Use This Work

Epoch’s data is free to use, distribute, and reproduce provided the source and authors are credited under the Creative Commons Attribution license.

Summary

Nvidia chips comprise most of the world’s AI compute, so most of our modeling focuses on Nvidia. Nvidia also has a relatively diverse customer base. We estimate Nvidia chip ownership for several hyperscalers and large cloud companies, one frontier AI model developer (xAI, now owned by SpaceX), and Chinese customers. For each of these, our approach differs somewhat:

• Hyperscalers: According to Nvidia earnings commentary and analyst estimates, around half of Nvidia’s flagship AI chip revenue in 2023–2025 came from the top hyperscalers (Microsoft, Meta, Amazon, Google, and Oracle). We estimate company-level shares using analyst estimates and evidence from hyperscaler capex figures.
• xAI: We estimate xAI’s compute ownership using operational capacity data from our Frontier Data Centers hub.
• China: We estimate Nvidia sales to Chinese customers based on reported revenues from their China-market chips (H20, H800, A800), as well as Nvidia disclosures about their China revenue, including regarding pre-export-control A100 sales. We do not explicitly model the number of smuggled chips to mainland China (more discussion below).

For AMD, disclosures and industry research coverage of customer shares are more limited than for Nvidia. We rely primarily on 2024 estimates that attribute most AMD Instinct GPU sales to Meta and Microsoft. Our extrapolation of these estimates to 2025 carries added uncertainty.

For custom hyperscaler chips (Google TPU/Amazon Trainium), we attribute ownership to the respective hyperscaler. We will revisit this allocation for future quarters, as Broadcom plans to begin selling TPUs to external customers such as Anthropic.

For Huawei, we attribute all Huawei Ascend chips to Chinese customers collectively, absent evidence of significant sales outside China.

The full code for our analyses is available here.

Note: this methodology was originally written in April 2026. While we plan to maintain and update this hub and methodology over time, descriptions of industry context may not reflect more recent developments.

Nvidia - Hyperscalers

The largest owners of Nvidia compute are the “hyperscalers”, meaning Microsoft, Alphabet (a.k.a. Google)¹, Amazon, Meta, and Oracle.

We rely on a mix of Nvidia disclosures, analyst estimates, and capex figures to estimate their Nvidia purchases. In short:

• Nvidia has disclosed multiple times through January 2026 that between 2024 and 2026 their top 4-5 customers made up approximately 50% of their data center revenue.
• We use analyst estimates to allocate this revenue to individual hyperscalers.
• We check for evidence from hyperscaler capex numbers and make slight adjustments to these analyst figures, e.g., bumping up Meta’s share of Nvidia in the past year.
• Finally, we take our estimates of each hyperscaler’s share of data center revenue and allocate Nvidia’s sales of non-China-spec AI chips to each hyperscaler in proportion, adjusting slightly due to hyperscalers paying lower prices than the average customer.

More details and a full list of our parameter values can be found in this sheet: Nvidia ownership shares

We then use these hyperscaler share allocations to model their overall chip purchases using Monte Carlo sampling. This builds on our revenue-based model of overall Nvidia chip sales, which is described here. Most of the explanation of how we implement our uncertainty modeling can be found in the code itself; the rest of this methodology focuses on our parameter choices (point estimates and uncertainty intervals) and the evidence behind them.

Nvidia - xAI

xAI is a frontier AI developer that recently merged with SpaceX. For now, we will continue to describe its compute as owned by “xAI” rather than “SpaceX” because xAI is more publicly associated with AI and because the xAI subsidiary is presumably still relatively operationally independent from the rest of SpaceX (by comparison, DeepMind and Google seem much more intertwined).

xAI owns and operates two large data centers in the Memphis, Tennessee area: Colossus 1 and Colossus 2. Note that Colossus 2 is jointly owned by xAI and outside investors; we attribute it to xAI for simplicity. From what we can tell, Colossus 1 and 2 make up nearly all of the compute capacity owned by xAI, though xAI reportedly also built a smaller data center in Atlanta with approximately 12,000 GPUs. xAI has also used cloud compute, at least historically. For example, Grok 2 was trained with Oracle-owned chips.

