OpenAI

OpenAI relies on cloud compute capacity rented from other companies; as of 2026, these were principally Microsoft, Oracle, and CoreWeave.

In 2025, OpenAI disclosed its total compute capacity over time in terms of electrical power capacity: at the end of 2023, it used 0.2 GW of compute, which grew to 0.6 GW at the end of 2024 and an estimated 1.9 GW at the end of 2025.³ We believe these gigawatt figures are most likely referring to the IT power (that is, peak power draw of IT equipment) of the data centers OpenAI uses.

We are not sure whether these power figures include the compute that Microsoft uses to run its own products running OpenAI models (e.g. API inference via Azure), though we think they probably do not. As a high estimate, the compute used to run Microsoft-hosted inference of OpenAI models may be up to roughly 25% of the size of OpenAI’s first-party compute fleet; more details can be found in this research appendix.

It is also debatable whether to count this Microsoft third-party compute as “OpenAI compute”. OpenAI allocates part of its own compute to inference in order to make revenue and gross profits, and third-party compute that generates revenue for OpenAI also fulfills this purpose. On the other hand, OpenAI likely makes much less gross profit from this hyperscaler inference than it does from its own first-party inference, e.g. historically OpenAI shared in only 20% of the revenue from Microsoft-hosted OpenAI products, likely significantly lower than its own gross margin. After a major renegotiation in April 2026, OpenAI no longer receives any revenue share from Microsoft-hosted OpenAI inference. We may track this as a separate compute usage category in the future, depending on data availability.

Anthropic

Like OpenAI, Anthropic uses AI compute capacity from external cloud partners, mainly Amazon and Google as of 2025 (adding SpaceX/xAI and CoreWeave in 2026).

As with OpenAI, there is a definitional ambiguity with “Anthropic compute”, because a significant share of Anthropic model inference is served by third-party APIs from Amazon, Google, and Microsoft. In terms of accounting, Anthropic and its cloud partners treat this differently than the comparable Microsoft/OpenAI arrangement. Anthropic pays for the compute used for these products, and books the entire revenue as Anthropic revenue, with the revenue Anthropic shares with the hyperscalers booked as an expense. Since Anthropic pays for all of this compute and books all of the revenue, we don’t attempt to exclude this third-party compute in our model. This simplifies the modeling, since the reporting on Anthropic’s compute spending likely includes this third-party compute.

As with OpenAI, Anthropic-related third-party hyperscaler compute is probably substantial but only a minority of the total. Anthropic reportedly shares around half of its gross margin on third-party inference with its cloud providers; if the overall gross margin is between 40% and 80%, this implies a 20% to 40% share of the total revenue from these products is shared with clouds. According to the same report, these revenue shares to cloud partners totaled $360 million in 2025, or 8% of Anthropic’s total 2025 revenue of $4.5 billion. This suggests that total revenue from hyperscaler Claude inference in 2025 was roughly $1 billion, or around 20-25% of Anthropic-related inference revenue. This means that inference compute for Claude on third-party APIs have made up roughly 10% of Anthropic’s total compute, since the majority of their compute expenses went to research and training in 2025.⁷

Google DeepMind

Google (here shorthand for all of Alphabet) is a major hyperscaler and probably owns more AI compute than any other firm in the world. Google has an AI lab division called Google DeepMind (henceforth “DeepMind”) that houses its frontier AI efforts along with many other AI research projects.

Here, we estimate how much compute is allocated to DeepMind, including inference of DeepMind-produced models, which may or may not line up with Google’s internal compute accounting. The definition here may be fuzzy, since it’s not clear how distinct DeepMind’s work is from e.g. ad recommender systems. Our general uncertainty about Google’s compute allocations is high enough to subsume this definitional uncertainty.

We estimate DeepMind’s compute as follows:

• We estimate Google’s overall compute, starting from estimates of Google TPU capacity and Google’s share of Nvidia GPU purchases (from AI Chip Owners), and applying a discount for deployment lags between sold and operational compute.
• We estimate what share of Google’s compute was DeepMind-related, versus compute allocated to external cloud sales or other internal uses. For 2025, we use Google’s disclosure that half of their ML compute was allocated to Google Cloud to help triangulate this share.

More details can be found in this Python notebook.

