policy barriers they are uniquely placed to address.

We believe that Europe’s current financial, political, and cultural climate does not enable or reward mission-driven entrepreneurship.

As the world becomes more uncertain, we are waking up to how innovation in defense has been left to wither on the vine. This stems from capture by incumbents, wishful thinking, and a dire lack of investment.

We’ve laid out our thinking on why investors have a responsibility to act, but this is only part of the picture.

The success of all investment is in part determined by wider context - whether it’s political, institutional, or regulatory. Innovation doesn’t speak for itself. These effects are particularly strong in a highly regulated setting, where governments are the only customer.

Download the report

Unfortunately, defense acquisition in much of Europe is fundamentally broken.

It relies on slow-moving systems designed for the peacetime purchase of hardware. It’s starved of funding by governments, who view defense as a painless source of savings. Meanwhile, the system has been captured by a handful of incumbents who lag the cutting edge.

We believe that openness is the best way to drive change. That’s why we’re sharing our internal assessment of where the system has gone wrong in Europe, based on our research, as well as conversations with former military officers and officials, executives at leading defense technology companies and investors. It dives into:

• How Western Europe is still approaching defense as if we are operating in peacetime;

• Our analysis of UK government defense purchasing data, revealing that despite public assurances to the contrary, our dependence on the primes has only deepened;

• How UK government attempts to engage startups have amounted to innovation theater, with little meaningful pull-through;

• The lessons we can learn from elsewhere in Europe;

• An initial sketch of some principles for reform.

We’re in the early stages of this work and we will continue to gather more examples of good and bad practice across Europe. We hope with time that the good to bad ratio will improve.

If this resonates with you, or you have feedback or stories you want to share, please get in touch with us. Similarly, if you are a founder working in defense or a related field, whichever stage of your journey you’re on - we always want to hear from you.

Before I joined the venture capital industry in 2013, I’d planned to train as a physician-scientist, with a focus on translational medicine. During my undergraduate summers of 2006-10 that I spent working on breast cancer metastasis at MIT, I saw the power of the Boston biotech ecosystem. This engine translated research breakthroughs from the lab into spinouts that commercialized transformational products at a dizzying pace.

When I moved to the UK in 2010 to pursue my PhD at Cambridge, I met the same range of world-class science produced by highly gifted and motivated researchers. But something was markedly different.

Despite a small number of success stories (SwiftKey, Arm, Solexa), almost no one I met had commercialized their research. Few even tried, despite the growing tech ecosystem an hour down the road in London. Potential breakthroughs in materials science, drug discovery, and artificial intelligence remained locked in the vaults of the ivory tower because of culture and bureaucracy.

Today's big win for UK university spinouts

Ever since, fixing the formation of spinouts from breakthrough university research has been close to my heart. It’s a mission that matters.

That’s why I’m delighted to see the publication of the UK Government’s independent review of university spinouts and the Treasury's acceptance of its recommendations today, which includes many of my proposed reforms.

Back in 2021, I launched Spinout.fyi, an open dataset on spinout deal terms crowdsourced from hundreds of spinouts. It painted a bleak picture of predatory deal terms, slow timelines, and an inexplicably secretive process. Many of the worst offenders were among the UK’s richest and best known institutions.

Our work was met with significant pushback from the universities sector, government, funding bodies, and the small world of spinout investors. Many previous campaigners for reform told me I was wasting my time. It was denounced by senior administrators and the Russell Group in national newspapers, even as more founders and academics came forward to corroborate our claims. With UK universities central to the government’s science and technology narrative, only a handful of people were willing to comment publicly two years ago.

It was only the slow grind of essays, several op-eds, debunkings, data (V1) releases (V2), case studies, and contributions to others’ work that led to this report existing in the first place.

That’s why today is such a significant day for Air Street and the entire spinout community. Among other measures, the report recommends:

• Significantly reduced and dilutable equity ownership by universities at inception;

• The adoption of a simple agreement to spinout, available publicly;

• Significantly faster deal times by moving away from ad hoc negotiations and committee-led decision making in which spinouts are powerless;

• Fairer distribution of IP among founding teams, so that PhD supervisors don’t walk away with disproportionate allocations;

• New transparency requirements around spinout deal terms, extending the spirit of spinout.fyi;

• Proof of concept funds for putative spinouts before they make a go/no go decision to launch.

These recommendations tally heavily with our original spinout reform proposals, which we reiterated subsequently many times, including to the inquiry team.

During the review, it was heartening to hear that many institutions now agreed their policies weren’t working and volunteered to change them. We hope it marks the start of a more forward-thinking, founder-friendly era in UK universities.

We also believe that today marks a significant victory for openness. Change doesn’t come from grumbling behind closed doors or snark on private WhatsApp groups. It comes from people sharing their experiences openly with real data-driven evidence, and being loud when the status quo just isn’t good enough. That philosophy underlines our approach to every policy question.

While today’s recommendations mark a significant step forward, we will continue to monitor these issues closely. Universities will need to live up to both the spirit as well as the letter of reform. For example, we hope the number of technology transfer offices that push for a 25% stake in life sciences companies remains vanishingly small, even if they technically have this leeway for IP-rich spinouts with long gestation periods.

If universities get this right, it’ll mark real progress towards the government’s ambition of global science and technology leadership. A wave of UK research commercialisation will also serve as a magnet for talented researchers and entrepreneurs across Europe, who don’t want to contend with unreformed systems.

I’d like to thank the review team and the UK Government for their thorough and fair-minded work and for being prepared to recommend radical change. Most of all, I’d like to thank the entrepreneurs and researchers, who’ve helped push for change. Whether you went public about your story, shared your experience through the database, or in the case of one friend, sued your university - you all contributed to today’s victory.

While the UK held its AI Safety Summit, primarily focused on long-term risks, there was a wider “AI Fringe” held in parallel. This brought together academia, industry, investors, and civil society together to discuss more immediate questions around applications and regulation.

I was invited to speak on a panel at the AI and National Security Symposium, hosted by Jonathan Luff of Epsilon Advisory Partners and Kevin Allison of Minerva Technology Policy Advisors - two firms focused on helping fast-growing technology companies navigate government and geopolitics.

Our panel, featuring speakers from academia and the intelligence community, covered a range of topics, including the difficulty of translating research breakthroughs into national security applications, the risks new technology brings, and how we can reduce barriers to entry for start-ups for start-ups. I also covered some State of AI Report 2023 highlights that were relevant for defense and national security - I’ve included the matching slides at the end.

I’ve tidied up the notes I made ahead of the event, including some material we didn’t have time to cover, in case they’re of interest to others. It was great to see such a packed room for an issue that has historically been under-discussed in the AI community. If you’re interested in any of the below, please don’t hesitate to get in touch.

Defense AI ecosystem

• Despite expressions of interest from many VCs after Putin’s invasion of Ukraine last year, the actual cheques are being signed by the same small, but dedicated, group of long-standing defense investors.

• While we can point fingers at LP restrictions or ESG mandates, many of the objections to defense investing would likely fall away if the commercial opportunity was big enough. Unfortunately, the market for defense acquisition is broken.

• Innovators are the victims of procurement systems designed for exquisite, manned hardware platforms that rarely require updating once they roll off the factory floor. This is unsuited to the AI age, where technology is developed and updated at breakneck speed.

• The small group of primes that benefit from the existing system are specialists in traditional hardware and aren’t structurally or culturally suited to software development or attracting top AI talent.

• Due to a lack of political will and institutional capacity, governments have often shied away from wholesale reform, in favor of creating innovation units or schemes (e.g. the Defence and Security Accelerator in the UK). These have a poor track record of supporting new entrants in winning substantial work and routinely trap them in a perpetual cycle of grant applications.

