The AI Chip Owners explorer builds on our AI Chip Sales explorer, which compiles our estimates of the total sales and deliveries of leading AI chips based on company earnings commentary, estimates from analysts and industry research firms, and media reports.

For the Owners hub, we then allocate these AI chips to owners using disclosures and earnings commentaries from chip vendors such as Nvidia, alongside analyst estimates and media reporting of major purchases. We rely significantly on the subset of analyst/industry research results that is shared with the public, so our work should not be viewed as a substitute for proprietary industry research.

More information can be found in the methodology.

The goal of this hub is to describe which companies control the world’s AI compute. We attribute ownership informally without strict legal/financial criteria. We do not recommend using our figures to understand assets owned by these companies from a financial perspective.

Some of the compute we attribute to an AI or technology company may actually be jointly owned by other players like the data center developer or external investors (often via a joint venture or special purpose vehicle). This is true of some of xAI’s data centers, and will be true of Meta’s upcoming Hyperion data center. Beyond joint ownership structures, an AI data center or compute cluster may be wholly owned by a tech company but also serve as collateral for debt financing.

However, when an AI lab or other company uses compute provided by a cloud company, we attribute that compute to the cloud company, not to the lab itself.

We want to estimate trends in AI compute, and how this compute is allocated, due to its strategic implications for AI progress and societal impacts. Our intended audience is primarily other researchers, policymakers, and the general public. This hub is not intended to serve as a tool for financial analysis of the companies mentioned.

H100-equivalent (H100e) compute capacity is total computing power measured in terms of the equivalent number of Nvidia H100 GPUs. We divide the peak number of dense 8-bit operations (FP8 or INT8) each chip can perform by the Nvidia H100’s spec, and then multiply this ratio by the number of chips. For example, since a TPUv7 can perform approximately 2.3× as many operations as an H100, then one million TPUv7s have a compute capacity of 2.3 million H100e. For chips that do not support 8-bit precision, we use their 16-bit (FP16 or BF16) performance instead.

The power figures are based on each chip’s thermal design power (TDP). It refers roughly to the maximum sustained power draw of the chip. Importantly, the total power draw of an AI data center is typically much higher than just the total TDP of the chips inside (roughly twice as high), due to overheads at the server, cluster, and facility level.

We don’t have a definitive, comprehensive breakdown of AI compute ownership, with a significant minority attributed to “Other” customers that are not among the largest hyperscalers and cloud companies, frontier AI developers, or Chinese customers.

Other categories of important compute owners include:

• Neoclouds: These are smaller, newer cloud providers that generally focus on AI. These are most likely the most important category besides the hyperscalers. Major AI neoclouds include Nebius, Nscale, and Lambda, and there are at least dozens more.
• “Sovereign AI” projects: This is Nvidia’s term for compute investments supported or facilitated by non-US governments to further their national interests. They made up around 15% of Nvidia’s AI sales in 2025. But these projects appear to be mostly owned by companies rather than states. This means that sovereign AI overlaps with neoclouds and the other categories.
• Enterprises and AI companies that buy their own compute to train or run AI models, e.g. to achieve greater control or data privacy.
• Scientific research organizations such as the US national laboratories that use large GPU clusters for traditional high-performance supercomputer tasks, in addition to AI/machine-learning research workloads.

Before xAI was acquired by SpaceX, xAI was the only major pure-play AI developer that owned most of the compute it uses. By contrast, OpenAI and Anthropic do not own the vast majority of the compute they use for training and inference. For now, their compute is currently provided by cloud companies (companies that rent out compute)—primarily Microsoft, Oracle (including “Stargate”), and CoreWeave for OpenAI, and primarily Amazon and Google for Anthropic. Frontier labs housed inside larger companies, like Google DeepMind and Meta’s AI divisions, presumably mostly use compute owned by their parent companies, though Meta has signed deals to rent cloud compute.

Both OpenAI and Anthropic are reportedly interested in investing in their own compute in the future, with Anthropic planning to buy a large number of TPUs in 2026.

No. The compute owned by Google (which is shorthand for Alphabet) is split across Google DeepMind (including research and training for Gemini models and other efforts), other internal uses at Google, and cloud compute for external customers. Similarly, the compute owned by Meta is not exclusively used by Meta’s frontier AI efforts.

