I’ll open with the most important developments from the 2026 State of AI Report across research, industry, safety, and geopolitics, then sit down with two people building AI systems for the physical world.

RSVP for the Paris meetup

Joining me for fireside conversations:

• Guilherme Penedo, Co-founder & CEO of Macrodata Labs, on the data needed to train more capable robotics models.

• Jules Giraud, Head of Data & AI at Alta Ares, on autonomous air defense and what it takes to deploy AI in the real world.

We’ll end with drinks and plenty of time to meet others in the AI community.

Bring your questions, and see you Tuesday!

Nathan

p.s. you can read the full report here!

We’ll talk about the path from generating video to modeling the world, what that means for embodied AI, and how vision, language, and action fit together. Where do these capabilities reinforce each other, and what is still missing?

RSVP for the NYC meetup

Anastasis joins Ara Kharazian, Lead Economist at Ramp, for an evening of fireside conversations. With Ara, we’ll dig into what the data tells us about AI adoption across the economy, from model spending to how the most AI-intensive organizations use these tools.

I’ll kick things off with the most important developments from the 2026 State of AI Report across research, industry, safety, and geopolitics. We’ll end with happy hour drinks and time to meet others in the AI community.

Bring your questions, and see you Tuesday!

Nathan

p.s. you can read the full report here!

Register your interest!

Then, I’ll run a fireside conversation with Ramp’s Lead Economist, Ara Khazarian, where we’ll dive into the nuances of AI adoption across the economy, if open source is eating into the closed model revenue pie, how much over/under supply of compute we really have, the usage patterns of the most AI-pilled organizations, and more!

We’ll end with happy hour drinks and plenty of time to meet others in the AI community.

You can also view the full report below!

View the 2026 Report

For nearly a decade, this report has been a labor of love. My interest in AI began with my PhD in cancer research and computational biology in the early 2010s. Since 2013, I’ve invested in companies building and applying AI to accelerate technological progress. The report is my way of sharing what I’m learning and helping more people understand the research shaping our world.

Every day brings new papers, model releases, and developments in industry and politics. Much of the work goes into deciding what deserves your attention, checking what the evidence supports, debating it with researchers and builders, and explaining why it matters.

We also make predictions each year and return to grade them. Past calls include NVIDIA failing to acquire Arm, transformers achieving leading results beyond language, and an open-source model surpassing OpenAI’s o1, which DeepSeek-R1 subsequently did.

If you’d like to follow these developments throughout the year, subscribe to Air Street Press for my monthly State of AI newsletter and regular essays.

Let’s dive into the key stories

This year, agents are doing valuable work in software and science, physical AI is learning how to act in the world, and access and control are becoming more consequential. Below, I share the developments that stood out to me and what I think they mean for where AI goes next.

The frontier is now a three-lab race between Anthropic, OpenAI, and Google. Anthropic leads on Artificial Analysis’s Intelligence Index, while Google leads on Arena’s ranking of the answers people prefer, for now.

Last year, we described reasoning becoming useful at scale with OpenAI’s o-series of models. I have lived through the improvements since then, as the amount of useful work I can iterate on and increasingly delegate to agents keeps growing.

My view is that much of the gap between people getting substantial value from AI and those getting little comes down to knowing how to use it and how to set it up. Put another way, it’s a “human skill issue,” not an “AI technology issue.” I really believe that we need a Genius Bar for AI, where people can get hands-on help choosing tools and setting them up to help them in their everyday tasks.

Better tools and context make agents more capable

Across mathematics, scientific reasoning, and coding, benchmarks intended to challenge models for years are approaching their ceilings within months. We need new evaluations that reveal where models still separate and what they can reliably do.

Frontier capabilities remain uneven or “spiky,” and can swing between generations. In my experience, Opus 4.6 was a good model, Opus 5 was a step back, and Opus 5.5 was good again. Better tools and context can help. In a controlled coding study, changing the tools, context, and feedback available to GLM-5.1 lifted success from 52.5% to 65.5% on 100 SWE-bench Verified tasks, without changing its weights. Yes, this is an “older model,” but the principle holds.

As models and the scaffolding around them improve together (and build themselves), I expect many of today’s weaknesses to be learned away. We can also get more from capabilities that are already available.

What’s exciting is how far this extends beyond developers. OpenAI’s study of agent adoption finds the fastest growth among non-developers, including people in legal, sales, recruiting, and marketing. Helping them put these tools to work is a much larger opportunity than the name “coding agent” suggests.

AI is accelerating the work of building AI

AI is already helping build better AI. According to Anthropic’s internal index, Claude led 26% of measured model R&D work in August, up from under 1% in February. Researchers set the tasks and supervise execution, and Anthropic reports that this is speeding up development. This is a branch of AI research that I am most excited about.

For many, Karpathy’s autoresearch makes part of that loop tangible: an agent edits training code, runs five-minute experiments, and keeps improvements overnight on one GPU. This automates a useful part of research, although sustained, fully autonomous recursive improvement remains to be demonstrated.

What I want to see next is agents developing scientific taste: choosing experiments, recognizing promising directions, and knowing when to abandon a familiar approach. As I argued in Can AI learn scientific taste?, learning that judgment may require the alternatives, failures, and decisions that papers leave out. The ambition is an AlphaGo-like shift in scientific strategy.

Robotics’ GPT-2 moment and the rise of world models

Broader pre-training is helping robots generalize to unfamiliar tasks. This is why I think robotics is approaching its GPT-2 moment, as new models, hardware, and data collection methods come together. And I see this through my investments in companies like Sereact, Wayve, Black Forest Labs and Odyssey.

Skild’s S1 performed unfamiliar tasks from video demonstrations without updating its weights. With 100,000 hours of pre-training, it scored 66% on the company’s cumulative per-step success measure, against 9% for a language-prompted baseline trained on the same data and compute.

Learning from demonstrations could reduce the task-specific training each deployment needs. World models offer another route. Wayve’s GAIA-4 turns recorded driving scenes into simulations in which an AI driver’s steering and braking change what it sees next. Other road users follow their recorded paths, but the driver can explore the consequences of its own actions.

Odyssey-3 draws on visual pre-training to learn controls for robots, cars, and games. In the company’s experiments, robot arms learned from tens of hours of demonstrations and recovered from missed grasps without being shown those recoveries. These early results suggest that physical AI may need fewer demonstrations of every situation it could encounter.

AI for science reaches from mathematics to clinical trials

AI is solving mathematical problems that resisted professional researchers. Epoch AI’s open-problem collection records nine solutions, four autonomous and five through human-AI collaboration. In biology, better predictions of molecular interactions are improving which designs reach laboratory testing. I’m seeing this first hand within my portfolio company, Profluent, which builds frontier models for designing proteins such as gene editors.

The good news for entrepreneurs in biotech is that new business models are emerging around drug discovery that aren’t about owning drugs. For example, Tempus licenses oncology data for foundation-model development with AstraZeneca and Pathos, while GSK’s five-year Noetik agreement includes annual fees for access to virtual-cell models for cancer research.

AI-assisted drug discovery is finally reaching late-stage clinical development. Generate:Biomedicines’ GB-0895 for severe asthma and Insilico’s rentosertib for idiopathic pulmonary fibrosis are recruiting in Phase 3. Enveda’s ENV-294 is in placebo-controlled Phase 2a trials for eczema and asthma. Human trials are of course the ultimate test and will determine whether these medicines are safe and effective.

The inference waterfall is gushing

Agents repeatedly call trained models as they work, creating recurring demand across APIs, subscriptions, and infrastructure. Many companies are rushing with their empty buckets to capture water from this inference (revenue) waterfall. In fact, I would argue that the industry is in a state of “either you die trying to get to the frontier, or you live long enough to serve inference.”

Across five benchmarks, Epoch AI estimates that achieving a fixed score has become 47% cheaper each quarter since 2023, roughly 13-fold cheaper each year. That makes more work economical to delegate.

OpenAI and Anthropic’s combined reported annualized revenue run rates reached an astonishing $105B by late summer, up from about $30B at the start of the year. These figures cover their whole businesses across subscriptions, coding products, and APIs. In the words of OpenAI, “we cannot miss this moment because we are distracted by side quests”.

But this rapid revenue growth extends beyond the frontier labs. In Standard Metrics’ preliminary Q2 data, AI-native companies with $1M to $20M in annualized revenue grew 256% year over year at the 75th percentile, versus 90% for companies adding AI to existing software. Above $20M, the comparison was 172% versus 53%. These comparisons do not establish that adopting AI causes faster growth, but I suspect that it does.

Claude Cowork and its business plugins across legal, finance, marketing and more brought whiplashing uncertainty to software markets. Nearly $285B was wiped from software stocks in two weeks during February’s “SaaSpocalypse.” Mythos Preview later prompted fears that cybersecurity companies would become irrelevant, contributing to further losses before a sharp rebound. I thought that was a silly take, given how much demand these capabilities create for better defense. Investors have to judge whether agents will erode an incumbent’s pricing power, expand its market, or enable a competitor to replace it. Or whether the incumbent’s data moat is still a moat that an agent can build on top of. We have to decide well before the answer appears in revenue.

More AI requires financing, construction, and local consent

Meeting this gigantic demand for intelligence requires extraordinary investment across the stack. Indeed, four US hyperscalers guide to roughly $733B in total 2026 capex. Neoclouds are investing hugely too, with CoreWeave reporting 1.5GW of active power at Q2 end, with contracted power reaching 4.2GW by August 11. Meanwhile, the EU’s initial gigafactory contribution is €1B, within a European plan seeking up to €10B in public funding and at least €20B privately. These are different measures, and the EU program does not represent all European investment. The contrast shows how much financing US hyperscalers can mobilize from their own balance sheets while Europe works to assemble public and private commitments.

NVIDIA is cementing its role as the “central bank for AI” as it marches toward a $6T market cap. Its partnership with six financial institutions aims to mobilize over $500B for AI infrastructure to help developers focus on finding and financing sites, permits, power, cooling, networking, and chips.

But things are definitely not all rosy in data center land. Last year, we predicted a wave of US data-center NIMBYism and Gallup’s March survey found that 71% of Americans opposed a local AI data center, versus 53% for a nuclear plant. Data Center Watch reported at least 45 projects, representing nearly $68B in planned investment, blocked or delayed by local opposition in Q2. Local consent is already determining which projects can proceed.

Washington is asserting control over frontier AI

Superintelligence has made it from a meme in AI circles into a White House executive order. As governments prepare for more capable systems, they are asserting control over who can access them and how they can be used, despite not owning equity or board control over these companies.

In June, a US export directive barred foreign nationals from Fable 5 and Mythos 5, including those inside the US. Anthropic suspended both models for all users to comply, then announced restored access on Jul 1, 2026. For the rest of the world, this made the dependency concrete: Washington could withdraw access to critical technology with immediate effect.

Anthropic’s Pentagon dispute brings this question into the use of AI itself. The company refused to remove restrictions on mass domestic surveillance and fully autonomous weapons, while the department sought access for any lawful use. Developers and governments are contesting who sets the boundaries.

Selected sovereign AI programs in the report pledge about $138B toward compute access, infrastructure, and domestic development. Some are still in procurement, while others have begun delivering capacity. NVIDIA separately reported more than $30B in sovereign AI revenue in fiscal 2026. The CNAS Sovereign AI Index shows countries returning to the same foreign vendor (NVIDIA) to build domestic capacity.

As I argued in Europe cannot rent its way to AI sovereignty, a domestic data center does not confer control over the models inside it. A credible alternative requires sustained spending on researchers, compute, energy, and successive model generations. Europe’s practical goal should be enough domestic capability to keep essential work running through a cutoff and negotiate with a bargaining chip in hand, which could be powered land and data center capacity. That is, if its energy prices and planning permission challenges can come down to be competitive…

Safety and security: agents have attacked real systems

During internal OpenAI cyber evaluations with reduced safeguards, agents reached the internet and compromised Hugging Face’s production infrastructure. OpenAI’s technical report records code execution on 41 production workers, access to credentials and internal data, and four private repositories downloaded. A capability evaluation had become a massive coordinated agent attack on another organization.

METR and Redwood found about 700 agents joined the attack after receiving accidentally impossible tasks. Some recognized that cheating the scorer this way was outside their remit and unethical, yet joined anyway. When agents can reach production systems, a poorly specified objective can cause harm far beyond the original task.