War as a catalyst for action

• Both the war in Ukraine and the recent Hamas atrocities in Israel have shown that war isn’t a historic question or something only people in other parts of the world have to worry about.

• While it’s been inspiring to see the democratic world rally in support of the Ukrainians, the attempt to meet some of these commitments have showcased the dire state of the European defense-industrial base.

• For example, Germany has so far only succeeded in delivering 10% of the tanks it promised to Ukraine, some of which were rejected due to serious technical issues.

• This means we need to significantly accelerate the rate at which we explore the adoption of new technology and avoid throwing up unnecessary obstacles. It’s true that any AI system comes with risks and challenges, but it’s important not to fall into the trap of comparing imperfect machines to perfect humans, and to avoid “AI exceptionalism”. We accept human imperfection in the field and in analysis, just as we accept conventional equipment can go wrong.

• One of the advantages of the national security apparatus is that it exists in a state of exception and has greater freedom to explore new capabilities at pace. We should take full advantage of this.

State of AI - highlights for defense

2023 was, of course, the year of the large language model (LLM), and OpenAI crushed all before it. The potential of LLMs to support intelligence analysis is obvious. Jonathan has written an interesting Substack advancing the possibility of the Foreign and Commonwealth Office fine-tuning an LLM using its archives to yield new insights. We’ve already heard that defense ministries around the world are exploring the use of AI in supporting strategy formation.

With this in mind, it’s been striking to see the performance of models, with even relatively small training datasets, on strategy-based tasks.

We’ve also seen striking advances in computer vision, with DINOv2 demonstrating the potential of models that haven’t been trained on manually labeled data to perform well on classification and segmentation tasks.

It wouldn’t be a discussion about defense without looking at drones and the progress of systems trained with model-free deep reinforcement learning in simulation and operated using just on-board compute and sensors is striking - especially as war takes place in an increasingly electromagnetically contested environment.

Although drones are the most high profile use of technology in Ukraine, they have been far from the only one. The Ukrainians have innovated to defend their country by reducing procurement times for privately built technology by 5x and by increasing capped profit margins on government contracts for private vendors.

The report isn’t all good news, however. Firstly, there is the very real possibility of AI being misused.

Leaps forward in progress, have unsurprisingly been associated with leaps forward in geopolitical competition.

And there are serious questions about democratic governments’ ability to get new technology into the hands of those on the frontline who need it most.

Introduction

The state is back. In recent years, calls for the government to play a more active role in the development of technology have grown and, perhaps surprisingly, industry has been leading the charge. There have been calls for greater investment in everything from semiconductors, quantum computing through to space and cloud infrastructure capacity.

This chorus reaches a crescendo when it comes to AI. As a force multiplier for intelligence, AI has the power to transform the economy. It also has clear geopolitical relevance amidst an intensifying arms race between the US and China. At the same time, as competition for talent and GPUs mount, the bar to entry is progressively being raised.

There is a long history of failed government technology investment, with projects that were fundamentally flawed in their rationale, design, and execution. This is particularly obvious in the various pushes for European technological sovereignty we’ve seen in recent decades.

At Air Street, we aren’t hostile to the idea of the government partnering with industry or investing in technology. We’ve even called for it ourselves in specific contexts. Our vision of European Dynamism is grounded in government and early-stage companies working together to tackle critical strategic challenges. At the same time, money should not be frittered away on low-impact grants or on vanity-driven big projects or institutes.

In this piece, we examine a range of examples, draw out the lessons, and identify where we believe that history is already beginning to repeat itself.

Confused rationale - “we need a European X”

Many major technology projects come unstuck when faced with the simple question of ‘why?’. This often turns out to be the misguided political belief that the absence of a major European player is proof of market failure.

The archetypal example is the late Quaero, a Franco-German attempt to build a competitor to Google. The project brought together a consortium of technology companies and public research institutes, with the aim of creating a multimedia search engine - allowing text, image, or video-based search. It would also allow people to search across a range of cultural works.

In 2005, President Jacques Chirac, the project’s main advocate, argued that Quaero was necessary as “we must take up the global challenge of the American giants Yahoo! and Google” and “defend the world's cultural diversity against the looming threat of uniformity … Our power is at stake.” Chirac then drew parallels with the “magnificent success of Airbus”. The exact threat Yahoo represented to cultural diversity or European power remained unspoken. Likely for good reason.

Similarly, the ongoing saga of Gaia-X is a case study in how a shaky strategic rationale can give rise to ill-conceived projects.

Gaia-X is an EU attempt to build a federated data infrastructure that will act as a competitor to US cloud companies. This infrastructure would be grounded in “European values”, giving people more control over their personal data. To stress the importance of sovereignty, the project’s supporters point to the US Cloud Act, which compels US cloud businesses to hand over data on European users to law enforcement when presented with a warrant.

These arguments ignore the prevalence of European cloud services, which already offer the option to store their data within European member states. They similarly overlook converging norms and regulation around data and privacy. As the European Centre for International Political Economy has pointed out, invoking the specter of snooping by US law enforcement is also disingenuous, and does not account for extant European safeguards or existing plans to enable smoother international transfers of data to support criminal investigations.

Unrealistic resource commitments - expense with lack of results

Technological innovation doesn’t come cheap and frequently requires sums of money that governments understandably feel unable to commit. This results in a funding death-zone, where projects receive enough funding to seem expensive, but rarely enough to ensure successful delivery.

Quaero, for example, received approximately €400 million in government and industry backing. This is a significant amount of money, but obviously wholly inadequate for the task of building a challenger to one of the most successful technology companies of all time.

In recent years, Europe has made a dramatic push on semiconductor sovereignty, pledging €43 billion of investment as part of the European Chips Act, in response to the US CHIPS Act. The plan intends to double the EU’s global market share in semiconductors from 10% to 20% by 2030.

This may sound like an ambitious project, but it actually only features €3.3 billion of direct investment - the vast majority of which has been diverted away from other research and development work. Considering the multi-billion dollar costs associated with starting individual semiconductor fabs, it’s implausible that this kind of funding will make a difference at a pan-European level. A 20% market share goal would likely require hundreds of billions of Euros to have even a prospect of success, according to industry insiders.

With both Quaero, the Chips Act, and many other European projects, it’s possible to argue that industry is paying for at least a proportion of the work, which means that taxpayers aren’t losing out. However, government attempts to spur business to invest in bad ideas distort existing markets, they lead to potentially productive investment being channeled into unproductive ends.

Government, industry, strategy - words that should be combined with caution

Even if there were clear rationales or adequate budgets in place for many of these programmes, it remains doubtful whether or not they could have been delivered successfully.

A combination of risk aversion and political pressure means that European innovation efforts are often divided up across a range of countries, private companies, and publicly-funded research institutions. Far from de-risking a project, this multiplies the potential points of failure.

In the case of Quaero, this meant that within a year, the French and the German teams began to clash. The French wanted to focus on the multimedia element, the Germans on a text search engine. This led to the teams parting ways and the German raising additional $165 million in funding for their own project, which was also later discontinued.

Gaia-X similarly brought together an unwieldy coalition of parties, with conflicting interests and objectives. Reports suggest fights between member states over control of the board, a proliferation of working groups, and disagreements about how much of a role non-European companies should be allowed to play in the project.

Once technology projects become a way of channeling money to incumbent businesses or vehicles for job creation, they inevitably become detached from their real purpose, which ought to be delivering an important or missing capability.

Returning to semiconductors, Chris Miller, a world expert on semiconductor competition, believes that the supply chain focus neglects how shortages stem from bad inventory management by European manufacturers, rather than genuine supply problems. Rather than fueling an international subsidy arms race that it has no chance of winning, Europe should be proud of how it already controls critical industry nodes, namely specialized chemicals and machine tools.