Meta Superintelligence Labs

Meta is one of the world’s largest hyperscalers, holding roughly 10% of the world’s total AI compute at the end of 2025 (not including its custom MTIA chips, which were relatively low volume in 2025).

Meta hosts a major AI division called Meta Superintelligence Labs (MSL) that works on frontier AI research and development, alongside basic AI research.¹⁵ Before 2025, MSL was called “Meta AI”, and before that “FAIR”. Here, we focus on estimating how much compute MSL has access to.

However, much of Meta’s compute is not used by MSL or for frontier AI. Meta also uses AI compute to train and deploy recommender systems to enhance its core social media business by recommending content and advertisements. Meta’s recommendation systems are powered by large transformer-based models trained on “thousands” of GPUs. The share of Meta’s compute that is allocated to MSL versus other uses like recommenders is the core source of uncertainty in our model for both 2024 and 2025.

We model MSL’s AI compute usage over time as follows:

• Estimate Meta’s deployed AI compute over time, starting from its compute purchases from AI Chip Owners and applying a discount for deployment lags between sold and operational compute.
• For 2025, additionally estimate how much external cloud capacity Meta rented, which may have been minor or negligible.
• Estimate how Meta’s compute was allocated between MSL and core business in both 2024 and 2025.

Evidence on the MSL/core business split

Meta strongly emphasizes the importance of recommender systems (RecSys) to its business, attributing the strong usage and revenue growth it saw in 2025 to trillions of daily AI-powered recommendations, suggesting that these systems are a highly lucrative multiplier on Meta’s over-$200B/year ads business. This would easily justify a large compute allocation, though this also depends on the returns to scale for training and inference of RecSys; the industry research firm SemiAnalysis believes that the marginal returns of scaling recommenders are attractive, and Meta has written about the gains of large-scale training of transformer recommenders.

Meta said in Q1 of 2025 (before Meta engaged in a hiring push for MSL) that the majority of its 2025 capital expenditures would go towards “core business”, including compute for recommenders, rather than generative AI. In other words, Meta’s capex breaks down into generative AI + AI for core business (e.g. recommenders) + non-AI capex, and Meta forecast that the latter two together would be over 50% of overall capex. This doesn’t give clear bounds on the share of Meta’s AI compute going to generative AI, absent a more precise breakdown of their AI share of capex.¹⁶ But it does mean that frontier AI has not squeezed out all of Meta’s data center investments for recommenders and other uses.

At the same time, LLMs and frontier AI have clearly been a major priority at Meta since at least 2023. In 2024, Meta released a frontier-scale model, Llama 3 405B, which was trained on 16,000 H100 GPUs, one of the largest known training runs at the time with around twice the training FLOP of GPT-4. Meta also announced two 24,000-H100 training clusters and, in late 2024, built a cluster of 100,000 H100 GPUs that it said would be used to train frontier models.¹⁷ This cluster alone was a substantial share, roughly 15%, of Meta’s overall AI compute at the time.¹⁸

In 2025, Meta pivoted hard towards frontier AI, engaging in a massive hiring spree to hire and poach top AI talent, including spending $15 billion to acqui-hire the CEO of the data startup Scale AI. Meta CEO Mark Zuckerberg also announced that he planned to equip Meta Superintelligence Labs with “industry-leading levels of compute and by far the greatest compute per researcher.”¹⁹

Overall, some industry analysts/commentators estimated in mid-2025 (in off-hand remarks on a podcast) that 50-60% of Meta’s GPUs went to recommender systems. Meanwhile, SemiAnalysis’ official models attribute roughly half of Meta’s compute additions in 2025 to MSL.²⁰

For 2025, we model MSL’s share of Meta compute as a lognormal variable with a 90% uncertainty interval of 33% to 80%, with a median of just over 50%. This is unfortunately pretty rough and “vibe-sy”, reflecting the fact that Meta clearly prioritizes both MSL and recommenders, but doesn’t disclose hard numbers on its compute allocations.

