Anthropic's $517B Compute Total Is Not an NVIDIA Contract

The $517 billion headline combines different kinds of compute#

NVIDIA's September 28 capital-return announcement is straightforward: its board added $150 billion to the existing repurchase authorization, leaving $235 billion authorized, and the company said it expects to execute the remaining program through fiscal 2028. The Anthropic headline is less straightforward. Reporting cited by Investing.com says Anthropic accumulated about $517 billion of compute commitments covering 14.8 gigawatts in the 11 months through August, versus an earlier investor disclosure of roughly $180 billion through 2029. That $517 billion figure is not an NVIDIA contract. Anthropic is deliberately spreading compute across several architectures and providers. Some contracts are clearly linked to NVIDIA systems. Others are explicitly tied to Amazon Trainium, Google TPUs, or CPU workloads. For NVDA options research, that distinction matters more than the aggregate headline. The useful question is not whether Anthropic is spending an enormous amount on compute. It is how much of that spending can be economically traced to NVIDIA, over what time horizon, and whether the options surface can isolate that channel from NVIDIA's separate buyback announcement and the many other events embedded in longer-dated volatility.

The cleanest way to test the Anthropic-to-NVIDIA transmission is to decompose the contracts before looking at NVDA options:

This decomposition changes the meaning of the headline. The reported $517 billion total is evidence of Anthropic's extraordinary infrastructure appetite, but it is not a clean measure of NVIDIA-addressable demand. Even the NVIDIA-linked contracts contain power, facilities, networking, cloud services, financing and other infrastructure economics that cannot be assigned wholesale to NVIDIA.

The opposite mistake would be to dismiss the Anthropic buildout because the total is mixed. The Microsoft partnership explicitly puts Claude workloads on NVIDIA-powered Azure infrastructure and describes an initial NVIDIA compute commitment of up to one gigawatt. Reuters reported that Nscale plans to deploy Vera Rubin systems for Anthropic, while Lambda's reported Texas project adds another NVIDIA-linked capacity path. The transmission exists. What the evidence does not support is converting every dollar of Anthropic's aggregate compute commitments into NVIDIA revenue.

• Amazon: More than $100 billion over ten years across Graviton and Trainium2 through Trainium4, representing large Anthropic compute demand that does not map directly to NVIDIA GPUs.
• Google and Broadcom: Multiple gigawatts beginning in 2027 on next-generation Google TPUs, establishing another major non-NVIDIA accelerator path.
• Microsoft Azure: $30 billion of Azure compute plus up to 1 GW additional capacity powered by NVIDIA Grace Blackwell and Vera Rubin systems, establishing a documented NVIDIA-linked compute channel where cloud contract value is distinct from direct NVIDIA revenue.
• Nscale: Reported $45 billion cloud-compute agreement deploying NVIDIA Vera Rubin systems inside a broader data-center service contract.
• Lambda: Reported $35 billion cloud-computing agreement at a roughly 350 MW Texas project, adding an NVIDIA-related infrastructure path without sufficient disclosure to equate contract value with chip sales.
• Akamai: $11.6 billion over seven years for CPU workloads on Akamai Cloud, demonstrating that compute commitments extend beyond accelerator spending.

NVIDIA's $235 billion authorization runs on a different clock#

The September 28 buyback announcement is a direct NVIDIA corporate action rather than a second-order customer-demand story. The board increased the authorization by $150 billion, leaving $235 billion authorized, with expected execution through fiscal 2028.

That does not mean $235 billion is being repurchased immediately. NVIDIA's July 26 Form 10-Q illustrates the difference between authorization and execution. The company repurchased $19.7 billion of stock during the second quarter of fiscal 2027 and $39.8 billion during the first half, while $99.3 billion remained authorized at July 26. NVIDIA also states that repurchases can depend on market conditions, operating requirements and other investment opportunities, and that the program can be suspended.