Red lights on the dashboard - UK science and technology policy

The AI age is fertile territory for incoherent and inefficient innovation projects. We’ve been closely following UK science and technology policy in recent years and are worried that the same traits are visible - even if the sums of money involved are smaller.

This manifests itself in sprawling lists of commitments and strategies that rarely channel enough funding to make a meaningful difference. This was at the heart of the House of Lords’ Science and Technology Committee’s 2022 criticism of the “profusion of sectoral strategies in areas such as artificial intelligence and life sciences”, which mean that delivery bodies “are being pulled in multiple directions, with insufficient resources to meet the demand”.

We see this in bigger commitments, like the UK’s £1 billion semiconductor strategy, which falls into the European trap of being simultaneously expensive but unimpactful. At the cheaper end of the spectrum, we saw a seemingly random £13 million commitment “to transform healthcare research” through a hodgepodge of research grants and surgical robots, unveiled at the same time of the UK AI Safety Summit.

Along with these kinds of random grants, the UK has also highlighted the danger of investing money in institutions without adequate accountability. In 2022, the Alan Turing Institute had a total income of £51 million (including £35 million of public money), but it would be hard-pressed to demonstrate that it has delivered anything approaching value for money. The UK Government has now moved to create an AI Safety Institute, which we will be monitoring with interest. The ATI experience shows that simple investment, without proper oversight or a clear political imperative, leads to stagnation.

While this is not the Safety Institute’s fault, we were dismayed to read that the Government would only be able to produce an “interim” State of Science report ahead of the follow-up AI Summit hosted by South Korea (already downgraded to a virtual event). Considering the speed at which the field is evolving, this is an unserious timescale and demonstrates the peril of governments being unable to adapt their usual ways of working to a new environment. Hopefully the Safety Institute will be given the freedom to operate more nimbly and ambitiously, so it does not risk becoming ATI mark II.

That’s not to say there’s been no progress. We were happy the government ignored political pressure and allowed Britishvolt to fail, rather than freeing up £100 million to be burned on a project with no prospect of success. Similarly, we’re glad that the prime minister did not jump on the Brit-GPT bandwagon. The idea of a “sovereign LLM” strikes us as a more expensive chatbot equivalent of Quaero.

This kind of discipline needs to be preserved and extended intelligently. It means clearing the barnacles off the boat by not throwing money at every potentially worthy cause or allowing the innovation agenda to be captured by the lobby of trade associations, research councils, and other insiders.

Government innovation - a checklist

As we stated at the outset, we are by no means opposed to all government investment or the concept of sovereign capability in technology. We believe there are some challenges that either private markets are unlikely to solve by themselves or are of sufficient strategic importance that we can’t just hope it will happen.

There are several key tests that policymakers should bear in mind when approaching these decisions:

• Relevance: is there a clear reason why the government needs to act here? Why won’t the private sector or research institutions be able to handle this?

• Novelty: is this investment creating a new capability or replicating one that already exists within another friendly power that we are able to access?

• “Insurance”: if this capability does exist elsewhere, how plausible is the scenario in which access is cut off? Would this investment meaningfully mitigate against this scenario? Were we to be cut off, what safeguards are already in place?

• Plausibility: is meaningful sovereign capability actually possible? Would the final “sovereign” capability actually be dependent on the very relationships we’re trying to de-risk?

• Commitment: are we actually able to spend the amount of money required to make this work in a meaningful way?

To illustrate this, we propose two sectors where the case is compelling.

Defense innovation is by definition dependent on the government. The Russian invasion of Ukraine has demonstrated the shifting nature of warfare and that European governments need to act. We also know that incumbent defense suppliers who specialize in creating exquisitely complicated manned hardware platforms are ill-placed to innovate in AI and software. Years of defence underfunding has also left the European defense-industrial base lacking in resilience. We’ve also seen warnings from US officials that there is a possible future where the US military may be too tied down in the South China Sea to adequately defend Europe’s eastern border.

While the US has been one of the main suppliers of traditional military aid to the Ukrainians, US companies are far from cornering the market in technology like drones. The amount of money needed to effectively partner with earlier-stage defense companies is also a fraction of the amount of money spent on hardware projects that frequently underperform or run billions over budget.

Another case is compute power for research and early-stage businesses. Considering very real capacity constraints and spiraling GPU costs, there is room here for the government to play a useful role. Despite being essential for cutting-edge AI research, the UK has fewer than 1,000 NVIDIA A100 GPUs available to researchers. This is less than a single research team at Meta. The UK has already recognised the problem, but has not currently attacked the problem with sufficient ambition. The 3,000 GPU cluster, with delivery by 2026, proposed in the UK’s Independent Review of the Future of Compute is inadequate for the task.

Closing thoughts

We will likely return to the role of government in the development of technology in future, as we continue to explore the foundations of European Dynamism. This will likely be more in education and removing obstacles to investment, rather than innovating in-house. After all, most successful technology companies were set up by software engineers or others with deep expertise. Jacques Chirac was no Bill Gates.

Europe has incredible potential, considering its political stability, world-leading universities, and existing talent. But it cannot and should not throw centuries-old economic truths out of the window to conjure up a simulacrum of technological sovereignty. Superpower status will not come via death by a thousand grants or attempting to recreate Silicon Valley in Strasbourg. It will come from what markets do best: discipline, picking battles wisely, and allowing bad projects to fail.

Update 2 Feb 2024: The House of Lords Communications and Digital Committee have published their report calling for Government to reject the safety panic + regulatory capture and focus on urgently upgrading the UK’s capabilities. Twitter thread + full report.

AI and the Future of Britain

The House of Lords Communications and Digital Committee is currently holding an inquiry into Large Language Models (LLMs) and the steps governments, businesses, and regulators need to take over the next 1-3 years to maximize the opportunities while minimizing the risks.

Alongside Peter Waggett (IBM), Zoe Webster (BT), and Francesco Marconi (Applied XL), I was invited to give evidence on the potential benefits of LLMs to the UK economy and the barriers to investment.

I’ve tidied up the notes I prepared ahead of time, grouped under the main themes of the session. There’s also some extra material in there that we didn’t have time to cover. You can watch the session in full here.

Discussion areas

Where is the LLM opportunity in the UK?

• The UK AI startups that have raised the most money have reflected either areas of the UK’s innate advantage or the effects of government policy. That’s why we have such high concentrations of investment in cybersecurity, the life sciences, and fintech.

• There are opportunities for the use of generative AI in all of these. I’m personally most excited by the life sciences - both in terms of the positive impact on humanity and the tie in with the UK’s strong research base.

• I’ve already come across businesses in the US that are using generative AI to support protein design, which could support the development of new therapeutics.

• There’s also the possibility of models being used to analyze chemical libraries to help us generate novel molecular structures that bind to certain proteins - making drug discovery faster and cheaper.

• Much of the expertise required to do this work is currently trapped in our universities, which is why I’m glad the UK government is pursuing its review of spinout policy.

• At the same time, we should avoid becoming too self-congratulatory. The UK is strong by European standards, but lagging global leaders:

  1. Between 2019-2023, the Bay Area alone saw $6B invested in Generative AI (excluding OpenAI), while London saw $365M. The vast majority of this fundraising occurred in the last 1-2 years.

  2. In the same time, $7.4B was invested in AI chips in China, $2.9B in the US, while all of Europe invested $446.7M combined.

What are the main barriers to investment/business uptake?

1. Enterprise adoption

• These are largely grouped around skills, privacy, and the need to finetune off-the-shelf models.

• You may want to train your model to respond to customer requests or mark-up documents in a specific way, or provide it with confidential or proprietary information that’s not in its training data, but want to keep it ring-fenced within your organization. This requires time and a degree of technical expertise.

• For example, in recent weeks, we’ve seen OpenAI introduce ChatGPT Enterprise, which incorporates enterprise-grade privacy: conversations are encrypted and the models don’t learn from usage.