We use estimates from our Frontier Data Centers hub of the collective compute capacity of Colossus 1 and Colossus 2 over time, as well as company statements on the chip splits in these data centers. These estimates are tied to specific dates, so we map them to calendar quarters. As of the end of 2025, Colossus 1 and 2 have approximately 550,000 H100-equivalents in compute capacity. This is around 3% of all estimated Nvidia AI chip sales since 2022, and a fraction (around one-seventh to one-quarter) of the Nvidia compute owned by each of the top four hyperscalers.

Because these figures are based primarily on when data center capacity comes online, they are likely lumpy and lagged relative to when xAI actually received these chips. This means that these data center-based figures are usually lagged estimates (that is, underestimates) of the chips xAI actually owns over time. In one case, however, we incorporate B200s delivered to Colossus 2 in fall 2025 shortly before we estimate they became operational.

Nvidia - Chinese customers

We attribute Nvidia sales to customers based in mainland China, building on work done in the AI Chip Sales explorer.

Nvidia chips designed for the Chinese market are assumed to have been sold to customers in China; this is a slight simplification (Nvidia sold approximately $500 million in H20s to a customer outside China in the last quarter in which H20 was sold), but one we believe has a negligible impact on the overall estimates.

For the H800s and A800s, which were more competitive with Nvidia’s flagships, it is less clear whether they were exclusively sold in China, but our estimates of H800 and A800 volumes are actually inferred from Chinese Nvidia purchases in the relevant time period, not from any H800/A800 revenue breakout (which Nvidia has not provided to our knowledge). In addition, Chinese customers likely bought between 10% and 20% of Nvidia A100s before export controls were first implemented in 2022.

Note that while Nvidia has now received US government licenses approving H200 exports to China, shipments have not yet begun as of the latest Nvidia fiscal quarter, ending in January 2026.

A longer write-up on our methodology for these estimates can be found here.

Note that our China ownership figures exclude unofficial exports to China in contravention of US export controls on leading US-designed AI chips. Prior work by Grunewald & Fist (2025) estimated that over 100,000 Nvidia A100s and H100s were shipped to China in 2024, though there is significant uncertainty in this figure. Several reports indicate that Chinese imports of export-controlled chips continued in 2025, including:

• Upwards of a billion dollars worth of black market chips sold during Q2 after the Nvidia H20 was banned
• A 2026 DOJ indictment accusing several employees of server assembler SuperMicro of shipping over at least $500 million and possibly up to $2.5 billion in Nvidia servers to China
• An allegation that Megaspeed, a Singaporean cloud provider, had tens of thousands fewer chips in their data centers than they ordered from Nvidia, which may have been redirected to China.

As with all of our ownership figures in this hub, we do not measure cloud compute sales. Some Chinese companies reportedly rent leading US AI chips via offshore cloud compute, which, as of early 2026, is allowed under US export controls.

Nvidia - CoreWeave

CoreWeave is the largest AI “neocloud” by a significant margin. CoreWeave recently cited a third-party estimate that they were larger than the next 15 neoclouds in the US and Europe put together. One industry analyst describes CoreWeave as a “neocloud giant,” alongside Lambda, Crusoe, and Nebius, and a cursory review suggests that none of these other neoclouds is close to CoreWeave in scale.¹⁵ As a publicly traded company focused almost entirely on AI compute, CoreWeave discloses a large amount of information about their AI compute capacity.

CoreWeave primarily uses Nvidia compute. We estimate their GPU fleet composition over time: CoreWeave directly discloses its deployed GPU unit counts and data center power capacity in annual S-1 filings, and has given updates on power capacity for every quarter of 2025. We derive or estimate the breakdown into specific chip types (allowing us to find the total compute capacity) using GPU power consumption data and our estimates of Nvidia’s chip sales mix.