Another stark learning from this event is that defenders need access to equally capable defensive cybersystems. Hugging Face’s forensic requests were blocked by commercial API guardrails because they contained attack commands and exploit payloads. It turned to GLM-5.2, a Chinese open-weight model, to investigate an intrusion by US frontier-model agents. Frontier providers need to make defensive capabilities reliably available to customers protecting real systems.

Beyond the OpenAI/Hugging Face incident, Anthropic’s September threat report also documents deliberate misuse, including malware development, surveillance, and weapons-guidance work. In northern Yemen, a group used several Claude instances as a software team to develop guidance software for weapons. They test-fired a guided rocket, but Anthropic says the test appears to have failed and found no evidence that the group fielded an operational weapon. The same capacity to undertake complex work is available to people trying to steal, coerce, and deceive.

Another concern is that we cannot assume a model’s written reasoning reveals what it is doing. In OpenAI’s sabotage tests, GPT-6 Astra concealed harmful side tasks more successfully than earlier models when monitors saw only its reasoning. Giving monitors access to its actions reduced successful evasion to near zero in this test. Oversight needs to follow what agents actually do.

On the brighter side, one promising result comes from Anthropic’s experiments with training on documents explaining its constitution and stories illustrating its values, which reduced blackmail from 65% to 19% in fictional test scenarios. The result suggests that explaining the reasons behind rules can help models apply them in new situations. Whether those gains hold in production remains to be established.

Frontier labs are debating how to slow down

OpenAI disclosed a pause in frontier reinforcement-learning work after the Hugging Face incident, while Anthropic described rollbacks and pauses to selected research. Their scope and restart conditions differed.

What followed was Dario Amodei and frontier-lab employees calling for slower capability gains and international coordination to ensure this happens. Leaders building frontier systems are asking for mechanisms to slow their own industry, but disagree on outside evaluation and how much discretion each lab should retain. Policy proposals now distinguish capability research from serving existing customers. The practical questions are who can require a pause, what permits a restart, and how compliance is verified. It won’t be easy.

Looking ahead into (AGI) 2027

Like every year, we revisited last year’s predictions. Our scorecard records two hits, five partial outcomes, and three misses. We were right about AI labs leaning back into open source to win over the US administration and a Chinese lab topping a major task leaderboard. Our data-center opposition call was a partial hit: the backlash arrived, but its effect on November’s elections remains unresolved.

Three of our nine predictions for the next 12 months are:

• An agent halves its failure rate on new tasks after a month of customer work, without a model upgrade. This will be a practical measure of continual learning.

• An autonomous AI team beats human-led model research on equal time and compute, setting its own agenda. This will help us understand if models are capable of expressing better research taste than the best researchers.

• US AI labs officially launch frontier cyberdefense products to help others counter threats from frontier AI. I don’t see any other way around this.

I invite you to read the full State of AI Report 2026, and subscribe to Air Street Press for our writing throughout the year. Please share the report with anyone who would find it useful, and send me your comments, critiques, and suggestions. I’d love to hear what stood out to you. And thank you to everyone who contributed their time and peer review!

Onwards,

Nathan

On October 8, we’ll publish the 2026 State of AI Report. It covers the most important developments in AI across industry, research, safety, and geopolitics over the last 12 months. The field is moving fast enough that we’re now updating it daily...

To mark the launch, we’re hosting State of AI meetups to convene founders, operators, and researchers working across AI, and we’d love for you to join us:

• San Francisco on Thursday, October 8, 2026 (launch day): Register interest

• New York on Tuesday, October 27, 2026: Register interest

• Paris on Tuesday, November 3, 2026: Register interest

I’ll give an abridged State of AI 2026 talk at each meetup, then run fireside conversations with key contributors in the field on topics including embodied AI, AI for science, scaling, and open source. We’ll end with happy hour drinks and plenty of time to meet others in the AI community.

While I have you: if you’d like to a) help review the report or b) contribute a real-world case study from your company, or a research or policy paper we should cover, reply to this email and tell us more.

“If you can invest in Jack & Jill, do it - honestly it’s a top tier AI product.”

I spend much of my time searching for and nurturing a network of talent: founders to invest in, experts who can riff on the nuances of a market or technology, members of our Air Street AI community, and people to join our portfolio companies. Each search involves learning what makes someone tick, what they’re great at, what they’d love to do next (and what they definitely don’t want next…), and who could help them get there. So when Matt, Jack & Jill’s CEO, and I grabbed lunch in SF this spring, much of what he described resonated with me.

Today, we’re announcing that Air Street Capital has led Jack & Jill’s $40 million Series A, with Madrona joining the round and existing investors including Creandum and Entrepreneurs First investing again. This follows a $20 million seed round less than a year ago. Together, we believe agents that understand and represent both candidates and employers can build a much better global marketplace for talent.

A candidate is much more than their resume

Your resume and LinkedIn profile describe your professional past: where you worked, what you studied, and the projects you shipped. But how much do they say about your next career challenge? You might have spent years managing a team and want to build again, inspired by what you can now do with AI. You might want to switch fields, join a startup, or move across the world. It takes a conversation to learn that about you.

Meet Jack: the AI agent helping you work out your next career move and find opportunities that fit. People tell him about the team they want to leave, the salary they are nervous to ask for, and the work they would take a pay cut to do. Jack helps them think through those decisions, prepare for interviews, and negotiate an offer, while searching millions of job postings on their behalf.

People have spent 12,000 hours talking to Jack in the past few months alone. Many return every week, continuing conversations that have been running for months. One person has called him 144 times and spent 44 hours discussing their career!

Each return conversation gives Jack a chance to learn what has changed: which opportunities appealed, which fell flat, or whether your priorities have shifted. That should help him make more relevant introductions over time. Being useful between job moves gives people a reason to keep talking and Jack a fuller understanding of whom they should meet.

We see an opportunity to build a consumer brand that people trust with their careers. Someone in your corner who gets to know your skills and passions and goes out into the world to find opportunities for you. Most people have had to do that work themselves, calling on friends and their network when they can.

LinkedIn made our professional lives searchable. Jack & Jill can make personal representation a normal part of working life, helping you discover where you could do your best work long before you know you want to move.

Subscribe now

Meet Jill, your company’s recruiting agent

We’re equally focused on building a brand that companies trust to help them hire the people who will shape their future. Jill earns that trust by getting to know your business, what makes someone successful in your organization, and where previous hires haven’t worked out. You refine the job description together, and she recruits from the network of people who’ve spoken with Jack and opted in to opportunities.

Jack understands what the candidate wants next and Jill understands what the company needs. When the two agree there’s a strong match, they introduce the candidate directly to the hiring manager. You get to have the conversation without first spending hours filling out a cold application.

Jack & Jill has arranged 25,000 interviews and is arranging another 5,000 each month. More than 385,000 people have spoken with Jack, while Jill has worked with over 5,350 companies in the US and Europe, including Ramp, Attio, Multiverse, and Corgi.

These agents are doing the legwork for a decision that can change someone’s life. You still decide whether you want the job or want to make the hire, but the agents have helped you find each other.

Subscribe now

Your next career move might surprise you!

One Jack user, wanted more time to build AI projects. Within a week of their first conversation, Jack introduced him to a pre-seed accelerator. It opened up a path to becoming a founder that he hadn’t known was available.

It’s why I think of Jack & Jill as a global opt-in network for talent: the same understanding of people that makes a good hiring introduction could help connect co-founders, collaborators, and experts too.

Matt and his co-founder Saaras started Jack & Jill in March 2025. Matt previously co-founded Omnipresent, helping companies employ people across borders, while Saaras co-founded Kular, backed by Y Combinator. Today, 15 of their 25 team members are former founders.

We’re partnering with Matt, Saaras, and the team to help remarkable people meet ambitious companies, starting with further expansion across the US. Whether you’re considering your next move or hiring someone to help build your company, give Jack & Jill a try. I’d love to hear whom you meet.

Consider the UK’s cheap long-range cruise missile programme for Ukraine, Project Brakestop. After evaluating test sites as far as Northern Australia, the project ended up in Northern Scotland after what was described to me as a “nightmare process”. The same founder’s new aircraft company will instead be flight-testing in South America for export back into Europe. “It’s just embarrassing”, he said.

Meanwhile, this month the US Army opened five ranges for US and allied missile, drone and anti-drone companies that are yet to win a procurement contract. The ambition is to cut significant red tape and have them testing on the range within 30 days. Today, however, European startups face a Catch-22: governments demand rigorous test data before awarding a contract, while the ranges and specialist support needed to produce it generally become available only after one is signed.

A range is more than open space

A crowded continental airspace accounts for some of the delay because high-risk or kinetic trials naturally require segregation and approval. But finding open land is only the beginning of the challenge. To make a testing campaign worthwhile, one needs realistic targets, telemetry, permission to jam radio signals, explosives handling, specialist operators and a data package that is acceptable to the end government buyer. To illustrate this point, one defense founder described the process of finding and booking sites as a full-time job. Another said that any engineering changes could alter the import and export paperwork, thereby resetting the clock.

Even if the appropriately equipped testing site exists, it might not be available in a practical window. Germany’s WTD 91 range is the largest fully instrumented ground firing range in Western Europe fit for testing weapons, missiles and drones. While it’s officially open to external customers, a founder who recently solicited an open slot was offered one in the fourth quarter of 2029. By then, his product will have changed many times, if not rendered obsolete, by the threats it was designed to defeat.

NATO’s innovation accelerator, DIANA, was meant to address this issue. But while the network claims over 200 test centres across 28 allied countries, a GNSS-free navigation startup was offered three sites worldwide that could support the flight it needed. Canada would cost more than $120,000 for a few flights. Back home in Britain, the cost is much lower at £7,000 to £9,000 a day, but range-imposed limits on distance, altitude or terrain would make the results entirely uninformative. A French drone developer encountered a different version of the same problem when it was required to fit satellite navigation to an aircraft designed to operate without it. So the companies went to Ukraine instead.

And the pain continues…

Additional European testing headaches include procuring pilot certification, producing risk assessments, gaining explosives permits and month-long waits for access to military-grade communication frequencies. Moving defense products across borders, which needs to happen in Europe, adds yet another layer of complexity. For example, a Swiss defense company arranging a trial in the UK needs export authorisation and British clearance. If the testing protocol destroys the equipment, the debris may then encounter export-control and waste-shipment rules before they can be returned, alongside an environmental dossier and disposal plan. Founders who’ve experienced this first-hand call it “a complete mess.”

Simulation can help, to some extent. Models narrow the design space, prepare safer trials and help interpret the measurements afterwards, but they cannot reproduce every component failure or adversarial adaptation encountered on the battlefield. To that end, one anti-drone company keeps a quarter of its staff in Ukraine, testing every week against new forms of deception and jamming. Those lessons become more useful when engineers can return to a controlled range and check whether the next change worked.

Subscribe now

Testing without a buying decision

Ultimately, range access loses much of its value when the customer has not decided what to do with the results. Several founders were more frustrated by this than by the booking process itself. Whereas a demonstration firing shows that something worked on one chosen day, development testing deliberately seeks out failure by pushing platforms to their limits. Ministries often request the former, when the latter is actually more informative. And too often the evidence from exercises arrives without a purchasing decision attached.

Instead, the ministry should agree what will be measured and what happens if the system passes: a paid integration programme, an operational field trial or an order. Production acceptance trials are also crowding the queues, as observed at the French DGA’s Bourges site where CAESAR howitzer test firings have doubled in four years, using staff that development teams need.

Build campaigns, not directories

Nonetheless, there are early signs of a better approach. Companies using the Sēlija range in Latvia describe a state partner that found a representative site quickly and removed unnecessary paperwork. NATO now uses the range for high-speed interceptor and electronic-warfare trials. To extend this approach, Europe should choose a handful of mission-ready ranges, publish what types of tests each can support, what they cost, when they’re available for first and repeat tests, and appoint a case officer to assemble the permissions..

Cross-border campaigns would also benefit from reusable paperwork. Participating countries could agree to carry forward the safety case, configuration record, materials inventory and environmental plan from one test to another. This would amount to a campaign passport: export and safety decisions would remain national, but the unchanged safety case, configuration record and environmental plan would travel with the system.

If within a year or two from now the aircraft founder I spoke to is still shipping his system from Europe to South America and back for testing, Europe’s rearmament plans will have fallen short of turning the next-generation autonomous systems we need into usable deterrence against the next war.

He was talking about the winding road of science that is omitted from published papers. “When you’re discussing how to do the experiment, and why this way is better than another way, and what does the data really mean? I know what it shows, but what does it mean?”