• OpenAI has also partnered with Scale AI to provide fine-tuning support for businesses using GPT-3.5, with the intention of extending this to GPT-4.

• Over the years, I expect we’ll see the emergence of a healthy market for fine-tuning support.

2. Investment

• Science and tech leadership ultimately doesn’t come cheap.

• Government subsidy isn’t a long-term viable alternative to private investment - the UK is not short of accelerators and support schemes.

• Unfortunately, it feels like the government has been doing everything it can to deter investment in recent years, with international confidence in the UK hitting historic lows in recent years.

• To change this, we need three different things:

  1. Political stability: as governments have changed in recent years, policy and priorities have varied significantly. Consistency, both on science and tech specifically, but economic policy more generally is essential.

  2. A welcoming immigration system: both in terms of actual policy and also tone. Talent needs to know that it will be able to move to this country, be made to feel welcome here, and not undergo a process that treats it with suspicion.

  3. Less hostility to the tech sector: proposals that would break end-to-end encryption in legislation like the Online Safety Bill, or amendments to the Investigatory Powers Act that would give the Home Office oversight over smartphone upgrades, undermine international attempts to paint the UK as a serious technology power.

Regulatory intervention

Note: read our essay in full here.

• At Air Street, we’ve previously expressed our support for the UK’s approach to AI regulation.

• We believe that AI is a general purpose technology, which means that AI risks will be context-dependent. As a result, we think the people who understand that context will be best-placed to identify and respond to the risks.

• Regulators like the ICO or the MHRA have existed for decades and navigated technological change. We would recommend helping them build out the expertise and capacity to respond to new challenges, rather than spinning up a series of new regulatory frameworks, many of which may be out-of-date quickly.

• We saw this with the EU having to rewrite its AI Act at the last minute to incorporate foundation models. Static regulation is a bad way to respond to rapid change.

• It’s important to avoid AI exceptionalism - we don’t normally heavily regulate things on the basis of hypothetical future risks, AI shouldn’t be any different.

• Premature regulation on the basis of safety fears is also bad for competition - that’s why big companies are happy to advocate for policies like licensing regimes, as they know they’ll hamper open source model providers.

Open vs closed source

The move away from open source

• I’m a great believer in open source and don’t believe that any of the frontier models that are likely to be open sourced in the near future are likely to be powerful enough to pose a serious safety risk.

• We’ve seen a relatively rapid move away from open source among the big labs (with Meta standing out as the obvious exception). While this has been framed as safety-motivated, I suspect commercial considerations are likely the main driver.

• Companies like OpenAI have invested huge sums of money in developing their technology and are now interested in commercializing it, rather than handing the instruction manual over to other people.

How can the government facilitate an open source ecosystem?

• The biggest single barrier to smaller players is access to compute power. The government has outlined plans to build out public cloud capacity, but the current plans aren’t nearly ambitious enough. We currently have fewer than 1,000 GPUs available to researchers.

• The Future of Compute Review recommended a 3,000 GPU target when some corporate labs already have 10x this capacity.

  1. By contrast, Anthropic has suggested the US should invest $4 billion over three years to build a 100,000 GPU cluster.

• Another obstacle is access to a high enough volume of training data. We could consider creating a national data bank.

  1. It could bring together data from the BBC, government departments, our universities, and other sources for values-aligned UK companies looking to build LLMs.

  2. But here, infrastructure investments are important - recall that DeepMind tried to revolutionize the NHS with AI and we ended up several years later with a task management app for clinicians.

• At the same time, we need to be clear-eyed about what’s possible - unless we see a sudden unexpected drop in compute costs or the emergence of a new less compute-intensive paradigm in AI research, the levers the government has at its disposal are only likely to make a difference around the margins.

• It’s essentially inevitable that only a few companies will create the most powerful models - in the same way that two companies design the operating systems used on the vast majority of the world’s computers. The competition will likely be much livelier in other parts of the value chain.

Liability

• Liability is unlikely to sit in one part of the value chain. It is obviously the responsibility of the developer to ensure that a model is trained on representative and ethically sourced data, that it’s resistant to adversarial attacks, and that it’s been subjected to appropriate auditing and testing before being released.

• At the same time, it is unreasonable to hold developers responsible for every downstream use of their system. If an operator finetunes a system badly or a user deliberately acts with malice - there’s only so much developers can do to prevent this from happening.

• There’s no other industry in the world that assigns all responsibility to the original creator of a piece of technology. If we don’t accept a degree of risk and personal responsibility, technological progress will stall entirely.

Introduction

Ahead of the UK’s AI Safety Summit, we’ve seen speculation about the guestlist mount. Initially, the UK suggested that only “like-minded” governments would be invited to attend, but it appears to have backed away from this position. It was reported last week that, over the objections of the EU, US, and Japan, China will likely be present “in some capacity”, potentially on the sidelines of the conference.

Before this apparent change of heart, there had been a coalition spanning the AI research and political worlds calling for China to be invited. For example, Tobias Ellwood, chair of the Commons Defence Select Committee, drew parallels with the international regulation of nuclear power, arguing that: “If we don’t have total buy-in from the start the dangers of AI over humans is given space to develop and any threat won’t be contained by geographical borders.”

In the AI community, Huw Roberts, a DPhil student at the Oxford Internet Institute wrote a letter to the Financial Times making a similar case around risk transcending borders. Roberts also argued that as China was ahead of the UK and US on AI regulation, its experience “would be invaluable for informing well-designed policy at the UK’s AI summit”. A number of participants at a recent workshop held by the Centre for the Governance of AI warned that the summit might be the only opportunity to involve China in governance discussions and that China was more likely to engage with frameworks it felt it had helped to create.

In our view, these calls for China’s inclusion at the summit are misguided for three key reasons:

• China is pursuing an approach to AI regulation that is motivated less by a sincere commitment to safety and more by political control;

• China has a long history of attempting to subvert multilateral institutions and technical standards;

• ‘Global governance’ is currently a pipe dream, so democratic nations should focus on reaching agreements among themselves.

Preserving the social order: China and AI regulation

From the moment the 2016 Lee Sedol-AlphaGo challenge match placed AI well and truly on the radar of the Chinese government, it has viewed AI through an intensely political lens. The People’s Liberation Army quickly held seminars to discuss its significance, and China raced ahead of the world in setting national strategies and passing AI regulation. We saw high-level governance regulations appearing as early as 2017, before a series of specific regulations appeared over 2020-2021 around online algorithms.

According to Matt Sheehan, a China specialist at the Carnegie Endowment for International Peace, this early focus on algorithms was not a coincidence. Sheehan argues that: “The first, overriding goal is to shape the technology so it serves the CCP’s [Chinese Communist Party] agenda, particularly for information control, and following from this, political and social stability.”

This prioritisation of ‘stability’ above all else is visible on even a superficial reading. For example, the 2022 Deep Synthesis Regulation is replete with references to “correct political direction”, “social morals”, the need to “accept social supervision”, and a prohibition on sharing “information inciting subversion of State power”. The latest regulations on generative AI drive home how content “shall reflect the Socialist Core Values, and may not contain: subversion of state power, overturning of the socialist system … as well as content that may upset economic order or social order”. In the same obvious category go the prohibitions on ‘false information’ and the requirements for people signing up to AI content generation platforms to supply their real names.

Even if there might be individual privacy protections, clauses around liability, or other individual points western commentators like individually, it’s impossible to separate these from the wider philosophy behind the regulation. We saw the same process unfold with China’s 2021-2023 “Big Tech'' clampdown (which also attracted western admirers), where the ‘public interest’ was used as a cover for a power grab. As data and privacy expert Jamie Susskind has observed: “When we start being starry-eyed about the Chinese model of enforcement, we've lost track of the fact that regulation isn't just supposed to rein in private companies, it's also supposed to limit the power of the state.”