Code for analysis can be found here.

2022–2024: S-1 Disclosures

CoreWeave disclosed their total deployed GPU count and data center power capacity at year-end in their S-1 filing with the SEC in March 2025. Throughout, we assume their reported power capacity refers to IT power (electricity consumed by servers and compute equipment) rather than total facility power. This is standard practice for data center compute providers whose disclosures are primarily meant to characterize their compute infrastructure. For example, Core Scientific reports their power capacity in terms of critical IT power.

2022: CoreWeave disclosed that they ended the year with 17,000 GPUs and 10 MW of capacity. The low implied average IT power per GPU (~588 W) indicates a mixed fleet of A40s, RTX cards, L40s, A100s, and other lower-power chips. We exclude 2022 from our estimates for three reasons: the GPU mix is unclear and hard to anchor, many of these cards have likely been retired, and even if still in service they would represent a negligible share of CoreWeave’s current compute capacity.

2023: CoreWeave ended the year with 53,000 GPUs and 70 MW of capacity. Even in the most conservative scenario — assuming all 17,000 chips from 2022 were still running — the implied IT power per GPU is 70 MW / 53,000 = 1,321 W, which points strongly to a high proportion of H100s, since previous Nvidia chips such as the A100 had significantly lower average power draws even accounting for server and other IT overheads. CoreWeave had also placed a $100M order for H100s in summer 2022 and made them available in Q1 2023. We treat all GPUs added in 2023 as H100s.

Rather than back-solving GPU counts from the raw 2023 MW figure, we apply our 2024-derived IT power assumption of 1,472 W per GPU (see below), which we have more confidence in. We also assume roughly half of the 2022 fleet was retired during 2023, freeing up approximately 5 MW of capacity. This implies 65 MW of gross new H100 capacity deployed in 2023, or approximately 44,200 H100s added.

2024: CoreWeave ended 2024 with 250,000 GPUs and 360 MW of capacity. Assuming no 2023 GPUs were retired, this means they added 197,000 GPUs and 290 MW of capacity, implying 1,472 W per GPU. We believe virtually all of these were H100/H200 for two reasons: Blackwell chips were not meaningfully sold by Nvidia until Q4 2024 and were not deployed by CoreWeave until early 2025, and our Nvidia chip sales estimates show 95% of Nvidia GPU units sold in 2024 were Hoppers. The H100 and H200 share the same 700 W TDP so we group them together. We use the implied 1,472 W as our H100/H200 IT power assumption going forward, equivalent to a GPU/IT power ratio of ~48%. As a cross check, we derive the H100/H200 IT power using the methodology from our Frontier Data Centers tracker: taking the H100/H200 server power of 1,275 W and applying a 1.14x factor to account for additional IT power overhead beyond the GPU itself. This gives 1,454 W, which closely aligns to the 1,472 W implied by CoreWeave’s GPU count and power-capacity disclosures.

2025: Quarterly Power Capacity

For 2025, CoreWeave disclosed their total operational data center power capacity in each of their quarterly earnings, but not GPU counts. Their cumulative capacity reached 420 MW, 470 MW, 590 MW, and 850 MW at the end of Q1 through Q4, implying quarterly additions of 60, 50, 120, and 260 MW.

To estimate GPU counts by generation, we use our Nvidia chip sales estimates as a proxy for CoreWeave’s chip mix in a given quarter, under the assumption that CoreWeave’s procurement broadly tracks the overall Nvidia sales mix for non-China chips. We apply a lag between when Nvidia books revenue for a chip and when CoreWeave deploys it, and model this lag as shortening over the course of 2025 for two reasons.

