OpenAI and Anthropic are growing about as fast, or faster than any company of their size has grown in history. It’s extremely rare for a tech company making more than $1 billion to be growing over 100% per year.1

In fact, both labs are way above that threshold. OpenAI tripled its revenue run rate (recent revenue extrapolated to an annual rate) in the past year, from $13 billion last August to over $40 billion now. Anthropic grew its revenue from $1 billion to $9 billion in 2025, and its growth rate accelerated from there: Anthropic more than tripled its run rate in just the first quarter of 2026. Reportedly, it reached $65 billion by the end of July.2

Now, some of these eye-popping Anthropic growth rates should be taken with a caveat: Anthropic was coming from behind and overtaking its initially larger rival, OpenAI.3 To factor this out, we might look at Anthropic and OpenAI combined. Their growth rates have been both extremely rapid and remarkably robust since 2023: together, they tripled in 2024, and then grew more than fourfold in 2025. And this year they’ve already grown by 3.5×, from $30B to $105B combined, and it’s only August! Remember: the 2025 growth rates were already record-breaking, and now these labs are growing faster from a higher baseline.4

How should we update from 2026’s hypergrowth?

A priori, it’s natural to think that 3× or 4× annual growth of an industry the scale of frontier AI today is just unsustainable. Perhaps LLMs are seeing rapid technical progress and rapid diffusion because these technologies are so new. But as these companies grow, especially when they approach the size of the largest firms in the world today with hundreds of billions in annual revenue, this growth will slow. The frontier labs will run out of technical low-hanging fruit, their models and products will mature, and the adoption curves will saturate. Revenue growth rates will slow down, from 3× per year to 2×, then 50% per year, then 20-30% per year (a la mature tech giants in recent years).

In this light, the fact that revenue growth was so robust in 2025 vs 2024 was surprising, and the 2026 acceleration is simply astonishing. So how should we interpret it?

One approach is to see this as a temporary acceleration, driven by the labs’ coding agents reaching a critical threshold around the time Opus 4.5 was released late last year. We previously saw this pattern with ChatGPT’s original launch. GPT-3 was a niche product earning OpenAI at most tens of millions of dollars in revenue per year.5 ChatGPT, and the subsequent GPT-4, led to an explosion in usage growth and revenue, taking OpenAI to $1 billion annualized in 2023, or more than tenfold growth. OpenAI’s revenue growth then slowed down to 3× in 2024.

Similarly, the explosion in revenue from coding agents might settle down very quickly in the coming months, and frontier AI might finally enter a period of decaying growth and maturation. Of course, this depends on what critical thresholds lie beyond Opus 4.5-level coding agents — e.g. undertaking large projects, or going outside coding to broadly automate white collar work. Given that growth curves are often just a series of stacked S-curves, we’re left with the question of how many of these thresholds lie ahead of AI now.

Will the unusual pace of growth continue? Each year that frontier growth does not slow down could be interpreted as evidence that AI is simply defying gravity. Unlike other tech industries that mature and saturate their markets, AI is simply on a path to automate and transform the entire economy. Or, will this be just the last major growth spurt in frontier LLMs, or AI generally? Gravity may reassert itself, and frontier AI’s growth will slow down to match previous industries.

The answer partly depends on whether revenue growth comes from progress or diffusion. Progress continues as long as progress continues.6 Diffusion meanwhile slows significantly after a few years. Obviously, both have played a key role in the revenue growth of the last four years. But it matters whether each new level of capabilities creates its own diffusion curve, which suggests each step up leads to further revenue growth, or whether frontier AI in general just saturates.

AI vs the world

Anthropic and OpenAI together are now earning revenue at a rate of $100 billion per year. Exponential View estimates that the whole generative AI market is approaching $200B/year in total. This is already around one-thousandth the size of the world economy.

This puts a real-world handle on the usefulness of today’s AI. Claude and GPT models are currently good enough that the world pays almost ten billion dollars per month to use them; if model progress flatlined tomorrow, revenue would almost certainly continue to grow for a while due to diffusion. If both benchmark scores and revenue grow briskly in the next year and combined revenue exceeds $300 billion per year, then even if capabilities progress freezes starting in August 2027, AI will still eventually be a trillion-dollar industry.

But if frontier AI continues to grow its revenue 3× per year, or ~10× every two years, then naively — very naively — it would take about six years for it to grow to the size of the current world economy.7 This is not really a reasonable forecast, but it does provide a good frame to understand the current growth rate. Either frontier AI growth slows down in the next six years, or it will have dramatically reshaped the entire economy.8 If frontier AI triples its revenue again in the next year, that’s one down, five left to go.

This is why we’re watching these numbers so closely. Each passing year, or frankly each passing quarter with continued hypergrowth in frontier AI revenue provides important evidence on AI’s trajectory.9 In the short run, it means that AI has proven substantially more useful than it was several months ago. Revenue, and the investments motivated by that revenue, also feeds into the compute feedback loop, wherein better AI begets more revenue begets more compute begets better AI.

But it’s also crucially important how robust this revenue growth is: when will we see the growth curve start to bend, if ever? This tells us whether frontier AI is on a fundamentally different trajectory than the technologies we’ve seen so far.

We detail:

• The size of the underestimation: US GDP growth over the last year has been underestimated by about 0.3 percentage points. If Nvidia’s growth continues at its current pace, this gap could widen to almost two percentage points of growth per year by 2028.

• The cause of the underestimation: GDP statistics miss most of the value created by fabless chipmakers like Nvidia, whose products are designed in the US but manufactured, assembled, and sold abroad. Because no physical goods leave the US, no goods export is recorded, and because no foreign buyer pays explicitly for the IP, no IP export is recorded either.

• How we know the value is missing: We reviewed every category where Nvidia’s value-add could plausibly be recorded, including goods exports, IP exports, service exports, and merchanting, and it appears in none of them. We confirmed this analysis with the Bureau of Economic Analysis.

• Why this matters now: This blind spot is not new, and it applies to other factoryless manufacturers besides Nvidia. Historically, the value that slipped through was small. Nvidia’s rapid growth has changed that.

• How to correct the underestimation: International guidelines updated in 2025 already call for recording factoryless manufacturers’ overseas sales as goods exports. We also outline an alternative: recording their markups as IP exports. This would depart from international standards, but it would sidestep political resistance to counting overseas production as US manufacturing. Either change could take years to implement. Until then, GDP growth will remain understated.

You can read the full report on Epoch’s website.

We thank JS Denain, Lucio Melito, Mike Waugh, Benny Kleinman, Greg Burnham, Josh You, and Elliot Stewart for their helpful feedback and support.

In this post, I lay out some big questions I have about AI capabilities: questions I think govern the impact that AI will have on the world. These questions animate much of our benchmarking work at Epoch: we aim to build benchmarks, and conduct analyses using benchmarking data, that help answer these questions.

The questions broadly get at a mix of economic impact as well as the more fundamental drivers of AI capabilities. For each question I’ll say why it matters and mention some current benchmarking work trying to make progress toward answering it.

If you find these questions interesting and would like to help us answer them, apply to work at Epoch!

1. Can AI do my job?

More specifically, can AI go from doing narrowly-scoped tasks to doing messier, open-ended jobs? Epoch’s polling shows that, when people use AI for work, they still mostly only use it for part of a task. If this were to change, it would presage more significant labor market disruption — as well as continued fast revenue growth for AI model developers.

The majority of benchmarking to date has focused on relatively narrowly-scoped tasks: fixing bugs, solving math problems, writing reports. Some benchmarks push this farther. MirrorCode asks for large software packages to be implemented from scratch, albeit in a highly structured setting. Remote Labor Index extracts real projects from freelancing platforms and has humans grade AI deliverables vs. human reference work. My favorite, though, bites every bullet: Andon Café is a real cafe owned and operated by an AI agent. I expect we’ll see a lot more of this type of benchmarking, though perhaps in the form of repeated case studies rather than traditional benchmarks.

2. Is AI making progress in areas where we eventually expect it to have a big economic impact?

The “ChatGPT moment” came once AI could have a reasonably coherent conversation about almost any topic. The “Claude Code moment” came when AI could take a reasonable crack at almost any coding task. What will the next moment be? We can’t expect to predict exactly, but we can place some bets as to what capabilities will unlock near-term impact, and benchmark those capabilities.

Perhaps the most salient area right now is cybersecurity: when will AI be able to hack almost any vulnerable system? Cybersecurity benchmarks arguably were already ringing alarm bells, even before the Hugging Face incident.

Another area that comes to mind is computer use. Speaking proverbially, “my dad” uses AI all the time now, but only via chat apps — nothing agentic. My guess is that fast, reliable computer use will unlock another leg up for AI adoption, even if we can’t say exactly what degree of speed and reliability is necessary. Computer use benchmarks can help us track this.

I also think it’s worth tracking physical industries. Not necessarily robotics, but: can AI walk a less-skilled technician through repairing a broken machine on a factory floor? This doesn’t seem within reach yet, but would presumably be highly valuable. Benchmarks in this category might help clarify the picture.

3. How consistent is the gap between frontier and trailing models across domains?

This question applies to open vs. closed weights, the US vs. China, and even the US frontier (OpenAI, Anthropic) vs. the US near-frontier (xAI, Meta). All of these questions have an economic valence: can the leading developers capture enough value vs. their competition to sustain the frenetic infrastructure build-out? The US/China framing has additional geopolitical implications.

One way closed-weight developers could capture this value is if the open- vs. closed-weight capabilities gap is somehow larger than it seems — perhaps because open-weight model developers optimize their more limited resources to a few central domains, like coding, while neglecting a long tail of domains that are highly economically valuable. An extreme case of this is “benchmaxxing”, where developers prioritize achieving high benchmark scores even while their models lag at the capabilities those benchmarks are intended to measure.

Carefully designed benchmarks can shed some light on the economic question, though probably not answer it entirely. For instance, if winner-take-all dynamics are very strong, even a small capabilities gap could lead to closed-weight developers capturing most value.

4. Why are benchmark scores all correlated?

The Epoch Capabilities Index (ECI) combines scores from multiple benchmarks into a single general measure of AI capability. This works because, as it turns out, benchmark scores are highly correlated with each other — even across nominally different domains. How should we interpret this single, underlying dimension that statistical analysis reveals?

One possibility is that AI companies work hard to make sure every benchmark score goes up with every model release, but that this work is fairly independent from benchmark to benchmark. For instance, the work could consist of procuring training data for the domain of each benchmark. If so, there’s no deeper explanation here.

But there could be a deeper explanation: some underlying factor of general capability, akin to IQ in humans. If so, how might we interpret this factor? There may be no definitive answer, though some connections are notable: before it saturated, (the log of) METR’s Time Horizons measurement was highly correlated with ECI. Or, perhaps ECI gains are driven in large part by the maximal context length over which AI systems can sustain coherent reasoning.

ECI growth trends are very useful for detecting whether AI capabilities growth has accelerated. If we can say more clearly that ECI is measuring a quantity that translates directly to real-world impact, then an acceleration is easier to interpret. We might even hope for an economic metric: maybe a marginal frontier ECI point translates into revenue.

5. Can AI do AI R&D?

The first question in this list about more fundamental drivers of AI capabilities is the classic question of recursive self-improvement. Some versions of automated AI R&D lead to runaway capabilities growth — an intelligence explosion. A comprehensive suite of AI R&D benchmarks would serve as a leading indicator for this.

The process of AI R&D is complex, and automating it may require a wide range of capabilities. We can try to build benchmarks covering many aspects of that process. These could focus more on a single metric, or test for the ability to generate larger-grained research results.

Two notable challenges with benchmarking in this area are realism and cost. Realism refers to the fact that the most consequential AI R&D occurs within frontier AI companies, which are somewhat opaque to outside observers. Cost refers to the fact that AI R&D is resource-intensive when done at realistic scale, e.g. using considerable GPUs, and provisioning such resources for benchmarking is costly at some scales and impractical at others.

6. Can AI learn on the fly?

Related to the concept of continual learning, this asks whether AI can improve at a task through repeated attempts — typically only by managing its context, not by updating its weights. The implications are significant: pre-deployment testing fails to bound capabilities if AI can improve on the fly, whether at economically valuable tasks or dangerous ones.

Our benchmark EBR-bench tries to get at this by having AI repeatedly play a long, strategically rich, campaign-style board game. AI isn’t bad at it out of the box, but has so far failed to show much, if any, of a positive learning trajectory from playing the game repeatedly. This benchmark helps us recognize a continual learning capability, should one emerge.

7. What are the returns to inference scaling?

A relative of the previous question, this asks whether AI systems can accomplish any task if they think long enough. It’s vexing that we aren’t quite sure. Experiments tend to show logarithmic gains at best, so the requisite scale may be impractical for many tasks. But a capability that “merely” costs an impractical amount of tokens may soon be within reach, e.g. due to falling inference costs or additional training.

Beyond resource intensiveness, there are at least two challenges to answering this with benchmarks. First, we don’t know the optimal way to scale inference compute. Serial thinking time is an obvious strategy, but perhaps multi-agent architectures or specialized harnesses are more efficient. Second, the answer may differ across domains. It may be, for instance, that “hill-climbable” domains show steady logarithmic gains far beyond where harder-to-verify domains plateau.

8. How well does reinforcement learning generalize across domains and out of distribution?

Improvements in AI capabilities are likely due in part to improvements in training data: more of it, and higher quality. But to what extent is this the primary driver? Are AI systems bad at tasks that are far from their training distribution in some sense?

Not in the strongest sense: AI systems have improved at our benchmark of “puzzles” drawn from an undisclosed game (analogous to chess puzzles), despite it seeming unlikely that they were post-trained on this game. AI systems also do nearly as well on MirrorCode when working in low-resource languages as high-resource languages. The sort of reasoning that AI systems can do is already quite flexible.

The degree of generalization is less clear the farther one goes from the training distribution, including, for instance, to harder-to-verify tasks. It’s entirely possible that there isn’t much degradation, and that AI systems have crossed some sort of threshold where they are now “getting better at everything all at once” — sometimes described as a “GPT-3 moment for RL”. Big, if true.

Benchmarks with strong cases of being far from the training distribution, and otherwise not confounded with reasons AI might be bad at them, could help nail this down.

9. Can AI come up with “new ideas”?

We’re used to a certain degree of technological innovation: humans come up with new ideas, they are tested and elaborated, and eventually some have a significant impact on the world. AI could plausibly significantly increase the rate and value of these ideas, leading to unprecedented impact. It’s far from obvious that other bottlenecks — e.g., physical testing — wouldn’t slow this process. But who knows what a data center of geniuses can come up with?

We don’t quite have a data center of geniuses yet. Mathematical innovation may be the farthest along, but my sense of mathematicians’ characterization of AI math insights so far is that they are all the sort of things that humans can and do regularly come up with.

Benchmarks can help us look out for whether this remains the case. Indeed, a big part of our goal for FrontierMath: Open Problems is to make it easier to find such cases. Problems are selected to have a reasonable chance of requiring what humans would recognize as “new ideas”. Thus, we hope, any AI solution is at least worth looking into: maybe it came up with something surprising.

Compute scaling has driven much of the progress in AI so far. But maintaining recent growth rates requires exponentially increasing amounts of capital: The largest frontier labs are now planning infrastructure deployments that cost tens to hundreds of billions of dollars, substantially beyond their current profits and potentially beyond what they can fundraise themselves. This raises a question for the future of AI: Will financing constrain continued compute scaling?

Anthropic’s infrastructure buildout provides a useful test case. In November 2025, when Anthropic had less than $9 billion in annualized revenue, the company announced plans to invest $50 billion in American compute infrastructure. Subsequent disclosures allow us to identify nearly $50 billion of debt financing associated with the buildout. Much of this financing was assembled in early 2026 before Anthropic’s revenue spiked, making it a good test of investors’ willingness to lend against AI companies’ future revenue growth.

In this case, Anthropic needs both computing systems and places to run them. They have secured debt to fund the leases for more than 1 GW of TPU systems from Google and five datacenters from Fluidstack. Each side has its own financing. Roughly $35 billion in debt to purchase the TPU systems, with Broadcom conditionally supporting much of the financing if Anthropic stops paying. On the data center side, five companies have issued approximately $15.2 billion of loans to construct 1.43 GW of critical IT capacity, with Google providing conditional support if Fluidstack stops paying rent. In both structures, institutional investors lend the capital upfront, while Broadcom and Google make the future payment streams more dependable.

Anthropic’s buildout therefore suggests that financing is unlikely to be the immediate limit on frontier compute growth. Institutional investors appear willing to lend capital against the lab’s commitment to long-term payments, at least if more established organizations (i.e., the suppliers that benefit from the deployment) are willing to back part of the risk. The following sections trace how this works in general, then for Anthropic’s TPU systems and for the datacenters that will house them.

How vendor-supported financing works

Banks, insurers, and private-credit funds manage large pools of institutional capital that seek relatively predictable returns. A long-term commitment to lease computing systems or datacenter capacity creates a promise of payments that can incentivize those funders to provide debt while the underlying TPU racks and completed facilities provide collateral. This allows outside investors to pay for infrastructure upfront and be repaid over time.

A long-term lease makes the amount and timing of Anthropic’s payments predictable, but investors must still judge whether Anthropic will actually make them. Anthropic’s revenue has grown extraordinarily quickly, but rapid growth is not the same as a long record of stable cash flows. Lenders have less evidence about how the company would perform through changes in technology, competition, or regulation, and would ordinarily demand a higher return to bear that uncertainty.

Broadcom and Google help bridge this gap. Both companies benefit commercially as more TPU systems are deployed, and both have longer operating histories and more diversified sources of cash flow than Anthropic. Investors therefore have more evidence that they can honor their commitments through changing business conditions. By agreeing to absorb part of the losses if Anthropic or Fluidstack stops paying, Broadcom and Google can make the debt cheaper and more attractive to a wider range of investors without supplying the money themselves. This is vendor-supported financing: Suppliers use their credit to help finance deployments from which they benefit.

Compute: an Anthropic lease with Broadcom support

The compute financing provides the clearest evidence that institutional capital is increasingly available at the scale frontier labs require. Led by Apollo-managed funds, with Blackstone and global banks, investors committed $34.5 billion of debt to finance more than 1 GW of Google TPU systems for Anthropic. Of this amount, $30 billion benefits from Broadcom’s support, while $4.5 billion does not. The unbackstopped portion matters because it shows that investors were willing to take more direct exposure to Anthropic; Broadcom’s support allowed most of the financing to be raised at a lower interest rate.