Even if we put all these concerns to the side and accept that the theory behind Chinese regulation could be interesting, we turn to the question of its implementation.

Authoritarian states have a long history of managing risk dismally. Bill Drexel and Hannah Kelley at the Centre for a New American Security have documented the CCP’s “disaster amnesia”, which results in the government burying bad news, suppressing death tolls from accidents, and rarely learning from mistakes. The suppression of early warning signs from doctors about Covid-19 combined with a disinformation campaign about its origins wasn’t an isolated example. It was preceded by a four-month cover-up of the 2002 SARS outbreak and multi-year suppression of reports of HIV-contaminated blood transfusions in the 1990s.

Considering the well-documented use of AI by the authorities in Xinjiang to enforce and deepen its repression of the Uighur population, it’s clear the CCP has no interest in abiding by the strictures on AI use it sets for the private sector.

China may well adopt the form of AI regulation, but magic words about responsible AI do nothing to bring the substance into being and we should seriously question whether they would have any interest in acting as an honest partner on AI safety.

“Cyber sovereignty”: China and international governance

China’s domestic attitude to technology is indicative of its international approach, with the government happy to violate agreements and attempt to subvert multilateral bodies. Were China to be a participant in discussions around global AI governance, we have no reason to believe that it would either comply with their outcome and every reason to believe that they would attempt to shape them in an authoritarian direction.

The recent past is littered with precedent. We see this, for example, in the World Trade Organisation, where China is still in clear violation of the many open market commitments it made as a condition of membership in 2001. There have also been more subtle efforts in less well-known international bodies. Perhaps the most striking is China’s multi-year effort to change internet standards. China has attempted to move internet standards away from multistakeholder bodies to the purview of the UN’s International Telecommunication Union (ITU), where only member states can participate in negotiations.

China unsuccessfully pushed for a former Huawei executive to be installed as the ITU’s Secretary General in 2022 and has used Huawei to advocate for a new internet protocol that would radically centralise the internet. The New IP, as it’s dubbed, would allow network operators to see the content of any information being shared, as well as identify the sender and receiver. Network operators would also gain the power to block delivery. Unsurprisingly, Russia, Iran, and Saudi Arabia are its biggest international champions.

China also attempts strong-arm tactics in the ITU’s daily operations. This includes forcing company delegates participating in study groups to take their phones into the voting booth to prove they’re voted ‘correctly’ or instructing them to deliberately obstruct the work of groups until they agree to Chinese proposals (e.g. on 5G standards).

Fortunately, many of these efforts have not been successful so far as democracies have rallied to block them, but we should not underestimate China’s determination to export its model of governance. We’ve seen it forge close ties with repressive East African states, providing money and expertise to governments looking to mimic its censorship of the internet and social media. It also established the World Internet Conference, in partnership with Russia and other authoritarian states, in an effort to legitimise its model of “cyber sovereignty”. Giving a flavour of the proceedings, the draft communiqué for the inaugural conference was slipped under delegates’ hotel room doors after midnight and anyone with suggested changes was given until 8am the next morning to supply feedback.

China’s philosophy of international governance is to take every possible opportunity, however clumsily, to legitimise and export its model of domestic repression - it would be naive to expect anything approaching good faith engagement in AI governance.

“Global governance”: frameworks or fan fiction?

This same naivety underpins many of the proposals for global AI governance. Based on the evidence we’ve seen, there’s little reason to believe that any kind of substantive global architecture is possible, or necessarily desirable at the current time, despite the growing range of frameworks emerging from researchers and entrepreneurs.

At a basic level, there seems to be little agreement about what we should be attempting to govern. Are we aiming to prevent catastrophic risk, seeking to establish global standards around equitable and sustainable AI use, or broaden access to AI? Or all of the above? These are all different remits that require different approaches.

The debate is further muddied by an alphabet soup of acronyms, with people variously reaching for the International Atomic Energy Authority, CERN, the Intergovernmental Panel on Climate Change, and others as inspiration. Beyond a conceptual fuzziness, the practicalities of then implementing any framework is usually treated as an after-thought.

To take a recent example, Mustafa Suleyman (CEO and Co-Founder of Inflection AI) and Ian Bremmer (President of the Eurasia Group) co-wrote an essay for Foreign Affairs outlining their views on governance. They propose a “technoprudential mandate”, which is designed to “address the various aspects of AI that could threaten geopolitical stability”.

However, ‘geopolitical stability’ is a slippery concept and is used to justify giving this hypothetical regime a seemingly limitless mandate, including: overseeing the entire AI value chain; improving US-China relations; governing open source AI (via online censorship if necessary); fighting disinformation and privacy violations; and convening ‘civil society’ and the tech sector. Bremmer and Suleyman lay out no path to how the necessary institutions would come into being, beyond a throwaway acknowledgement that “none of these solutions will be easy to implement”. That’s to say nothing of the desirability of placing China at the heart of this Leviathan.

Perhaps the most rigorous attempt to explore different models comes in the form of a recent paper from Lewis Ho, a researcher at Google DeepMind, written with collaborators from OpenAI and a range of leading universities. Lewis and his co-authors essentially accept that different priorities require different institutions, and each has its costs, implicitly suggesting political choice is required. They propose four options and are upfront about the scoping challenges and issues around incentivising international participation. As a result, they stop short of endorsing any one of these approaches and acknowledge that while more international cooperation is needed, we are not ready to commit to a model.

Considering this total lack of agreement on the right starting point for governance among even AI experts in democratic nations, inviting a motivated adversary with a clear philosophy seems reckless. In this context, a forum of “like-minded” nations is exactly what’s needed.

Unlearning the lessons of the past?

As the UK Government eyes potential rapprochement with China, it’s vital we don’t ignore the flashing red lights on the dashboard. These include:

• Tens of millions of pounds flowing from Chinese state institutions into UK universities;

• UK universities and research institutes unintentionally hiring staff from Chinese defence conglomerates and conducting research sponsored by Chinese ICBM manufacturers;

• UK universities self-censoring on subjects like Tibet or the treatment of the Uighurs;

• The People’s Liberation Army hiring ex-Royal Air Force pilots to train its military;

• Widespread Chinese control of the UK’s civil nuclear sector.

Last year, Parliament’s Intelligence and Security Committee painted a damning picture of how “China’s size, ambition and capability have enabled it to successfully penetrate every sector of the UK’s economy, and - until the Covid-19 pandemic, Chinese money was readily accepted by HMG with few questions asked”. Few in the British political class like to remember the official discussion of the “golden era” in UK-China relations as recently as 2015.

While some in the tech sector may not like the rhetoric of an ‘arms race’, renaming it doesn’t make it go away. As we’ve seen in the case of the ITU, when democracies fight back against interference, they can be successful. But when short-term economic gain or idealistic hopes of the global community win out, we find ourselves in much more treacherous waters. If democracies delegate the nascent and confused AI governance debate to an amorphous global forum, there is a high risk that it will be hijacked. We may well end up discovering that the road to “AI sovereignty” is paved with good intentions.

Introduction

This autumn, the UK will convene governments and technology experts for a global summit on AI safety. While the exact agenda and guestlist is still to be revealed, the government clearly hopes that it will position the UK at the centre of the international debate on AI governance.

The event is likely to bring renewed scrutiny to the UK’s “pro-innovation” approach to AI regulation. Since the UK first published its white paper, the tone of the debate has shifted markedly. We’ve seen letters signed by AI luminaries warning of a potential “extinction-level” event, while other parts of the world have pursued stricter approaches to short-term regulation. China’s national AI law will resemble elements of the EU’s sweeping AI Act, while US states are developing a patchwork of uneven, but often tough, regulations.