First, the high capital cost of holding GPUs idle makes a long lag increasingly implausible as deployment scale grows. In Q3 2025 CoreWeave added 120 MW of capacity, and in Q4 2025 they added 260 MW — at $30–40k+ per GPU, pre-purchasing a full quarter’s worth of chips before the data center is ready to receive them would mean holding billions of dollars of assets in storage with no return. We therefore model CoreWeave as timing their GPU purchases to land close to when power capacity is available, rather than accumulating large advance inventories.

Second, we assume CoreWeave’s installation process became faster and more streamlined as 2025 progressed. On their Q4 2025 earnings call, management noted that after resolving delays flagged in Q3, they deployed over 50,000 Grace Blackwells “within weeks of receiving access to the requisite data center infrastructure.”

We model a median lag of one quarter for Q1 and Q2 2025, which is consistent with observed deployment timelines: CoreWeave made GB200 NVL72 generally available in February 2025, aligning with Nvidia first booking meaningful B200 revenue in Q4 2024, and deployed GB300 NVL72 in July 2025, aligning with B300 sales beginning in Q2 2025. The one-quarter lag is further supported by CoreWeave’s Q1 and Q2 2025 earnings presentations, in which they indicate a 3 month lag between purchasing GPUs and making revenue from them. We shorten our median lag to 0.8 quarters for Q3 2025 and 0.4 quarters (~5 weeks) for Q4 2025.

To the best of our knowledge, CoreWeave does not purchase or host China-spec chips such as the H20, H800, and A800. Thus, we exclude these chips from the sales mix and renormalize the remaining GPU types — H100/H200, B200, and B300 — to 100% before applying their IT power-weighted shares to CoreWeave’s quarterly MW additions. We weight by IT power rather than unit count because chips with higher IT power per GPU account for a disproportionate share of capacity for a given number of units.

IT power per GPU is estimated as follows:

The H100/H200 figure is derived directly from CoreWeave’s disclosed GPU count and power capacity. For B200 and B300, we estimate IT power by using the power capacity of each GPU’s most common server spec — GB200 NVL72 and GB300 NVL72 — and applying a 1.14x overhead for other IT power costs. This follows the methodology of our Frontier Data Centers tracker; more details can be found in this sheet.

We model uncertainty in these estimates using Monte Carlo simulation with 10,000 draws, capturing three sources of uncertainty: (1) Nvidia unit counts per chip per quarter from our chip sales hub; (2) the lag between procurement of GPUs and deployment; and (3) the B200 and B300 GPU/IT power ratios. The 90% confidence intervals on total GPUs added in 2025 run from approximately 237,000 to 258,000, with H100e ranging from approximately 508,000 to 556,000.

This brings CoreWeave’s estimated cumulative fleet to approximately 497,000 GPUs at the end of 2025, or roughly 781,000 H100-equivalents [90% CI: 758,000–805,000].

Sanity Checks

Revenue: Our GPU count estimates imply approximately $4.93B in 2025 revenue [90% CI: $4.65B–$5.23B], assuming H100/H200s rent at $2/hr and B200/B300s at $3/hr, 80% utilization, and linear ramp-up of newly added GPUs within each quarter. CoreWeave’s actual 2025 revenue of $5.1B is reasonably consistent with our estimates.

Power to Compute Capacity: In our Frontier Data Centers hub, we estimate that OpenAI Stargate Abilene data center has 590 MW of IT power capacity and ~510,000 H100e as of March 2026, which is 864 H100e per MW. Applying this ratio to CoreWeave’s end-of-2025 deployed capacity of 850 MW implies 734,000 H100e, which aligns with our median estimate of 781,000 H100e.

End-of-2025 GPU Inventory

The figures above reflect CoreWeave’s deployed capacity. However, CoreWeave likely holds additional GPUs in inventory at the end of 2025, which are purchased but awaiting power capacity coming online in early 2026. To estimate these, we model the GPUs CoreWeave would have procured for their Q1 2026 deployments.