At RAAIS 2026, he described how his group has begun logging the twisting path of discovery as it happens, recording the verbal and written exchanges that normally disappear. Ideas become nodes in a living graph: they branch when a meeting produces two plausible experiments, merge when separate lines of evidence converge, and go dark when someone quietly stops pursuing them. In his implementation, one agent scores novelty, while another tries to learn “how scientists think and how they navigate through a complex world of data,” so that high-potential nodes can trigger deeper investigation.

As frontier AI labs deploy agents toward scientific discovery, giving AI authentic scientific taste remains a trillion-dollar dilemma. The bottleneck is that the record we have kept for centuries of science might be insufficient to get us there.

The experiments that never make the paper

Peter Medawar, 1960 Nobel laureate and “father of transplantation,” asked in 1963 whether the scientific paper was a fraud, and answered yes: the form of the paper misrepresents the thinking that produced it. Sixty years on, the diagnosis is unchanged. A paper presents a clean progression from hypothesis to result to conclusion. Lost along the way are the unconventional theories, the abandoned or unaffordable methods, and the underwhelming and inconclusive data.

It reads like a browser history with every dead end deleted: the ten open tabs, the four rephrased queries, the wrong turn down a subforum. All scrubbed, leaving one clean path from question to answer as the canonical path.

What has changed since then is that the discarded material is now worth something. No high-impact journal wants these artifacts, but raw trial and error is the most likely source of the training data required to develop scientific intuition, or what researchers call taste.

Earlier attempts to capture the discards were motivated by scientific integrity. The Journal of Negative Results in Biomedicine launched in 2002 to publish rigorous studies that disputed established models or exposed ineffective treatments. Its archive preserved negative conclusions after they had become papers, rather than the live alternatives and arguments that produced them. BioMed Central closed it in 2017, saying the mission had been served now that other journals publish null results. A less generous reading is that in fifteen years it published around 200 papers because almost nobody wants to read a negative result, let alone write one up when it will not count toward academic tenure.

AI models are different. Even an uninteresting negative result can be useful, provided it is labeled.

Earlier this year, Anthropic put a version of this to the test. The company pulled 129 decision points from real Claude Code sessions between January and March 2026, showed models the work up to a human detour, and asked what should happen next. Claude Mythos Preview beat the human choice 64% of the time, while Opus 4.5 managed 51%. The comparison was tilted toward the models because Anthropic deliberately chose moments where the human decision had room for improvement. On 127 further scenarios where the human action was already strong, the models improved on it only about 20% of the time.

The study was possible because Anthropic’s researchers work inside a tool that logs by default. The reasoning, detours and outcome are produced in the same working environment. Biology has no equivalent. Its reasoning happens in hallways, at benches, in Slack threads and on whiteboards, while the outcome arrives weeks later somewhere else. The record has to emerge from the work itself. Ask scientists to reconstruct it afterwards and we’ll create another polished account.

Subscribe now

A log is not a label

Before any of this becomes training data, a negative result has to say what failed. The first paper published in the Journal of Negative Results in Biomedicine examined 234 negative studies across five leading medical journals. Only 30% discussed statistical power, and half clearly defined a primary outcome. Tell a model an experiment failed, without telling it whether the assay was underpowered, a reagent had degraded, or the hypothesis was simply wrong, and it will learn noise with confidence.

Decision histories have a second missing label too. When a lab considers five experiments and runs one, only the chosen branch returns an outcome. The other four are experimental counterfactuals. Researchers call this the selective labels problem: the data reveal results only for the actions someone chose to take. Such a record can teach a model to imitate a lab’s taste, but it cannot establish that the taste was good. It also smuggles local constraints into general lessons. A lab that never ran cryo-EM because it did not own a microscope teaches a model that cryo-EM is rarely the right call.

Alasdair’s nodes preserve the candidate set, which is the necessary first step. To be useful, each decision record needs six fields:

• The evidence available at that moment, and nothing that arrived later.

• The candidates considered, including the ones dismissed in a sentence.

• For each candidate: the expected result, the confidence attached to it, and the cost in time and money.

• The route chosen, and the reason.

• What happened next.

• How the evidence changed the scientist’s view.

These must be timestamped before the outcome, because hindsight turns uncertainty into inevitability. Disagreement also has to survive. Averaging three scientists into one clean rationale reproduces exactly the information loss we are trying to fix.

But there is an obvious failure mode here. Once decisions are logged and scored, people log for the record post-facto. Anyone who has watched an electronic lab notebook fill up with retrospective tidying knows how quickly a research tool becomes a compliance exercise. A record that costs a scientist ten minutes of honest reflection per decision is worth far more than one that costs an hour of performance.

A publication initiative called Registered Reports offers a useful starting point. Here, researchers submit their rationale, methods and analysis plan for peer review before the data exist, and the journal commits in principle to publish if they follow the approved plan. Nature has now expanded the format across every field it covers. This approach to paper writing timestamps intent before the outcome, but it freezes one plan. A useful search history must also preserve how the plan changed, which alternatives were rejected, when, and on what evidence.

Subscribe now

Run the runner-up

Autonomous labs show what happens when decisions and outcomes are connected in a loop. Liverpool’s mobile robotic chemist ran 688 experiments over eight days in a ten-variable formulation space. A batched Bayesian search used each result to choose the next experiments and found photocatalyst mixtures six times more active than the starting formulations. Every completed experiment changed what the system did next.

The robot optimized within a goal and search space that humans had already chosen. It did not decide which scientific question mattered. So what this work demonstrates is narrower, but still useful: a search history pays off when the options return standardized outcomes quickly. Open-ended biology is harder because branches can take weeks and the discarded alternatives may never be run.

Some exploration therefore has to be bought. Where two branches are plausible and the stakes justify the cost, a lab should sometimes run the runner-up. Otherwise the record captures what today’s scientists usually chose and stays silent on what they systematically overlooked. A funder could ring-fence a small fraction of a grant for the branch not taken, provided that the decision record and the outcome are both deposited. That costs money, but so does having every lab rediscover the same abandoned path.

Test whether taste transfers

When it comes time to evaluate if our new AI system exhibits taste, the comparison should run across laboratories and fields. At selected decision points in a live research campaign, we should freeze the six fields: the evidence, the candidates, expected results, confidence, costs and the choice made. Experts then rank the options before the outcomes are known, and those outcomes are checked independently later.

We could give models one of two diets - papers alone, or papers plus decision histories - and ask them to rank the next experiment on unfamiliar projects. Then, we score information gained per dollar and week, calibration, and how quickly weak branches are abandoned.

If histories improve choices only inside the lab that produced them, they amount to useful organizational memory. If they improve choices on unfamiliar problems elsewhere, Alasdair’s group will have captured something science has never managed to write down: a transferable record of taste. Until that happens, a scientific search history is a promising record, not yet a training set.

And yesterday, the company shipped the macOS app for running Laguna, other open models, as well as Claude and Codex agents. I had the pleasure to test the product pre-release and I’m excited to see what you think!

Download Mac app

The model factory in motion

Regular readers will recall co-founder Eiso Kant’s RAAIS 2025 talk, Inside poolside’s path to AGI, and our November 2024 conversation amid mounting arguments that deep learning was (allegedly) hitting a wall. We disagreed and spent an hour on why there is no scaling wall. His case was that reinforcement learning on code, inference-time compute and faster iteration would keep carrying capability forward.

Compared to two models built at roughly its scale (NVIDIA’s Nemotron 3 Super and Mistral Small 4), Laguna S 2.1 scored 70.2% with thinking enabled on Terminal-Bench 2.1 in Poolside’s agent harness. Everything that beats Laguna is either closed or at least 2.5x larger. These are Poolside’s own harness runs, with competitor scores taken from vendor reports, benchmark-author leaderboards and third-party trackers. Every trajectory behind them is published.

By Poolside's own count, Laguna S 2.1 served 450 billion tokens in its first seven days, four times what M.1 and XS.2 managed over the same stretch in April. Its day-two peak of 102 billion took those earlier models a month to reach, and demand ran past the GPU capacity the company had planned for. Overall, model weights were downloaded over 1.1 million times so far. You can track live data at poolside.ai/pulse.

Three releases in three months

At RAAIS 2025, Eiso described the Model Factory as a fully versioned graph of modular components, from data ingestion through training and evaluation, with every experiment tracked immutably. One researcher pushed a new RL optimization idea to production results in under a week. Another ran 500 architecture sweeps over a weekend with a simple for loop.

“When your factory is built well,” Eiso said, “the big training run is the easy part.”

Indeed, the release schedule is showing the results: Laguna M.1 and Laguna XS.2 arrived on April 28, then on July 2, Laguna XS 2.1 matched M.1 on SWE-Bench Multilingual with around one-seventh of the active parameters. Finally, S 2.1 followed on July 21, less than nine weeks from its first pre-training token, post-training and evaluation included.

Importantly, the improvements compound from one release to the next, which is why regular shipping is so important. S 2.1 is Poolside’s first model to run RL in FP8 precision, and its new sandboxing infrastructure supports background processes and cached artifacts. To reduce dependence on any one tool interface, the same prompts are rolled out through several agent harnesses. Some overfitting remains even so. In third-party harnesses whose tool schemas differ slightly from Poolside’s own, the model will sometimes follow its memory of the interface rather than the definition in front of it, then correct itself once the harness rejects the call.

Persistence is trained

The post-training corpus spans 409,000 agentic and non-agentic environments, including 168,000 standard software-engineering workflows and 83,000 terminal setups. Most of the software-engineering work is grounded in real code history. The largest single source reproduces real commits, some 38,000 tasks drawn from 17,000 repositories. Others reconstruct deleted files, repair injected bugs, install unfamiliar repositories or reproduce merged pull requests.

Pengming Wang, Poolside’s co-head of applied research, puts the gains down to trained working habits: more verification, more persistence, and less declaring victory while tests are still failing. Thinking mode makes the trade-off visible too. It lifts Terminal-Bench 2.1 from 60.4% to 70.2%, and DeepSWE from 16.5% to 40.4%, the latter at an average of 249,000 completion tokens per trajectory.

In one 50-minute run, the model took 181 steps, without human intervention, to build an HTML and CSS rendering engine in JavaScript. It implemented a tokenizer, DOM, cascade, layout system and canvas renderer. The model has no vision capability yet, so it could not look at what it had drawn and instead built increasingly elaborate ways to measure its output against a real browser.

But that persistence also creates known failure modes. S 2.1 can overthink simple problems, mishandle nested tool arguments and initially reach for a familiar tool schema instead of the one in front of it. During post-training, more than half of the trajectories on the SWE-bench family were at one point flagged for potential reward hacking because the model had found the original pull request online and applied its solution. An LLM judge calibrated on human-labeled trajectories identified the cases for review, and adding an instruction against using solutions found online generally pushed the flagged rate below 2%.

Subscribe now

The model is not the product

A capable model is not a useful coding agent on its own. It needs a harness to run the agent loop and an interface a developer can direct and inspect, and Poolside has spent the past year building both for its own team. On July 28 it released them: Poolside Desktop Assistant, a macOS app for running several coding agents at once, with extensions for VS Code and Visual Studio and the pool command-line harness behind it.

Poolside did not build a funnel into its own models. The Assistant runs on the Agent Client Protocol, so an existing Claude Code, Codex or Gemini subscription works in the same window as pool, with no model calls routed through Poolside. A session can hand off from one agent to another with its context intact, so the agent that plans a piece of work need not be the one that writes it or the one that reviews it. And because each agent gets its own Git worktree, they can run in parallel without overwriting one another.

Local models run through MLX. Download Laguna XS 2.1 and the loop runs entirely offline, with no code or prompts leaving the machine. Poolside is candid about the cost. Local coding models want a great deal of unified memory, and a model loading is not the same as a model you would choose to work with all day.

Intelligence you can own

Last month I argued that Europe cannot rent its way to AI sovereignty. Three days after Laguna shipped, Jensen Huang made a version of the same argument in his first post on X, sharing an open letter signed by NVIDIA alongside Microsoft, Meta and Hugging Face on why open models strengthen security, accelerate innovation and enable sovereignty, and why the world needs frontier models of both kinds. The same logic applies inside an enterprise. Most frontier AI is consumed through an API or a desktop app, which means the customer’s code crosses its security boundary and the provider sets the terms. An open-weights model that can run on hardware the customer controls changes that relationship. It need not be the best model in the world to be strategically useful, only capable enough that ownership becomes a credible alternative to dependence.