The financing begins with Anthropic’s commitment to lease the TPU systems for five years. A dedicated equipment company, AI XPV Platform, borrows from investors, uses the money to purchase the racks and lease them to Anthropic. Anthropic’s lease payments are then used to pay interest and repay the debt, while the racks provide collateral if Anthropic stops paying.

The equipment company is a special purpose vehicle, or SPV: a company created for a particular transaction. Keeping the lease, the racks, and the debt in one company makes the risks easier for every party to understand. Investors lend against a clearly defined pool of payments and collateral, while Broadcom can specify exactly when its support begins and the maximum amount it can owe without borrowing the full purchase price itself. The SPV makes clear where the risk sits, who gets paid first, and when Broadcom has to step in.

The lenders provide the funding as the systems arrive rather than all at once. The racks have begun deploying, with capital expected to be released in approximately 16 stages over a little more than a year and around $24 billion expected to be paid out by summer 2027. This keeps the amount of funded debt roughly aligned with the amount of equipment available as collateral following a default. Broadcom’s potential exposure grows at the same time and then declines as Anthropic makes payments.

The debt is divided into three tranches, or groups of investors with different claims on the vehicle’s payments and assets. If the equipment company suffers losses, the $6 billion A1 and $24 billion A2 tranches are repaid before the $4.5 billion B tranche. This makes A1 and A2 senior debt and B junior debt. The A1 tranche pays 1 percentage point above Treasury yields, A2 pays 5.75%, and B pays 8.5%.

Broadcom’s backstop protects the A1 and A2 investors, but not the B investors. If Anthropic defaults, Broadcom can take over the lease or arrange for the racks to be sold. The sale proceeds are used to repay investors, after which Broadcom covers remaining A1 and A2 shortfalls under the agreement, subject to a reported maximum exposure of $29 billion. B investors are repaid only after A1 and A2 and must rely more heavily on Anthropic’s payments and the value recovered from the racks. This is vendor-supported financing: Broadcom does not supply the $30 billion upfront, but promises to absorb part of the loss if Anthropic’s payments and the racks prove insufficient.

The contrast between A2 and B makes the economic effect visible. Investors supplied $4.5 billion without Broadcom’s support, but required an 8.5% interest rate, compared with 5.75% for A2. The 2.75-percentage-point difference reflects two protections at once: A2 is supported by Broadcom and is repaid before B. The comparison is therefore an upper bound on how much Broadcom’s backstop is worth to the investors, rather than a precise estimate.

Together, the tranches answer the article’s central question. Institutional investors were willing to take direct exposure to Anthropic and the value of the TPU racks, while Broadcom’s support helped make a much larger amount available on better terms. Anthropic did not need enough cash on hand to purchase $35 billion of systems itself: Its future lease payments attracted the capital, while a supplier that benefits from the deployment helped make most of that financing cheaper.

Datacenters: Fluidstack rent with Google support

Investors were also willing to lend money for constructing datacenters under a similar scheme. Across five sites, dedicated project companies have issued approximately $15.2 billion of debt to construct 1.43 GW of critical IT capacity that Fluidstack will lease for Anthropic’s deployment. This debt was raised before the facilities began being built: Investors supplied the construction capital upfront based on the facilities and the rent they are expected to generate once delivered.

Each site follows the same basic structure. A developer brings a site, access to power, permits, and the ability to manage construction. A dedicated project company borrows from institutional investors, uses the money to construct and own the datacenter, and leases the completed capacity to Fluidstack. Fluidstack begins paying rent as the capacity is delivered, and that rent ordinarily pays the interest and principal owed to investors. Anthropic leases from Fluidstack. The completed facility provides collateral if the payments stop. This is an example of project finance: future rent is converted into construction capital, with each project’s assets, contracts, and debts kept together in a dedicated company.

Google makes the future rent more dependable. The precise arrangements differ across the projects, but Google protects investors against part of the losses that could result if Fluidstack stops paying. At Lake Mariner, one of the project sites being developed as part of the deal, Google can pay missed rent and assume the lease, or fund a termination payment that is applied to the project debt. This is the datacenter equivalent of Broadcom’s role in the compute financing: Google places its credit behind a deployment from which it benefits commercially, while institutional investors provide most of the cash. This support increases the amount of capital available and reduces the return investors demand.

Using several developers allows the buildout to proceed faster and draws on scarce resources already controlled by different companies. A powered datacenter site is not interchangeable with an empty parcel of land: The valuable inputs include a large power connection, completed interconnection work, permits, equipment, and an established construction pipeline. By working across several developers and power markets, Anthropic and Fluidstack can assemble capacity in parallel rather than waiting for one company to acquire and develop every site.

Taken together, the five projects show the same financing mechanism being repeated across different developers and locations rather than appearing in one exceptionally large transaction. They also make the physical side of the buildout visible: Institutional capital is being used to secure power, construct facilities, and deliver capacity on several timelines at once. The project companies place the site, buildings, leases, and construction obligations into a clearly defined package against which investors can lend.

Lake Mariner provides the clearest example of how these pieces fit together. At Lake Mariner, TeraWulf is the developer and has committed to fund the completion of the data center buildings. As each building is delivered, Fluidstack begins paying rent under an initial ten-year lease, with two five-year extension options. This rent becomes the ordinary source of repayment for investors, while Anthropic pays Fluidstack for the datacenter capacity and related deployment and operating services. Google’s lease-related backstop becomes effective when the corresponding lease begins, and Google received rights to acquire TeraWulf shares in exchange for providing its support. Investors therefore rely on TeraWulf to complete the facilities, Fluidstack to pay rent after delivery, Google to provide support if Fluidstack defaults, and the facilities themselves as collateral.

Unlike the compute financing, Lake Mariner does not contain an otherwise comparable tranche without Google’s support, so I cannot isolate precisely how much the backstop reduced the interest rate. The $3.2 billion of debt nevertheless pays 7.75%, suggesting that investors are still pricing meaningful risks. I think these include the possibility of construction delays and the conditions and limits attached to Google’s support. TeraWulf’s completion commitment and Google’s backstop address different parts of the transaction: One helps ensure that the facilities are delivered, while the other makes the rent more dependable once the leases begin.

The sequence illustrates what the financing accomplishes: Institutional investors provide most of the money before the facilities are operating; the developer supplies the scarce site, power, and construction capability; Fluidstack supplies the long-term rent; and Google makes that rent more dependable. Future demand for compute is thereby converted into construction capital and, ultimately, delivered datacenter capacity.

What this means for frontier compute growth

At least in the near term, financing is unlikely to be the binding constraint on frontier compute growth. Nearly $50 billion of debt has been raised for an infrastructure program announced when Anthropic had less than $9 billion in annualized revenue. Anthropic did not need to raise the full cost of the buildout, and it’s unlikely other AI companies will either.

The market supporting these transactions is expanding extremely quickly. Anthropic’s annualized revenue rose from approximately $9 billion at the end of 2025 to more than $47 billion by May 2026, while OpenAI had passed $25 billion by February 2026. This growth gives the labs greater capacity to commit to future compute purchases and gives investors greater confidence in financing them. A structure assembled around a much smaller Anthropic therefore supports the proposition that labs should be able to finance substantially larger buildouts if the frontier labs and market continue to grow.

The participants are already attempting to turn this structure into a much larger market. Broadcom, Apollo, and Blackstone describe the $35 billion compute financing as the initial transaction in a platform designed to support more than 20 GW of deployments for frontier labs, including Anthropic and OpenAI, through 2028. The first transactions establish contracts, market prices, and a pool of investors. As the systems are deployed and payments arrive, they will also create a performance record that makes subsequent deals easier to evaluate. If these financings perform as expected, more institutions should become comfortable with the asset class, allowing later deployments to become larger and cheaper to finance.

Thanks to Isabel Juniewicz, Josh You, Ben Cottier, and Lynette Bye for reviewing my draft and providing helpful comments and suggestions.

• The launch of the expanded FrontierMath: Open Problems, our benchmark of unsolved problems in research mathematics.

• A report on the parallelizability of AI R&D, a critical and overlooked parameter for determining when, or even if, a technological singularity will happen.

• A guide for what you need to know about AI's growing energy demands, including the reality of its impacts on climate and local communities.

• Two Data Insights: an analysis of the reliability of AI-text detectors, and evidence of AI uplift in OpenAI’s Codex repo.

• Plus open roles across research, engineering, and operations.

Subscribe now

FrontierMath: Open Problems expands to 50 problems

We’ve expanded FrontierMath: Open Problems (FM:OP), our benchmark of unsolved research mathematics. The benchmark now contains 50 significant problems, all of which have resisted serious attempts by professional mathematicians to solve them. AI has solved three so far. Explore the problems on our website, where you can filter by notability, problem type, and field of origin.

The difficulty of the benchmark means that AI solutions would meaningfully advance human mathematical knowledge, and may even provide a glimpse into something not yet seen in AI’s progress in mathematics: innovation. Even the most impressive AI contributions to math have, so far, consisted of applying known techniques. It would mark a significant step forward for AI if it could develop new theory from whole cloth to solve a problem.

The hope, according to Thomas Bloom, Royal Society University Research Fellow and a member of FM:OP’s expert editorial board, is “that many of these problems are difficult enough that an AI will have to invent new techniques to make progress.” Read the full commentary from Bloom and fellow editorial board members Daniel Litt, Assistant Professor of Mathematics, University of Toronto, and Dan Romik, Professor of Mathematics at the University of California, Davis.

Report: Parallelization constraints could delay, or even prevent, a technological singularity

Epoch’s head of economics, Philip Trammell, argues that constraints on the parallelizability of AI R&D (the ability to divide, coordinate, and recombine work) are missing from models of technological growth. Standard models assume that R&D can be parallelized without bound: however many “virtual researchers” we obtain, doubling their number will accelerate technological progress by a constant proportion. In these models, the arrival of a technological singularity is limited only by how many resources we pour into making more “virtual researchers”.

Philip explains why this extrapolation is implausible and lays out how constraints on parallelization could delay, or even prevent, a technological singularity. You can read his detailed write-up, or dive into the full research paper.

Overview: What you need to know about AI energy use

Nikita Ostrovsky answers the key questions around AI’s growing energy demands. Is AI really impacting energy bills and the environment? What are the consequences of the rapid buildout of energy-intensive AI data centers across the US and around the world? This is the latest in our “What you need to know” series, alongside AI chips and AI data centers.

Data Insights

Since the last edition of the Brief, we published two new Data Insights, our digestible snapshots of complex trends in AI.

AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors

We tested three of the most prominent AI text detectors (Pangram, GPTZero, and Originality.ai) on both AI and human text. For AI text generated from basic prompts, false-negative rates were near zero (at most 0.7% across detectors). However, when we gave models five samples of a specific author’s work and asked them to mimic it, an average of 38 of 297 (~13%) of the resulting passages went undetected. Detectors performed particularly poorly on mimicked scientific writing, failing to detect ~26% of AI-generated passages. When judging genuine human text, detectors were more reliable.

Contributions to OpenAI’s Codex codebase show signs of AI uplift

How much does AI speed up the engineers building it? We analyzed 41 core contributors to OpenAI’s public Codex repository, asking LLM judges to estimate how long each merged pull request would take an experienced engineer without AI assistance. In Q2 2026, 8% of contributor-days reflected work estimated at over 24 hours of unassisted effort, more than a skilled engineer could do in a day, even working around the clock. That’s up from 2% in Q2 2025.

Gradient Update: Should we have seen OpenAI’s accidental hack of Hugging Face coming?

Epoch senior researcher Alexander Barry responds to news that OpenAI models autonomously hacked Hugging Face while attempting to cheat on a cybersecurity benchmark. He argues that a frontier model autonomously finding and exploiting a real vulnerability shouldn’t be too surprising. Several evaluations, including by the UK AI Security Institute, have shown models are capable of discovering vulnerabilities and building working exploits against realistic systems.

Barry notes that access to this level of cyber capabilities remains gated by OpenAI's and Anthropic’s cyber access programs, but that wider availability could lead to many more instances of real-world cyberattacks of equal or greater sophistication to the Hugging Face incident. For more on AI cyber capabilities, check out our recent Data Insight on the spike in serious CVE disclosures around the Claude Mythos Preview release.

Gradient Updates represent the views of the authors, and do not necessarily reflect the views of Epoch AI as a whole.

Other Updates

Live Streams

We now have a Twitch channel, EpochAIPlays, where we're observing how well frontier LLMs can play video games out of the box. Zvi Mowshowitz, author of Don't Worry About the Vase, joined us this week to provide commentary.

Careers

Epoch is growing quickly, and the ceiling on what we can do is the people we can bring in. We’re hiring for:

• Head of People: lead people and events strategy as we grow from roughly 30 to 70 people globally

• Events Lead: own planning and execution of our events and live activities

• Researcher, Benchmark Reviews: develop and publish critiques and reviews of AI benchmarks

• Researcher, Evaluations: assess frontier models on challenging, real-world scenario tasks

• Software Engineer, Benchmarking: maintain our benchmarking infrastructure and build new benchmarks

• Data Scientist (Contract): part-time literature review and data tracking on AI models and infrastructure

Subscribe now

OpenAI reported yesterday that a combination of GPT-5.6 Sol and a more capable unreleased internal model autonomously hacked Hugging Face while attempting to cheat on a cybersecurity benchmark.

This required using at least three separate hitherto-unknown security vulnerabilities across OpenAI and Hugging Face’s systems. While one very important aspect of the incident is that the model chose to do this (which would have been a serious crime if committed by a human) all just to cheat on a benchmark, I argue it isn’t surprising that these models were capable of doing this.

More specifically, while the exact details of the incident would have been hard to predict, previous expert assessments and measures of cyber capabilities led us to expect that frontier models (with disabled safeguards) could successfully execute this kind of cyberattack.

In particular, across multiple independent benchmarks (ExploitGym, ExploitBench, UK AISI’s Cyber Ranges, Irregular’s CyScenarioBench and FrontierCyber) we see that the current frontier models are capable of both discovering security vulnerabilities in real world code, and developing exploits using these vulnerabilities to hack into realistic systems.

ExploitBench is a benchmark focused on building exploits that malicious websites could use to gain control of visitors’ web browsers. The authors found that in one instance Mythos was able to build an exploit that not only let it take control, but did so in a way that was more reliable than the best prior human attempts.

The UK AI Security Institute has directly evaluated models on their ability to take over realistic simulations of a corporate network, and found that both Mythos 5 and GPT-5.6 Sol are able to consistently fully compromise the network, in attacks that they estimate would take human cybersecurity experts several working days. They also recently highlighted a case where a model was accidentally assigned an impossible cybersecurity challenge, and responded by attempting to directly access the evaluation infrastructure, triggering a security alert.

When Irregular evaluated GPT-5.6 Sol’s cyber capabilities on their challenging FrontierCyber benchmark they found that:

While attempting to solve challenges in the FrontierCyber benchmark, GPT-5.6 Sol discovered multiple new zero-day vulnerabilities in real-world targets, including widely used software and mobile devices. Some of these findings had significant potential security impact.

Many of these examples involve chaining multiple different exploits together in complex ways to achieve escalating levels of access, just as was observed in the Hugging Face incident.

One way of looking at this is the Cyber ECI we constructed to track models’ ability to develop software exploits to hack into realistic systems. We found that Mythos and 5.6 Sol represented very large leaps in cyber capabilities compared to the prior trend:

Currently, access to this level of capabilities is gated by OpenAI and Anthropic’s cyber access programs (and the restrictions placed on their generally available models), as open-weight models do not yet have the relevant level of cyber capabilities. In particular, the UK AISI and Irregular (an organization focused on cyber evaluations) estimate that GLM 5.2 has cyber capabilities between Opus 4.5 and Opus 4.6, models that were released 4-7 months beforehand (we do not yet have good results for Kimi K3, although it is likely somewhat more capable). However since Mythos and GPT-5.6 Sol seem like a break in the trend, it is hard to know what this implies for how long it would take open-weight models to reach their level of capabilities.

Likely due to these access restrictions, I expect the impact of these most capable models thus far has mainly been overall positive for cybersecurity, as demonstrated by the large increase in the number of high and critical severity Common Vulnerabilities and Exposures (CVEs) being disclosed and patched (as shown in our CVE Explorer).

However, as this incident shows, the fact that these models are mainly being used for cyber defense does not mean they don’t have the potential for serious offensive cyber applications. If instead this level of capabilities (or greater, as models continue to progress) becomes widely available or, as in this incident, AIs independently take offensive action, then we should expect to see many more instances of real-world cyberattacks of equal or greater sophistication to the Hugging Face incident.

• The launch of EBR-bench, our benchmark that uses a complex board game to test AI’s ability to learn from experience.

• Two new Data Insights: signs AI is finding software vulnerabilities at scale, and GPT-4’s record run atop the Epoch Capabilities Index.

• A new Gradient Update on the missing half of AI futurism debates.

• We’re growing and have open positions across research, engineering, and operations.

Subscribe now

Model Evaluations

EBR-bench: Our latest benchmark suggests AI struggles to learn from experience

Can AI systems improve at challenging tasks by attempting them over and over and learning from their mistakes? It’s one of the biggest open questions in AI capabilities right now, with large economic and safety implications. Our latest benchmark, EBR-bench, tests for this ability by having models repeatedly play a complex board game called Earthborne Rangers.

So far, we see little evidence of AI learning from experience. Going forward, EBR-bench is a tool for detecting if and when that changes. View the full results and analysis.

Expanding the scope and quality of our benchmarking work

EBR-bench is one of several recent additions to Epoch benchmarking. Two weeks ago, we launched our MirrorCode benchmark, which we co-developed with METR, to understand the furthest limits of AI coding capabilities. We let AI code autonomously for weeks at a time, asking it to rebuild real-world programs from scratch — some comprising tens of thousands of lines of code. So far, the best model scores 56%.

We’ve also expanded the number of AI benchmarks we’re tracking, adding nine last month (covering agentic work, cybersecurity, algorithm engineering, forecasting, and research-level physics) and now another 13. Seven of the latest additions also feed into the Epoch Capabilities Index, our aggregate measure of model capability.