For 2024, there is less available commentary from Meta executives or third-party analysts, so uncertainty should arguably be greater. But Meta’s aforementioned disclosure of a 100k H100 cluster they were using for Llama 4 by the end of 2024 is useful because it anchors their frontier AI efforts at over 100k H100e, compared to ~500k to 800k in total. (It’s not clear whether the two previously-disclosed 24k H100 clusters were expanded or merged into this 100k cluster.) So a greater-than-20% share is very likely, but Meta’s pivot towards MSL in late 2025 suggests that Meta AI may have had a lower share of AI compute relative to ads in 2024 than in 2025. So we choose a lognormal 90% uncertainty interval of 25% to 80%, which decreases the median somewhat from 51% to 45%.

2025 cloud compute purchases

In addition to the compute Meta owns, Meta has signed several large deals to purchase cloud compute from Google, Oracle, CoreWeave, and Nebius starting in late 2025. These deals were probably not active, at least at large scale, by the end of 2025, but it’s possible that Meta was buying cloud compute at an annualized rate of single-digit billions by that time. (On the flip side, Meta may begin selling some of its compute capacity on the cloud starting in 2026.)

The publicly reported deals in 2025 were with CoreWeave ($14.2 billion through end of 2031 for a $2.3B/year run rate), Google ($10 billion over six years, for a $1.6B/year run rate), Oracle (reportedly a $20 billion deal with unspecified duration, not officially confirmed), and Nebius ($3 billion over five years, or $0.6 billion/year). These add up to around $7-8B/year if the Oracle deal is also over six years, though these public reports may not be exhaustive.²¹

CoreWeave’s CEO said that the $14.2 billion deal would come online “in 2026”, and its 10-Q form in Q1 2026 suggested that the capacity was still ramping up during the first quarter: its top two customers made up 45% and 20% of its $2 billion in revenue that quarter, while no other customer made up more than 10%. Since these top two are almost certainly Microsoft and OpenAI respectively,²² this implies that Meta was spending at an annualized rate of at most $800 million (10% of an annualized $8 billion run rate) in early 2026, or at most around one-third of the full run rate of the cloud deal.

Extrapolating from the CoreWeave example, we model a fair chance (25%) that none of Meta’s purchased cloud compute capacity was active at the end of 2025; if some capacity was online, Meta’s spend rate was probably less than $3.5 billion, or half of the eventual run rate of these reported deals. The spend rate distribution is converted to H100e using typical cloud contract prices from 2025 (see below). For simplicity, this compute bucket is entirely attributed to MSL.

For more details on the model, see this Python notebook.

SpaceXAI

SpaceXAI, previously known as xAI before its acquisition by SpaceX, is a large frontier AI developer.

SpaceXAI uses AI compute from two sources:

• Most of its compute is owned by SpaceX, primarily in the Colossus 1 and Colossus 2 campuses near Memphis, Tennessee.
• xAI has historically used (and probably still uses) some cloud compute capacity, with Oracle as a confirmed partner as of 2024.

Beginning in 2026, SpaceX started renting out part of Colossus to other companies such as Anthropic, Google, and Reflection. SpaceX also acquired Cursor in 2026, with the Cursor and SpaceXAI teams co-developing models. Going forward, we will likely group these two teams’ compute together, while subtracting out the SpaceX-owned compute rented out to other parties.

Colossus capacity

We take estimates of Colossus 1 and Colossus 2 operational compute capacity over time from the AI Data Centers explorer. We assume that 100% of this was used by SpaceXAI/xAI before SpaceX’s compute partnership with Anthropic began in May 2026. Colossus 1 and 2 almost certainly make up the vast majority of SpaceXAI’s owned AI compute capacity. For example, SpaceX’s Q2 earnings deck discloses a nameplate compute capacity of 1.0 GW as of March 31, 2026, while its S-1 filed in May (pg. 76) states that Colossus 1 and 2 “collectively provide approximately 1.0 gigawatt of compute power”.

While some of our compute estimates from AI Data Centers are uncertain, our Colossus estimates are relatively confident because many of them are grounded in specific statements from SpaceX, now a publicly traded company. However, there is still some uncertainty about when phases of Colossus 1 and 2 came online:

• In 2024, only Colossus 1 was operational. Colossus 1’s first phase of 100,000 Hopper GPUs was completed and operational in September, and we estimate that the second phase of another 100,000 Hoppers was online by February 2025. For the end of 2024, we interpolate some share of this second phase as operational. We model this share as likely between 40% and 95% (since a simple linear interpolation between these dates would imply the second phase was ~70% complete).
• By 2025, Colossus 1 was complete, with 200,000 Hopper GPUs and 30,000 Blackwell GPUs, for ~276,000 H100e online. These chip counts are well-grounded by company statements. Meanwhile, we estimate that Colossus 2’s first phase of 110k Nvidia GB200s was online by October 2025, and that the second phase of another 110k Blackwell GPUs came online by April 2026. Our satellite-based model estimates that Colossus 2’s capacity was still at the initial 110k GPUs as of January 2026.²³ Accordingly, here we model that Colossus 2 likely still consisted of its first phase at the end of 2025, with some chance that the second phase was partially completed.