With the conversation becoming more feverish, the UK government has slowly backed down from the framing, if not the substance, of its approach. The idea of the summit was born in this period. In May, events took a turn for the surreal, when the AI Minister wrote into The Economist to dispute their characterisation of the UK’s approach as “light-touch” - a term they’d in fact borrowed from the government.

In the past, Air Street hasn’t shied away from challenging AI policy when we think it’s been wide-of-the-mark or lacking in ambition. On this occasion, however, we believe that the UK’s proposed approach is correct, particularly given current technical capabilities and considerable future uncertainty. We are concerned that the present media and political conversation, along with heavy lobbying, may trigger an undesirable change of course.

We realise that there is sincere and often passionate disagreement in the AI community on a range of safety questions. We firmly believe that AI should be developed responsibly, robustly, and safely. Two years ago, the State of AI Report was pointing to the big labs’ small safety teams and limited investment in alignment research. While we are not looking to relitigate the academic AI safety debate, it should come as no surprise that as a firm that invests in AI-first companies, we are sceptical of the most pessimistic narratives. This perspective shapes the argument that follows.

In this essay, we’ll lay out the strengths of the UK’s approach to AI regulation, the danger of the alternatives, and some initial thoughts on where the conversation needs to go next.

What the UK gets right

The point of regulation is not to eliminate risk by reducing the probability of adverse outcomes to zero. Regulation starts from the basis that legal goods or services have an intended use. Enlightened regulation understands that this intended use could be good for an individual or wider society, allowing the upside to be maximised while setting proportionate standards. It accepts that restrictions can come with trade-offs, and there is a balance we need to get right.

We routinely allow or only lightly control a wide range of potentially dangerous items that we use in our everyday lives. We accept that they are essential for a range of human tasks, the number of bad actors is usually small, and that most people exercise personal responsibility in their use.

The UK has applied this cool-headed and proportionate approach to AI regulation, preparing regulators for potential future action without prematurely driving up compliance costs for industry. This incremental approach has also prevented the UK from overcommitting to initiatives that would’ve been rendered obsolete by future technical developments, like Finland’s national AI course.

While there is widespread agreement that AI is a general purpose technology, the UK is unusual in acting as if this assumption is actually true. In practice, this means accepting i) that AI risks depend largely on context, ii) that some contexts are significantly higher risk than others, and iii) the people who understand those contexts will be best-placed to respond to the risks.

Instead of starting from a blank page, this approach acknowledges that sectoral regulators, many of which have existed for decades, have grappled with new technology and complex ethical issues before. The Medicines and Healthcare products Regulatory Agency (MHRA) knows how to set minimum standards for medical devices. Similarly, the Information Commissioner’s Office (ICO) has adapted to deal with individual privacy in the platform era and has investigated the use of facial recognition technology.

Both organisations have updated their standards and methodologies as the technology landscape has shifted - sometimes supported by legislation where necessary, but not preemptively. In their use of regulatory sandboxes, UK regulators have often been particularly forward-looking in allowing firms to explore how innovation runs up against existing frameworks. For example, the FCA’s sandbox has been particularly successful in supporting the emergence of a new generation of fintech companies.

The UK approach also accepts that sectoral regulators shouldn’t be abandoned in the face of new challenges. We see this in the planned creation of a central risk function to identify potential threats that might require action from the central government. Similarly, responding to the rapid progress in frontier models, the government has made a considerable £100 million initial investment behind the Foundation Model Taskforce, which has a clear remit around safety.

By treating AI as a general purpose technology and harnessing existing subject matter expertise and legislation, the UK’s approach is a sensible way of balancing safety and innovation. Instead of treating AI as something society needs to be shielded from, it operates on the basis that we shouldn’t squander future economic growth or leaps forward in our scientific knowledge. It may be that as capabilities develop further, we might need to revisit or tighten regulation. Nothing the government has announced so far prevents this from happening if need be. Its critics, however, disagree, and approach AI as an entirely novel challenge.

If all you have is a hammer…

Underlying many of the arguments that the UK regulation is “falling behind” is the assumption we need to “keep up”, in other words, that more regulation is better. While critics of the technology sector rightly highlight ethical lapses and failures, they spend significantly less time thinking through the potential adverse effects or the challenges of implementing their own proposed interventions.

Regulation is assumed to be a cost-free action implemented by entirely rational actors. As the political philosopher Chris Freiman argues: “Perfect states beat imperfect markets, but that doesn’t establish the superiority of state solutions any more than finding that omnivorous non-smokers have lower rates of cancer than vegan smokers establishes the superiority of an omnivorous diet. We should compare like-to-like.” The downside of some of these implementations is outlined in more detail below.

The call to ‘keep up’ is often mashed together with one or two other forms of argument.

One is simply rebranding long-standing critiques of the tech sector as “AI harms” that merit AI regulation. It’s perfectly legitimate to believe that social media platforms don’t do enough to fight disinformation or that we should extend medical device regulation to cover wellness apps. These, however, are arguments for regulating app stores or social media platforms, not for introducing an entirely new regulatory framework for AI.

The other approach is an “AI exceptionalism” that simply takes it as read that anything AI-related should be regulated in a fundamentally different way.

For example, a number of critics have suggested that not regulating more is tantamount to deregulation, which as far as we can tell, is a novel standard.

Similarly, we see the spectre of ‘gaps’ in sectoral regulation being raised. Regulation has always evolved as new challenges or issues that we couldn’t predict have arisen - there’s no reason why this wouldn’t be the case with AI. It’s also why we are not sold on the need for measures like an “AI ombudsman” or additional statements of AI-specific rights. With the advent of personal computing, the UK didn’t create a Government Office for the PC or pass a precautionary General Computing Regulation. Instead, we expanded and flexed existing consumer standards, and introduced legislation (e.g. the Computer Misuse Act) in response to specific real-world challenges when they arose.

The UK government’s existing framework (visualised below) even contains specific provisions to monitor for new risks. You may worry that this monitoring operation won’t be well-resourced or have a clear enough remit, but these concerns would presumably apply equally to a hypothetical specialist AI regulator or ombudsman.

Source: AI White Paper

The danger of a changing course

You may wonder why we feel the need to respond to these criticisms of the UK’s approach far more stridently than the government has to date. This is because we believe that abandoning this approach so that we can be seen to “keep up” could have serious long-term consequences.

We’ve seen a number of warning signs in the real-world about what happens when regulation goes wrong - both in the UK and internationally.

When you’re draining the swamp, you don’t ask the frogs for an objective assessment

The UK’s missing infrastructure, housing shortages, and lack of lab space can all be tied back to a broken planning system. Sam Dumitriu has documented how a combination of incredibly burdensome assessment processes (running to tens of thousands of pages), along with an intricate web of ministerial policy statements that are out-of-sync with legislation, have produced an impenetrable bureaucracy. Added to this, stakeholder groups and lobbyists have endless opportunities to jam up the process through consultations and legal challenges.

You could argue that the construction of roads has nothing to do with foundation models, but unfortunately, we see many of the same traits seeping into more established UK tech regulation.

The powers being requested by the Digital Markets Unit that sits within the Competition and Markets Authority are a warning sign of what the future could hold. If granted, they would give a UK regulator unprecedented reach. These include the abolition of the right of appeal against decisions; routine regulatory intervention in product decisions; the right for the regulator to alter remedies up to ten years after imposing them; and mandatory arbitration. The last point is particularly ripe for abuse. It could easily result in a world where lobbying leads to newer industries being made to give ground to rent-seeking incumbents (e.g. digital platforms in favour of traditional news platforms), with the consumer an after-thought.

Considering the UK’s track record in large-scale, top-down regulation - it’s difficult to see such an approach playing out differently when approaching a vastly more complicated field in the midst of rapid evolution.

There’s no such thing as a regulatory superpower

The EU’s AI Act has been examined and debated exhaustively and we do not propose to rehash these arguments in full. There are, however, a few specific points that drive home the advantages of the UK’s more flexible approach.