Most of the world’s AI compute is owned by the “hyperscalers”—usually describing Google, Microsoft, Amazon, Meta, and sometimes Oracle. Note that the hyperscalers do not own all the chips they use: Meta has signed cloud deals with Google and CoreWeave, and Microsoft acquires some compute from neoclouds that they then rent out to their own customers.

Hyperscaler compute usage for a hyperscaler mostly falls within three broad categories. The relative ranking of these three categories is quite unclear to us.

First, renting out AI compute on the cloud to AI companies and other customers. Hyperscaler cloud compute is the primary source of compute for most pure-play AI labs, e.g. OpenAI and Anthropic. Amazon, Microsoft, and Google are the top three cloud companies overall, and almost certainly the top three AI clouds as well. Providing cloud compute is also the primary use of Oracle’s AI chips. Meanwhile, Meta does not have a cloud division.

Second, powering internal AI labs for training models, research, and inference. The four largest hyperscalers all have internal AI divisions working on large-scale foundation models. The internal labs at Google (Google DeepMind) and Meta (MSL/FAIR) appear to be much more advanced than the labs at either Microsoft or Amazon, and presumably use much more compute.

Third, powering AI products or AI enhancements to products. This is significant for both Meta and Google who use AI compute for content/ad recommender systems as well as AI features like AI overviews in Google Search. Many recommender algorithms are now powered by large-scale transformer models like Meta’s GEM. While these companies don’t disclose hard numbers, these AI systems are likely highly lucrative and compute-intensive for Google and Meta, with both companies attributing recent growth in their traditional advertising businesses to AI advances. This category also includes product features such as AI features in Google search and a wide variety of other internal uses.¹

The hyperscaler clouds all host their own model APIs for models from e.g. OpenAI and Anthropic as well as open-weight models.

This is a bucketed category for the AI chips that we infer were sold to companies based in the People’s Republic of China. For example, this is true of the vast majority of the Nvidia and AMD chips designed around US export controls (e.g. the H800, H20, and MI308X), and of Huawei Ascend chips. It does not mean Chinese government ownership of compute.

Our estimates of Chinese compute do not include the smuggling of export-controlled chips. See our methodology for more discussion on potential volumes of these imports.

This is outside the scope of the current phase of this project as of early 2026. A company headquartered in one country may build AI data centers elsewhere; this is fairly common for the US hyperscalers. Our AI Data Centers hub shows the location of many large AI data centers, though the total coverage is currently still a minority of the estimated global AI compute stock. And most chips owned by Chinese companies are likely physically located in China (though they can also procure offshore cloud compute).

Following our AI Chip Sales methodology, our headline numbers primarily come from recognized revenue by the relevant chip vendor (e.g. Nvidia, AMD, and Broadcom for Google TPUs), though in some cases (like xAI’s Colossus) we use operational data center capacity as a proxy metric. These vendors recognize revenue when they deliver hardware to their immediate customers, which can be server assemblers rather than the company actually deploying the chips. So there may be a time lag for when the server is assembled and delivered to the final customer that is not accounted for in our figures. This is on top of the time lag between hardware delivery and when chip clusters are installed and operational in data centers, which can take weeks or months.

On the flip side, our headline figures do not count unsold inventory held by Nvidia and others. Further discussion of inventory volumes can be found here.

We estimate chip ownership using Monte Carlo models incorporating uncertainty in both the total volume and computing power of AI chips, and in the case of Nvidia and AMD, each company’s share of those chips. For some owners and time periods, we have significant uncertainty about how many chips they acquired (for example, most hyperscaler shares of Nvidia pre-2024, and Oracle’s pre-2025); we do our best to widen the uncertainty in share parameters when we are extrapolating or operating from limited evidence. Our headline figures are the median predictions of these models, and we sample from our models and report the 5th and 95th percentiles as confidence intervals. More details can be found in the methodology and in our code notebooks.

When comparing confidence intervals, one should consider whether the uncertainty may be correlated. For example, if our median estimate is that company A has more Nvidia compute than company B, but their Nvidia confidence intervals overlap, that does not necessarily mean it is plausible that company B has more Nvidia compute than company A. This is because the confidence intervals account for uncertainty in both the total number of Nvidia chips and each company’s share of Nvidia, so when holding the total stock of Nvidia chips constant, it is possible that the CIs will no longer overlap.

When CIs are not available for a figure, this means that we do not have a probabilistic model for that figure, not that we are completely confident in our reported estimate.