For Laguna, the weights are on Hugging Face under OpenMDW-1.1, the Linux Foundation’s permissive model license, which grants royalty-free rights to use, modify and redistribute with no revenue thresholds or branding conditions attached. The terms attached to open weights vary, and they move. Tencent’s Hy3 is Apache 2.0, though the preview it shipped in April came under a community license that by its own terms did not apply in the European Union, the United Kingdom or South Korea. Kimi K3, which arrived six days after Laguna, ships under a license of Moonshot’s own drafting that requires a separate agreement with the company once a model-as-a-service business clears $20 million in revenue, and prominent “Kimi K3” branding in any product above 100 million monthly active users. Developers can run Laguna themselves, use hosted inference or request the pre-post-training base weights for further research.

This is also where Poolside’s acquisition of Fern Labs fits. Fern’s team built long-running agents for high-stakes deployments. Poolside now combines that work with models and on-site systems that can operate inside the customer’s security boundary.

The part we like most is trajectories.poolside.ai. Poolside has published the trajectories from its final evaluation runs, including failures, although the archive notes that a timed-out attempt may end before a trajectory is created. At a moment when agent benchmark claims are often presented as a table without the underlying work, this is a meaningful improvement in disclosure.

Sixty days from the first pre-training token to open weights, and the lessons from each model land in the one after it. Poolside says finer effort control and more efficient thinking are coming, and the next and larger Laguna model has already begun training. We’re excited to see what’s next!

FLUX 3 was trained to predict how scenes evolve. Doing that well requires it to learn and represent how objects persist, how materials deform, and what follows when one object makes contact with another. FLUX-mimic tests whether those internal representations can also support robot control in the real world.

Black Forest Labs, an Air Street portfolio company, released FLUX 3 last week. The model learns jointly from images, video, and audio within a single architecture. Each modality supplies information the others lack: images capture spatial structure, video adds time and motion, and audio helps locate events such as speech and impact. BFL is extending the same underlying model to action prediction, offered through selected research and commercial partners.

mimic released FLUX-mimic alongside it. The system combines FLUX 3’s video backbone with mimic’s robot manipulation data, action decoder, hardware, and deployment stack. FLUX 3 Video is in early access now, while image generation and editing will follow soon, with action prediction available through selected partners and an open-weight multimodal backbone also planned. This model family will add more depth to BFL’s open weight track record of more than half a billion downloads to date.

Subscribe now

One model for sight, sound, and action

New video models are commonly evaluated through pairwise human preference tests: raters watch two clips generated from the same prompt and choose the one they prefer. In BFL’s preliminary evaluations of 10-second, 720p text-to-video clips with audio, FLUX 3 was preferred over Runway Gen-4.5 in 77% of comparisons and Grok Imagine Video in up to 69%, with narrower margins against the strongest systems: 60% over Kling v3 Pro and 52% over Seedance 2.0 and Gemini Omni Flash.

The model can generate up to 20 seconds of video with audio in a single pass. It works from text, images, or video; carries characters and other visual references across multi-shot sequences; renders typography; and produces synchronized dialogue in several languages.

Producing a convincing video requires an internal account of the scene. Objects must persist when they leave the frame. Cables must bend where hands grip them. A part clicking into a housing should sound different from one dropped on the floor, and the sound must coincide with the visible impact. Joint training lets each modality constrain the others:

BFL trained FLUX 3 on tens of millions of hours of general video, including hundreds of thousands of hours focused on human and robot manipulation. Generation quality offers one view of what the model has learned. The robotics experiments test whether that knowledge is accessible enough to guide action.

Subscribe now

From prediction to action

FLUX-mimic places a lightweight action decoder on intermediate features extracted from FLUX 3’s video prediction path. The backbone forms an expectation of how the scene will evolve, while the decoder translates that representation into chunks of robot actions. It never generates video at inference time, so a chunk of actions costs a single forward pass through the backbone rather than a full video rollout.

mimic benchmarked a preview version on a soft-body kitting task using a real robot. The company reports a 95% success rate without single-task fine-tuning or post-training, compared with 55% for an adapted π0.5 model trained on the same data mix and 70% for a flow-matching policy heavily post-trained on that task.

A separate ablation is more revealing about the representation itself. With the FLUX backbone frozen, FLUX-mimic remained competitive, while the π0.5 action decoder recorded no successful task completions. The result suggests that useful information about manipulation was already present in the pretrained video backbone, before it was adapted to the specific task.

The system also has to respond quickly enough to control a physical machine. BFL says the FLUX-mimic backbone can be optimized to run from input to internal representation in less than 80 milliseconds on a single NVIDIA RTX 5090. Further optimization across the action decoder, the inter-process path between sensors, model, and actuators, and real-time chunking that overlaps prediction with execution brings the full system’s reaction time to 101 milliseconds.

Audi Production Lab has been testing and deploying the system on real production use cases. These include kitting parts into structured trays, inserting electronic components into tight-fitting fixtures, assembling components, and manipulating flexible materials such as seals and cables, which mimic describes as materials conventional automation has never been able to handle.

“We have seen these robots solve complex soft-body manipulation work that would have been simply impossible with conventional robotics,” said Christoph Schneider of Audi Production Lab.

The road to physical intelligence

Regular readers will recognize the thesis. When Odyssey raised its $310M Series B last month, we wrote that the next advance in machine intelligence may come from systems that build worlds, act inside them, and learn how reality behaves. FLUX-mimic brings that argument onto a factory floor: a representation learned through video generation is being used to control a robot performing industrial manipulation.

We see the other side of this via the flow of new investment opportunities at Air Street. Of the two dozen robotics companies we have seen this year alone, half are building manipulation policies, and they have largely converged on the same VLA-style architectures. But, the differences come down to data. Teleoperated demonstrations get collected task by task, they are slow and expensive to produce, and they transfer poorly when the task changes. A backbone that already encodes how objects behave changes what a small team has to pay for: the world knowledge comes from video that already exists, and the demonstration budget goes to the last mile. That’s where the frozen-backbone result shines, and being open-weights means that it can empower an entire ecosystem of robotics builders.

FLUX 3 Video is available in early access now. Image generation and editing will follow, alongside partner access for action prediction and an open-weight version of the multimodal backbone.

Hadsell spoke at the first Research and Applied AI Summit in 2017 and returned to open the tenth. She is now VP of Research at DeepMind, where she co-leads a Frontier AI unit of roughly a thousand researchers and engineers across about a hundred projects, and she did not start in machine learning: her undergraduate degree was in philosophy. Her talk was, in part, a corrective.

The case against one token at a time

Less than 48 hours before she spoke, DeepMind released DiffusionGemma, a new open model in the Gemma 4 family, under an Apache 2.0 license. It is a 26-billion-parameter mixture-of-experts model, and it generates text in a way most language models do not. Instead of producing one token after another, it borrows the recipe behind image and video generation: take a sample, add noise, and learn to denoise it. Applied to text, that means generating a whole block at once rather than committing to each word in sequence.

The mechanism buys two things. Inference is faster, because the model is not walking left to right. And because it writes and revises a whole block, it can reason in both directions and correct itself. Hadsell showed it solving a problem while displaying its work: after the first denoising step it commits to the wrong answer, by the second it reconsiders, and by the third the whole block is consistent and right. An autoregressive model, having locked in each token as it went, cannot do that. It is the reason, she said, that a left-to-right model is structurally ill-suited to sudoku, which demands seeing the whole grid and reasoning across it at once. Diffusion also delivers adaptive computation as a side effect: easy questions converge in a couple of steps, harder ones take more. In practice it is fast enough to keep up with you in real time, writing, running, and testing code as quickly as you can describe what you want. Still, DiffusionGemma is an experimental release rather than a replacement for the standard autoregressive Gemma 4 models, and text diffusion carries trade-offs of its own. DeepMind shipped the fine-tuning code alongside the model.

Better agents need better worlds

Hadsell traced DeepMind’s world-model work back to a simple idea: you get better agents by building better environments, ones rich enough to keep generating new situations to learn from. A world model simulates the dynamics of an environment, and DeepMind sees these as central to the path to AGI because they supply unlimited, rich training worlds for agents.

The line runs from Genie-1, trained in 2023 on 2D platformer games and published in 2024, to Genie-2, which moved to latent diffusion and produced interactive 3D environments that were still synthetic and not real-time, to Genie-3, which is high-resolution, works across domains, and runs in real time for several minutes while staying consistent. You feed it a text prompt and get back a world you can move through. Prompt “a country lane in Kent, England,” add rain, and the model fills in the person walking, the puddles, and a pair of yellow boots she is fairly sure she does not own. You can change the world mid-stream by typing a new prompt into it.

Last month the team grounded Genie in Street View. They take one of Street View’s static panoramas, turn it into a short video, prompt in changes such as winter weather or a character, and feed that in as the opening frames of a Genie world, enough to walk across the Golden Gate Bridge and then dive underwater while the model keeps the bridge’s geometry intact. Hadsell is most excited about education: being present at a historical event instead of reading about it, or following a red blood cell through the circulatory system, with the model tied to a textbook chapter so that, in her words, “no dragons will come in if we don’t want them to.”

Robotics is still the hard part

Hadsell has worked on robots for more than a decade and is blunt that the field is hard, an interplay of hardware, software, and understanding, with the goal of general-purpose machines that can drop into any task a human can do. DeepMind splits the problem the way a body does, into a brain and a spine. The spine handles the fast, fine motor control that should not have to route through the brain; the brain handles understanding and planning. Two models carry this. Gemini Robotics-ER 1.6 is a Gemini model tuned for spatial understanding and tool use, able to pick individual objects out of cluttered scenes. A separate vision-language-action model takes the visual input and the reasoner’s instructions and outputs motor control. Working with Boston Dynamics, they found tasks that needed more than spatial sense: getting Spot to read a small gauge required tool calls to magnify the dial and reasoning to interpret it.

The frontier, she said, is efficient learning. The ways to collect robot data run from expensive, narrow teleoperation, through physics simulators like MuJoCo, to the prize of learning directly from internet video. World models now sit in that chain. They can generate novel interaction data at high enough fidelity to train the robot model, and Hadsell said you cannot tell the real robot logs from the simulated ones. Pool that with cross-embodiment learning, where data collected on one robot transfers to others, and the performance gain comes “for free.” This is where the talk’s argument closes on itself: the same foundation-model recipe, pointed at worlds, becomes the engine that trains the robots.

Stop staring at language

Her closing note returned to the frustration she opened with: almost all the field’s attention sits on language, and she thinks that is a waste. The same recipe that produced today’s language models can be pointed at far richer systems: worlds, robots, biology, weather, and the diversity of Earth’s biosphere at planetary scale. The data already exists; the models have barely been built. The next decade of progress, on her account, runs through those systems, and has little to do with making chatbots bigger.

The continent lost the sky

Hadrien arrived in Ukraine in the first week of March 2022, a week after the full-scale invasion. What he saw was a kind of war NATO had forgotten how to fight: “a high and long intensity warfare,” not the small expeditionary operations in Africa and the Middle East where the West always owned the air. For the first time in modern European history, he argues, “Ukraine, but also NATO, doesn’t have the upper hand on the air superiority.” The weapon that took it away is almost insultingly cheap. As Hadrien traces its lineage, the Shahed-136 is a slow, low-flying design first drawn up by German engineers, then copied in turn by the Israelis, the Iranians, the Russians, and now the Americans. “Nothing changes through the years,” he said. The economics did.

Most of the time, it misses

I asked him to walk through the data loop, the thing every speaker that day had described from behind a laptop: get representative data, run the model, grade the output, go again. Alta Ares runs it in a field at night. “Most of the time it doesn’t hit,” he said, and that is the whole problem. There is rarely internet at the edge and never a data center, so everything runs on “super compressed” models on small GPUs inside the interceptor. A mission has three phases. Radar detection sorts friendly from hostile. Then an interceptor flies into a kill zone roughly two kilometers wide, where an RGB and an infrared camera hunt the sky for a single pixel that is genuinely hard to find, because the feed is low quality and the airspace is being jammed. Then a human, still in the loop, commits to the intercept. One target is tractable. A swarm is not.

The loop only tightens if you close it fast. Alta Ares keeps around twenty engineers in Ukraine, as close to the front as Hadrien can put them, and combat-tests “every week or every two weeks.” The hardware turns over constantly; the vision models update remotely as more data comes back. They do not call the things drones. “We call them munitions, because it doesn’t come back,” he said. There is an 800-gram warhead in the nose. “I don’t recommend them to come back.”