Firstly, let’s focus on the difficulty that top-down systems face in adapting to fast-changing fields. The original text of the Act was prepared over 2019-20, before the explosion of interest in foundation models. As a result, they were entirely absent from the original draft. The EU’s protracted regulatory process meant there was time to incorporate them at the last minute. They are unlikely to be so lucky with future breakthroughs. These amendments, however, were done in a slapdash way. As a result, the EU’s consistent ‘risk-based’ philosophy is now unevenly overridden by specific rules for foundation models and general purpose AI systems (the technical distinction between the two not being clear, even to experts in the field). This means many providers of foundation models will struggle with the burden of being doubly-regulated.

Secondly, we have the challenges of enforcing this style of top-down legislation. In the Parliament’s AI Act proposal, member states will be required to designate one market surveillance authority, essentially forcing them to create dedicated AI regulators. This means that existing agencies, with appropriate subject matter expertise, in deeply complex areas like health, will not be responsible for interpreting or implementing the Act in their own domains. There is of course the possibility of poaching existing experts from other regulators to staff a new AI regulator, but this has its own knock-on effects; the supply of experts in medical devices, for example, is not limitless.

Considering the huge regulatory burden the EU will be taking on, these new regulatory bodies will need significant funding and expertise. With a small handful of exceptions, there is little sign that member states are rising to the challenge, with Alex Engler at the Brookings Institute arguing that the obvious lack of preparedness is “certainly a cause for concern”.

Thirdly, there is the risk of the Act further concentrating power. For example, a number of open-source foundation model providers are likely to be subject to the same weight of compliance as Big Tech. Considering the already small number of open-source models that have emerged from non-profit initiatives, this is a recipe for entrenching the dominance of a handful of big companies. Some of these requirements are so onerous that even Big Tech may struggle to meet them. As a recent paper from Google DeepMind’s Harry Law and Sébastien Krier notes, “some transparency requirements in the AI Act are either technically impossible to comply with, or very difficult to implement … [and] may not address the risk of harmful models diffusing”.

This is the natural endpoint when regulation is designed to address often nebulously-defined harms, without sufficient thought being given to wider market dynamics. We saw exactly the same misguided philosophy behind the EU’s General Data Protection Regulation (GDPR). Welcomed by Big Tech, its blunt, one-size-fits-all approach reduced competition and entrenched the power of the very companies it was targeting, while polluting the internet with cookie banners that make many websites unusable.

You may well think that these anti-competitive effects are a price worth paying, but supporters of the Act rarely acknowledge the existence of the trade-off. When a number of European companies signed an open letter warning about the consequences of the Act for European technological sovereignty, Dragoș Tudorache, an MEP leading on the Parliament’s draft dismissed this as an “aggressive lobby” that was undermining Europe’s “undeniable lead” on regulation.

While a number of countries have adopted data protection models inspired by GDPR, does anyone believe that this has made the EU richer or more powerful? Similarly, arguing that the AI Act has given the EU geopolitical sway over China due to overlaps with their planned national AI regulation would stretch credulity. Shaping innovation born in other parts of the world is not a meaningful substitute for scientific and technological breakthroughs of your own. As Emmanuel Macron has acknowledged, “the US has GAFA … we have GDPR”. Being a “regulatory superpower” is a poor consolation prize.

Refining and reinforcing the UK’s approach

The above is not to argue that the UK’s approach is perfect or that there’s no room for improvement.

For example, we believe there are genuine concerns around resourcing. While we don’t support giving regulators a blank cheque, we believe that sectoral regulators confronting AI-issues for the first time will need to be able to draw on dispassionate expert advice.

While technology companies, civil society organisations and other NGOs have their role to play, they also have their own agendas, and external stakeholders can’t act as a substitute for in-house expertise. Otherwise we risk recreating the “stakeholderist” nightmare we see in other UK regulation.

We can already see the most established companies trying to position themselves as trusted partners to governments. The Frontier Model Forum, unveiled in July of this year, brought together the biggest companies operating in the sector to shape best practice and engage with governments. Unfortunately for smaller businesses or those working on open source projects, the Forum appears to be a closed shop. Unless you “demonstrate commitment” to safety (as defined by the Forum) and have already deployed a frontier model (as defined by the Forum), you’re not invited.

Avoiding this kind of regulatory capture will undoubtedly require significant resourcing, as well as an overdue reappraisal of how the civil service compensates staff from technical background. The Advanced Research and Invention Agency (ARIA) has been rightly exempted from the service’s inflexible (and often stingy) payscale. Building real AI expertise in government will require similar creativity.

Beyond capacity, there are other issues where the UK may start having to draw clearer lines. We feel that the white paper is optimistic about the ability of sectoral regulators in their current form to grapple with liability. While this concern remains hypothetical for the moment, we think that it may become real sooner than the white paper’s authors have anticipated. A failure to take a consistent position across government risks storing these problems up for the future.

Closing thoughts

Imperfections and potential adaptations aside, we should use the UK’s AI safety summit as an opportunity to showcase the strength of our approach, not to panic about whether or not we are falling behind.

We can demonstrate that there’s nothing about our approach to short-term risks that stops us shaping early norms around longer-term international governance. This could include supporting the creation of an evaluation ecosystem, granting additional resources to open source projects, or normalising red-teaming.

If anything, our more flexible approach makes it easier to respond to these emerging norms. We can even point to the agreement UK Prime Minister Rishi Sunak secured from OpenAI, Google DeepMind, and Anthropic to give the government early access to their most powerful frontier models. He achieved this without having to pass a single sentence of legislation. Maybe not being a regulatory superpower comes with its upsides after all…

Introduction

Duncan Robinson, The Economist’s Bagehot columnist, recently identified the ‘Reform Fairy’ as the new spirit fluttering over British politics. Robinson described it as the “fundamentally unserious” belief that it is possible to “improve public services without spending political or financial capital”. The Reform Fairy’s closely-related cousin is the ‘AI Fairy’: the mythical creature that will apply the latest breakthroughs in cutting-edge research to decades-old public services through an as-yet undefined process.

AI policy across the political spectrum has been focused on either inputs or the hypothetical regulation of outputs. This means we have long lists of small government spending commitments for a range of startups, labs, and technologies and calls to license large language models like nuclear technology from the opposition. When it comes to identifying priorities, procuring technology, and overhauling existing practices, beyond variations of “embracing the opportunities presented by AI”, the picture becomes murky.

Opening the recent Future of Britain conference, former UK Prime Minister Tony Blair provided a stark outline of how the British state was operating on an unsustainable footing. He was followed by a range of speakers covering topics from health to geopolitics, who outlined the role technology could play in addressing today’s challenges. At Air Street, we believe that AI-first companies will be at the forefront of delivering that change and have the potential to remake the public sector in the same way that they are already rewiring the private sector.

We’ve seen helpful interventions around how to make this promise more tangible from the Tony Blair Institute, Onward and other organisations that we have supported and championed in the ecosystem. Nevertheless, the public and political conversation is still far from where we need it to be if we are to get technology out of the labs and into the hands of those on the frontline. In short, we risk squandering a generational opportunity.

AI isn’t an upgrade, we need to replace an operating system that’s broken at its core

Science and technology leadership requires us to overhaul the 19th century institutions that 21st century founders have to navigate. We can see this really clearly in two areas of policy.

The application of AI will one-day significantly reduce costs across the public sector - whether that’s replacing failing tank programmes with cheap drones or a preventative approach to healthcare that reduces the number of seriously and expensively ill patients. It is, however, unserious to believe that these benefits can be realised overnight.