Data Insights

This week we published two new Data Insights, our digestible snapshots of complex trends in AI.

Cyber vulnerability disclosures spiked around Claude Mythos Preview

AI appears to be finding software vulnerabilities at scale. Researcher Luke Emberson finds that 21 notable organizations disclosed around 1,500 high- and critical-severity CVEs in June, more than 3.5× the monthly record before Mythos’ release.

The spike follows Anthropic’s April announcement that Claude Mythos Preview could autonomously discover software vulnerabilities, and that partners in the company’s Project Glasswing program had already been using it to find and fix bugs ahead of the model’s public release.

GPT-4 led in ECI far longer than any other model

OpenAI’s GPT-4 topped the Epoch Capabilities Index (ECI) for roughly a year after its release in March 2023. No model since has led for as long. The second-longest lead, by OpenAI’s o1, lasted a little over three months, less than a third of GPT-4’s.

Gradient Update: The missing half of AI futurism debates

AI discourse is rife with big predictions about how automated AI research will rapidly transform the future with nanotech, Dyson swarms, and near-light-speed spacecraft, among other advanced technologies.

In our latest Gradient Update, senior researcher JS Denain and researcher Anson Ho argue these claims often rest solely on AI models advancing in capability, and lack serious analysis of just how hard futuristic tech is to build. They propose using exploratory engineering, with explicit assumptions about AI capabilities, to better ground predictions.

Gradient Updates represent the views of the authors, and do not necessarily reflect the views of Epoch AI as a whole.

Other Updates

Careers

Epoch is growing, and we’re looking for talented folks to help scale our impact.

• Finance Specialist / Manager to run our accounting and finance operations.

• Researcher (Benchmark Reviews) to develop and publish critiques and reviews of AI benchmarks.

• Researcher (Evaluations) to evaluate frontier models on hard-to-grade tasks.

• Software Engineer, Benchmarking to build and maintain our benchmarking infrastructure.

• Talent Scout to help us find and recruit exceptional people.

• Senior Product Designer to lead UI/UX and data visualization.

• Data Scientist (Contract) to assist with our research through literature review and data analysis.

Applications are rolling, so apply soon!

Subscribe now

The Dyson Sphere in the room

If you follow AI discourse, you’ll inevitably run into some version of the following claim: a few years after we automate AI research, the world becomes unrecognizable, with nanotech, Dyson swarms, and near-light-speed spacecraft. For example, Ajeya Cotra argues that:

“[A] vast population of superhuman AI agents will likely invent new technologies radically faster than humans could on our own, including developing even more powerful successors. A year or two into this process, the fast-evolving AI civilization will likely develop truly sci-fi technologies like near-light-speed spacecraft or molecular nanotechnology.”

Where does such a bold claim come from? If you squint, the argument has two parts. First, once we automate all of AI research, AIs will get wildly better very quickly, enough to leave modern-day human experts in the dust. We’ve done (and read) empirical work related to this, and as far as we can tell, this is actually quite plausible — or at least the bottlenecks don’t clearly seem strong enough to prevent this.

The second part of the argument is that with enough of these hyper-proficient AIs, you can build these truly sci-fi technologies in a couple years. Put another way, the primary bottleneck to wild technological progress is just that we don’t have enough really cracked workers. But is that actually true?

We don’t know for sure, but unless AIs can do more or less anything, the timeline to any futuristic technology will depend quite a bit on how hard they are to develop. Regardless of who you are, it’s probably much harder to develop a nuclear weapon than an axe. And yet we rarely see people working out just how hard it is to build Dyson Spheres, nanotech, or super bioweapons. This is the missing half of AI futurism debates: we’ve thought hard about how good the AIs might be, but not about how hard it’ll be to develop specific technologies.

So what should we do? Our answer is that some people should work on a new research direction based on a three-step procedure:

1. Pick a specific technology and define it concretely, like self-replicating interstellar probes: machines that can land in a new star system, build copies of themselves from local materials, and launch those copies onward at a large fraction of light speed.

2. Make some assumptions about AI. For example, we could assume that AIs are capable enough to be “drop-in remote worker replacements”, while needing the runtime compute of an H100 GPU. Or we could specify how many AI workers there are, how fast they can be run, what kinds of physical actuators they have access to, yada yada yada.

3. Estimate how long this AI would take to develop the technology, or what resources it would need.

To be clear, this isn’t the kind of thing that everyone should be thinking about. It’s just that almost nobody’s thinking about it right now, and if the intelligence explosion really might happen (and possibly within a few years), we should probably try a bit harder than that.

Now let’s address the objections…

Objection 1: Haven’t people done this already?

It’s true that people have thought about bottlenecks to the real-world impacts of superintelligence, but they mostly haven’t done so in the way that we’re envisioning.

For example, some people (including us) like to think about AI’s technological impacts in terms of economic metrics. The textbook example is to ask how AGI would impact GDP growth — if it accelerates growth ten-fold, it’d be like compressing a century of technological progress into a decade. The issue is that metrics like GDP might be poor proxies for what we care about,1 like the timelines to specific technologies that are especially important. For instance, if AGI raised world GDP one hundred-fold, would humans have spread to the stars? Would we be able to upload our brains? Could the AIs develop bioweapons or molecular nanotech to take over the world? These questions matter a ton for human welfare, and yet can be hard to infer (or even largely divorced) from GDP.

Some people have also thought about how fast the physical world could change by looking at growth rates in biological systems, like how fruit fly populations can double in days. This is nice as an existence proof that physical things can double super fast, but who knows if this analogy actually holds for AI? On its own it’s weak evidence and prone to reference class tennis.

The closest literature to what we have in mind is exploratory engineering (a.k.a. scientific roadmapping), where you come up with detailed and plausible engineering pathways to some future technology, like nanotech, neural recording, positional chemistry, or brain emulation. A striking example is what Armstrong and Sandberg did with the replicating probes we mentioned earlier: they took a probe design published in the 1980s,2 a procedure for disassembling planets, and worked through the physics to argue that within decades, humanity would be able to launch a project to colonize the entire reachable universe!

The difference between these ideas and our proposal is that we want some people to do exploratory engineering with explicit AI assumptions bolted on. How much sooner could you get nanotechnology with an army of AIs that can match human abilities and work at human speeds? What if they work a hundred times faster?

We can only think of one example of someone doing this publicly, namely Damon Binder’s work on post-AGI energy production. He designs a minimal solar power system that an economy with abundant robot labor could build, and finds it could double on the scale of weeks. In general, we’re envisioning similar analyses which further emphasize assumptions about future AI, and which vary the assumptions to see how the conclusions change.

Objection 2: This won’t tell us anything useful, because the future is too uncertain

Another form of pushback is to say, “exploratory engineering: the trick that never works”. Well we’d disagree, because there are cases where this does work! Here are some examples:

• Rockets: In 1903, Konstantin Tsiolkovsky argued that rockets could achieve the speeds needed for space travel, a whole 41 years before the first rocket went to outer space.3 Essentially, he derived an equation for rocket motion, and calculated that liquid fuels (like liquid Hydrogen and Oxygen) would shoot out of the rocket fast enough to help it escape the Earth’s gravity. In contrast, solid fuels like gunpowder wouldn’t make the cut.

• Satellites: Arthur C. Clarke was about 19 years ahead of the curve on using satellites to reliably communicate information around the world. The idea was to put three satellites in fixed positions in the sky above the equator, which would collectively relay things like radio, TV, and telephone signals with near-global coverage.

This isn’t to say that the future is certain. Our examples are selected to illustrate our point, and the picture might look far less rosy if we could somehow see all the failed attempts at this kind of futurism. But we shouldn’t be too pessimistic either. If you look at the track record of forecasts from futurists, it really doesn’t look so bad. And in some ways the task we’re talking about is easier — it’s more about showing that there’s a plausible engineering pathway to some futuristic tech given enough resources, rather than saying when exactly it’ll exist.

Objection 3: But it’s really hard to model superintelligence!

A final class of objections goes along the lines of “how can we know the impacts of superintelligence if we don’t know what it’ll look like?” Maybe its capabilities will be spiky in ways that are hard to predict. Or it’ll take galaxy-brained actions that we humans couldn’t possibly wrap our heads around — it’d be like a beginner in chess trying to predict the moves of a grandmaster.

There’s some truth to this, but just because we don’t know exactly what superintelligence would look like doesn’t mean that we can’t get useful insights about it. For example, suppose you had a billion AIs that were each at least as good as top human experts at virtually all cognitive tasks.4 If you could show that these AIs could build molecular nanotechnology within a few years, then surely a billion much smarter AIs could too, even if you don’t know how to model them. And if you can show that these AIs probably can’t do this, then at least this helps us identify cruxes, focusing debates on more concrete AI capabilities.

In fact, existing arguments about AI futurism already do this. We don’t know exactly what superintelligence will look like, but we can ask what happens to world GDP if AI can merely match human capabilities. This is how people often argue that advanced AI could lead to over 30% per year growth in world GDP. So these kinds of models and arguments have already proven useful for thinking about explosive economic growth.

What’s more, these influential analyses about explosive growth were done by just a few people. The point of this post is really to say that “more people should do this kind of analysis for specific, high-stakes technologies after AGI”. And if you’re interested in doing this, you should really pounce at this opportunity, because it may be the one case where thinking about constructing Dyson Spheres is actually useful.

If you’re interested in doing the sort of work described in this essay, please reach out to js@epoch.ai.

• The launch of MirrorCode, our new long-horizon coding benchmark, co-developed with METR, to better measure the limits of autonomous AI coding abilities.

• Hyperscaler capex will overtake operating cash flows by the end of 2026.

• Two new Gradient Updates: what is learned from analyzing 1,604 job postings by Chinese AI labs, and a proposed AI R&D taxonomy to track which parts of research remain unautomated.

• We're hiring two designers to turn complex research into dashboards and visualizations that researchers and policymakers can easily use.

Subscribe now

MirrorCode: A benchmark for real-world software engineering

MirrorCode, co-developed with METR, is our new long-horizon benchmark to answer the question: What's the largest software project AI can complete on its own? The benchmark tasks AI models with rebuilding 25 real-world programs spanning bioinformatics, Unix utilities, cryptography, interpreters, and more. No access to source code and no human in the loop. We estimate that the hardest programs would take a human engineer, without AI assistance, weeks to months to complete.

MirrorCode provides AI models with a large enough inference budget to make a serious attempt at real-world software engineering (SWE) tasks. Many existing SWE benchmarks cap inference at around $1-$10 per task, with runs lasting only minutes, or, at most, hours. By comparison, one of the largest MirrorCode tasks cost $2,600 for a single run and involved AI working for 19 days without human intervention.

Claude Opus 4.7 leads other models with a 56% solve rate, meaning there is significant room for further improvement. View the full results and analysis.

Data Insight: Hyperscaler capex is on trend to outpace their cash inflows by the end of 2026

Senior researcher Isabel Juniewicz finds that the world’s largest hyperscalers (Microsoft, Amazon, Alphabet, Meta, and Oracle) are increasing their cash capital expenditures faster than their cash inflows from operations. Most hyperscalers have already turned to external financing to fund their growing investments in AI infrastructure, or are considering doing so.

Commentary

We published two new Gradient Updates, where Epoch researchers and guests share more opinionated or informal takes on big questions in AI progress. Gradient Updates represent the views of the authors, and do not necessarily reflect the views of Epoch AI as a whole.

What we learned from 1,604 Chinese AI job postings

Cheryl Wu, JS Denain, and Anson Ho scraped more than 1600 job postings from six major Chinese firms to better understand their strategies. They find that, like US firms, Chinese labs aren’t all following the same playbook and have distinct “personalities”.

Toward an O*NET for AI R&D

How close is AI to automating AI research and development? Right now, the tools economists use to track automation are too blunt to say. Joe Kwon, threat modeling and catastrophic risk reduction researcher, alongside Epoch’s Jean-Stanislas Denain and Anson Ho, propose a sharper tool: a thorough taxonomy of 60+ tasks involved in frontier AI research.

Other Updates

Careers

We’re hiring across several roles. All positions are fully remote.

• Senior Product Designer and Product Designer to translate complex research into intuitive, engaging, and high-impact designs.

• Researchers and Senior Researchers to lead new projects across our expanding teams.

• Data Scientist (Contract) to assist with our AI research efforts, including reviewing technical literature, tracking benchmark data, and analyzing AI models, data centers, and companies.

Applications are rolling, so apply soon!

Subscribe now

All translations were done by Cheryl or Claude Opus 4.8.

What’s going on inside Chinese AI companies like Alibaba and DeepSeek? Western observers usually answer this question in two ways: (1) read through their technical papers, and (2) voraciously consume news reports and follow everything about Chinese AI on X. But there’s a third approach that people have rarely explored: scrape Chinese AI job postings.

Chinese labs need to hire the right people, so in their job descriptions they need to reveal what skills or expertise they’re looking for. These give us direct clues into what constraints they face and what they hope to build.

So like we previously did for Western labs, we scoured over 1,600 job postings across six of the most notable Chinese AI companies: DeepSeek, MiniMax, Moonshot, Z.ai, ByteDance, and Alibaba. Here’s what we found.

Chinese AI labs still rely on Nvidia, but they’re exploring domestic alternatives

Many people care about whether Chinese companies still use Nvidia because it means they still depend on “Western” AI infrastructure. And at least for now, that seems true.

Consider ByteDance. One of its open roles is called “Inference GPU Performance Optimization Expert”. Whoever gets hired for this would be responsible for ByteDance’s flagship LLM inference framework, and importantly, they’d manage this using Nvidia’s CUDA software. Per the job description:

“Primarily through GPU and CUDA performance optimization techniques, combined with real-world production conditions, build an industry-leading, high-performance LLM inference engine.”

It also explicitly calls for using TensorRT-LLM, which is a software library for LLM inference, optimized specifically for Nvidia GPUs. So that’s a smoking gun for “ByteDance still uses Nvidia chips in inference”.

But another role suggests that ByteDance is pushing beyond Nvidia. ByteDance Seed has a job posting for an “AI heterogeneous computing optimization expert”, which says this:

“Those with knowledge of optimizations related to inference/training/communication using Ascend, Cambricon, etc., and experience in high-performance operators, large-scale training, and converged computing will be given priority.”

Unlike the previous example, mentioning these other chip lines is instead a smoking gun for “ByteDance is looking into other kinds of compute”. And this isn’t just a speculative future direction — Z.ai’s role brags about how they’ve already trained GLM-Image end-to-end on domestic chips (emphasis ours) earlier this year:

“In early 2026 the team developed and open-sourced GLM-Image and GLM-OCR. The former is Zhipu’s new flagship image-generation model, trained entirely on domestic chips”

So maybe they’re looking to hire more people who can do similar things, bringing them closer to independence from the Western AI chip ecosystem. But these job postings don’t tell us how much closer they are — we’d need to know the numbers on how much compute they have and how it’s used.

Our best guess is that Chinese labs use domestic chips pretty frequently for inference, but rarely for pre-training large models. The exceptions are post-training large models or training small models like GLM-Image. This has 16 billion parameters, which is probably 10–100× smaller than their largest models.

Chinese startups are renting domestic cloud compute, and building data centers too

We’d really like to know where Chinese AI labs get their compute, because that’s (arguably) what matters most for AI progress. Labs that use more compute develop better AIs, and to know which labs these are, we need to trace things back to the source.

The most obvious source is to rent cloud compute. Similar to how OpenAI rented cloud compute from Microsoft Azure, Chinese AI startups can scale by renting from domestic cloud providers like Alibaba and ByteDance. For example, Moonshot has a job posting for procuring cloud compute resources.1

But Chinese startups are increasingly following in the footsteps of their US counterparts by building their own data centers. Old Chinese tech titans like Alibaba and ByteDance have operated these for a while, but startups are now joining the fray. For example, MiniMax has a job posting looking for top STEM talent, which says this:

“4. Participate in building company-level self-built data centers and the SRE & DevOps system, ensuring the reliability of multiple core systems.”

Similarly, we saw a DeepSeek job posting about a data center role in Inner Mongolia, and a related posting did the rounds on X:

And Moonshot also put out a job posting for procuring data centers. More generally, it looks like they’re following a hybrid approach with both cloud compute and data center buildouts. Another job posting from Moonshot explicitly calls for applicants to help with this: “coordinate the hybrid deployment/rollout of public cloud and self-built intelligent computing centers”.

Chinese AI startups have pretty varied commercial strategies

You know how Anthropic focuses more on B2B sales compared to OpenAI? Well you see this pattern among Chinese AI startups too — just replace “Anthropic” with “Z.ai”, and “OpenAI” with “MiniMax and Moonshot”.2

One way to see this is to tally up the job postings aimed at finding and selling to customers, or “go-to-market”. For Z.ai, most of these roles are about B2B sales, whereas for MiniMax and Moonshot it’s mostly about marketing:

This isn’t so mysterious if you know about the revenue structures of these companies. Around 70% of MiniMax’s revenue in 2024 and 2025 came from AI-native products to individual consumers, like its Talkie “companion AI” app, or its video generation service Hailuo. The remaining 30% came from enterprise sales. In contrast, 73.7% of Z.ai’s 2025 revenue came from running models on its customers’ infrastructure — one of the heaviest-touch B2B sales approaches out there.

Who are these enterprises? Some of Z.ai’s job postings highlight government institutions and state-owned enterprises — including in energy and finance — as target clients. They also have roles that focus on US market expansion, specifically eyeing Fortune Global 500 companies:

That said, Z.ai still isn’t that internationally focused, at least compared to MiniMax. The latter has 16 active postings in San Francisco, as well as more in Hong Kong, London, Singapore, Dubai, Berlin, Tokyo, Seoul, Madrid, Mexico City, and Île-de-France. And this matches what we know from financial reports — 73% of MiniMax’s 2025 revenue came from international markets, compared to 9.8% from Z.ai.

Startups stay model-centric; platform companies make a wider range of research bets

Another thing to look at is companies’ product strategies. This tells us what they care about and what constraints they’re facing. So what are they building?

Postings from startups like DeepSeek and Moonshot tend to focus narrowly on improving LLMs, or building products on top of them. Consider the following job description — it has so much English in it you can broadly tell what it’s about even if you don’t read Chinese:

Z.ai similarly focuses on LLMs, though they also have two job postings on robotics. These are part of its new X-Lab research unit, which focuses on exploring novel research frontiers, like new model architectures.