External cloud capacity

xAI partnered with Oracle in 2024 to train Grok 2 on 20,000 H100s. There’s a good chance that xAI still used this cluster through 2026 — xAI was only founded in 2023, and multi-year contracts are common for large AI cloud deals.

SpaceX’s S-1 also contains some information that can be used to place rough bounds on its AI cloud compute spending. On page 112, they disclose that:

• GPU depreciation increased by $1.673 billion in 2025 vs 2024.
• “Infrastructure and cloud computing R&D expense” increased by $1.44 billion.
• “Infrastructure and cloud computing cost of revenue” increased by $412 million.

The latter two buckets add up to ~$1.8 billion in “infrastructure” and “cloud computing” spending. However, this $1.8B combines cloud compute spending with “infrastructure”, which presumably means operating costs and depreciation on non-GPU AI infrastructure like networking and facilities (since GPU depreciation is a separate category). In addition, this is only the increase in 2025; the 2024 baseline is not confirmed.

2024 baseline

SpaceX doesn’t disclose its “infrastructure and cloud computing” spend in 2024, but it did confirm (page 117) that its AI R&D spend in 2024 was $1.176 billion. Note that SpaceX’s “AI” segment actually includes both the xAI lab and X (the social media network formerly known as Twitter). Its AI cost of revenue was ~$1.7 billion in 2024, but this can be presumed to mostly be serving costs for X traffic, since xAI’s inference compute expenses were likely smaller than its R&D compute expenses in 2024.

This $1.176B in turn is an upper bound on xAI’s R&D cloud compute spend, since it would also include R&D spending for X, non-compute AI R&D expenses like staff compensation and data, and non-cloud compute R&D spend from Colossus 1 depreciation. Colossus was $6 billion in capex in 2024, so on a six-year depreciation cycle ($1 billion per year) with the depreciation clock starting in early September, xAI may have booked roughly $300 million in Colossus 1 R&D expenses in 2024.

Given these exclusions, even though the $1.176 billion in AI R&D spending in 2024 doesn’t include inference cloud compute, it is still probably higher than xAI’s total cloud compute bill that year. This is consistent with the size of the Grok 2 cluster: 20k H100s rented for $2/hour would cost $350 million per year, so xAI’s cloud spend run rate was probably in the $300M to $1B range at the end of 2024.

2025

SpaceXAI increased its “infrastructure and cloud-computing” expenses in 2025 by ~$1.8 billion. A large part of this $1.8 billion was likely the “infrastructure” side (i.e. depreciation on non-GPU equipment), since they reported an increase in GPU depreciation of $1.673 billion, and non-GPU equipment is a large minority of the capital costs of building an AI data center (the S-1 does confirm on page 91 that AI-related cloud computing spend was “higher” in 2025, without any numbers). Given the rough 2024 baseline of $300 million to $1 billion, this suggests a plausible spend rate of $1 billion to $2 billion/year for 2025, with a fairly strict upper bound of ~$3 billion. At $2/hour, these roughly correspond to 57k and 114k H100e, respectively.

Since we’ve established that cloud compute is likely only a small minority of xAI’s overall compute, it is not necessary to decompose cloud spending into a rigorous estimate of annual spending and price per H100-hour. For simplicity, we model a single H100e distribution of 20k to 60k H100e in 2024, and 30k to 120k in 2025, for xAI’s non-Colossus cloud compute capacity.

Appendix: Frontier developer compute in 2026

While we do not have a formal forecast for Anthropic’s and OpenAI’s compute as of the end of 2026, several lines of evidence point to continued rapid growth in their compute fleets.