To take the example of healthcare, we used to hear the EY valuation of NHS data at £9.6 billion bandied around UK policy circles as proof of the NHS’s world-beating longitudinal datasets. This headline figure didn’t account for the significant amount of cleaning work required to make this data remotely usable for AI workloads. The majority of primary care (GP) practices in the UK still use paper records and 100 million of them are still largely sitting in old-fashioned medical libraries in the middle of hospitals. Past attempts at a national integrated electronic health records system have ended in disaster. Anecdotally, we hear from AI researchers working on research with the NHS that the first months of any project are consumed by data permissioning, acquisition, evaluation and cleaning work.

Meanwhile, day-to-day provision is little better. When DeepMind first started working with the NHS in 2015, they quickly pivoted away from their initial plan to focus on AI research. Why? The urgency of remedying a situation of staff relying on antiquated shared computers, pagers, and fax machines to communicate with each other was far greater than advanced diagnostic AI. We still see regular news stories of staff being found sharing patient data over WhatsApp, due to the inadequacy of approved communications systems.

In the defence world, as we see the Ukrainians wreak havoc with 3D-printed bombs and home-made drones, the UK struggles on with institutional structures and procurement processes from a different century. The delays and failures of the General Dynamics Ajax programme are well-known, but more concerning is the culture that allowed it to happen. Over 172 pages, the government’s independent review of the programme details institutional warfare across the Ministry of Defence’s various science and tech bodies, a lack of accountability, and how concerns from those on the ground were (sometimes willfully) ignored. Beyond Ajax, the MoD’s most recent Command Paper, in common with many government technology strategies, is sharp in its analysis of the importance of AI, but blurrier on actual action, rarely straying beyond repeated expressions of the need to partner more with industry.

None of these environments in their current form sound like they are ripe for smooth digital transformation.

Reform and not instead of investment

In her book recounting her work chairing the UK’s Vaccine Taskforce (VTF), Kate Bingham writes with barely concealed exasperation about having to construct a 100-page business case for the Treasury, outlining the cost-benefit ratio of discovering a Covid vaccine. She similarly faced repeated, time-consuming requests for information from the National Audit Office. In the end, she concluded that: “If I had ever been asked the question ‘What is the biggest threat to the success of the VTF?’ the honest answer would have been ‘Large parts of the rest of Whitehall’.”

This combination of overbearing process and penny-pinching has precedent in the UK, even in the context of national crisis. The UK lost valuable time in its efforts to develop the jet engine in the run up to World War II, due to reluctance by the Ministry of Air to devote even minimal funding to the project. Similarly, Tommy Flowers, the inventor of Colossus, the groundbreaking programmable computer that decrypted German messages, largely had to self-fund his work. Even after the war, the government failed to compensate him fully for the initial materials he bought.

We risk seeing exactly the same process play out in the AI race - only this time, we’re facing better-funded and equally or more-skilled competition across national boundaries.

Unfortunately, the scale of investment needed here is significantly greater than the sums required for the development of the jet engine. This funding is needed both at an individual department level (for example, reversing the collapse in NHS tech investment), but also on major national priorities.

As we’ve argued previously, moves to rectify the UK’s inadequate next-generation infrastructure like the National Compute Strategy are promising first steps, but the resource commitments don’t match the scale and urgency of the task. A national compute cluster of 3,000 GPUs in several years time will set us behind where most corporate labs are today. While the government operates on borrowing rules that don’t distinguish between investment and day-to-day spending, this seems unlikely to change at a time of national stringency.

Even if we do successfully end up developing some of these capabilities, we have to then avoid the trap of governing them through the same failed processes. For example, if the UK one day does build out sovereign compute capability that researchers or promising start-ups could tap into, we can’t govern accessing rights using the same Treasury criteria Kate Bingham fought against. We would need to operate it with the philosophy of the new San Francisco Compute Group, with its single form application, even if this would inevitably mean some capacity being used by unsuccessful projects. This would mean a radical shift in mindset.

The owners of stagecoaches aren’t going to build the railways

Proper investment, however, is only one piece of the puzzle. There’s no point increasing investment only to throw good money after bad, with the same coalition of incumbents, outsourcing companies, and legacy IT suppliers.

If the Prime Minister is serious in his aspiration to build the ‘Unicorn Kingdom’, we need to become significantly more energetic about bringing challengers into government. A 2015 analysis of feedback from SMEs trying to sell to the UK Government listed a number of killers: overly prescriptive requirements, poorly written specifications, cost and time, bureaucracy, and the perceived unfairness of the process as their most common challenges. This tallies with similar research from the Startup Coalition, which laid bare how businesses increasingly didn’t want to work with the government, or how when they did, they couldn’t find a way in.

At the moment, if you run a health startup, there is no central portal through which you can sell to the NHS. Instead you have to hope an individual hospital trust is prepared to take a chance on you. If they are, then they’ll also need to have the right infrastructure, and be willing to work through hurdles of bureaucracy required to make a partnership possible, all while trying to deliver a standard of care amid record waiting lists.

Oliver Lewis, co-founder of Rebellion Defence, described the MoD as “one of the worst customers in the world”, warning that: “It takes Herculean patience, specialist knowledge and expertise, trusted networks back into the system and, critically for tech start-ups, significant patient capital. Typically, that is a level of venture funding that is available only in the United States, even if British and European investors are close followers.” The US Department of Defense has already started exploring specific measures to help startups navigate what they term “the valley of death”. This new attitude is beginning to bear fruit, but we have not seen the same energy in the UK. This needs to change.

Failure to reform procurement means attempting to reform the state with our hands tied behind our backs. As Camilla de Coverly Veale at the Startup Coalition has pointed out, challengers provide a way to move away from bloated, ‘bespoke’ solutions. To this, we’d add that earlier-stage AI businesses have to stay at the cutting-edge of the field if they’re to stay alive, move much more quickly, and embrace community norms like interoperability (as opposed to the vendor lock-in strategies of incumbents).

Realism requires courage, courage requires trade-offs

It’s both easy and incredibly lazy to blame everything on ‘the blob’, ‘bureaucrats’, or non-specific Whitehall forces. There’s definitely more that could be done to bring in relevant expertise and a raft of individual process reforms that others have laid out in more detail. Reform’s latest report details plenty of official “eye-rolling” in response to transformation efforts from the centre. However, the fiscal fantasies of the Reform and AI Fairies, as well as the procurement horror show have another shared root.

This is a cultural aversion towards risk that comes as much from politicians as it does from officialdom. If a procurement team feels like it will be punished by its superiors and hung out to dry by their minister in the press, they will of course always opt to add in that one extra check. Similarly, the longer politicians insist on maintaining the fantasy tax-and-spend debate Duncan Robinson alluded to, the longer it will take to reach an honest account of what we need to do to enable the next generation of public services. It’s simply not good enough for elected politicians to act as if they don’t run the country and these forces are beyond their control.

These conversations would only mark the beginning of the cultural shift we’ll need to enable AI-first transformation. As the intense public reactions to attempts to rationalise GP records and health data, or past attempts to involve the private sector in service delivery, have shown, we have plenty of future difficult conversations to come. Vehicles like ARIA will be able to shield the most critical research from short-term political impulses, but there are some fights that future governments will need to be prepared to have in public. As Nathan has said before, “there’s no world where we can build high-quality, technology-first public services that don’t require any money, data, or private sector involvement”.

At Air Street, we often talk about how these discussions aren’t luxury debates about future technology - there’s a wider, urgent point about national security. In a 2008 essay about the hidden costs of additional checks, Y Combinator co-founder Paul Graham warned that: “Checks instituted by governments can cripple a country's whole economy. Up till about 1400, China was richer and more technologically advanced than Europe. One reason Europe pulled ahead was that the Chinese government restricted long trading voyages. So it was left to the Europeans to explore and eventually to dominate the rest of the world, including China.” We have a chance to stop history repeating itself in reverse - let’s not waste it.