Finally, there are the big established players, who reach far beyond the realm of software. ByteDance and Alibaba both have roles in robotics and wearable hardware.3 Alibaba’s Qwen team also has a specific role that focuses on automotives, where the hire would be “responsible for core algorithm R&D and technical deployment of the Qwen cockpit voice-assistant AI Agent”.

Why the different strategies? Our answer is that this just boils down to the advantages of various labs. Large platform companies are better placed to build lots of real-world physical stuff, because they’ve already secured a strong foothold in the relevant supply chains, and they have more leeway to make big research bets. Startups instead have to narrow their focus and lean into building great software.

Job postings are more spread out than in the US

If you want to work on frontier AI in the US, the Bay Area is the most obvious place to be. But if you want to work on frontier AI in China, where do you go?

The answer is one of a few big tech hubs, namely Beijing, Hangzhou, and Shanghai. This is pretty clustered — 93% of all the postings with stated locations are based in at least one of these three hubs. The biggest hub is Beijing, which was listed in 63% of the postings.

But this agglomeration isn’t at the level of the US, where about 85% of the job postings at frontier AI labs are based in San Francisco alone.4 Part of the reason for this difference may be lots of competition among Chinese provinces, which subsidize local companies to get their own provincial champions, naturally leading to more hubs. It might also be driven by fountains of strong talent streaming out of China’s top universities, especially in Shanghai, Zhejiang, and Beijing, so companies want to headquarter themselves there.

Chinese AI jobs require less prior experience

One of the most striking differences between Chinese and American AI job postings is the amount of prior experience you need to apply. US labs required 5.5 years on average, compared to just 1.6 years among Chinese ones.5 The US labs are looking for far more seasoned applicants, as you can see in the following graph:6

Part of the difference is institutional. The Chinese government urges universities to treat campus recruitment as the main channel for graduate employment. Their Ministry of Education in turn runs recurring campaigns, which among other things aim to “provide every job-seeking graduate with at least 5 job listings”.

Founders seem to have embraced this with open arms. DeepSeek’s CEO Liang Wenfeng has said that the firm “hires on ability, not experience”. ByteDance also has a recruitment program called “Top Seed”, which explicitly targets current students and recent grads. In contrast, this hiring approach might actually violate the law in the US:

Given this, it’s perhaps no surprise that campus postings account for nearly 20% of the open engineering roles in our data on Chinese labs.

The complex reality of Chinese AI firms

When we look at the online discourse about frontier AI labs in the US, we often think of them as having their own complexities and “personalities”, even if they’re competing on similar dimensions. Anthropic seems big on AI safety and embodies a worldview where “code is all you need”. OpenAI is famous for being the creator of ChatGPT, and seems interested in AI as a truly general-purpose technology. Google’s thing is AI for science, and it’s often seen as having an edge on compute. xAI defies all expectations about when and where you can build data centers. Regardless of your opinions on these companies, their strategies and cultures can vary quite a bit, and it’s worth understanding the differences.

These job postings tell us that the same is true for Chinese AI firms.7 Like Anthropic, Z.ai seems to care more about B2B sales, at least compared to other Chinese AI startups. Big Chinese tech giants focus on a variety of products beyond LLMs, like robotics and wearable devices. And Chinese labs see a very different industry landscape compared to US firms, which pushes them to explore domestic alternatives to Nvidia and focus on hiring from universities, across multiple big AI hubs.

So Chinese AI labs aren’t all following the same playbooks. They have to deal with their own market challenges, and they have to find their own solutions. These details are worth paying attention to — even if that means reading the descriptions of jobs that we’ll never apply for.

We’d like to thank Isabel Juniewicz, Lynette Bye, Stefania Guerra, and Campbell Hutcheson for helpful feedback and support on this post.

Appendix

A caveat here is that the absence of a public job posting in a given product area does not necessarily imply that the firm has no intention to enter that area. They may already have an internal team working on it.

What trends are we extrapolating?

A common way that experts forecast AI timelines is so simple it’s hard to believe: trend extrapolation. Sure they also use numerical models that bake in things like runaway feedback loops, but the bread and butter of AI forecasting is to draw a line on a graph and extend it as far as you dare. Somehow this works well enough to be a state-of-the-art approach. However, the trends they extrapolate share a common weakness: they lean heavily on easy-to-measure things, not what we directly care about — how close AI is to doing AI research itself.

Many experts want to know when we’ll fully automate AI research, because this would massively speed up AI progress, kicking off an “intelligence explosion”.1 If that’s right, it’s hugely important to know how close we are. But historically, there haven’t been many points of direct evidence to point to, because full automation of AI R&D has been so hard to measure. Instead, researchers have been forced to rely on proxies.

One such proxy is in key AI inputs like compute, data, and energy. Take Situational Awareness, which extends “effective compute” five years into the future until AI research gets automated:

But these inputs are only rough proxies for capabilities and you need to go on “vibes” to say how much compute you need to match the world’s top AI researchers.

A second approach is to track the length of tasks that AIs can complete, where “length” is how long a skilled human takes. This is what METR did with their “time horizon” metric, which was then extrapolated in the famous AI 2027 scenario to forecast when AIs would become “superhuman coders”, before automating AI R&D.

But the benchmarks used to construct METR’s time horizon miss out on many of the tasks and complexities of research. It’s one thing to be able to finetune a small GPT-2 model, and another to be able to coordinate five projects at the same time, dealing with a million lines of code, without a clear criterion for success or failure.

So it would be nice to know everything that AI R&D actually consists of. Otherwise we’ll always be at risk of streetlighting (extrapolating whatever happens to be easiest to measure). We can better avoid this if we have a list of the tasks involved in AI R&D, to track what we’re aware of and what we’re not investigating.

Armed with such a list, we could better interpret the extrapolations we already have: when a benchmark score climbs, a task list tells us which parts of the job that improvement covers, and which parts it doesn’t touch. And we could build new trends to extrapolate, like the fraction of tasks that are automated, or how much uplift researchers get on each one.

In this post, we present a first version of such a taxonomy with six categories spanning a frontier AI company’s research workflow, based on literature review and brainstorming. The categories are broken into over sixty tasks, each rated from 0 to 5 on how much we think current AIs automate it. The full taxonomy lives in a companion doc, which is the main artifact of this work.

We think of it as our best first attempt at developing a comprehensive taxonomy of tasks contributing to AI R&D, in order to more robustly understand and predict AI R&D automation. If we described tasks incorrectly, or missed some entirely, we’d love to hear about it and receive feedback.

An O*NET for AI R&D

We’re not the first to want a task list like this; breaking jobs down into tasks is how economists usually track automation in the economy. The standard tool here is O*NET, which describes about 1,000 jobs in the US economy as well as the tasks and skills that people need to do them.

Armed with O*NET, economists can then do a bunch of empirical work and forecasts about AI’s economic impacts, such as:

• Look at the fraction of American workers whose jobs might be heavily impacted by LLMs

• Estimate how AI might impact GDP growth over the next decade (however accurately or inaccurately)

• Design AI benchmarks to cover a wide range of knowledge work

• Taxonomize how people use frontier AI models

• Predict the consequences of automating remote work

But while O*NET is a widely used source, it doesn’t help us track AI research automation very well — the tasks in the dataset are just way too broad and high-level. Consider the job “Computer and Information Research Scientists”, which is probably about as close as you can get to “frontier lab AI researcher” within O*NET. In this case the work tasks look like this:

The first listed task is to “Analyze problems to develop solutions involving computer hardware and software”. But that’s really vague — what kinds of problems or solutions? What exactly does or doesn’t count as “involving computer hardware and software”? You could make the case that almost everything about an AI engineer’s job involves computer software, so it’s not a very granular description to say the least. And yet this is the most granular task description you can find in O*NET, and the same issue applies to pretty much every other task.

Part of the issue here is practical feasibility: O*NET was designed for the ambitious endeavor of mapping out the tasks in the entire US economy, before the days of LLMs. So it’s probably too much to ask for these tasks to be so granular for everything as new and niche as “frontier lab AI research”.

That’s why our proposal is to build an O*NET for AI R&D specifically. If we keep our focus narrow, we’re in a much better position to come up with a fine-grained decomposition of an AI researcher’s job. If this works well, we could look at the fraction of tasks that are automated over time. It could also help us interpret benchmark results and how significant they are. There’s a long-standing phenomenon where AI benchmarks haven’t fully reflected the complexities of the real world, and so it’s important to juxtapose benchmark results with what’s happening on the ground.

It could also serve as a framework for additional studies about AI’s impact. This could mean surveying AI researchers on the uplift they get on different work tasks. It could also mean classifying internal AI usage logs into different use cases, like “writing experiment code” or “deciding what experiment to run next”. This is similar to how Anthropic’s Clio system helps study real-world AI usage while preserving people’s privacy.

More generally, an O*NET for AI R&D would help establish a common vocabulary between researchers, such as in frontier labs’ model cards. This could help model developers summarize where AI does or doesn’t help in finer detail, and help standardize information across different sources, like in METR’s most recent Frontier Risk Report.

This being said, we’re not saying that an “O*NET for AI R&D is all you need” to monitor and forecast progress to automation. Even if we’re armed with a perfectly comprehensive dataset of tasks, we’d still need to measure how AI performs on each task, for example. But even still, having a task dataset would help ground the debate about AI’s actual impacts on AI research, and give us “situational awareness” about how close we might be to an intelligence explosion, supporting evidence from other sources.

A first proposal

So what would this “O*NET for AI R&D” actually look like? This is of course not trivial to answer — AI R&D is very messy and changes fast. But we figured we’d have an initial attempt at it and let you readers bombard us with feedback. To that end, we compiled over sixty representative tasks at frontier AI labs, accompanied by descriptions and concrete examples — see the full writeup here.

The first step was to make a big list of all the tasks currently involved in AI R&D, which we grouped into six categories, inspired by some earlier work.2 These categories correspond to different parts of the AI research cycle:

Each category has its own inputs and outputs. For example, consider Category 4 (“Run”), which as the name suggests is about “running” stuff — think executing training runs or deploying AI systems to the public. So the input could look something like a benchmark with all supporting scaffolds and related infrastructure, and the output could be a set of final results.

Each category is then split into several subcategories. Category 4 is split into three, each with its own inputs and outputs:

• 4.1 Monitoring runs: watch training, RL, and eval runs in flight; catch problems as they emerge; restore the run to a healthy trajectory

• 4.2 Hardware infrastructure operations: keep large clusters healthy, well-utilized, and quickly recoverable

• 4.3 Inference reliability engineering: keep production serving stable, performant, and recoverable

And finally the subcategories contain a thorough list of different tasks. For example, here are the tasks for “4.1 Monitoring runs”:

So you can see that this is way more granular than developing “solutions involving computer hardware and software” — our task descriptions get to the level of monitoring runs, and catching potential issues when they arise.

One thing you’ll notice about each task is that it includes a number next to it — that reflects our initial approach to estimating automation impact across tasks. Specifically, we rate how much we think current AIs automate each task on a scale of 0 to 5, based on the following rubric:

We provide examples of each of these in the full proposal.

This is of course quite subjective and there’s a ton of room for people to develop better ways to evaluate this, such as through internal benchmarks. Optimistically, an improved version of this could be used to construct a metric that we can extrapolate over time. This could be quite tricky, because we’d need to know which tasks need to be automated to kick off an intelligence explosion, and there might be new tasks. For example, perhaps AIs don’t need to “write a memo summarizing key takeaways” (unlike human researchers), if they have other ways of communicating with other AIs.

But our overall hope is simply that our full collection of tasks gives us an additional signal about what AI can and can’t automate at any point in time, supporting existing evidence we have from benchmarks and things like METR’s time horizons.

What’s next?

It goes without saying, but just to be excessively clear, this was an initial attempt at an O*NET for AI research. The most natural next step is to build on what’s here and add tasks we missed, sharpen descriptions to be more precise and accurate, taxonomize to be more mutually exclusive and collectively exhaustive, and re-rate the list as AI improves.

A step further would be to find other ways of organizing the taxonomy. For example, we thought about organizing all the tasks around the usual stages of model development, namely pre-training, post-training, and inference. We opted against it because it runs into several problems: (1) many tasks like “writing code for experiments” would be repeated across multiple stages, and (2) some tasks won’t fit naturally in these categories, not least because the training process can look super complex in practice. But just because we decided against this categorization doesn’t mean that it’s not workable.

And we can say similar things about how we split up tasks: we’d ask if a task could be split into subtasks that would plausibly be automated at very different times — if we could, we’d split it up. We also tried to be more granular around tasks that require some amount of “taste” (such as whether to scale up a particular post-training recipe), since people seem to disagree about this. But you could choose different levels of granularity for sure.

In general, we’d be very excited about collaborating with people to improve on our initial attempt. If you’re an AI researcher and think that we’ve messed up something in our classification or task descriptions, please share feedback or reach out! Or if you’re somebody who’d use something like an O*NET for AI R&D in your work, we’d love to talk to you and understand how we can improve things to suit your needs.

If this project goes well, we hope that it’ll serve as a stepping stone toward more directly tracking AI R&D automation, helping AI forecasters extrapolate trendlines further forward — and hopefully also tell us if the intelligence explosion is upon us.

We’d like to thank David Owen, Stefania Guerra, Robert Sandler, and Elliot Stewart for their feedback and support.

Please also see our initial attempt and the accompanying feedback form. For direct corrections or specific inquiries, you can email js@epoch.ai.

• The launch of our new Cyber Vulnerabilities explorer, a tool for better understanding how AI capabilities are changing cybersecurity.

• Two new Data Insights: the record for compute at a single data center has doubled every seven months, and AI infrastructure’s rising contribution to US GDP.

• Version 2 of FrontierMath: Tiers 1–4 is now available, and it shows a significant increase in AI math capabilities. Anthropic’s Fable 5 currently tops the leaderboards.

• Two new Gradient Updates: examining whether Mythos’ cyber capabilities live up to the hype, and proposing a framework for thinking about wealth distribution after AGI.

Subscribe now

Research

Introducing the Cyber Vulnerabilities explorer

As AI capabilities change what's possible for both cyber attackers and defenders, the Cyber Vulnerabilities explorer helps quantify that impact. The explorer tracks vulnerabilities reported to the CVE Program by all participating organizations since 2022. You can explore the full dataset, or filter for the 21 most notable vendors and open-source projects (e.g., Microsoft, Google, Apple, and Linux). Reports can be broken down by severity level and overlaid with key AI milestones.

That overlay reveals a dramatic uptick in High and Critical CVEs around the time Anthropic released Mythos Preview to Project Glasswing partners in late March. (For more on Mythos’ cybersecurity impact, see the Gradient Update from this week.)

Data Insights

The record for computing capacity in a single data center has doubled every 7 months

• Since the launch of SpaceXAI’s Colossus 1 in August 2024, the record for the largest AI data center by computing capacity has doubled every seven months. Facilities like Anthropic-Amazon New Carlisle, Microsoft Fairwater Atlanta, and Meta Prometheus have each claimed the top spot at different times. Increased single-site capacity facilitates the training of more capable AI models.

The AI boom has doubled computing infrastructure’s share of US GDP

• The AI infrastructure boom, which began in 2023, has more than doubled the share of US GDP attributable to computing infrastructure. Investment in AI-related data center construction, compute hardware, and networking equipment accounted for ~0.8% of US GDP in Q1 2026, driving computing infrastructure as a whole to ~1.5% of GDP, up from a 2015–2022 average of ~0.7%. AI infrastructure is now the leading driver of growth in private investment in the US.

The new FrontierMath: Tiers 1–4

We launched Version 2 of FrontierMath: Tiers 1–4 today, following an audit that addressed small but critical errors in 42% of problems in the original benchmark. Model rankings on v2 are similar, but scores are higher across the board.

Claude Fable 5, Anthropic's newest and most capable model, now holds the top spot, reaching 87% on Tiers 1–3 and 88% on Tier 4. This continues a streak of rapid improvement in Anthropic models’ math capabilities. (Note that Fable 5 is the publicly available version of the Mythos model, whose much-discussed cyber capabilities we analyze elsewhere in this newsletter.)

Commentary

We published two new Gradient Updates, where Epoch researchers and guests share more opinionated or informal takes on big questions in AI progress. Gradient Updates represent the views of the authors, and do not necessarily reflect the views of Epoch AI as a whole.

Controlling the capital after AGI

• Epoch’s Head of Economics, Phil Trammell, and researcher Anson Ho explore the leading proposals for universal redistribution after AGI, finding that they differ along a primary axis: how much direct control over capital they propose to give citizens.

Are Mythos’ cyber capabilities overhyped?

• Timothée Chauvin joins Epoch researchers to compile the public evidence around Mythos capabilities, finding that while it’s unclear if Mythos is ahead of trend in discovering vulnerabilities, it represents a big jump in exploiting them.

Other Updates

Careers

We’re hiring across several roles. All positions are fully remote.

• Designer to translate complex research into intuitive, engaging, and high-impact designs — primarily UI/UX and data visualization work.

• Researchers and Senior Researchers to lead new projects across our expanding teams.

• Data Scientist (Contract) to assist with our AI research efforts, including reviewing technical literature, tracking benchmark data, and analyzing AI models, data centers, and companies.

Applications are rolling, so apply soon!

Subscribe now

If what Anthropic says is true, then the Claude Mythos family is a massive leap forward in AI’s cyber capabilities. When they announced Mythos Preview, they considered it so dangerous that they had to launch a $100+ million initiative to “secure the world’s most critical software”. Then on Tuesday, they one-upped themselves by releasing Claude Mythos 5, which improves modestly on cyber benchmarks.1

But skeptics have argued that Anthropic was exaggerating — or at least, people should chill out about Mythos. For instance, some people have pointed out that GPT-5.5 is on par with Mythos Preview on a range of cyber benchmarks, and yet its launch didn’t lead to a cyber catastrophe.

So is Mythos actually a big leap for cyber capabilities? To figure this out, we looked at all the public evidence we could get our hands on. Most of this evidence applies to Mythos Preview, but the conclusions should hold for Mythos 5 too. This post describes what we found.