OpenAI’s president, Greg Brockman, testified that the company would spend $50 billion on compute in 2026, roughly triple what it spent in 2025. This is somewhat faster than OpenAI’s growth in compute spending between 2024 and 2025, and if the incremental compute spending is more cost-effective due to improved chip efficiency over time, this suggests that OpenAI’s compute capacity will more than triple in 2026.

Anthropic is also aggressively ramping up its compute in 2026. Notably, in May 2026, Anthropic reached an agreement with SpaceX to rent the entire Colossus 1 data center (consisting of over 220k Nvidia GPUs, or ~280k H100e), and part of Colossus 2, paying $1.25 billion per month ($15 billion per year annualized) for this capacity. Additionally, we estimate that the Project Rainier campus in Indiana increased its compute capacity by around 200k H100e in the first half of 2026. Overall, Anthropic reportedly planned to spend around $6 billion on compute in Q2 2026 alone, or close to what it spent in all of 2025, indicating a very aggressive ramp in compute spending.

SpaceX is also rapidly scaling up its compute fleet. SpaceX’s total compute fleet in Colossus 1 and 2 now stands at around 1.4 million H100-equivalents. As mentioned above, Colossus 1, and possibly some portion of Colossus 2, is now being used by Anthropic, which implies around 1.1 million H100e in capacity available to SpaceXAI, SpaceX’s internal AI lab, as of mid-2026, not including external cloud capacity.

SpaceX has also signed deals with Google to rent out 110,000 GPUs (presumably Blackwell GPUs, for around 280,000 H100e) for $920 million/month beginning in October 2026, as well as a smaller $150 million/month deal with the AI model startup Reflection AI. This will further subtract from the capacity available internally to SpaceXAI, but we also forecast that Colossus 2 will grow to around 1.8 million H100e by early 2027.

Appendix: deployment lags between chip sales and data center deployments

Deployment lags between when chips are sold (revenue recognized by the chip vendor, e.g. Nvidia or Broadcom) and when they are deployed and available in a data center factor into the compute models above in two places:

• They are used to adjust estimates of Google’s and Meta’s owned compute stocks (which come from estimates of the respective companies’ purchases of Nvidia, AMD, and Broadcom chips) into deployed compute capacity.
• Less significantly, they are used to estimate the mix of OpenAI’s chip fleet over time between Nvidia GPU generations.

Our AI chip sales estimates are primarily based on recognized revenue from major chip vendors such as Nvidia, Broadcom, and AMD.²⁴ These companies recognize revenue after hardware is delivered, but even after AI chips or servers have been delivered to a customer (which may be an intermediate customer such as a server ODM/OEM rather than a cloud company), some portion may be in storage, in transit, or being installed in a data center.

Detailed empirical evidence of these deployment lags is lacking. At a high level, there are two main sources of lag: server assembly and data center installation.

For data center installation, one useful reference is CoreWeave, which has disclosed that its hardware goes through a “~3 month installation time” between delivery and monetization. By late 2025, they had decreased this deployment time to “within weeks”.

Note that typical deployment lags may be misleading in a few ways. First, there can be systematic, industry-wide factors delaying the installation of AI chips, like technical problems with a new chip generation or infrastructure bottlenecks like power delays. For example, many Nvidia GB200 deployments were reportedly delayed for months in early 2025 due to rack heating issues, even after cloud companies received the servers. This means the ratio of deployed compute to sold compute was unusually low in the first half of 2025, though we’re not aware of any comparable issues with the GB300 or other modern AI chips right now.

If industry-wide deployment bottlenecks are persistent, they will eventually impact upstream chip sales. So when a chip company like Nvidia reports healthy growth in sales, we can retroactively judge that the industry probably did not experience severe deployment bottlenecks in the recent past. Customers may still temporarily ramp up purchases of chips they can’t install in order to secure future supply, but this will eventually become irrational or financially unsustainable.

Additionally, applying a typical/median deployment lag to past AI chip sales to estimate the deployed stock ignores the variance in deployment lags. Some chips may take longer to deploy than others. In an extreme example, imagine that two-thirds of AI chips wait in storage for an entire year, while the rest are installed almost immediately. If chip sales are growing rapidly, applying a uniform one year delay will underestimate the deployed chip fleet, because the one-third of chips that are installed right away will make up a disproportionate share of the deployed fleet at any given time.