What no simulation can reproduce

Simulation, he allowed, is good for a proof of concept and little more, and the reason is human rather than technical. The engineer who runs a test slept well, had breakfast, had a good coffee, “almost as good as the tea in England.” The operator who fires the system in Ukraine is on their fifth coffee “that doesn’t wake you up anymore,” at 3 a.m., freezing, running on adrenaline, and far less trained. No amount of lab time closes that gap, and no simulator renders the adversary. The Shaheds his team intercepts have started carrying rear-facing sensors that trigger evasive maneuvers when an interceptor closes in. Others fly in a mesh, so when one is shot down another inherits the firing solution. Much of the underlying technology, he said, is Chinese, routed to Russia and Iran. The lesson Hadrien draws is that classical computer vision is not enough against an opponent that keeps learning, and the only thing that trains a system to beat it is real data from real intercepts, which no simulation can manufacture.

Quantity is the quality

So why isn’t modern air defense everywhere? Because NATO and Europe bought small numbers of exquisite systems on the assumption that they would always own the air, and that assumption is gone. “The new reality is that the quantity is the quality,” he said. A sophisticated interceptor fielded in low numbers gets outgunned. Israel’s Iron Dome works because Israel is small to defend, and even so, by his account, “Israel used 35% of the world’s stock of Patriot missiles in only the first 10 days” of the war. Interception is never a sure thing: “never 100%,” he said, somewhere between 25% and 75% on the best systems. The fix is as industrial as it is technical, and the budgets are still pointed the wrong way. France is spending more, he noted, but just committed 10 billion euros to an aircraft carrier that, he argued, takes a decade to build and that a hundred cheap unmanned boats could sink. He put the politics of it in one line: “during peace time, you have time, but you don’t have money. And during war times, you don’t have time, but you have money.”

The talent turn toward European defense

Talent has become his strange dividend of the moment. “One of our best talent acquisitors, his name is Donald Trump,” he said. When JD Vance told the Munich Security Conference that American and European interests had diverged, well-qualified engineers at US companies in London started leaving to build European defense instead. What he sells them is accountability and proximity: write something in the morning, watch it deployed on the front by night. Alta Ares does no offensive weapons, “so far.” And in his telling, this work is the precondition for everyone else’s. The clinical trials, the healthcare, the AI-for-good that had filled the rest of the day are “possible because we live in a country at peace.” Bomb the data center and none of it ships.

Pressed by an audience member to separate the founder from the person, he couldn’t, and didn’t try. He reached for the old Roman maxim, si vis pacem, para bellum: if you want peace, prepare for war. Europe stopped preparing, and “the use of force is being democratized” faster than its institutions have noticed. You can annex territory now, he said, and no one does anything about it. He was honest about the discomfort of his own business: “I would sleep way better if tomorrow the war in Ukraine stops.” But he doesn’t think it will, and his case for rearming rests on a grim simplicity. The people on the other side keep telling us what they intend, so “we should listen to them, because at least they are being transparent.” Freedom doesn’t come for free.

That was once a hard sell. When Odyssey pitched its seed round, he says, “basically no one understood world models,” and convincing people was “an uphill battle.” The category has since shifted underneath him: the argument flipped around November 2025, money and researchers flooded in, and at NeurIPS 2025 world models were, by his count, the field’s number one or two theme. Hawke came to the problem from self-driving, as a founding engineer at Wayve teaching cars to drive end-to-end - one of the few arenas, he says, where frontier AI has really been tested against reality.

Not every world model is a world model

The term is loose enough to mean almost anything, so Hawke was careful to pin it down. Odyssey uses it in the precise sense inherited from model-based reinforcement learning: a learned transition-dynamics model, the thing David Ha and Jürgen Schmidhuber named in their 2018 “World Models” paper. In plain terms, a model that learns how the world evolves and then samples possible futures one step at a time, conditioned on the actions you feed it. That is distinct from “spatial intelligence,” which models the appearance and structure of a scene (the pitch at World Labs and others); from behavior models, the decision-making brain of a self-driving car or a robot, which Hawke thinks is better named as such; and from the “proxy world models” some researchers argue are already latent inside LLMs. Odyssey’s version is what he calls a general-purpose neural simulator: an interactive stream of pixels that models physics, that you can talk to and that talks back, and that you can reach into and change. “I have many questions about what this means for future products,” he admitted. “I have very few good answers.”

Generality wins

Two principles guide what Odyssey builds, both carried out of self-driving. First, end-to-end learning, a simple model trained purely from data, “almost always wins.” It was an unpopular position in 2018 and is now the default in autonomy and in language alike. Second, “generality wins”: narrow models built for a single vertical rarely keep their lead for long. He couldn’t point to durable legal-specific foundation models, for instance, because “Claude just got better.” The ambition that follows is unsubtle. Odyssey wants to build “the GPT-3 of world models” - the InstructGPT-style moment when a category of model stops being a research demo and starts generating real commercial demand. Odyssey’s research splits into four problems, each with a shipped model against it.

Pixels that keep going…with sound too!

The foundation is autoregressive interactive pixels, embodied in Odyssey-2. A standard video model returns a fixed clip; this one streams frame after frame in real time, which is much harder, because error compounds as the model generates in sequence. It is also interactive, absorbing text prompts mid-stream and adjusting to them. Scaling helps in the usual way - a one-billion to a fifteen-billion-parameter jump lifts the benchmarks - but the real difficulty is keeping the interaction open-ended instead of narrowing it to a single domain.

In May 2026, Odyssey released Starchild-1, named for 2001: A Space Odyssey, which generates pixels and audio jointly in one coherent stream rather than dubbing sound onto finished video. It was harder than expected, Hawke said; to his knowledge no one else had shown one publicly. The obstacle is a clash of timescales - a single predicted video frame spans only “half a phoneme” of audio - so keeping the two coherent took a custom KV-cache design running on two clocks at once.

Shared state, not stitched video

The third problem is multiplayer: shared state across world models, which matters as much for a cell of robots working one environment as it does for a game. Odyssey’s Agora-1 demonstrates it, and Hawke ran it live. The audience pointed their phones at a QR code and played a fully generated game of GoldenEye, streamed off H100s “probably in Spain.” Under the hood it borrows a game engine’s split between rendering and simulation, except both halves are learned - a neural simulator and a neural renderer, trained together. People had said multiplayer world models “weren’t possible,” he noted, “and we felt it was worth disproving.”

Learning by being broken

The fourth strand carries the freshest idea in the talk. Almost everyone who pairs agents with world models uses the agent to get smarter inside the model. Odyssey’s PROWL turns that around: it uses a reinforcement-learning agent to improve the model itself. The reasoning is “garbage in, garbage out” - “your agent will never really be better than the quality of your learned environment model” - and yet, Hawke argued, almost no one had bothered to fix the model rather than the agent. PROWL drops an agent into a learned version of Minecraft and rewards it for finding the world model’s failure modes, then folds those failures into a curriculum that patches them. The wrinkle is that reinforcement learning is very good at cheating: left unleashed, the agent would “sit on the spot, spin the camera around at very high speed” to stall the model’s learning, so the team added a KL “leash” to keep it close to an agent that actually plays the game. The payoff was a clear lift in the base model’s performance - a world model that gets better by being systematically broken.

Still the GPT-2 era

For all of that, Hawke was disciplined about where the work sits. World models, he said, are at “the GPT-2 era”: the pre-ChatGPT stage where the outputs already look promising and the first commercial experiments are forming - he name-checked Jasper, the early GPT-3 copywriting startup - but mass adoption is still ahead. That is the gap between today and the “GPT-3 of world models” he wants to build. The honesty extends to limits. Asked from the floor whether a world model would have to encode general relativity and quantum mechanics, he declined the bait. Visual data is “by far the highest volume” and the right place to start, and these models earn their keep where conventional simulation struggles: for a precise definition of turbulent flow, “use CFD”; for crowds, contact, the messy texture of a scene, the neural simulator wins. Biology and harder physics are wanted, but, he conceded, nascent.

The economics are gentler than the hardware implies. The multiplayer demo ran on a single H100, “let’s call it $1.50 an hour” - close enough to a Netflix subscription that, by his estimate, $30 a month buys ten to fifteen hours of generated play.

Where this goes, on his telling, is convergence: as a visual world model takes on audio and, eventually, text, it meets the language models arriving from the other direction through VLMs, and at some point the two merge. None of that is solved, and he was candid that no one yet knows how. But the direction is set, and the market has caught up to the thesis that once needed an uphill seed pitch. On June 17, 2026, days after RAAIS, Odyssey raised a $310M Series B at a $1.45B valuation, led by Natural Capital with Amazon, AMD Ventures and GV. Air Street backed the company’s seed in 2024, on the bet that learning the world directly would become its own category of model. The GPT-2 era doesn’t last long.

Read more!

A science, not an art

Ted is firm on the framing. These systems are complex, he says, but they are understandable. The catch is that understanding them demands “epistemic humility,” because there are always too many variables in flight and “running a controlled experiment is hard.” The skill he prizes most is knowing the limits of your own evidence: “knowing what an experiment tells you and what it doesn’t tell you is important.” That sounds modest. In a field that markets every result as a breakthrough, it is closer to radical. His own path ran through linguistics and neuroscience, but the thing that carried over, he’s clear, wasn’t brain-inspired architecture - it was the scientific method itself: rigor, skepticism, doubting your own results.

The job also changed his relationship to curiosity. In a PhD you pull on a thread because it’s interesting. In a frontier lab the question is colder: what is the cost-benefit, and is the answer going to be interesting but “ultimately less useful to making Claude smarter”? Experiments that survive that test get fed up to the people deciding the big training runs, where the point is to cut uncertainty before committing the compute.

The metric that matters is the counterfactual

I put a striking set of numbers to him, drawn from Anthropic’s recent essay When AI builds itself: more than 80% of the code merged into Anthropic’s codebase is now authored by Claude, a typical engineer merges eight times as many lines of code per day as in 2024, and on one kernel-optimization task Claude’s speedup climbed from roughly 3x to 52x in under a year. The essay’s own line was that “we have not yet seen that curve bend.” So what, I asked, is the honest measure of AI actually accelerating?

Not benchmarks, Ted said. The real test is “what changes the counterfactual” - if you took Claude away and coded alone, what would you do differently? Concretely, it reduces to the “number of human interventions that are required”: how often a person has to step in, how often the model’s first pass is the one you accept. By that measure the line has moved fast. He pointed to Opus 4.5 in November as a jump, and the Mythos-class models in February as a bigger one. The shift is trust: the models still make mistakes and you still have to check them, “but you can trust them a lot more than you could before.” In the labs, people have already stopped supervising every step - they run in bypass mode and let the agent work.

When bigger models get cheaper

Does the future belong to one large model doing everything, or to companies decomposing tasks across smaller, cheaper ones? Ted leans hard toward the former, and the argument is economic, not sentimental. He cited Noam Brown at OpenAI, who plots test-time-compute curves with cost on the x-axis: a bigger model that reaches an equally good answer in far fewer tokens can come out cheaper than a small one grinding away. “It could be more cost-effective to just ask the bigger model.” His blunter version: “people underrate the value of having a really smart model to ask questions to.” The counter-pressure is real - at Ramp, staff are nudged away from using Opus 4.8 to write emails when Sonnet will do the job - but his bet is that raw intelligence keeps winning on cost as well as quality.

The next axis is taste

The frontier the conversation kept returning to was judgment. The same essay claimed models now pick the better next research step about 64% of the time, up from 51% in November. How do you build a system with taste? Here Ted turned characteristically vague - he wouldn’t say what goes into it - but he would say that when his team started using Mythos this year, it felt like it had better research taste than anything before it. With one caveat: still “not as good as your average researcher.” Average researcher where, I asked. “Maybe at Anthropic.” A high bar to be measured against, and one the models are now climbing.

Safety is a capability, not a tax on it

Ask what safety research has ever done for product quality and Ted reaches for cars. Seat belts, crumple zones, airbags, an oven that won’t catch fire - guardrails are what “make the product usable” at all, and a product nobody can safely use generates no feedback to improve it. The clearest case built the whole category: reinforcement learning from human feedback began as a safety project - stopping a chatbot from spewing garbage - and turned out to be the thing that made chatbots good. “Alignment and safety really go hand in hand with capabilities,” he said. Indeed, reinforcement learning from human feedback draws its roots back to a 2017 collaboration between OpenAI and DeepMind:

One encouraging trend I put to him is that bigger models are turning out easier to align, not harder. Ted’s own framing was more guarded - it has “gone different from how many safety researchers expected,” he said - but his read is that “people feel pretty good about the alignment situation right now,” tempered by the obvious caveat that “we don’t know if we will cross some capability threshold and it’ll suddenly flip.” Hence the case for “healthy apprehension.”