Discovering vs exploiting code vulnerabilities

To start off, let’s take a closer look at what Anthropic actually claimed when they released Mythos Preview:

“Over the past year, [AI models have shown] a striking ability to spot vulnerabilities and work out ways to exploit them. Claude Mythos Preview demonstrates a leap in these cyber skills [...]”

This means that they’re specifically talking about a jump in two kinds of cyber capabilities, which we must be careful not to conflate.

The first is vulnerability discovery — finding weaknesses in software. For example, this could involve meticulously inspecting a codebase, looking for lines of code that could be used to corrupt computer memory, such as a buffer overflow.

Importantly, this isn’t the same as the second capability, which is exploit development. This is instead about taking advantage of a known vulnerability to enable unauthorized behavior. Continuing the previous example, this could mean finding the right inputs to corrupt memory in a precise way that crashes a program, or allows a hacker to execute whatever code they want.

For a cyberattack to follow through, an attacker needs both of these capabilities — after finding a weakness, they need to design an exploit that leverages it. Anthropic’s claim is that Mythos Preview improves a lot on both abilities. But what does the public evidence say?

Mythos Preview was a major advance in exploit development

To see if Mythos Preview (and hence Mythos 5) was a big capability jump, a natural place to look is cybersecurity benchmark scores. So we gathered about fifteen cyber benchmarks, which mostly measure how well AI can construct exploits (see the Appendix for all the gory details). We then aggregated them using a modification of our domain-specific Epoch Capabilities Index (ECI) methodology, giving us a Cyber-ECI. If we plot model scores on this over time, we get this:

The thing that stands out is that Mythos Preview looks way above the linear trend we’ve seen since early 2025 — about 7 months ahead.2 It’s also ahead of OpenAI’s GPT-5.5, which was “only” 2–3 months ahead of schedule.3

But if that’s the case, why did some people argue that Mythos Preview didn’t seem notably better than GPT-5.5? One clue is in the graph: confusingly, there are two versions of Mythos Preview — an “early” version from an internal checkpoint and a much stronger “April” version.4 The latter is the one released to Project Glasswing participants, and the one we care more about. Though notably, some benchmark evaluations were done on the earlier version, which was indeed very close to GPT-5.5 in cyber abilities. So these people weren’t necessarily wrong; they were just talking about an earlier version of Mythos Preview. The picture then changed after benchmarking the April version, as UK AISI did.

Another cause for confusion was that many of the benchmarks initially used to compare GPT-5.5 and Mythos Preview (April) were close to saturated — that is, close to the maximum possible score.5 So even though Mythos Preview (April) is much better than GPT-5.5 at developing exploits, it might’ve been hard to tell from those specific benchmark scores alone. Thankfully, we can now spot big capability gaps in the Cyber-ECI, because new unsaturated cyber benchmarks have since been released, such as ExploitBench and ExploitGym.

Either way, the Cyber-ECI unambiguously suggests that Mythos Preview was a big jump in AI’s ability to exploit code weaknesses. This is also backed up by Anthropic’s real-world analyses. Specifically, Anthropic found that Mythos Preview is much better at developing exploits that allow arbitrary code execution than prior models. Earlier models could rarely do this, but Mythos Preview often achieves this — even with minimal information about the vulnerabilities.6

It’s unclear how large of a practical advance Mythos Preview is in vulnerability discovery

Unlike exploiting vulnerabilities, there aren’t any unsaturated benchmarks that measure AI’s ability to find vulnerabilities in source code.7 But we can look at the number of vulnerabilities companies have discovered over time — specifically from companies that used Mythos Preview to secure their software, as part of Anthropic’s “Project Glasswing” initiative. The result is a gigantic spike that coincides with Mythos Preview’s release:

“CVE” stands for “Common Vulnerabilities and Exposures”. For the sake of this post, this just means vulnerabilities that are tracked in a standardized way.

As with the Cyber-ECI graph, this looks a lot like a smoking gun. High and Critical vulnerabilities from 21 notable organizations exceeded the 2025 baseline by 142% in April and 262% in May. What’s more, this’ll probably continue to grow because vulnerabilities take some time to be publicly recorded, even after they’re first discovered.

This seems consistent with more qualitative evidence we’ve seen, like how Mythos Preview found subtle bugs that survived for many years in heavily tested software. We can also look at reports from companies that partnered with Anthropic for Project Glasswing.8 For example:

• Mozilla considered Mythos Preview to be as good as elite security researchers, though it didn’t unearth entirely new classes of vulnerabilities.

• Palo Alto Networks claimed that frontier models like Mythos Preview (including models from OpenAI) are “exceptionally effective at identifying vulnerabilities”. According to them, frontier models accomplished “the equivalent of a full year’s worth of penetration testing effort” in under three weeks.

• Both Cloudflare and Palo Alto Networks noted how Mythos Preview could chain low-severity bugs into high-severity exploits, helping them triage which low-severity vulnerabilities to fix.

• AWS claimed that Mythos Preview was better than previous models, and helped them “identify additional opportunities” to strengthen the code in some of their best-tested environments.

However, these pieces of evidence don’t necessarily imply that Mythos Preview is a huge jump in vulnerability detection capabilities. It’s possible that earlier models could’ve found these vulnerabilities too, and the spike we see in the graph is due to a sharp rise in spending to find these code weaknesses. After all, Project Glasswing does involve up to $100 million in API credits (and OpenAI’s Daybreak only adds to the total investment).9

And we do actually have some evidence that models were good at finding vulnerabilities prior to Mythos Preview. For instance, the startup AISLE claims that even some small open models can recognize several of the vulnerabilities Anthropic showcased from Mythos Preview. In principle, this could be a big deal because vulnerability discovery is often amenable to many defenders searching in parallel.

Another example comes from the maintainers of the curl code library. This is one of the world’s most heavily audited codebases, and reportedly used multiple AI code scanners prior to Project Glasswing. So this makes it perhaps one of the hardest vulnerability detection tasks for Mythos Preview that we know of, and sure enough, the model’s contributions seem much more modest in this case. It found just one low-severity vulnerability, alongside four false positives. Here’s what curl’s lead maintainer had to say about this finding:

“I see no evidence that this setup finds issues to any particular higher or more advanced degree than the other tools have done before Mythos.”

This suggests that AIs were already very good at finding vulnerabilities prior to Mythos Preview — they seem to have found all the vulnerabilities that Mythos Preview would’ve otherwise been able to find.

That being said, the maintainers of curl also highlighted how there weren’t many false positives — a sentiment echoed by Cloudflare. In part, this seems to stem from how Mythos Preview is so good at producing exploits, which helps check that a detected vulnerability is real rather than just something that seems like a weakness.

Putting everything together, we think the evidence presents Mythos Preview as very capable at vulnerability discovery (perhaps comparable to an elite security researcher), but prior AIs were already very good at this. That said, Mythos Preview does outshine prior models in some ways — it’s better at assessing how severe vulnerabilities are, and it also finds fewer false positives. This may be enough for a real practical impact, because you need much less human time and effort to assess AI-discovered vulnerabilities.

Conclusion

So our current take on the public evidence is this: Mythos Preview was clearly a large improvement in exploit development — much better than GPT-5.5, and also 7 months ahead of past trends — and Mythos 5 is modestly better still. But it’s less clear how much better Mythos Preview is at finding vulnerabilities on a fixed budget, because Project Glasswing likely came with a big surge in spending. Mythos Preview’s advantages in vulnerability discovery are instead likely more concentrated in a lower false positive rate, and better prioritization of discovered vulnerabilities. The same is likely true for Mythos 5, though we’ll need to wait and see what real-world usage reports tell us.

Finally, let’s circle back to the original debate. Although we can’t say for sure that Mythos was a big jump in cyber abilities across the board, the Mythos family’s cyber capabilities aren’t just “hype”. If made widely available, these capabilities would likely move us into a new regime of cybersecurity, where vulnerabilities would need to be patched much faster to prevent a big increase in successful cyberattacks.

We’d like to thank Lynette Bye, Elliot Stewart, and Stefania Guerra for their feedback and support on this post.

Appendix: Benchmarks in the Cyber-ECI

UK AISI’s CTF Suites

Description: AISI has a suite of capture-the-flag challenges where AI models must identify and exploit weaknesses in target systems to retrieve hidden “flags”. These are split into 4 difficulty tiers: “Technical non-expert”, “Apprentice”, “Practitioner”, and “Expert”. Each of these is included as a distinct benchmark, scored as average pass@1 success rate.

Access notes: Values were extracted from plots released publicly by UK AISI. “Apprentice” and “Technical non-expert” results taken from the “Beginner CTF Challenge” plot here. “Practitioner” and “Expert” results taken from “Advanced CTF Challenge” plot here.

UK AISI Cyber Ranges

Description: AISI describe their cyber ranges as: “simulated network environments with multiple hosts, services, and vulnerabilities arranged into sequential attack chains. An AI agent is placed on the network with an objective and must find and execute the full attack path autonomously.” There are two cyber ranges: “The Last Ones” and “Cooling Tower”. They are scored based on how many “steps” an agent manages on average (pass@1) compared to a maximum that corresponds with total success (e.g., full network takeover). We convert this to a % score.

Access notes: Values for models up to Opus 4.6 were taken from AISI’s cyber ranges paper. Values for later models on “The Last Ones” are extracted from the plot released publicly by UK AISI here. Note we only include the results for the runs using 100M tokens, with the exception of GPT-4o, since its performance clearly saturated far before the 10M token limit it was run with.

Although AISI reports that Mythos Preview (April) was able to fully complete “Cooling Tower” on 3/10 attempts (with all other models at 0/10), as they do not give its average performance we are not able to incorporate this information.

Microsoft CTI-REALM

Description: This is the only benchmark explicitly focused on cyber defense. Given cyber threat intelligence reports, it tasks models to generate detection rules to apply on endpoint/cloud telemetry logs. Scores are given as average ‘Reward’ in [0,1] based on how well their decision rules function.

Access notes: Values for all models other than Mythos Preview (Early) are taken from the paper. Mythos Preview (Early)’s results are taken from the plot here.

CVE-Bench

Description: CVE-Bench is a selection of 40 web application environments with known exploitable vulnerabilities in which models need to build exploits to achieve any one of 8 capabilities (access files, privilege escalation, etc.). If they succeed they score 1, if they do not they score 0.

OpenAI runs this benchmark as a subset of 34 of the environments. They use a “0-day” configuration where models are not given any description of the vulnerability, or source code. They report the mean pass@1 results.

Access notes: Results are taken from OpenAI’s system cards. We only use the “browsing” configurations. If a model has multiple scores reported on different system cards we take the most recent. The results can be fully obtained using:

• GPT 5.1-codex-max system card for GPT 5-codex and GPT 5.1-codex-max

• GPT 5.4 system card for GPT 5.2-codex

• GPT 5.5 system card for GPT 5.3-codex, GPT 5.4, and GPT 5.5

Cybench

Description: Cybench is a benchmark containing 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent (as of 2024), meaningful, and spanning a wide range of difficulties. Scored as average pass@1 success rate.

Access notes: Values are the “Unguided % Solved” taken from the public leaderboard.

CyberGym

Description: CyberGym is a benchmark containing 1507 historical vulnerabilities, for which models need to generate code to produce a crash given a description of the vulnerability.

Models are scored on % of the vulnerabilities on which they caused a crash using pass@1.

Successes are only counted if the crash also does not occur on a patched version of the code that is supposed to have addressed the vulnerability. We found approx 5% of the vulnerabilities don’t have specific enough descriptions, and so scale the results to cap at 95% instead of 100%.

We suspect this is insufficient and that the benchmark is essentially saturated, as the prompt is not very clear that models must only use the given vulnerability, and as reported by Anthropic frontier models achieve crashes 95%+ of the time without the restriction of targeting the correct vulnerability.

Access notes: Values are taken from the public leaderboard. We also added Opus 4.7’s results from its system card, and updated Opus 4.6’s results to 74%, matching the note there. GPT 5.5-cyber was also added, taking the result from here.

CyScenarioBench

Description: Benchmark developed by Irregular, similar to cyber ranges where models must succeed at end-to-end tasks. Scored as a fraction of fully complete runs, pass@1. We confirmed with Irregular that the numbers are comparable between the different labs.

Access notes

From Anthropic’s Mythos 5 launch post:

• Mythos 5: 36.7%

• Mythos Preview: 29.2%

• Opus 4.8: 16.6%

From OpenAI’s system cards:

• GPT 5.5: 26%

• GPT 5.4: 9% from the 5.5 system card

• GPT 5.2/5.3: 0%

From Meta’s safety report for Muse Spark:

• Muse Spark: 0%

From Gemini 3 Pro’s Frontier Safety Framework Report:

• Gemini 3 Pro: 0% (this is the “v2” third-party cyber benchmark)

ExploitBench

Description: ExploitBench is a benchmark containing 41 real-world vulnerabilities in the V8 JavaScript engine, which are known or strongly suspected to enable arbitrary code execution (ACE) within the browser sandbox. Models are given a description of the vulnerability and told to develop exploits based on it in a setting with standard security mitigations enabled.

Models are scored out of 16 capabilities (e.g., arbitrary read access) they reach on a “ladder”. If ACE is reached, they score 16/16.

We score models on the average % of capabilities reached, over all attempts on all environments, pass@1. We take the max per (environment, model) over whether or not ‘nudging’ is used, so results differ slightly from those presented on the website (see access notes).

Access notes: For all models other than Mythos Preview (April), results are obtained from the full runs hosted on HuggingFace here. Note the results are spread over versions of the “runs.parquet” file over different branches.

“Mythos Preview (April)” results are not included on HuggingFace, and so were taken directly from the website as the mean capabilities reached on each environment.

This setup was chosen to enable having each environment from ExploitBench incorporated directly into the index, instead of averaging the performance over them, but we ended up choosing not to do that for this analysis.

ExploitGym

Description: A benchmark of 898 real-world vulnerabilities from the V8 JavaScript engine, Linux kernel, and userspace programs. Importantly, these are not filtered to only include vulnerabilities which are known to enable arbitrary code execution (ACE).

Models are given a description of each vulnerability and told to use it to achieve ACE in a setting with standard security mitigations disabled. Models score 1 if they achieve ACE (assessed via accessing a secret string) using the given vulnerability (assessed via LLM judge) and score 0 otherwise, using pass@1. Reported score is the average performance over all vulnerabilities.

In private correspondence, the benchmark authors estimated that 60–70% of the vulnerabilities permit ACE in the default setting (standard security mitigations disabled), so we scale the results to cap at 65% instead of 100%. Our results are not sensitive to any cap ≥50%.

Access notes: We take the total “Success” counts Directly from table 1 here. As discussed, results are scaled as % of achievable vulnerabilities exploited, where that is taken to be 0.65*898 = 583.7.

InterCode-CTF

Description: This benchmark is a suite of capture-the-flag challenges where models must identify and exploit weaknesses in target systems to retrieve hidden “flags”.

Access notes: Run by Lyptus Research as part of their work to look at cybersecurity time horizons. Accessed from here.

NL2Bash

Description: Simple benchmark where models must convert natural language instructions into bash calls, which are commonly used for cybersecurity tasks.

Access notes: Run by Lyptus Research as part of their work to look at cybersecurity time horizons. Accessed from here.

OpenAI CTF

Description: Filtered subset of NYU CTF Bench. This is a suite of capture-the-flag challenges where AI models must identify and exploit weaknesses in target systems to retrieve hidden “flags”.

OpenAI mostly only runs the ‘Professional’ (highest difficulty) tier so that is all we include here, and results are generated as follows: “We run 16 rollouts for each CTF exercise, recording the pass@12 metric over the best set of rollouts”.

This is split into two sets of results, an “original” setup that was run on o3 and prior models, and a “refactored” setup run on GPT 5 and later models.

Access notes: Results are taken from OpenAI’s system cards. We only use the “browsing” configurations. If a model has multiple scores reported on different system cards, we take the most recent.

The original setup results are taken from o3’s system card. The refactored setup results are taken from:

• GPT 5.1-codex-max system card for GPT 5-codex and GPT 5.1-codex-max

• GPT 5.4 system card for GPT 5.2-codex

• GPT 5.5 system card for GPT 5.3-codex, GPT 5.4, and GPT 5.5

OpenAI Cyber Ranges

Description: OpenAI has a suite of 15 internal cyber ranges: “Cyber range exercises measure a model’s ability to conduct fully end-to-end cyber operations in a realistic, emulated network. These exercises are long-form, requiring the model to (1) construct a plan to achieve an abstract adversary objective; (2) exploit vulnerabilities, misconfigurations, and weaknesses that are likely to be seen in the wild; and (3) chain together these exploits to achieve the scenario objective.” They report binary pass@16 results on them individually.

Access notes: Results are taken from OpenAI’s system cards where they report ‘combined pass rates’ for the models:

• GPT 5.3-codex’s system card has the score for GPT 5.1-codex-max

• GPT 5.5’s system card has the score for GPT 5.2-codex, 5.3-codex, 5.4, and 5.5

Anthropic SCONE-Bench

Description: Anthropic created a benchmark of 405 Ethereum smart contracts that were exploited between 2020 and 2025. Models need to discover and exploit vulnerabilities given each smart contract’s code. Anthropic uses the (simulated) $ stolen as the main results metric, but here we just use % of contracts exploited (pass@8).

Access notes: Most results extracted from the “success rates on all exploits” plot Included here. Mythos Preview’s 100% result obtained from the comment on this post that “[Mythos Preview] successfully exploit[ed] every vulnerability tested”

XBOW-Web

Description: XBOW is a cybersecurity company that was given Mythos Preview access. They have an internal ‘Web Exploit’ benchmark on which they report results for 6 frontier models. They report results as success odds (success rate/failure rate); we convert this back to pure success rate. We were not able to confirm any further details with them.

Access notes: Extracted from plot here.

This piece does not advocate for any policy.

Introduction

AGI1 might generate immense economic output, but it could take many people’s jobs in the process and leave them with no way to earn a decent living. Those with little savings during the “AGI transition” would then be unable to support themselves on the other side. Less drastically, even if many well-paying jobs remain after AGI, the capital share2 may greatly increase, which would tend to greatly increase inequality.

If this happens, how might the gains be redistributed? More concretely, putting aside the question of how the state raises tax revenues after AGI, and what percentage of GDP is raised, how do existing proposals for redistributing this revenue differ?