CoreWeave has guided to ending 2026 with approximately 1.7 GW of capacity, implying roughly 850 MW of additions over the course of 2026. To estimate Q1 2026’s share, we use their capex guidance: full-year 2026 capex of $30–35B and Q1 2026 capex of $6–7B, implying Q1 accounts for approximately 17–23% of annual capex. We model this share as uniform over that range, giving a Q1 2026 MW addition of approximately 145–200 MW.

Importantly, we model Q1 2026 capacity coming online gradually throughout the quarter. With a procurement lag of 0.4 quarters (~5 weeks) and assuming a linear ramp in capacity additions, CoreWeave would have pre-purchased only the fraction of Q1 2026 capacity coming online within 5 weeks of year-end — approximately 40% of Q1’s total MW addition. We assume no capacity retirements in 2026 given continued strong demand for older GPU generations.

This adds 79,000 GPUs to CoreWeave’s end-of-2025 inventory, predominantly B200s and B300s given the Q4 2025 Nvidia mix. Combined with the 249,000 from 2023–2024 and ~247,000 deployed in 2025, this brings our median estimate for CoreWeave’s total GPU stock at end of 2025 to approximately 520,000 GPUs [90% CI: 507,000–534,000], or roughly 852,000 H100-equivalents [90% CI: 811,000–866,000].

Nvidia - Other/unlisted owners

We estimate that the five hyperscalers own around 60% of total Nvidia AI compute (measured in H100 equivalents), with xAI owning at least 3%. CoreWeave, which we track separately above, is likely the sixth-largest owner. The remaining Nvidia compute — roughly a quarter of the total — is distributed across several other owner categories, though we do not know of a comprehensive breakdown.

Neoclouds, of which there are at least dozens, may own a majority of the remaining Nvidia compute.

Another large customer category for Nvidia, likely overlapping heavily with neoclouds, is “Sovereign AI”. This segment made up 15% of the company’s data center revenue in 2025. “Sovereign AI” is a marketing term without a clear definition, but in practice, it refers to compute buildouts to enhance a non-US nation’s AI capabilities or independence. These projects are generally championed by national governments but owned by corporations, meaning that many of these projects are actually owned by non-US neoclouds. Nvidia has said that this category is “driven primarily by customers based in Canada, France, the Netherlands, Singapore, and the UK”. We don’t know for sure that there is zero overlap between “sovereign AI” and hyperscaler compute located outside the US, though Nvidia’s description of customer locations suggests that this overlap is small at best.

Finally, Tesla disclosed in a 2025 investor presentation that it had 120,000 H100-equivalents of AI training compute capacity as of the end of September 2025. This capacity is likely mostly or entirely Nvidia GPUs. While Tesla specified training capacity (we take this to mean all R&D compute) as opposed to inference, Tesla’s main AI product is autonomous driving, and its AI inference is mostly done on board its vehicles using custom hardware, not leading datacenter chips. Tesla also doesn’t specify how it converts compute to H100-equivalents; we take their conversion at face value.

Nvidia-owned inventory and shipment lags

At any given time, Nvidia holds significant volumes of hardware as unsold or undelivered inventory: at the end of 2025, its total inventory was worth $21 billion (approximately 30% of its recent quarterly sales). Our topline ownership figures don’t include this inventory.

Our ownership figures also don’t account for shipment lags at the server/distributor level. This means we may be overestimating hyperscaler purchases for any particular quarter. Under generally accepted accounting principles, Nvidia recognizes revenue when it delivers hardware to customers, not when it receives purchase orders. However, the first layer of customers are often server OEMs, which assemble the chips into servers, rather than the company that will ultimately deploy the chips. Nvidia’s customer attribution to the hyperscalers likely accounts for this intermediate step, but at a lag; i.e., the company reports an estimate of how many of the chips sold in a given quarter to server assemblers will eventually end up at the hyperscalers. The time lag between when the server is assembled and delivered to the final customer is not accounted for in our figures. This is in addition to the time lag between revenue recognition and when chip clusters are actually installed and operational in data centers, which can take weeks or months.