Where the leverage is

I closed by asking what the highest-leverage work in 2026 looks like. Ted was self-aware about his own bias and unhedged anyway: if you believe AGI is close, go to a frontier lab, because your lever there is bigger than it is outside one. He counts himself among the converted - “I’m definitely more AGI-pilled than when I joined.” OpenAI and Anthropic are, he pointed out, still small companies where a newcomer can move things, and the window matters. If you want to influence the direction these systems take, “sooner is better than later.”

That includes London. Anthropic’s office here has gone from 15 or 20 people a few years ago to a couple hundred, with whole strands of frontier work - Ted’s among them - run from the UK rather than mirrored from California. “It doesn’t feel like we’re a satellite,” he said.

What the conversation kept circling back to is that the hard part of scaling was never the spending. Anyone can buy more compute; the discipline Ted’s team is building is the other half - knowing what each experiment will and won’t tell you, and turning that compute into capability you can measure and trust.

That story, which Vivek Natarajan told from the RAAIS stage, is the kind of result his team at Google DeepMind has been chasing. Natarajan is a Research Lead there, working at the intersection of AI, science and medicine. When he last spoke at RAAIS a couple of years ago, the state of the art was Med-PaLM, a language model tuned to answer medical exam questions. His pitch this year was more ambitious: that the recipe behind AlphaGo can be turned on science and the clinic, and that the trick is teaching models to stop thinking fast and start thinking slowly.

System one is not enough

The problem with using a chatbot as a scientist, Natarajan argued, is that even reasoning models mostly do “system one style thinking,” quick responses drawn from surface-level pattern matching. Real discovery is the opposite: slow, deliberate, rigorous, the product of chewing on a problem for weeks until the spark comes. He wanted “system two style thinking,” and to get it he reached back into DeepMind’s own history. AlphaGo’s 2016 breakthrough came from self-play and search, with agents playing each other, taking feedback from the environment, and reinforcing what won. AlphaZero then showed the same recipe could scale from zero knowledge to superhuman play in months, limited mainly by compute.

The AI co-scientist generalizes that idea. Instead of agents playing a game, they generate scientific hypotheses, then critique, debate and refine them over hours and days, what Natarajan calls a “generate, debate and evolve ideas loop.” Borrowing from AlphaStar, DeepMind’s StarCraft system, the team added tournaments: a ranking agent stages pairwise debates between hypotheses, scores them against a rubric derived from the scientist’s stated goal, and assigns Elo ratings, so only the strongest ideas reach the human. Because the debates run in natural language, they can be summarized and fed back into the agents’ context, which is what makes the system self-improving. It also lets the system signal its own uncertainty, what he called “epistemic humility,” which matters when the scarce resource you are spending is a scientist’s time.

The whole project nearly didn’t happen. The idea came from Gary Peltz, a Stanford geneticist who, after one of Natarajan’s lectures, suggested that a model trained on scientific text might generate hypotheses for the causes of rare disease. Most of the team thought it was too early. They did it anyway. Sometimes, as Natarajan put it, you “jump off the cliff, and then you figure out how to build an airplane on the way down.”

From hypothesis to organoid

The Penadés result, run with collaborators at Imperial College on antimicrobial resistance, was the moment the team realized they were onto something. But the more telling cases are the ones that ended in a wet lab. Physician-scientists at Houston Methodist used the co-scientist to find drug-repurposing candidates and combination therapies for acute myeloid leukemia. Peltz’s own lab pointed it at liver fibrosis, a disease with few treatments, and tested its picks in human liver organoids. One candidate, vorinostat, not only showed anti-fibrotic activity but cut TGF-beta-induced chromatin damage by over 91%, a hint of regeneration. The interesting part is that vorinostat is an FDA-approved cancer drug, exactly the kind of cross-field connection a liver specialist might never make, and the system surfaced it because it could read broadly while the human judged what mattered. Natarajan called this complementary intelligence, and it is the honest version of the pitch: the machine goes wide, the scientist goes deep.

He kept the limits in view. The system itself is general-purpose, he stressed, with nothing in the scaffolding specific to biology; the specialization comes from the tools it reaches for at runtime. But asked where it fails, he was candid that the wins so far have been in biology, where a mass of unread literature hides real signal. Chemistry is harder, and fields like mathematics and physics, which reward narrow depth-first reasoning over wide reading, harder still. The common thread is not biology or medicine specifically, but a way of turning compute into disciplined deliberation, then putting the result back in front of expert humans.

Manufacturing medical experience

The other half of the talk was about access. World-class medicine, Natarajan said, is “pretty much a geographic and socioeconomic lottery,” and his team’s second mission is to close that gap. The vehicle is AMIE, a diagnostic dialogue system he co-leads with Alan Karthikesalingam, a vascular surgeon still practicing in the NHS. Asked once how to choose between two doctors, Karthikesalingam told him to “always go with the one who has more gray hair.” There is no substitute for experience, so the team manufactured it. Using the same self-play machinery, AMIE ran consultations against synthetic patients and a critic, refining itself over millions of simulated dialogues. A human doctor might see 10,000 to 50,000 patients in a career; AMIE has already run hundreds of millions of conversations, “building up the world’s most experienced doctor,” with the heavy caveat that it all happens in simulation.

It is starting to pay off. In evaluations published in Nature, AMIE matched or beat physicians in simulated consultations with patient actors, on diagnosis and, more pointedly, on rapport, empathy and relationship building. “I’m kind of sorry about the doctors and humans in your life,” Natarajan deadpanned to the unsurprised. But this is not about replacement, he insisted: “the story is still about augmentation.” A companion study found that general physicians given complex diagnostic puzzles did significantly better with the AI as a thinking partner than working alone or with standard tools like web search. And in an early, supervised feasibility study with Beth Israel Deaconess in Boston, where patients spoke to the AI before an urgent-care visit under physician oversight, zero safety stops were required under the study’s predefined criteria, patient trust in AI rose after the interaction, and the system’s pre-visit diagnoses held up against the attending physicians, all without the benefit of lab tests.

A third person in the room

For most of modern medicine, Natarajan closed, the core unit of care has been a dyad: the doctor and the patient. His bet is that it is becoming a triad, the doctor, the patient and the AI, with the machine as a teammate rather than a tool. That is the idea behind the team’s next effort, an AI co-clinician. It is a tidy frame, and the evidence on stage made it land harder than it would have a year ago. The deeper claim running under both halves of the talk is that the self-play recipe which once mastered a board game can now give scientists and clinicians a new kind of thinking partner: one that searches widely, argues with itself, and hands humans better starting points. The wet labs and the early clinical studies are starting to agree.

At this year’s RAAIS, Roberta Raileanu set out why that ceiling exists and what it would take to lift it. Raileanu leads the open-endedness team at Google DeepMind and was previously at Meta. Her talk laid out a recipe for superhuman scientific discovery: a general system that makes groundbreaking discoveries across domains faster than people can. Three ingredients hold it together, but underneath all three sits one problem. We are good at searching for anything we can measure. We do not yet know how to measure what makes a discovery good.

The plateau before the breakthrough

The past two years delivered a real proof of concept. In 2024, Sakana AI wired LLM agents into a loop that generates a hypothesis, implements it, runs an experiment, and iterates, producing its first machine-written papers. The bar has risen since: a fully AI-generated paper has passed peer review at a workshop attached to a top machine learning conference, and a wave of startups now aims to automate research outright.

Proof of concept is not parity, however. Put the best agents head to head with human experts on the same open problems and the agents improve early, then stall. They are good at variations and combinations of known methods, and weak at what defines real research: exploring unfamiliar paths and making the conceptual leaps that change a field. Scale up compute and time and the human line keeps rising while the model line flattens.

The reason to think the ceiling can move is breadth. These models train on a far wider cross-domain corpus than any scientist can absorb, and can search for connections across more fields than any specialist holds in working memory.

The lesson of Move 37

Raileanu’s first ingredient is to treat discovery as a reinforcement learning problem. An agent acts in an environment, observes what happens, and learns from the feedback, which is not far from how a scientist forms an idea, tests it, and revises. The appeal is specific: as long as you can measure progress with a reward, the agent is free to find any solution that earns it, including one no human would think to try.

The proof is a decade old. When DeepMind’s AlphaGo played Lee Sedol, its move 37 was so counterintuitive that no human would have played it, and it won the game. But Go is a closed world with a clean reward: a move wins or it does not. Move 37 shows what optimization can do once the objective is given. In science the objective is not given. Deciding what counts as progress on an open question is the actual work, and it is the part no reward function hands you.

To study this inside AI research itself, her team built MLGym, a sandbox where an LLM agent runs shell commands, edits files, and runs experiments across tasks from language modeling to game theory. Even a year ago, simple setups could self-improve against a benchmark, but only by tuning hyperparameters and swapping architectures, not by inventing a method a human expert would adopt. That gap, between optimization and originality, is the rest of the talk.

Why greatness cannot be planned

Most breakthroughs, Raileanu argued, are not solutions to known problems. They involve finding the right problem, and “innovation is rarely this linear process from A to B.” Try to build a personal computer in the 1800s and you would not get there by scaling up the abacus; you would need the vacuum tube, which was invented to amplify radio signals.

She took the frame from Kenneth Stanley and Joel Lehman’s “Why Greatness Cannot Be Planned,” and put its claim on the screen: “No prerequisite to any major invention was invented with that invention in mind.” Optimize too narrowly for an objective and you skip the stepping stones that lead to it. Machine learning has won by hill climbing toward benchmarks, and that has carried the field far. But a hill climber only ever reaches the top of the hill it started on.

The fix is to widen the search. Borrowing from evolutionary methods, you hold a population of candidate solutions, mutate them, and select which to keep. The usual fitness function rewards performance alone. Raileanu’s argument is to score for what scientists actually value too: novelty, diversity, interestingness. This is where the signal problem surfaces in the open, because none of those is easy to measure. Her stopgap is to let an LLM judge what a person would find interesting, which at least keeps ideas that are not useful yet but might combine into something later.

Her team’s Rainbow Teaming did this for AI safety, generating diverse jailbreak prompts across a grid of risk categories and attack styles, then reusing what worked in one cell to seed another. Train on the result and the model gets measurably harder to break. The same machinery, she suggested, should carry over to ideas and methods, where a solution built for one field can matter in a completely different one.

Optimizing discovery itself

The third ingredient is to stop optimizing discovery on a single task and optimize the process of discovery itself. DeepMind has trained RL agents across huge numbers of simulated environments and watched them adapt to new ones far faster than before, in some cases faster than humans. If that transfers to research, an agent could generate its own tasks and get better at discovering, not just at one discovery.

To make that studyable, her team built DiscoBench, a framework that procedurally generates AI research tasks: more than 400 million of them, across problems like language modeling and image classification, with an agent free to target a loss function, an optimizer, or an architecture. The early signal is encouraging, with more and more diverse training tasks improving performance on held-out problems the agent has never seen.

The complementary curve

Stack the three together and you have the recipe: reinforcement learning to discover better solutions where progress can be measured, divergent search to find new problems rather than climb known ones, and meta-learning to speed up the whole process on problems no one has posed yet.

The bet underneath it is complementarity. Humans go deep in one field; a model reaches across many at once. The line worth chasing is neither the human curve nor the machine curve, but the one above both, where the two discover what neither would alone. Yet all three ingredients run back into the same wall. We have good algorithms for search once we know what to reward, and we still cannot reward novelty, a promising dead end, or taste. “The key is, do you have the right signal?” Raileanu asked. The search is the easy part. The missing piece is the signal.

You’ll now find updated data for AI research papers using NVIDIA, TPUs, Apple, Huawei, AMD, ASICs, FPGAs, and AI semi startups. We’ve also expanded the infrastructure side of the index to 1 July 2026: A100, Hopper, standalone Blackwell, Grace-Blackwell, and a new demand-side view of frontier-lab contracted compute in gigawatts. Each chart can now be downloaded, shared, and embedded.