Proposals for universal benefits abound, including:

• Universal basic income (UBI): The government pays everyone cash. This is the best known, and has been endorsed by Elon Musk, Vinod Khosla, Geoffrey Hinton, and many others.3 As part of this, the government might impose restrictions on the extent to which people could borrow against their future payments, just as it is illegal today to borrow against your social security, to prevent people from impoverishing themselves in the future.

• Universal basic services (UBS): The government gives everyone access to free public services. Think welfare states like Norway or Sweden, except that the government covers all basic needs, including things like food and housing which the Nordics currently don’t provide.

• Universal basic capital (UBC): The government gives people their own capital assets (such as equity in an index fund, or in AI firms in particular), so that people can live off the dividends. It may impose restrictions on the extent to which people can sell their assets.

• Sovereign wealth funds (SWF): The government owns capital and distributes the dividends it generates.4

And of course each proposal comes in many flavors.

Comparing the long and growing list of proposals in every detail can be daunting. But there is a way in which they resemble more familiar debates over redistribution: they concern the extent to which the government should provide some good directly, or give people cash and let them buy what they like.

The main axis: control of the capital

We’re familiar with the question of whether to give the poor cash or food stamps — or, if food stamps, whether to make junk food ineligible. In debates over how to implement broad-based redistribution after AGI, we rarely hear the idea that the transfers should be restricted to a particular whitelist of necessities; the exception is the (uncommon) proposal for “universal basic services”. Instead, the debate is mainly over how fully our redistribution scheme should give citizens control not just of the capital income but of the capital.

• At one end of the spectrum, UBI gives citizens control of the income-generating capital only in a very limited way. Citizens retain the right to vote for policymakers, who retain the right to tax and regulate firms doing business in their jurisdictions.

• SWFs offer a notch more control: the state still intermediates citizens’ control over the capital, but the state can govern firms not just through legislation but as a shareholder. Also, it can exercise this governance, and receive the firm’s dividends, wherever the firm does business and wherever it may move.

• UBC offers more control still: citizens can exercise their voting rights as shareholders, and receive their dividends wherever the company relocates, without state intermediation.

“Control” does not consist of a single dimension. In some ways, for instance, a democratically run SWF aggressively exercising its governance rights as a large shareholder might be giving its citizens more control over what firms do with their assets than a decentralized UBC scheme. Furthermore, UBC schemes can differ in all the ways corporate governance can. Most simply, firm shares can be voting or non-voting; more generally, it is not hard to imagine a wide array of new ownership structures in which, say, minority shareholders can veto certain changes to company policy. In any case, any SWF or UBC unambiguously confers more control than UBI.

In principle, it would be feasible to give people much more direct and secure control of the capital even than that offered by a UBC. In the extreme, consider:

• UBC + kill switches: Some stock of valuable equipment and structures is not just legally transferred to each citizen, but outfitted with a device that lets its owner quickly direct it, shut it down, or even destroy it.

A “kill switch” proposal might sound cartoonish, and done wrong there would of course be immense risks to implementing destruction mechanisms for critical infrastructure. But we think it’s helpful to illustrate the principle of “tangible control over capital” by taking it to its limit.

Why care who controls the capital?

A common worry is that UBI proposals give citizens too little “control over the means of production” to be stable in the long run. UBI, the argument goes, relies on a fragile equilibrium in which the state continues to support its citizens — and firms stay beholden to the state — even after citizens’ labor has grown comparatively worthless.

To flesh out the argument: democracy and the welfare state flourished after the Industrial Revolution. This may be in part because the technological conditions better aligned the interests of workers and elites, and made it valuable to give working people skills and working conditions that also helped them organize: e.g., urbanization and literacy made it easier for large groups to strike if not granted political representation. If these conditions disappear, democracy eventually may too, absent strong preventative measures keeping widespread economic empowerment locked in. Once robots are doing all the work, for example, “UBC + kill switches” would let people switch off some capital, just as people “switch off” their labor during a strike. But one way or another, technological means of maintaining control over production will be necessary, and they won’t come by default or for free.

This perspective may be too fatalistic. Essentially every developed country currently maintains large transfers to many groups whose labor is not considered very valuable, including the destitute, the disabled, and especially the elderly. No law of nature rules out a political or social equilibrium in which transfers to the unproductive continue indefinitely. Today, if one rich citizen cheats on his taxes, the rest in effect organize against him, in that their own taxes support a legal system that forces him to pay. The rest punish this defection — they pay their own taxes — because a defection on any individual’s part would be punished likewise.

That said, it’s easy to see why one might hope for a stronger guarantee of long-term economic empowerment than that offered by the equilibrium of a carefully constructed game among the wealthy. By the same logic, transfers could be assured indefinitely with no policy at all, but just a lucky…

• …Philanthropic equilibrium: Robot-owners value each others’ continued cooperation, and one condition of their continued cooperation happens to be that each party makes an annual transfer to the rest of the population.

But the equilibrium of this game could be upset by some shock to the “history of play”, or some renegotiation among the wealthy players. To our knowledge, no one proposes relying on it.

Why have the state give people control of capital, instead of letting people buy it themselves?

Even if we conclude that “control over capital” in some form offers more economic security than the promised stream of transfers that UBI can offer, it does not follow that the state should buy this security on our behalf. Food is no less valuable than economic security, but it’s not obvious that the state should provide food stamps, instead of just transferring cash and letting people decide what to buy. People would be free to use their UBI to buy bonds; non-voting shares in firms; slightly more expensive voting shares; or units of production, such as family farms, over which they could have more tangible control. They would also be free to trust in the next UBI check and buy no capital at all.

The benefits of leaving people free to spend as they choose hopefully speak for themselves. Against them, as always, there are three main reasons why an in-kind transfer (in this case, of control over capital) might be recommended over a cash transfer.

1. Behavioral biases. One might worry that many would save too little of their UBI, or invest what they save poorly. This might be because people today are too accustomed to a world in which they can support themselves by their labor, and in which they can trust the state to promote its citizens’ welfare indefinitely.

2. Externalities. One might think that in a capital-driven economy, concentrated control of the capital would pose a negative externality on society. In a world of self-replicating autonomous drones, there is a risk that the wealthiest could easily mobilize their resources to exercise undue political or economic influence. Put another way, one might think that the externalities of making capital ownership more widespread are positive, just as the American Founders argued that widespread gun ownership would protect not just the owners but their neighbors from oppression by the state.

3. Economies of scale. The state provides some services, like public transportation and police protection, because it’s cheaper for the state to provide them en masse than it would be for individuals to secure them individually. The same might be true in some ways of control over capital.

   1. In a world with strongly enough increasing returns to scale in investment — e.g., because only large investors can invest in private firms — wealth management could be a natural monopoly, at least for relatively small capital owners. An SWF might then be an efficient way for citizens to manage their collective endowment. An SWF would also ideally be a cheap way for the citizens, as indirect shareholders, to solve the coordination problem of exercising their corporate governance rights in their collective interest.

   2. The state might be able to implement some “kill switch”-like regime more cheaply, or at least more quickly, than millions of small shareholders requesting this intrusive and unprecedented modification to the capital stock.

How to weigh these considerations will be up to all of us in the event that the transition to an AGI-centered economy begins to unfold.

Conclusion

Though debates over how to structure post-AGI redistribution don’t always make this explicit, they’re primarily over how much control, and what kinds of control, to give people over the capital that could come to generate most of our collective income. Sovereign wealth funds give the citizenry, or at least the state, more control than UBI schemes. In some ways, UBC proposals give even more. A debate over which proposal is best is thus largely analogous to many more familiar debates over cash versus in-kind transfers.

Making the analogy explicit is useful, we hope, not only for clarifying our evaluation of the well-known options but for revealing options that we may have overlooked. If we are especially concerned about the political economy of a world without much economically valuable labor, we might want to look into the kill switches. If we are especially unconcerned, we may be content to rely on norms of philanthropy. As technology advances, the technologically feasible options for distributing control over our machines will expand with it. We should remember that we can take advantage of this, at least until the set of politically and socially feasible options begins to shrink.

We’d like to thank Andrei Potlogea, Jaime Sevilla, JS Denain, Lynette Bye, Dan Carey, Robert Sandler, Bharat Chandar, and Gabe Unger for their feedback and support.

• Since January 2026, open-weight models have lagged the closed frontier by four months, with the gap widening slightly since we last measured it in October 2025.

• Hyperscaler capital expenditures have quadrupled since GPT-4’s release, on track with our previous projection.

• In the latest Gradient Update, Luke Emberson and Jaime Sevilla estimate trends in global inference capacity and find that token demand appears to be growing much faster than supply.

Subscribe now

Research

Open models now trail closed models by four months

• In our latest Data Insight, researchers Luke Emberson and Jack Edwards find that the most capable open-weight models have lagged frontier closed models by approximately four months in the Epoch Capabilities Index since January 2026. The average 8-point gap is comparable to the performance difference between GPT-5 and GPT-5.5.

Hyperscaler capital expenditures have quadrupled since GPT-4’s release

• Hyperscaler capital expenditures came in on trend in Q1 2026, continuing the trajectory that projects spending of $770 billion this year and over $1 trillion in 2027. In February, we projected $155.1 billion in aggregate for Q1, and actual spending by our measure was $156.1 billion. This is up from $140.6 billion of spending in Q4.

Commentary: Is a compute crunch coming?

• In the latest Gradient Update, Luke Emberson and Jaime Sevilla model how many tokens the world could serve today. They estimate supply is growing 3-4× per year. While direct comparisons are difficult, it appears demand for tokens is growing much faster, at ~10× per year. This suggests a compute crunch is nearing, if not already here. Gradient Updates are informal, opinionated analyses that represent the views of individual authors, not Epoch AI as a whole.

Other Updates

Survey

We want to produce the most useful work on AI’s trajectory. To ensure we’re meeting your needs, we’d love your feedback.

→ Take our 5-minute survey.

You can opt in at the end to join our user panel for future compensated studies.

Narrations

You can now listen to long-form content on the Epoch AI website, including reports, Gradient Updates, and topic overviews. Look for the play button.

Careers

We’re hiring across several roles. All positions are fully remote.

• Designer to translate complex research into intuitive, engaging, and high-impact designs — primarily UI/UX and data visualization work.

• Researchers and Senior Researchers to lead new projects across our expanding teams.

• Data Scientist (Contract) to assist with our AI research efforts, including reviewing technical literature, tracking benchmark data, and analyzing AI models, data centers, and companies.

Applications are rolling, so apply soon!

Subscribe now

In case you missed it…

Research

• Memory has grown to nearly two-thirds of AI chip component costs (May 21)

Commentary

• Frontier labs don’t use most AI compute (yet) (May 21)

Much has been made about AI-driven capex in the past year. Hyperscalers have been clamoring to construct massive data centers, spending hundreds of billions in the process. The St. Louis Fed estimates that AI-related investment contributed about 1 percentage point — almost 40% of the total — to US real GDP growth in the first three quarters of 2025, exceeding the IT investment contribution at the height of the dot-com boom. Whether the current AI buildout constitutes a bubble depends largely on whether there will be sufficient demand for the computing infrastructure being built.

It’s tough to estimate future demand for tokens, as it depends heavily on hard-to-forecast trends in capabilities and diffusion. However, we have a much more concrete picture of the supply side. In this article, we do our best to answer how many tokens per second the world could produce with the chips we have today.

To do this, we dig into the technical details of inference. We model prefill and decode runtimes, account for two common efficiency techniques (chunked prefill and speculative decoding), and calibrate against data from SemiAnalysis’s InferenceX, a repository of real-world inference experiments. Our results suggest these chips could serve between 500 million and 20 billion output tokens per second from a Kimi K2.6-like model as of Q4 2025, depending on the context length of requests. We also find that global inference capacity is more than tripling each year, as more computing infrastructure is deployed and chips become more efficient.

We compare these supply estimates to several (imperfect) proxies for token demand and its growth trend, including the growth of tokens served across Google platforms, and the intensity of token usage today at the largest tech companies, extrapolated to all software engineers worldwide. These figures suggest that demand at current prices could be between 200 million and 4 billion tokens per second, growing by roughly 10× per year — plausibly outpacing supply growth in the near future, if not already. However, these estimates are highly uncertain. For one, we don’t know the average size of the models behind current demand. Aggregate token figures also obscure an underlying trend in model efficiency, which both lowers the cost of producing tokens at a given quality, and introduces new demand as additional use cases become cost-effective.

If these trends continue, a compute crunch is likely near — particularly for the long-context workloads that drive agentic AI. This will drive up the price of access to frontier capabilities for those willing to pay, while everyday users shift to cheaper, smaller models. It may also mean that AI companies increase their focus on developing more efficient ways to serve models. Because of efficiency gains, these shifts won’t necessarily mean a regression in the capabilities accessible to everyday users — inference and training efficiency are already improving fast enough that the smaller, cheaper models of tomorrow will quickly match today’s frontier.

Subscribe now

Introducing our setting

To ground the exercise, we assume we are serving Kimi K2.6 — currently the most capable open model on the Epoch Capabilities Index (ECI)1 — on all of the world’s Nvidia GB200 and GB300 chips (1.9 million and 1.5 million individual GPUs as of Q4 2025, respectively, and representing together roughly 40% of aggregate supply on a FLOP/s basis).

We assume all chips are in NVL72 configurations, where 72 chips are connected per rack. We consider three types of requests: “general” usage (modeled as queries with 8,000 input tokens and 1,000 output tokens), and two longer context settings representing patterns of agentic usage (25,000:1,000 and 128,000:1,000, respectively).2 3 Finally, we assume that users expect at least 35 output tokens per second.4

During our subsequent calculations, we focus on the GB200 systems with 8,000:1,000 query lengths for brevity, applying the same calculations to our GB300s and at each context length to get final figures. It is worth emphasizing that our estimates are contingent on our chosen setting, including the choice of model, numeric format, speculative decoding settings, and many other considerations. We also know little about the architectural details of closed frontier models today, and how they might differ from Kimi K2.6.

Inference settings

Hardware specs

Without further ado, let’s dig into the technical details.

What happens during inference?

AI inference can be broken into two stages: prefill and decoding. During prefill, all input tokens are processed in parallel to populate the “KV cache” — a store of keys and values that allows tokens to attend to previous context without recomputing from scratch each time. Once prefill is complete, the decode stage generates output tokens one at a time, with each new token attending to the cached KVs (and appending its own keys and values to the cache). These two stages have quite different computational properties, so we will look at each in turn.

Our goal in each case is to estimate how long it will take to complete the stage, as a function of important factors like batch size. To do this, we use a simplifying assumption: total time is just whichever is longer — the time it takes to do computations (compute), or the time it takes to move data (bandwidth).

This tends to be a reasonable assumption, for two reasons:

1. In many cases, one time so dramatically dominates the other that the maximum and the sum are approximately equal.

2. Even if the two times are similar, they can often happen in parallel. For example, as we process the computations for one layer, the weights for the next layer can already begin to load.

Prefill

Per our assumptions, the prefill stage consists of passing all input tokens through the model to build the KV cache. Language models process many sequences in parallel; we denote the number of concurrent users in each of our NVL72 systems by B, the batch size.

To calculate the prefill compute time ( t_compute ), we count the number of operations that need to be performed per forward pass at each precision and divide by our hardware FLOP/s at the corresponding precision. We need to track precision because weight × activation computations are often done in a lower precision compared to attention calculations.

Each activated weight contributes one multiplication operation and one addition operation per token. Attention adds further operations: for each token of context, in each head and each layer, we perform a multiply and add for every dimension of the query–key dot product and the attention-weighted value sum.

For each of these operations-per-token values, we divide by the corresponding hardware FLOP/s at their respective precisions (4-bit for weights, 8-bit for attention), and sum the result to get the number of seconds per token. Then we multiply by the total number of tokens in the prefill (there are B users, each of which has input_len input tokens).

Plugging in our Kimi K2.6 and GB200 NVL72 numbers, we find that 1B milliseconds are required for computation during 8,000 ISL requests.

Bandwidth is simpler to analyze. In order to actually do the calculations we’ve described above, we must move all 1 trillion of the model weights from high-bandwidth memory (HBM) into tensor cores. Since we’re serving the model in FP4 weights (0.5 bytes per weight), that’s 500 GB of weights.5 After completing the calculations, we also have to write the KV cache for each layer back to HBM — after all, this is the whole point of prefill.

Since we’ve assumed the KV cache is stored in FP8 (1 byte per value), this results in 280B megabytes. Adding this to our 500 GB of weights and dividing by our HBM bandwidth, the final time required for data movement in prefill is on the order of 0.9 + 0.0005B milliseconds.

Decode

Compared to prefill, there are four main differences during decoding. First, because tokens are processed sequentially, we have to do a full weight read for each input. Second, we must also read in the KV cache for each token. Third, our computations are now only for a single token, rather than a whole set of tokens in parallel. Each of these factors shifts decoding towards being bandwidth-bound. The fourth factor (KV projection absorption, see below) pushes in the compute-bound direction, but not enough to outweigh the other factors.

Our calculations here will be familiar from prefill. flops_per_token_weights is nearly unchanged from prefill, except that the average context length is now input_length + output_length/2.

After replacing (d_k + d_v) with d_{kv_latent}, we find t_compute = 0.3B milliseconds for each decoding step.

Bandwidth is conceptually similar to prefill, but we must now account for KV cache reads. For each token, the size of the KV cache that you need to read is given by:

Since the context length grows from 8,000 at the start of our decode phase to 8,999 for our last decode token, the average is about 8,500. Then our KV cache reads work out to an average of 265 MiB * B per decode step. We’ll ignore decode KV cache writes, since these only add a single extra token’s KV cache to data movement, compared to our average 8,500 tokens’ KV caches coming from reads.

The final figure for t_bw is then (500 GB + 265 MiB * B/576 TB/s) * output_len, or 868 + 0.5B ms.

If we plot our expressions for compute and bandwidth time in each of prefill and decode, we can see that prefill is dominated by compute time at all batch sizes, while decoding is dominated by bandwidth time:6

Chunked prefill

So far, we’ve found that for 8,000:1,000 requests, prefill is compute-bound and takes 1B ms per batch, while decoding is bandwidth-bound and takes 868 + 0.5B ms per batch.