Beyond inventory, we are not aware of Nvidia itself owning and deploying its own chips (which would be excluded from its sales data) on a large scale. Nvidia has a major AI research and model development arm that works on the Nemotron language model family as well as robotics models. We know that the company uses neocloud compute for at least some of these workloads. Nvidia also has an AI cloud division that reportedly runs on rented capacity from Oracle and neoclouds. We also aren’t aware of any Nvidia-owned data centers or clusters approaching the scale of frontier data centers.

Custom chips (TPU/Trainium)

AMD

AMD’s flagship AI chip line is the Instinct GPU family. Based on industry research reports as well as announced partnerships, the largest customers of AMD’s Instinct GPUs are Meta, Microsoft, and Oracle. We are not aware of reports of any major Instinct purchases from Amazon and Google.¹⁹

We estimate the allocation of Instinct GPUs across Meta, Microsoft, Oracle, and a remaining “Other” category for 2024-2025. We also attribute sales of their Chinese-market GPU, the MI308X, to Chinese customers generically. We obtain total AMD AI GPU shipment volumes and average prices by chip type from our Chip Sales Explorer.

For 2024, we use estimates from the research firm Omdia of MI300X purchases by customer: Omdia estimates that Meta and Microsoft respectively purchased approximately 50% and 30% of all MI300X in 2024, while Oracle bought just over 10%.²⁰ We aren’t confident in these estimated shares, and we treat them as high-end estimates given a prior that 80% market share for two buyers is unusually high. The MI300X chip made up the majority of Instinct sales in 2024; most of the remainder was the MI300A, which was primarily tailored for scientific computing, e.g., at US national laboratories. We allocate MI300X sales to Meta/Microsoft accordingly, while assuming they did not buy any MI300A.

For 2025, we make two carve-outs before allocating units to customers. First, in line with our 2024 estimates, MI300A shipments are assigned entirely to “Other.” Second, AMD disclosed $390 million in MI308X sales to China in 2025. We convert this to units using the MI308X ASP of $12,000 and assign those units to a separate “China” category, removing the associated revenue from the pool available to other customers.

For the remaining data center GPU types sold in 2025 (MI300X, MI325X, MI350X, and MI355X), we extrapolate Omdia’s estimated Meta and Microsoft shares from 2024 to 2025 as high-end estimates, while factoring in significant uncertainty about how the sales split may have changed in 2025. For example, we model Meta’s 50% share with a confidence interval of [30%, 60%] — we don’t have enough information to pin down this range more precisely, but we are fairly confident that Meta and Microsoft remain the largest customers of AMD’s data center GPUs. Meanwhile, in 2025, Oracle announced plans to build two large AMD clusters of 50,000 MI450X GPUs by Q3 2026 and 30,000 MI355X GPUs with an unclear timeframe. This appears to be broadly consistent with a persistent ~10% share, though this also carries significant uncertainty.

A full list of parameters and values can be found in the code notebook for our AMD model and accompanying data sheet.

Huawei

For Huawei, we use our estimates of total shipments of Huawei’s Ascend AI chips from our AI Chip Sales explorer, which are synthesized from analyst estimates and media reports. More details can be found here. We attribute all Huawei AI chips to Chinese companies collectively, rather than to specific companies. Estimating Huawei volumes for specific customers is left for future work. We also assume Huawei was able to sell its full production — however, this may not be correct.