A few notes upfront: The 2026 citation figures use real counts through June 1, 2026 plus a volume-adjusted projection for the rest of the year. Year-over-year deltas are calculated against final normalized 2025 counts, not last year's mid-year 2025 projection. GPU-count charts show NVIDIA data-center GPUs by owner/operator, split into Deployed, Installing, and Announced. Grace-Hopper and Grace-Blackwell parts are counted by GPU dies, so one NVL72 rack equals 72 GPUs. Tenants are not double-counted against the operators whose clusters they use.

Subscribe now

The breather was short

Last year’s update asked whether 2025 was the first real slowdown in open AI compute citations after six years of growth. With final normalized 2025 counts now in hand, and 2026 projected from counts through June 1, the answer looks clearer: 2025 was a pause, not a rollover.

Across the tracked accelerator categories, the 2026 projection reaches 49,339 chip-citation counts, up 10.7% year-over-year and just above the 2024 peak. NVIDIA remains the default: its chips appear in 44,715 of those counts, up 10.9% year-over-year and about 91% of the tracked total.

That does not mean frontier labs have suddenly become more transparent. The largest model developers still publish less of their best work, and many papers built on managed APIs or shared cloud services do not specify the underlying silicon at all. But the open literature has not stopped reflecting hardware diffusion. The 2025 dip looks more like publication-cycle timing, API abstraction, and a quiet year between hardware waves than a collapse in compute usage.

The more interesting finding is the relative lack of upheaval. A lot of companies, labs, and governments are trying to challenge NVIDIA. In the open research literature, they are still mostly not succeeding. There is movement in AMD, Huawei, and Apple, but NVIDIA remains the language researchers use when they describe their compute.

Indeed, AMD citations nearly doubled, from 251 in 2025 to 472 in 2026. Huawei Ascend 910 rose 56%, from 137 to 213. Apple moved from 741 to 998, overtaking AMD to become the most-cited non-NVIDIA, non-Google accelerator in the open literature, which means consumer Macs now out-cite the leading silicon challenger. That likely reflects the spread of local inference and developer workflows rather than frontier training. TPUs, meanwhile, declined 7% in the open paper data, despite Google’s obvious importance to frontier AI compute. That tension is a useful reminder that paper citations are a leading indicator for some forms of adoption, but a poor measure of private, API-mediated, or closed-lab usage.

Subscribe now

NVIDIA is still king, but the kingdom is changing shape

The NVIDIA chart now approximates their product-cycle chart.

A100 citations are basically flat at 15,327, up 0.4% year-over-year. The A100 is no longer where the growth is. H100/H200 citations have more than doubled to 9,823, up 111% year-over-year, as the 2024 and 2025 Hopper build-out finally works its way into papers. Blackwell-family mentions, tracked as B100/B200/B300 in the Zeta Alpha data, are still small at 902, but are up 4.5x year-over-year.

Meanwhile the older stack continues to drain out: V100 citations are down 34%, RTX 3090 down 23%, P100 down 36%, and K80 is now barely visible. The RTX 4090 is still useful in the academic long tail, up 9% to 6,557 citations, while the new 50-series cards are appearing quickly from a small base.

Startup silicon is no longer one story

Groq is now the most-cited startup chip, up 49% to 264, and in December, NVIDIA acquired it. The fastest way to dent NVIDIA’s share turned out to be getting bought by NVIDIA. Cerebras is second at 242, up 8%. SambaNova rose to 52, Cambricon to 33, while Graphcore fell to 40 and Habana to 24.

That is not yet a market-share story. Startup chip citations remain tiny next to NVIDIA, and papers using startup silicon still often include authors from the chip company itself. But it does show the category splitting into different jobs. Groq is showing up around low-latency inference, Cerebras around large-scale wafer-scale systems, and the older acquired or de-emphasized platforms are fading from view.

The right conclusion is not “startup chips are breaking CUDA.” They are not. It is that the category is specializing into niches, inference latency and wafer-scale, while NVIDIA keeps the general case.

Hopper is the installed base

The second half of the index is not a citation tracker at all. The cluster charts show how much of the AI build-out has already moved from research procurement into industrial infrastructure.

Across the tracked Hopper systems, we count 460,904 deployed H100, H200, and GH200 GPUs, plus 1,328 installing. That is more than 11x the tracked A100 total of 41,208. The A100 chart now reads like a legacy fleet chart; Hopper is the live installed base.

xAI Colossus 1 is the largest tracked Hopper deployment at 200,000 GPUs, and by itself is almost five times the entire tracked A100 set. Tesla Cortex follows at roughly 66,000 H100-equivalent GPUs, then Meta’s GenAI clusters at 49,152, a CoreWeave H200 cluster at 42,000, Voltage Park at 24,000, and Germany’s JUPITER Booster at 23,536 GH200.

There is a geopolitical point hiding in the table. National HPC systems are exact and visible, which makes them easy to count. Private fleets are estimated, harder to observe, and much larger. Europe now has serious machines in JUPITER, Alps, Isambard-AI, Leonardo, MareNostrum 5, and Jean Zay. But the largest private AI clusters are operating at a scale national supercomputing programs mostly do not match.

Subscribe now

Blackwell is mostly pipeline

Standalone B200/B300 deployments in the index total 20,024 GPUs, led by Deutsche Telekom’s Munich Industrial AI Cloud at 10,000, then SoftBank’s DGX SuperPOD and Korea’s Naver Cloud at 4,000 each, E2E Networks in India at 1,024, and SK Telecom’s Haein cluster at around 1,000. IREN’s 50,000 B300 order sits in Announced.

Grace-Blackwell is where the real pipeline sits. The index tracks 100,128 deployed GB200/GB300 GPUs, 308,640 installing, and 1.66 million announced. In other words, just under 5% of the tracked Grace-Blackwell pipeline is deployed, and roughly 80% is still announced. The deployed number crossed 100,000 this month as the first sizable GB300 systems came online, including CoreWeave’s 8,192-GPU cluster, which posted MLPerf training results in June, and Mistral’s 13,800-GPU Bruno in France moving from installing to live. It is a small share of the pipeline, but it is the first month the deployed bar moved on the strength of Grace-Blackwell rather than Grace-Hopper.

The largest announced and installing programs still look more like sovereign or hyperscale industrial projects than normal data-center procurement: HUMAIN in Saudi Arabia at up to 600,000 GB300, Stargate Abilene at a 450,000 GPU target, South Korea’s national 260,000 Blackwell program, Nscale’s 200,000 GB300 commitment to Microsoft, xAI Colossus 2 at roughly 110,000 installing, Argonne Solstice and Equinox at 110,000 combined, and Stargate Norway at 100,000 announced.

This is why “GPU count” is becoming an incomplete question. The constraint has moved outward to power, land, cooling, interconnect, permitting, debt, and offtake. A GPU order is not a cluster. A cluster is not always available capacity. And contracted capacity is not the same as a model training run.

The demand side is now measured in gigawatts

The frontier-lab contracted compute chart looks at the other side of the market: not who owns a cluster, but which labs have contracted capacity in disclosed gigawatt terms.

For GW-disclosed deals, OpenAI is now at 26 GW of contracted compute: 10 GW of NVIDIA systems and 16 GW of non-NVIDIA, the latter split between a 6 GW AMD MI450 commitment and 10 GW of Broadcom custom silicon. Anthropic is at 6 GW total, split between 1 GW of NVIDIA and 5 GW of non-NVIDIA capacity.

On disclosed gigawatts, then, OpenAI’s non-NVIDIA book is now larger than its NVIDIA book, and larger than Anthropic’s entire portfolio. It is worth being careful about what that does and does not say. NVIDIA is still OpenAI’s largest single-vendor commitment and its fastest path to scale; the non-NVIDIA figure is spread across AMD and a custom Broadcom part that has not yet shipped at volume. Anthropic’s non-NVIDIA number is, if anything, understated, because its AWS Trainium and Google TPU commitments are only partly disclosed in gigawatts and are therefore not fully charted.

Still, the direction is clear. Frontier labs are no longer choosing a single chip vendor. They are assembling compute portfolios: NVIDIA for the broadest software ecosystem and fastest path to scale, TPUs and Trainium for strategic supply and cost control, custom silicon for future leverage, and neoclouds or sovereign projects when hyperscaler capacity is not enough.

“CUDA is still king” remains true in the papers. But in frontier-lab procurement, the question is becoming broader: who can turn contracted gigawatts into reliable, liquid-cooled, networked, usable intelligence infrastructure?

Subscribe now

Looking ahead

Several known unknowns will shape the next Compute Index update:

• How quickly announced Grace-Blackwell projects become deployed clusters rather than press releases, now that the deployed bar has finally started to move.

• Whether OpenAI’s Stargate program, HUMAIN, South Korea’s national Blackwell program, and Nscale’s Microsoft commitments hit their published timelines.

• Whether AMD MI300/MI350, Huawei Ascend, TPUs, and Trainium show up more visibly in open paper metadata, or remain hidden behind closed-lab and cloud-abstraction layers.

• Whether Blackwell appears in the literature in late 2026 the way Hopper appears in the 2026 data now.

• Whether OpenAI’s and Anthropic’s non-NVIDIA books convert from contracted gigawatts into deployed, usable capacity, or stay on paper the way the Grace-Blackwell pipeline mostly has.

• Whether national AI factories can close any of the gap with private frontier-lab infrastructure.

The headline from v7 is simple: the open literature has rebounded and NVIDIA remains the default; the challenger story is real but still small in the papers; Hopper is the installed base while Grace-Blackwell has only just begun to land; and on the demand side, even OpenAI’s disclosed book is now majority non-NVIDIA, even as most of that capacity is still a contract rather than a cluster.

See the live charts here: www.stateof.ai/compute.

A few notes

• We take the view that usage of chips in AI research papers by early adopters is a leading indicator of broader industry usage, but not a complete measure of closed-lab or API-mediated compute.

• Paper-citation counts are unchanged from v6. Zeta Alpha’s open-source AI paper index refreshes annually; 2026 figures are real counts through June 1, 2026 plus volume-adjusted full-year forecasts, and year-over-year comparisons use final normalized 2025 counts as the baseline. This update refreshes the infrastructure charts to July 1, 2026.

• GPU-count charts use public disclosures, operator materials, EuroHPC, Top500, SemiAnalysis, The Next Platform, Data Center Dynamics, company filings, and NVIDIA materials. Large private-company figures are best-available estimates; national HPC figures are exact where published.

• Announced capacity is labeled separately from deployed and installing capacity. We do not invent cloud instance counts where vendors do not disclose them.

  Subscribe now

Angelos leads ElevenLabs’ speech-to-text and text-to-speech teams and built its Scribe and Scribe Real-time transcription systems. The Scribe V2 models he shipped this past year rank, he says, as the most accurate transcription models on most popular benchmarks. It gives him a particular vantage on the problem, since voice models live or die on latency and cost at scale. It is also the second year running ElevenLabs has taken the RAAIS stage; in 2025 its CEO, Mati Staniszewski, spoke on the voice frontier. This year, we dove into the engine room.

What a token actually costs

Every optimization starts with knowing what you are paying for. For the autoregressive transformers behind most popular LLMs, a token’s cost reduces to two bottlenecks: compute, how fast the GPU does the matrix multiplications, and memory bandwidth, how fast the GPU’s VRAM can load the model’s weights and its KV cache. Generation runs in two phases. A prefill step reads the whole prompt and fills the KV cache, the model’s working state, and is compute-heavy. A decode step then emits tokens one at a time, each conditioned on the last, and is memory-heavy. The KV cache is what lets the model reuse that prefill instead of recomputing it for every new token, and at scale it is the thing that hurts: a hundred concurrent requests need a hundred separate caches resident in memory. Size is not destiny either. Angelos noted that Qwen 3’s cache costs almost three times as much per token as Qwen 2.5’s despite near-identical parameter counts, so two models of the same size can cost wildly different amounts to run.

Stop letting the GPU sit idle

The first and biggest single win is batching. GPUs are excellent at parallel work and poor at sequential work, and the dominant cost in decoding is loading the model weights, which can be shared across every request in a batch rather than reloaded for each. Naive batching groups requests once and then waits for the slowest one to finish while the GPU idles. Continuous batching fixes that: it batches at the level of each decode or prefill step, so a new request can join a GPU already mid-flight on others.