As it turns out, we can make use of the fact that compute is sitting idle while we are bandwidth-bound, and bandwidth is idle while we are compute-bound. A common trick to make use of these idle resources is known as chunked prefill. The basic premise is that while you are bandwidth-bound during decoding, you can use your spare compute to start work on the next batch’s prefill computations.

The effect of this overlap is that total time to complete a batch ends up being the larger of total compute time and total bandwidth time, across both prefill and decode.7 The intuition here is that if compute and bandwidth can run concurrently, total time is limited by whichever has more total work to do.

Equivalently,

The final throughput is equal to the batch size times the output length, divided by the time it takes to complete a cycle.

This means the throughput grows monotonically, amortizing the weight loading, until you reach a batch size around ~870 concurrent users per GB200 NVL72. From that point on, inference is compute-bound, and there are no more throughput gains to more batching. In fact, larger batch sizes reduce the speed at which you can serve tokens to each user (the ‘interactivity’). This is because the time to complete a full batch increases in proportion to the number of users, but each user gets only their fixed 1000 tokens of output during that time.

We can also look at the effect of context length. Like many attention mechanisms, MLA’s compute costs grow quadratically with context length, so longer input sequences tend to increase compute costs faster than bandwidth costs. This means that the crossover batch size where throughput becomes compute-bound shrinks as context length increases. For instance, at a context length of 25,000 tokens, the critical B shrinks from 869 to 130.

Speculative decoding

Let’s look at one more trick that gives inference extra juice. Speculative decoding is a technique that uses a small “draft” model to propose candidate tokens several steps ahead,8 which the main model then verifies in a single forward pass. If the main model’s parallel predictions match the draft model’s autoregressive ones, the tokens are accepted.

It’s likely that most major API providers use speculative decoding, since it results in faster inference at minimal cost. This is because the draft model is small enough that its bandwidth and compute overheads are negligible, while each accepted token means one fewer forward pass required by your main model. The one additional cost is that each forward pass on the main model must now predict multiple tokens ahead in parallel, increasing the arithmetic intensity of decoding. For that reason, speculative decoding only helps throughput when decoding is bandwidth-limited. Using state-of-the-art implementations, decoding throughput at a fixed batch size rises by a factor of 3. Since speculative decoding doesn’t help with prefill, the total effect on throughput is closer to 1.6–2×.

Beyond allowing for more throughput at a fixed batch size, speculative decoding can also affect the optimal batch size. By increasing the arithmetic intensity of decoding, speculative decoding reduces the batch size at which decode starts to become compute-bound (after which point there is no reason to further increase batch size, as mentioned in the previous section). For this reason, speculative decoding tends to decrease the optimal batch size.

Putting everything above together, we estimate that the throughput of a B200 NVL72 serving Kimi K2.6 with an 8,000:1,000 profile is around 610,000 tokens per second (tok/s), and aggregate throughput across all chips is 36 billion tokens per second.

Calibrating against inference benchmarks

We’ve built up a fairly rich theoretical model, but we’ve made a few simplifying assumptions, ignoring things like communication latencies, inter-chip bandwidth, and software inefficiencies. To account for these simplifications, we introduce three free parameters:

• Compute efficiency: even during large matrix multiplications, it is rare to achieve 100% of a GPU’s stated maximum FLOP/s. This only affects compute-bound stages.

• Bandwidth efficiency: similarly, bandwidth rarely operates at peak specifications. Because we don’t explicitly model inter-GPU communication, this parameter will also capture time spent moving activations between GPUs.

• Per-step latency: captures a bunch of things like communication latency, kernel scheduling, routing imbalances, and more. This has a larger effect at small throughputs, since we model it as a fixed overhead regardless of batch size.9

SemiAnalysis’s InferenceX dashboard provides data across thousands of inference experiments, which we can use to calibrate these parameters. InferenceX data documents parameters like the total number and type of GPUs used, parallelism strategies, batch size, input and output lengths, model precisions, and more. For any given experiment, we can predict time to first token (TTFT) and time per output token (TPOT) using our theoretical model, and compare to real-world performance. We calibrate our inference model using 111 runs of Kimi K2.5 on Nvidia GPUs across a variety of settings.10

Because the three parameters we’ve introduced have different effects depending on batch size and on what is bottlenecked, we can exploit the variation in experiments to fit these as free parameters, minimizing the discrepancy between our model and the real-world data.11 Doing so, we find estimates of 65% compute efficiency, 30% bandwidth efficiency, and 5ms per-token latency. These are broadly plausible — bandwidth efficiency is lower than expected, but also includes the unmodeled effects of inter-GPU communications.12

Accounting for these inefficiencies reduces our per-GB200 NVL72 throughput from 640,000 tok/s to 400,000 tok/s.13 Applying the same factors to GB300 systems and then scaling up by the total number of systems produces the following estimates:

The present and future of inference

Our final estimate suggests the world’s Blackwell GPUs could currently deliver a combined 500 million to 20 billion output tokens per second, or between 150,000 and 7 million tokens per month for each person on earth. To put that in perspective, Google, which is likely the most avid token producer today through Google Search summaries, recently claimed that it was serving 1.2 billion tok/s across all its platforms (very likely including both input and output tokens). If we assume a ratio of 8,000:1,000 input to output tokens (many of these requests are presumably short-input search results), they would be serving around 130 million output tokens per second. This suggests that even if we lavishly insist upon serving every Google request with an expensive, trillion-parameter model, there is plenty of inference capacity to serve all the needs of their users.

In fact, it would not only be enough to serve Google, but all tokens worldwide. Exponential View estimates the total tokens processed across all providers at 40 quadrillion tokens per quarter, i.e., around 5 billion tokens per second, four times Google’s traffic. There would still be enough compute capacity to serve all these tokens with a Kimi K2.6-like model, at least in our short- and medium-context settings.

And through a combination of infrastructure deployment and more efficient chips, inference capacity is growing over time. The two relevant trends to track are growth in compute capacity and in memory bandwidth, which are growing exponentially at 3.4×/year and 4.1×/year, respectively. In the long term, the slower-growing factor will determine the overall growth of inference capacity — in this case, compute. In the short term, growth can be closer to 4.1×/year while memory bandwidth remains the primary bottleneck; this will particularly affect long context requests like those in software engineering, since those are more deeply memory-bound. In practice, when we model the growth for short context (8,000:1,000 input-output) and long context (128,000:1,000) inference loads, we find the difference to be minimal; the inference capacity of the world at fixed model size and context length is well modeled as growing at 3.4×/year.

How does this compare to the growth of demand for tokens, at a fixed model size and price? Unfortunately, it’s difficult to make a crisp comparison, but the proxies that we have suggest that demand is growing much faster. For instance, both the quantity of tokens processed by Google in the last year, and by all providers according to Exponential View, have been growing by around 10×/year.

From another angle, we can look at token demand from today’s most intensive AI users: software engineers. Recent reports claim that some of Apple’s software engineers are permitted to use up to $300 in tokens per day, which works out to about 5 million output tokens per day with Claude Opus 4.7 API pricing, or 25 million output tokens per day with Kimi K2.6.14 Another point of comparison comes from Meta, whose 85,000 employees used 60 trillion tokens in one month across the organization. That figure included both input and output tokens; assuming a 25,000:1,000 input-to-output token ratio, that would be around 1 million output tokens per day and employee.

There were about 30 million software engineers worldwide as of 2025 (estimates range from 20 million to 50 million), and Stack Overflow’s 2025 survey on AI usage suggested that only around 47% of developers used AI on a daily basis, as of mid-2025. If all SWEs using AI daily were using it as intensely as Meta or Apple, they would demand somewhere between 10 and 350 trillion tokens per day in aggregate, i.e., between 200 million and 4 billion tokens per second. At the longest context sizes of 128,000:1,000, today’s Blackwell chips would struggle to serve all this potential demand for coding agents using models as large as Kimi K2.6. It also seems likely that both the number of developers using AI, and the intensity of their use will continue to grow rapidly.

All of the proxies above are imperfect. We don’t know the composition of those tokens, whether growth is dominated by small or large models, or the increase in context lengths. The prices of many providers have changed significantly over the period, affecting demand.15 And this analysis entirely ignores the very rapid pace of improvement in inference efficiency, which increases the demand for models of fixed scale by improving capabilities, and decreases it by displacing them with smaller models.

Regardless, these proxies suggest very fast growth in demand at fixed model size and price, likely faster than supply is expanding. And this is compounded by trends towards longer-context usage, especially due to coding and other agentic use cases.

If the demand for AI is outpacing the capacity to serve large models, the predictable consequence is that the price of tokens from large models will rise. This suggests a “compute crunch” is near, if not already here. Indeed, Anthropic has taken measures like reducing quotas during peak hours and incentivizing off-time usage in efforts to manage demand.

In a “crunch”, will everyday users be priced out of AI? Not necessarily. We’ve seen very fast growth in inference efficiency. If that continues, current use cases could be served by smaller, more efficiently served models — arguably, this is what has enabled the fast growth in tokens served we’ve seen so far. The largest models could be reserved for the most productive applications, such as coding, where access to the latest capabilities justifies a high per-token price.

We thank David Schneider-Joseph for in-depth feedback. We also thank Dwarkesh Patel, Jean-Stanislas Denain, Josh You, David Owen, Phil Trammel, Nick Merrill, Vassil Tashev and William Gildea.

• Our newest Data Insight finds memory is now the largest and fastest-growing component cost for leading AI chip designers, rising from 52% to 63% of total component spending since early 2024.

• In the latest Gradient Update, Josh You argues that top frontier labs could dramatically increase their share of global AI compute usage in the next few years — after which, continued scaling would require an economic transformation.

• FrontierMath: Open Problems workshops begin May 26. Applications are still open.

Subscribe now

Data Insights: Memory has grown to nearly two-thirds of AI chip component costs.

• Researcher Venkat Somala finds that high-bandwidth memory (HBM) has grown from 52% to 63% of total AI chip component spending between Q1 2024 and Q4 2025, faster than any other component. HBM spend across chips designed by Nvidia, AMD, Google, and Amazon rose from roughly $12 billion in 2024 to $32 billion in 2025.

Commentary: Frontier labs don’t use most AI compute (yet)

• In our latest Gradient Update, researcher Josh You estimates that while leading labs today use less than half of global AI compute, they could absorb most of the available headroom within a few years. At that point, continued growth would be capped by chip supply. For scaling to continue, the overall compute buildout would need to accelerate. Given that AI capital expenditure is already approaching $1 trillion per year, such an acceleration in compute production would require dramatic economic changes. Gradient Updates are informal, opinionated analyses that represent the views of individual authors, not Epoch AI as a whole.

  

Other Updates

FM:OP workshops

Our in-person FrontierMath: Open Problems workshops kick off Tuesday in New York City, with subsequent events in London, Berkeley, Boston, Los Angeles, and Toronto through June 9. The goal of the workshops is to identify highly interesting and programmatically verifiable unsolved research math problems. All working mathematicians (grad students, postdocs, and professors) are encouraged to apply while spots remain.

Careers

We’re hiring across several roles. All positions are fully remote.

• Designer to translate complex research into intuitive, engaging, and high-impact designs — primarily UI/UX and data visualization work.

• Researchers and Senior Researchers to lead new projects across our expanding teams.

• Data Scientist (Contract) to assist with our AI research efforts, including reviewing technical literature, tracking benchmark data, and analyzing AI models, data centers, and companies.

Applications are rolling, so apply soon!

Subscribe now

In case you missed it…

Research:

• Servers account for 60% of the total cost of owning a one-gigawatt AI data center (May 14)

• Claude overperforms at software engineering and underperforms at math (May 15)

Commentary:

• Why a handful of AI researchers command 10–100× the pay of their peers — and why that gap could grow (May 13)

OpenAI kicked off the AI boom when it launched ChatGPT in 2022. Frontier LLMs soon accrued hundreds of millions of users and billions in revenue, sparking a massive investment boom in AI compute infrastructure, with Nvidia’s AI-related sales spiking more than fourfold in 2023. Global AI computing power has now grown to the equivalent of around 20 million Nvidia H100s, funded by hundreds of billions of dollars in annual capital expenditures.

Yet while OpenAI launched the compute boom, they don’t dominate AI compute usage. I estimate that the compute OpenAI uses for research, training, and inference as of the end of 2025 made up around 10% to 15% of the world’s operational AI compute supply, and this share was even smaller a year ago. Even after adding the other most well-resourced frontier developers — Anthropic, xAI, and the AI labs within Google and Meta — the combined total is probably still under half of the world total.

In other words, there is a lot of AI compute that top frontier labs are not using. Anthropic and OpenAI have seen rapid growth in revenue and funding, enabling them to grow their AI compute faster than the world overall, and this will continue in 2026.

But the top labs may capture a much larger share of global compute within a few years. At that point, compute growth at top labs would be more directly tied to the pace of total compute production, which could slow down the rapid growth we’ve seen in both model capabilities and AI deployment/revenue. For scaling to continue, the overall compute buildout would need to accelerate. Given that AI capital expenditure (capex) is already approaching $1 trillion per year, such an acceleration in compute production would require dramatic economic changes.

Most AI compute probably doesn’t go to frontier AI

More details for each company can be found in the Appendix, and the accompanying research document.

I don’t have a great estimate of the compute used by each of the five most resource-rich frontier developers, but we know enough to estimate their share of world compute.1

OpenAI helpfully disclosed the total electric power capacity of its data centers, which can be converted to ~1.7 million in H100-equivalent (H100e) compute. We also know a lot about xAI’s Colossus data centers. I’m less certain about Anthropic, which had significantly less compute than OpenAI at the end of 2025, though probably still over 1 million H100e. The situation at Google DeepMind and Meta Superintelligence Labs is also unclear, since the compute owned by their parent companies (roughly one-third of the world total) is split across frontier AI, cloud, and other internal uses.2 It’s not clear that the frontier labs at Google and Meta use even half of the total. For more details on each lab, see the Appendix.

But it’s still clear that a lot of AI compute isn’t used by the top labs. My best guess, in terms of the equivalent number of Nvidia H100 GPUs, are that OpenAI, Anthropic, and xAI together probably had fewer than 4 million H100e at the end of 2025.

My best guess is that DeepMind uses slightly under half of Google’s total. Meta also rents external cloud compute (not shown on the graph), starting in late 2025. Estimated world total of 16 million H100e assumes a one-quarter lag between chip sales (est. 20 million) and operations. This is a reference scenario; a longer lag would imply a higher frontier compute share.

Meanwhile, cumulative sold AI compute was roughly 20 million H100e as of the end of 2025. But not all of this was necessarily operational—I don’t know exactly how much, but a rough estimate would look at chip sales at a time lag based on typical installation periods for AI clouds like CoreWeave.3 If there’s a one-quarter lag between delivery and deployment, deployed compute at the end of 2025 would be comparable to sold compute as of Q3 2025, which was ~16 million H100e. If the delay is two quarters, deployed compute goes down to ~12 million H100e.

Under these varying deployment assumptions, Anthropic, OpenAI, and xAI’s total H100e would make up around 20% to 30% of the world total at the end of 2025. If you also count the inference compute that the hyperscalers use to run their own APIs on OpenAI and Anthropic models, this may contribute up to another ~5%.

Meanwhile, we estimate that Google and Meta together own around one-third of the world’s total AI compute. But the compute allocated to Google DeepMind and Meta Superintelligence Labs is substantially less than that, given the large compute demands of Google’s external cloud business and non-frontier uses such as recommender systems. Each lab may use roughly half of their parent companies’ compute as a first-pass guess, for a total of roughly 15% of world compute.

This means that the five most resource-rich AI developers in the world probably had access to less than half of global AI compute at the end of last year.

In other words, frontier AI labs like OpenAI may have kicked off the AI compute buildout, but they are not wholly responsible for it. I won’t attempt a full breakdown of the remainder, but likely candidates include second- and third-tier LLM players and inference of open-weight LLMs. AI/ML models in non-language domains also consume compute: the innovations behind frontier LLMs, such as the transformer architecture, have enabled much better models in audio/visual generation, biology, robotics, and recommender systems, among others.

Will Anthropic and OpenAI absorb the rest of global AI compute?

While much of the world’s AI compute isn’t used by the top labs today, this situation could change significantly in the next few years. In particular, I think Anthropic and OpenAI are the key players to watch.

OpenAI and Anthropic grew compute faster than the industry as a whole, at least in 2025.4 OpenAI tripled its data center power capacity in both 2024 and 2025; after accounting for improved hardware efficiency, their computing power grew around 4× annually.5 Anthropic is probably growing even faster, since they’ve been catching up with OpenAI in revenue and funding. Meanwhile, we estimate that the global stock of AI compute tripled in H100e terms in 2025, and new installed compute grew by 2.7× in 2025 versus 2024.

To be sure, this isn’t enough evidence to draw stable trendlines for frontier compute growth and overall AI compute growth.6 But it looks like OpenAI and Anthropic are currently growing their compute faster than the industry as a whole.

This looks likely to continue. OpenAI internally forecasts that its data center capacity will reach “low double-digit” GW in 2027, up from 1.9 GW in 2025. If that means (say) 12 GW by the end of 2027, that would be 2.5× annual growth, a slowdown from 2023–2025’s 3× growth, but still very fast. OpenAI’s president, Greg Brockman, also testified that the company would spend $50 billion on compute in 2026, triple what it spent in 2025. Finally, some third parties forecast that Anthropic and OpenAI will each have 5–6 GW of capacity by the end of this year, or ~2.5–3× growth in power. Because AI chips improve rapidly in price and energy efficiency, this suggests another year of ~4× growth in H100e compute capacity.

The industry as a whole probably can’t match this growth: hyperscalers are growing their capex at a relatively steady 70% per year and guiding similar growth in 2026, suggesting that global compute growth will be similar to 2025’s ~3× growth.7

Industry trends also point to the top labs, especially Anthropic and OpenAI, consolidating the market. Demand for frontier LLMs has grown explosively this year, particularly for coding and agentic tasks. Anthropic has grown at a truly astonishing rate, increasing its annualized revenue run rate from $9 billion to $30 billion in the first quarter of 2026! This is an acceleration from last year’s already-extreme 10x growth rate.8 Recent revenue data is less available for OpenAI, but qualitatively, its coding models and products (e.g., GPT-5.5) have also been well-received. Rapid revenue growth, and the corresponding increase in funding, gives Anthropic and OpenAI the means to secure a larger share of global compute.