While there are some reports of Huawei marketing Ascend chips in moderate quantities (“low thousands”) to international customers, two features of US semiconductor export controls make it unlikely that Huawei has sold large volumes of AI chips to non-Chinese customers to date:

• Customers in most countries, besides China and other sanctioned US rivals like Russia, Iran, and North Korea, are not subject to export controls on leading US-designed AI chips. Huawei AI chips are not competitive in these markets, given they lag behind US alternatives in both performance and cost-efficiency.
• The US government has accused Huawei itself of violating US export controls on semiconductor manufacturing equipment, warning that customers of Huawei Ascend chips may be subject to “substantial criminal and administrative penalties”. This legal risk could be a major deterrent to international customers considering Huawei Ascends.

Appendix: Nvidia hyperscaler revenue share uncertainty

We have uncertainty in both overall Nvidia chip sales and each hyperscaler’s share of those chips. In our model these uncertainties are modeled as independent, absent a clear reason to think they are correlated.

The overall share of Nvidia revenue that goes to the top 4-5 hyperscalers is close to ground truth due to the Nvidia disclosures, but their wording admits uncertainty: they usually describe this share as “approximately 50%” or slightly above or below 50% or similar. Nvidia describes these numbers as estimates on their part, but Nvidia presumably has excellent visibility into their largest customers.²¹ In addition, these share disclosures are not made in every quarter (and sometimes omit Meta), which requires some interpolation.

So assuming approximately 10% uncertainty in each direction (e.g. point estimate of 50% to a credence interval of [approximately 45%, approximately 55%]) seems reasonable.

How would this translate to uncertainty in each company’s individual share of revenue? If these shares are independent, there’s a closed form solution, because the variance of the sum of independent variables is equal to the sum of the variances of each variable.

Let S be the total share of four hyperscalers. If each company’s share Xᵢ is independent with equal variance σ², then:

Var(S) = Var(X₁) + Var(X₂) + Var(X₃) + Var(X₄) = 4σ²

The standard deviation of S, σ_S, is the square root of Var(S) and is proportional to the width of the CI. So σ_S = 2σ, or equivalently σ = σ_S / 2: each company’s absolute uncertainty is half the uncertainty in the sum.

Because each company’s share is also about one-fourth of the total (on average), the relative uncertainty is larger: approximately double the overall relative uncertainty in S. So if the total hyperscaler share has ±10% relative uncertainty, individual company shares have roughly ±20% relative uncertainty on average.

This assumes that each company’s share is independent (uncorrelated). If their shares are positively correlated, then uncertainty in each company’s share is closer to the overall uncertainty in hyperscaler share (i.e. less uncertainty), since their values vary jointly. If they’re anti-correlated, then uncertainty in individual company shares should go up. Arguably, uncertainty in company shares should be anti-correlated, given the Nvidia information about overall hyperscaler share.²² If we learned that (e.g.) Amazon’s share was surprisingly low, we may want to bump up Microsoft’s share. But we don’t model any correlation, partly for simplicity, and partly because the analyst estimates provide independent grounding for individual company shares, versus merely inferring point estimates and uncertainty from the overall hyperscaler share.

Appendix: Nvidia direct capex model

Stepping beyond the black-box approach, is it plausible that Nvidia spending is approximately one-third of Meta’s and Microsoft’s capex? Should that be higher, or lower?

Here we attempt to model Nvidia chip purchases based on capex numbers. Let’s discuss Meta, which spent $72 billion on capex in 2025, as an illustrative example. How might this translate to their spending on Nvidia hardware?

This is the product of the following parameters. Each listed parameter multiplies the previous value, so some implicit qualifiers are omitted for brevity:

• AI-related share of capex: Some of Meta’s capex goes to traditional non-AI data centers and other assets. Before the “AI boom” starting 2022-2023, their capex was approximately $20-30 billion a year. So AI-related capex may have been approximately $50 billion or approximately 70% of the total.
• Nvidia platform share (share of AI capex going to Nvidia-based data centers): Likely around 80-90%. This excludes data centers powered by Meta’s custom chips (MTIA) and AMD GPUs (as we discuss in the AMD methodology, Meta is reportedly the largest customer of AMD AI chips).
• Share of capex that goes to IT equipment: This is probably roughly 70%, but a range of sources report 65-80%.²³ This also depends on the hyperscalers’ business model/accounting methods, i.e. whether or not they own the data center facilities in addition to the IT equipment (or vice versa).
• Share of IT spending that accrues to Nvidia as revenue: This is likely a large but not overwhelming majority, e.g. approximately 75% to 85%. Nvidia sells the GPUs and most of the networking (at least for Blackwell) that goes into an Nvidia-based data center. However some of the value is captured by server OEMs that sell server racks directly to customers, by other networking vendors like Arista and Broadcom, by memory companies who sell memory/storage (distinct from the HBM inside GPUs), and others. Some bill-of-material analyses point to Nvidia servers being 75-80% of the capital cost of a GPU cluster, but this needs to be adjusted in two directions because not all server costs accrue to Nvidia, but Nvidia also sells cluster-level networking.
• A “timing adjustment” to capture the fact that non-IT data center capex (e.g. building construction) happens in advance of IT equipment purchases. If Meta is rapidly scaling its AI data centers, non-IT expenses will be overrepresented in any given year.

The diagram below illustrates these using several point estimates. We also build a simple Monte Carlo model based on these parameter values, yielding a point estimate that Meta purchased around 11% of Nvidia’s data center equipment in 2025 [CI: 8% to 14%]. For Microsoft, this share is 18% [CI: 15% to 22%]. This corroborates the information from analysts, but the uncertainty is quite high, and our parameter value choices are not necessarily well-grounded. So, for now this model should be viewed more as rough, suggestive evidence that corroborates evidence from analysts, and provides a directional guide to how Microsoft and Meta’s shares of Nvidia revenue might have changed when analyst estimates become outdated.

More on the timing adjustment

While we won’t model the timing adjustment in full detail, this is largely a function of growth rate in capex and data center construction timelines. For illustration, imagine a hyperscaler with the following properties:

• They grow their AI capex by 70%/year for an extended period (this is roughly true of Meta’s overall capex from 2024-2026; Meta forecasts capex of $115 billion to $135 billion in 2026, the midpoint of $125 billion is a 70% increase from $72 billion in 2025, which was up approximately 85% from 2024).
• Capital spending on a data center takes place over a two-year span, with construction spending happening in year 1 and IT spending in year 2.
• 80% of AI capex goes to IT equipment.

Then for every $100 worth of data centers (measured in terms of total lifecycle cost) the hyperscaler constructs in year 1, it constructs $170 worth of data centers in year 2. In year 2, they spend $80 on IT equipment for the data centers constructed the previous year, and 20% * $170 = $34 on construction for data centers that will be finished in year 3. This means that in year 2, the hyperscaler incurred $80 + $34 = $114 in capex, and the IT share of capital expenses is ($80 / $80 + $34) = 70.1%. The baseline lifecycle IT share was 80%, so the timing adjustment is 70.1/80 = 0.876x.

Some more illustrative values can be found below, courtesy of Claude Opus 4.6. Lower lifecycle IT share, longer construction timelines, and higher capex growth rates demand higher timing adjustments, though most plausible parameter values land in the range of a 70% to 90% timing multiplier.

As with the IT share of capex, the non-IT timing adjustment will also vary with a corporation’s accounting standards and business model. Oracle in particular states that “the vast majority of our CapEx investments are for revenue-generating equipment that is going into our data centers and not for land, buildings or power that collectively are covered via leases. Oracle does not pay for these leases until the completed data centers… are delivered to us.”

Cumulative ownership by chip type

Cumulative estimates of AI chip ownership over time, broken down by owner and chip type.

Cumulative ownership by designer

Cumulative estimates of AI chip ownership over time, aggregated by chip designer/manufacturer. Includes power estimates.

Quarterly acquisitions by chip type

Estimates of AI chip acquisitions per quarter, broken down by owner and chip type.