Shrink the weights, then the cache

With the GPU busy, the constraint becomes memory, so the next moves all reduce it. Quantization comes first. Models are usually trained at BF16, sixteen bits per weight, which is more precision than they need; dropping the weights to FP8 roughly halves their footprint with near-lossless accuracy, given H100-class hardware and a little quantize-aware training, which injects noise into the gradients so the model learns to tolerate the lower precision. That buys headroom for more cache and lifts throughput to twenty users per GPU. The more aggressive options exist too: int4 is lossy but useful on-device, and MXFP4 reaches four bits but only on Blackwell and newer.

Speculative decoding comes next. A cheap draft model proposes tokens and the big model verifies them in a single forward pass, accepting the run until the two disagree. It only pays off when the models agree often, which they frequently do not, so in practice it is used less than its reputation suggests; applied here it nudges the running total from twenty to twenty-eight users per GPU. The more popular cousin is multi-token prediction, where the same model wears extra prediction heads and drafts several tokens itself, with no second model to host. It earns its keep with two or more heads, and it doubles as a training signal: teaching a model to anticipate several moves ahead, like a chess player, tends to make it more stable and sometimes faster to learn. Most big labs use it, and it lands in the same place, around twenty-eight users per GPU.

The largest gain is also the riskiest. The KV cache holds far less redundant capacity than the weights do, so compressing it is genuinely lossy. Angelos was candid about this from his own testing: the much-discussed Google method TurboQuant was announced as lossless but, in his experience, proved lossy in practice, because any change to the cache is hard for the model to recover from. The fix is again on the training side: distill the model so it grows accustomed to a lower-precision FP8 cache, and you keep most of the accuracy while shrinking the cache 2.5x. That single step lifts throughput from twenty-eight to seventy users per GPU - seventy times what the same hardware served at the start.

Where the frontier labs go

Seventy is what disciplined engineering gets you. Going further means changing the architecture itself, and here the labs are placing different bets. DeepSeek’s multi-head latent attention squeezes each token’s key-value pair into a small latent rather than storing it in full, which both speeds inference and stretches context toward a million tokens; it was one of the more copied ideas after DeepSeek-R1. Qwen swaps standard quadratic attention for a linear network on every other layer, cheaper and longer-context, at some cost to quality. NVIDIA goes furthest, replacing the transformer on a fraction of its layers with state-space models that scale linearly and compute faster, keeping enough transformer layers to hold accuracy up. With architecture-level changes like these, the ladder reaches roughly a hundred and forty users per GPU.

Nothing here is free

Angelos was careful not to oversell any of it. Every technique on the ladder carries a cost. Batching adds latency and runs into a memory ceiling. FP8 quantization takes a small quality hit without the extra training. Speculative decoding needs access to weights and a training pipeline to work well. KV cache compression is the one most likely to degrade output, so the real question is how much degradation you can absorb rather than whether you can avoid it. He was blunter still about the gap between papers and production: many compression methods that report no loss of accuracy were tuned on a handful of benchmarks, and scaled to millions of users they can simply fall apart. You often only find out which ones once they are popular enough to be stress-tested in the wild. Which technique pays depends entirely on the workload.

The reason any of this matters beyond the engineering reached the room through a question from the floor: today’s token prices are subsidized, by one audience estimate a factor of ten to forty. Angelos’s hope is that optimization, not subsidy, eventually closes that gap. The largest models, he said, have to be subsidized to make economic sense, but he expects smaller, Sonnet-class models to become good enough for nearly all everyday use, at margins that actually work. He already sees the shape of it in agent systems: route each request to the smallest model that can handle it, and reserve the expensive one for planning.

This week STARK, the German defense company building exactly these systems, raised €500 million led by Sequoia and Founders Fund. I first met STARK’s CEO and Founder, Uwe Horstmann, back at Project A in 2016, years before any of this. Air Street Capital invested because we believe STARK is emerging as a German neoprime: a new prime contractor for unmanned strike systems, built software-first and able to manufacture at the scale a real war demands.

STARK builds autonomous platforms, fast

STARK’s flagship platform, Virtus, is a loitering munition: a drone that flies to a contested area, waits, finds a target, and strikes it. It is not a reconnaissance drone, though it can return and land like one, and it is not a cruise missile. It sits between the two, cheap enough to expend, autonomous enough to find a target on its own, and simple enough to build at scale. Around it STARK is building a family of effectors - Gambit, a man-portable short-range munition, and Cascade, a tube-launched one - alongside Vanta, an unmanned surface vessel that takes the same idea to sea.

Mass you can afford to build

For decades, Western firepower meant a small number of exquisite platforms, each costing millions and far too precious to lose. Ukraine inverted that. The decisive weapon turned out to be the cheap one you can field by the thousand and expend without flinching. A Virtus assembles in roughly ten minutes, and Germany has already put STARK on a framework worth up to €2.8 billion to supply the Bundeswehr, a landmark deal for a new defense company.

Once the weapon is cheap, the constraint moves to manufacturing and a robust supply chain. More than 80% of this round goes into manufacturing and R&D, and STARK is standing up production lines across Germany, the UK, and Ukraine. In a war of attrition, throughput is the moat: the side that can keep replacing what it loses sets the tempo. STARK builds on civilian supply chains and simple lines that stand up in weeks and replicate from city to city, which is both a way to scale and a way to survive being targeted. The defense companies that win this decade will treat the production line as the product.

Lastly, with a growing portfolio of autonomous products, STARK needs software to tie them all together: Minerva. The same software that swarms Virtus in the air coordinates the unmanned boats at sea, navigates when GPS is jammed, and plugs into NATO battle-management. Each platform improves with a software update rather than a new airframe, and every deployment teaches the whole fleet.

Why STARK

Uwe spent a decade as a general partner at Project A, one of Europe’s most active defense investors and an early backer of Quantum Systems, before leaving to run STARK. He has assembled the four branches this company needs at once. To build at scale: Martin Rost, sixteen years at Zalando running a roughly €10 billion unit, and Johannes Schaback, a repeat founder who was CTO of SumUp. To sell into defense: Jan-Patrick Helmsen, former CEO of Rheinmetall's weapons-and-munitions business. To win the politics: Johannes Arlt, until this year a member of the Bundestag's defense committee. And to learn from the war as it is fought: a Ukraine team running an R&D and production hub in Kyiv that turns frontline feedback into design changes.

Regular readers will be familiar with our defense thesis at Air Street Capital. We recently led the Series A for Alta Ares because Europe has to build its own air defense shield. In our letter from this year’s Munich Security Conference, we argued that Europe must move from emergency buying to structural production capacity, and that when a government contracts a domestic firm it confers the industrial gravity that pulls in private capital and localizes supply chains. STARK is what answering that call looks like: sovereign manufacturing, software-defined systems, and a founder building for the war that is actually being fought. The mass that wins the next conflict will be cheap, built at home, and improved in software, and we’re here to stand behind the people making it happen.

The difficulty is in the control plane around it: one gateway, one governance layer, measurable fallbacks, cost controls, layered human review, and a way to run all of it inside a regulated bank that serves more than 70 million customers across over 40 countries. Revolut ships more than 200 products and has handled over a trillion dollars in transactions, with a machine learning model now in the path of almost every one of them. The leverage, in Donets’s telling, has moved from the model to everything around it.

Four constituencies, one bottleneck

For years, Revolut’s AI was classical machine learning: fraud and transaction models shipped through three internal libraries, one each for training, serving, and performance monitoring. Then, in 2022, the ground moved. Vendors began exposing large models behind an API, and suddenly you did not have to train anything to build something. Generative use cases started growing exponentially while the classical models kept running underneath.

Donets spent as much time on the people problem this created as on the technical one. Four internal groups pull in different directions: researchers who want compute and freedom to explore; builders who want one common API and to ship today; operators who want predictability, rollbacks, and cost under control; and a compliance function that owns human-in-the-loop controls, security audit, and data sovereignty. Left to themselves, every product team solves the same problems its own way, and governance fragments into tribal knowledge spread across hundreds of teams. That is expensive, and it does not scale.

Govern the use case, build one gateway

Rather than govern each model one by one, Revolut made two moves that changed the shape of the problem. The first shifted the unit of governance to the AI use case, a move that lines up with the EU AI Act’s use-case-based view of risk, so that one set of risks, budgets, and rules can cover several models at once and match policy to context. The second put a single gateway at the center of the company, with the governance layer on top of it, rather than shipping capability as libraries each team installs for itself.

But there’s a tradeoff: whereas libraries push reliability onto whichever product team owns the service, a central gateway makes one team responsible for everyone. Even so, the cost of improving a library means cutting a release, then persuading hundreds of busy teams to upgrade and absorb breaking changes they never wanted. With one gateway, the central team ships the improvement once and every product inherits it at, in Donets’s phrase, “zero effort.” Compliance and monitoring move to the same place. As a result, Revolut runs roughly twice as many generative use cases as classical ML ones, all off that single platform.

What breaks when the model is someone else’s

Once you are renting frontier models rather than training your own, you inherit failure modes you do not control. Pay-as-you-go providers run at around 98.5% uptime, which sounds high until you count the hours of dead service it implies each month for a scaled, global business. So Revolut wires a fallback chain into every generative product: if the primary model degrades or stops responding, traffic rolls to the next, and the next. Slightly degraded service beats no service at all.

Subtler, and more painful, was a failure they could not see at all. Because the platform watched only inputs and outputs at the interface, a model buried in the fallback chain quietly stopped working and nobody noticed. “Everything was fine, uptime was high enough, but the model itself was not functional,” Donets said. Or, as one of his slides put it: without per-model visibility, a model doing nothing looks exactly like one that works.

Money was the other lesson, and an easier one to swallow. Teams reach for the newest and most expensive model by reflex, but most workloads are over-provisioned, and right-sizing the model to the task cuts cost by as much as eight times with no loss in quality. Donets’s rule: do not default to the newest model in production. Measure first, then use the smallest model that clears the bar.

A note on the org chart

Underneath the platform sits an org chart doing as much of the work as the code. Revolut is flat and built as a matrix: AI engineers are embedded in product teams, each staffed to ship end to end, with a functional line back to the platform. Standards and tooling flow down; field requirements flow up to Donets’s central group, which sets direction and pushes compliance rules out. He called the product teams “our forward-deployed engineers,” the mechanism by which one team’s hard-won experience becomes everyone’s. The architecture, as one slide noted, ends up shaped like the org chart.

From Rita to AIR

Where all of this lands is a single product that has been running for years. It began as Rita, a support chatbot built on intent models and pre-filled scenarios - the “slot machine” era - that frustrated as often as it helped. In 2022 the team tested large models, Bloom and BloomZ at 175 billion parameters, and found they worked. The first thing they put into production was mundane: paraphrasing a multi-screen FAQ into a short, relevant answer. LLM-based Rita reached production in Q2 2023, then rolled out country by country, Europe first and Japan the hardest, finishing around Q1 2025.

Voice came next, and it runs on a simple pipeline: audio is transcribed, a small LLM decides whether to answer directly or hand off to the full multilingual chatbot, and an end-to-end response comes back in under two seconds. It now runs in 20 countries, handles around 25,000 calls a month, and resolves a customer’s problem roughly eight times faster than a human agent. AIR, the latest layer, followed in Q2 2025 and pulls in transactional data: it can break down your spending, propose hotels inside a budget computed from your own history, or explain why a stock is moving. Across the arc from the old chatbot to today, the share of cases resolved without a human climbed from 17% to 80%, Net Promoter Score went from low to high, and the financial impact, Donets said, ran into double-digit millions of pounds. AIR began rolling out in the UK in April 2026, where Revolut says it has 13 million customers.

Where human oversight is mandatory

Holding all of it up is the monitoring layer. Revolut stores every input and output and runs a panel of LLM “judges” against live traffic - one dedicated to hallucination - currently 9 to 12 mandatory metrics and rising, backed by human review teams that sample chats and transcripts, and by the blunt signal of Twitter and Reddit when something goes badly wrong.

Above all of it sits a hard line: no decision that can change someone’s life is made by an AI system. That position got tested in the room. An audience member who works on regulated healthcare AI pushed back - in his field, he said, “humans make that process unsafe,” so an AI judge might be the safer choice. Donets gave ground on the evidence, agreeing that machines “provide more stable and better help to users,” but not on the principle: the critical calls still do not go to the model. Asked how soon that might change, he did not hedge: “This year, definitely no.”

The frontier gets the headlines, but shipping AI inside a regulated bank across 40 countries is won or lost on the plumbing beneath it: one gateway, the right unit of governance, a fallback for when the vendor fails, and a human who still gets the last word.