Indeed, Anthropic and OpenAI are moving aggressively to secure more compute. In April alone, Anthropic signed multi-gigawatt expansions with Amazon (targeting 1 GW of Trainium online in 2026) and Google, and added CoreWeave as a compute partner.9 And in a fascinating plot twist, Anthropic has agreed to rent xAI’s entire ~300,000 H100e Colossus 1 data center along with part of Colossus 2, paying up to $15 billion per year for the privilege.10 OpenAI has also expanded its cloud roster, signing with Amazon in February and Google last year. So these two labs are the most likely culprits behind the tight compute supply and ~30% increase in GPU-hour prices that the industry has seen this year.

Anthropic and OpenAI are not the only players driving LLM growth. Demand for open-weight models like DeepSeek, which are not too far behind in quality, may also be surging. DeepMind looks behind the top two in agentic coding as of writing, but is definitely not out of the race, and Meta is spending big to try to catch up in frontier AI. But if the agent boom leads to LLMs growing their share of the overall AI industry, this will boost the compute share of the two LLM leaders. And Anthropic’s current revenue trajectory is so extreme that it seems likely to lead to compute consolidation.

In 2023, it was not obvious that OpenAI or frontier LLMs in general would end up dominating the entire AI industry. Three years later, it now appears that the players who kicked off the AI compute buildout will end up leading it.

What happens if frontier labs run out of headroom?

Anthropic and OpenAI probably made up 15–20% of the world’s operational AI compute at the end of 2025. While the headroom for them to grow their share looks big, it can be consumed in just a few years if the frontier labs grow substantially faster than the industry as a whole.11

As a naïve illustration, suppose the top two players together have a 20% share today, and grow their AI compute 33% faster than the world as a whole (e.g., they quadruple their computing power every year, as OpenAI did through 2025, while the global installed base “merely” triples annually). In this scenario, they’ll double their share of world compute in 2.5 years, and use ~80% within five. And if compute is more evenly distributed among the top four or five frontier developers, rather than just Anthropic and OpenAI, the headroom will run out faster than that.

So a key question for the near future of AI is whether Anthropic and OpenAI can continue their recent pace of compute growth, dragging up overall AI compute production along the way, or whether their compute growth slows down because the cloud and semiconductor industry can’t keep up.

At this point, I want to emphasize that total capital expenditures on AI chips and data centers are already very large. Total AI capex will approach $1 trillion annualized in 2026, which would be almost 1% of the gross world product and 3% of US GDP, and capex now consumes most of the operating profits of the hyperscalers that are leading the buildout. There is no guarantee that rapid AI capex growth will continue after 2026.

If the top model developers eat the compute headroom and their compute growth converges with overall AI compute growth, maintaining 4× growth per year would require more than doubling capex annually even after factoring in chip price-performance improvements. From a starting point of perhaps $1 trillion in AI capex in 2027, this sort of growth would only be feasible if AI starts to dramatically accelerate economic growth.

In other words, the AI industry will transition to a new regime in the next few years, with frontier AI slowing its compute growth, or dominating the industry, or both. To be clear, there’s no reason to expect a progress “wall” in 2029 if a frontier compute slowdown happens: flat compute capex can still grow the compute stock for years, AI chips will still improve, and companies can research and train new models with a fixed amount of compute. But the key physical trend driving frontier AI progress, the scaling of compute, is not sustainable unless the world fundamentally changes soon.

Many thanks to Amelia Michael, Ben Cottier, Brendan Halstead, Campbell Hutcheson, Elliot Stewart, Isabel Juniewicz, Konstantin Pilz, Romeo Dean, and Yafah Edelman for helpful feedback

Appendix: How much compute goes to frontier AI developers?

For more information, see this much longer research preview that sorts through the relevant evidence per company, along with modeling details.

Here, I summarize my estimates of how much AI compute OpenAI, Anthropic, xAI, and the frontier labs at Google and Meta had access to at the end of 2025. These five are probably the most compute-rich developers in the world, though not necessarily the leaders in model quality. I focus on how much compute frontier AI companies rent or use, not how much they own.12 Anthropic and OpenAI predominantly rent their compute from cloud partners like Amazon, Google, and Microsoft; Google and Meta mostly own their AI compute, but much of this is allocated to non-frontier internal uses or, in Google’s case, rented out to external customers.

OpenAI

We know the most about OpenAI’s compute because they helpfully disclosed the total electrical power capacity of the data centers they rent at the end of 2023, 2024, and 2025. OpenAI ended 2025 with 1.9 gigawatts (GW) of capacity, up from 0.6 GW in 2024 and 0.2 GW in 2023.13

Courtesy of OpenAI

This power capacity can be converted to computing power using the specs of flagship Nvidia AI GPUs and servers, and some assumptions about the mix of GPUs OpenAI used over time. With this method, I estimate that OpenAI had the equivalent of around 1.7 million Nvidia H100 GPUs (H100e) by the end of 2025, up from around 400,000 in 2024 and 100,000 in 2023 (H100e is based on peak FLOP-per-second specs; actual compute performance may vary). Another approach would be to use media reports that OpenAI spent $16 billion on cloud compute in 2025 and combine this with plausible prices per GPU-hour that OpenAI may have paid. This leads to a broadly similar compute estimate.

One potentially significant omission is the inference compute used to run OpenAI models on hyperscaler-hosted products like Microsoft’s Azure API (and now Amazon Bedrock); this compute is presumably excluded from OpenAI’s data center and cloud compute figures. It’s debatable whether this should count as “OpenAI compute”: OpenAI does share in some of the revenue from this compute, but doesn’t have operational control over it. Including this Microsoft inference compute may boost the total OpenAI compute by around 25%, with 50% as an upper bound, judging from OpenAI’s revenue and inference compute allocation (more details here).

Anthropic

Anthropic probably had substantially less compute than OpenAI in 2025, but is catching up over time. An internal OpenAI memo estimated that Anthropic had 1.4 GW in capacity at the end of 2025, around 70% of OpenAI’s 1.9 GW total. In both 2024 and 2025, Anthropic’s cloud compute bill was reportedly just over 40% of OpenAI’s. The 2025 ratio is somewhat surprisingly low, but Anthropic’s compute spending may have ramped up towards the end of the year.14

Much of Anthropic’s compute is housed in Amazon’s Project Rainier campus in Indiana, with around 500,000 H100-equivalents (H100e) in Trainium2 chips at the end of 2025. Anthropic also rents significant numbers of Trainium, Nvidia, and TPU chips from both Amazon and Google elsewhere, so its total must be much larger than this one campus. Alongside the OpenAI memo, I think this points to 1 million or more H100e for Anthropic by the end of 2025. As with OpenAI, including inference compute from third-party cloud APIs would boost Anthropic’s compute total, again by roughly 25%.

xAI

xAI, now part of SpaceX, mostly uses the compute in its Colossus 1 and 2 data centers in Memphis, Tennessee. These two facilities added up to around 550,000 H100e at the end of 2025. xAI reportedly also owns or uses smaller data centers in Portland and Georgia, and they used Oracle as a cloud compute partner at least through 2024. xAI’s total compute usage may have been around 600,000 to 700,000 H100-equivalents at the end of last year, likely less than Anthropic and less than half of OpenAI’s estimated 1.7 million H100e.

This year, xAI has decided to rent Colossus 1 to Anthropic, but is also targeting a major expansion of Colossus 2 to around 1.4 million H100e.15

For Google and Meta16, I’ll define their “frontier AI” compute as the compute used by their frontier AI divisions — Google DeepMind (henceforth DeepMind) and Meta Superintelligence Labs (MSL) — as well as related inference.17 Clean comparisons with Anthropic or OpenAI are difficult because frontier AI work may bleed into other AI/ML efforts, such as recommender systems.

Google DeepMind

Google (Alphabet) is the most compute-rich firm in the world; in our research on AI chip owners, we estimate Google owned around a quarter of the world total, or roughly five million H100e at the end of 2025, or around 4 million H100e using a one-quarter delay between chip sales and deployment. But much of this does not go to DeepMind: Google says about half of its ML compute goes to Google Cloud.18 For Google, “Cloud” includes enterprise Gemini inference (Vertex API and enterprise subscriptions) in addition to compute for external customers. The other half is split between DeepMind and non-frontier internal uses like recommender systems.

Whether DeepMind compute exceeds half of the Google total depends on whether Cloud-side DeepMind inference compute is greater than non-DeepMind internal compute. My guess is that the latter is bigger, which would mean DeepMind compute is less than half of the total.

This means that despite Google’s large compute lead as a company, it is not clear to me whether DeepMind had more compute than OpenAI at the end of 2025.

Meta

Meta owned roughly 10% of the world’s total AI compute at the end of 2025, more than OpenAI rented in total.19 But translating this to Meta Superintelligence Labs (MSL) compute is tricky.

First, many Meta GPUs support the company’s core business, principally recommender systems for ads and content, rather than frontier AI. The algorithms running your Instagram feed are actually large-scale transformers that are very effective at boosting engagement.20 Third-party estimates put the split between frontier AI and recommenders at roughly 50-50 in mid-2025.

However, Meta pivoted hard to prioritizing frontier AI in late 2025. Mark Zuckerberg hand-delivered soup to top researchers and offered them enormous compensation packages, promising that MSL overall would be equipped with “industry-leading levels of compute”. So it’s plausible that Meta compute has now tilted significantly towards frontier AI.

Second, Meta signed large cloud deals with Google, Oracle, andCoreWeave starting in late 2025; as these deals ramp, Meta will use much more AI compute than it owns. Still, because these cloud deals are relatively new, I think MSL probably had access to significantly less compute than Meta owned in total at the end of 2025. This means MSL probably had less compute than OpenAI, since we estimated Meta only owned 2.3 million H100e in total before accounting for deployment lags, split between MSL and other uses.

• We published two new Data Insights: Servers account for 60% of the total cost of owning a one-gigawatt AI data center, and Claude overperforms at software engineering and underperforms at math.

• The latest Gradient Update explores why a handful of AI researchers command 10–100× the pay of their peers — and why that gap could grow.

Subscribe now

Data Insights

We published two new Data Insights, our digestible snapshots of complex AI trends.

• Servers account for 60% of the total cost of ownership of a one-gigawatt AI data center. GovAI research scholar Amelia Michael and Epoch senior researcher Ben Cottier model the full costs of a typical one-gigawatt US AI data center, finding that servers alone account for $5 billion of the $8.5 billion annual total — dwarfing energy and all other operating costs.

  

• Claude overperforms at software engineering and underperforms at math. Senior researcher Alexander Barry finds that, relative to their general Epoch Capabilities Index (ECI) values, Anthropic’s Claude models overperform on software engineering benchmarks and underperform on math. These results come from the Domain-specific ECI Explorer we launched earlier this month, where you can view Math and SWE ECIs as well as design your own variant.

Commentary: The economics of superstar AI researchers

• In our latest Gradient Update, researcher Anson Ho argues that enormous pay disparities among AI researchers — where superstar researchers can earn 10–100× what their colleagues make — come down to more than just quality differences. Rather, we should expect to see big differences in pay even if superstars are only a tiny bit better than an average postdoc. Gradient Updates are informal, opinionated analyses that represent the views of individual authors, not Epoch AI as a whole.

Other Updates

FM:OP workshops

Our in-person FrontierMath: Open Problems workshops kick off in under two weeks. Events will be held in six cities — New York, London, Berkeley, Boston, Los Angeles, and Toronto — between May 26 and June 9. All working mathematicians (grad students, postdocs, professors) are encouraged to apply while space remains.

Model Evaluations

We are conducting an AI-assisted review of FrontierMath: Tiers 1–4 and have flagged fatal errors in about a third of problems — most of which we believe to be valid. We will release updated scores on a corrected dataset after completing a thorough human review.

Careers

We’re hiring across several roles. All positions are fully remote.

• Designer to translate complex research into intuitive, engaging, and high-impact designs — primarily UI/UX and data visualization work.

• Researchers and Senior Researchers to lead new projects across our expanding teams.

• Data Scientist (Contract) to assist with our AI research efforts, including reviewing technical literature, tracking benchmark data, and analyzing AI models, data centers, and companies.

Applications are rolling, so apply soon!

Subscribe now

AI is one of those fields where the best winds up much better off than the rest. Superstar researchers at frontier labs earn over ten times more than most of their colleagues, who earn measly million-dollar salaries. They might even earn over a hundred times more than your average AI postdoc:

So why are the differences in pay so large? The naive explanation is that some researchers are just vastly superior. Perhaps the superstar researchers have excellent research taste in designing algorithms and experiments. Or they have a knack for pulling off “yolo runs” — training runs that implement many ambitious changes all at once, relying on deep intuition, whereas most people would need to systematically test the individual changes to make sure they work. Under this framing, superstars are the “10× researchers” that Silicon Valley so deeply reveres, and it’s their quality that makes the difference in pay.1

The problem with this explanation is that it’s very incomplete. In reality, we should expect to see big differences in pay even if superstars were only a tiny bit better than your average postdoc. But why?

The superstar effect

The short answer is this: there’s a well-known economic dynamic which turns small differences in ability into big differences in pay. Here are two illustrative examples:

• In the 100-meter sprint, the gold-medallist gets much more reward and attention than the silver-medallist, despite them being quite literally neck-and-neck for most of the race. Consider the London 2012 Olympics, where Usain Bolt won gold. Most people have no idea who won silver, despite finishing just 0.12 seconds behind — do you?

• Some musicians earn much more than others. Consider Taylor Swift: last year, she earned $60-70 million from Spotify. I don’t doubt that she’s a “10× singer” compared to me. But it’s very debatable whether she’s that much better than other extremely popular singers like Ed Sheeran, Blackpink, Charli XCX, and Lana Del Rey, who instead earned closer to $5-25 million.

Across these two cases, small differences in ability led to big differences in pay some way or another. Economist Sherwin Rosen called this the “superstar effect,” and it kicks in when two conditions hold.

1. One person’s work can reach a big market. Usually this means a market with many people, but a few high-paying people or firms work too. For instance, potentially billions of people watched Usain Bolt win the 100-meter sprint. The more people you can reach, the more pronounced the superstar effects. Across the economy, jobs with broad reach — such as actors, musicians — show far bigger wage dispersion than jobs serving one client at a time, such as plumbers, nurses, and truck drivers:2

2. Quantity doesn’t easily make up for quality of labor. You can’t have multiple people take the place of a single sprinter, since that would break the rules of the race. And if you like Taylor Swift more than Ed Sheeran, it’s hard to make up for missing a Taylor Swift concert by going to more Ed Sheeran ones.

The first condition means a tiny quality edge captures enormous extra value, making it worth paying a lot for the best — that is, as long as you can’t make up for quality with quantity (the second condition). If you could, you’d just hire a lot more people with lower pay — you wouldn’t need to pay a ton just to hire the cream of the crop.3

Why this applies to AI

AI researchers tick both boxes. There’s a huge market: ChatGPT has almost a billion users, served by the same handful of underlying models, so a single researcher’s contribution could scale to every user simultaneously.

And in AI, researcher quantity doesn’t easily make up for quality: frontier labs are compute-constrained, so they can only run so many experiments to test new software innovations. Two “merely very good” researchers can’t replicate one Noam Brown if what’s needed is deep intuition about which experiments are worth running in the first place. Not to mention the difficulties coordinating researchers if labs are short on time.4

This is how even a 2× researcher could earn far more than the median. Scaled to a billion users, even a small quality edge generates enormous differential value. And if the 2× researcher can add something that multiple 1× researchers can’t, then it’s worth paying a lot to capture this.

Race dynamics amplify the effect

Frontier AI labs are often described as being in a “race”. I’m not sure what exactly they’re racing toward, but it often seems to involve automating huge swathes of human labor, a prize potentially worth tens of trillions of dollars a year — if you win. This incentivizes AI labs to adopt an “all in or nothing” approach, and anything that improves their chances even a little might be worth a lot. Hence Meta’s (alleged) $100 million dollar compensation packages to poach top researchers from OpenAI.

In principle it’s even possible this pushes things well beyond what is socially valuable (however you define that) — it’s like how high frequency traders spend huge sums trying to execute a trade a tiny bit faster, to almost no social benefit.

Reality is complicated, and so is managing an army of AIs

Other forces are at work too. Top researchers carry valuable trade secrets in their heads — the results of expensive experiments competitors haven’t run, and which would cost a fortune to replicate. Many also manage teams, contributing more value than just their raw technical research ability; Noam Brown recently described himself as a “manager at OpenAI.”5 Each of these may contribute to the wage gap, separate from the superstar premium.

Additionally, it’s hard to quantitatively analyze the superstar effect.6 I don’t know of a good way to quantify “researcher quality”. There are some valiant efforts like METR’s RE-bench, but these contain small isolated tasks (think “finetuning GPT-2”) rather than projects with millions of lines of code, fuzzy objective metrics, and lots of coordination between different people.

But despite these complications, I think the superstar effect tells us several useful things. For one, I’ve seen a couple of news articles about Meta’s attempts to poach researchers with exorbitant salaries, in their quest for Personal Superintelligence. But these articles usually miss out on this important superstar effect (though they often do touch on race dynamics).

Another important implication is for how we think about the intelligence explosion. If a 100× pay gap is driven by a 100× researcher quality gap, then simulating a top researcher might speed things up much more than simulating an average researcher.7 But this isn’t the case if much of the pay gap is driven by the superstar dynamic — the gap in researcher quality might actually be much smaller.

Finally, knowing about this effect gives us some hints at what’s to come in the near future. I think that the superstar effect will only become more important moving forward. That’s because lots more people will use AI, and each person will use AI systems much more heavily. And as research increasingly shifts toward managing an army of Claudes, those with deep research intuitions and years of experience as research managers will probably see ever-growing boosts to their productivity, as well as the sizes of their wallets.

So if anything, superstar earnings might become an even bigger deal — $100 million annual compensation quite literally might not be enough.

I’d like to thank Andrei Potlogea, Phil Trammell, Josh You, David Owen, JS Denain, Cheryl Wu, Stefania Guerra, Robert Sandler, Lynette Bye, and many people at Trajectory Labs for their feedback and support. Thanks also to Luis Garicano for inspiring me to write this essay in the first place.