For options analysis, the distinction is important. An authorization changes the capital-allocation envelope. Actual repurchase timing determines when cash is deployed. Neither fact, by itself, establishes a price floor, support level, dealer positioning or a directional options conclusion.

The buyback and Anthropic stories therefore reach NVDA through different mechanisms. The Anthropic buildout concerns the multi-year demand and infrastructure path for accelerated computing. The buyback concerns NVIDIA's use of internally generated capital over a fiscal-year horizon. Combining them into one "positive NVIDIA catalyst" would discard the timing information that options are supposed to help organize.

The visible NVDA volatility curve cannot be labeled an Anthropic premium#

A September 28 research snapshot from Perspicium still used an NVDA reference price near $225.02 and showed implied volatility rising from 27.01% for the September 30 expiration to 29.86% for October 16 and 36.02% for November 20. The January 15, 2027 expiration was at 35.64%, while December 17, 2027 was at 38.92% and January 21, 2028 at 38.77%.

Those values are useful as a baseline, not as proof of a post-announcement repricing. The accessible page during this research run did not provide a clean standardized before-and-after snapshot that isolates the September 28 buyback announcement. Longer expirations also contain many additional NVIDIA, macroeconomic and AI-infrastructure events. A higher long-dated IV therefore cannot be attributed to Anthropic's compute commitments merely because Anthropic's contracts are long dated.

The same caution applies to the vendor's expected-move figures. Perspicium derives its expected ranges from at-the-money options and presents them as one-standard-deviation windows. That produces a consistent way to compare expirations inside the same source, but the result is a priced distribution window rather than a forecast of where NVDA will trade.

There is also a basic horizon problem. Some Anthropic agreements extend for seven or ten years, while currently listed NVDA expirations cover only a portion of that infrastructure cycle. Even an expiration inside NVIDIA's fiscal-2028 buyback window still combines the repurchase program with earnings, product cycles, customer spending, regulation, macro conditions and future competitive developments. The option cannot identify which fraction of its implied variance belongs to one long-lived customer relationship.

The more informative test starts after the headline#

The next useful observation is a standardized NVDA options snapshot after the market has had time to incorporate the September 28 capital-return announcement. The comparison should use the same data source and the same IV methodology across expirations.

If near expirations change materially while longer maturities remain comparatively stable, that would be consistent with the market concentrating more uncertainty around the immediate corporate-action window. It would not prove the buyback caused the change. If the entire curve reprices, the interpretation becomes harder because the same session may contain broader market, semiconductor or AI-demand information.

A second test belongs outside the options chain: continue decomposing Anthropic's compute commitments as providers disclose hardware, capacity and contract economics. A new multibillion-dollar Anthropic agreement only becomes strong NVIDIA evidence when its transmission path identifies NVIDIA systems or economically defensible NVIDIA content. An AWS Trainium expansion, a Google TPU agreement and an NVIDIA Vera Rubin deployment should not be treated as interchangeable simply because all three increase Anthropic's compute capacity.

That is the central research boundary. Anthropic's reported $517 billion compute stack is large enough to matter to the AI infrastructure cycle, but its headline value is not an NVIDIA order book. NVIDIA's $235 billion repurchase authorization is directly attributable to NVIDIA, but it is an authorization executed over time rather than a same-day cash deployment. NVDA options can help compare the uncertainty attached to different horizons, but they cannot convert those two facts into one clean catalyst without a before-and-after surface and a more precise map of Anthropic's NVIDIA-linked capacity.

What would change this interpretation?#

A stronger conclusion would require two kinds of evidence. First, a consistent post-announcement NVDA options surface could show whether the September 28 event produced a measurable change in near-term or long-dated implied volatility relative to the prior-session baseline. Second, provider disclosures could reveal how much of Anthropic's future capacity actually uses NVIDIA architectures and when that capacity is scheduled to come online.

Until those observations exist, the defensible research opportunity is the separation itself: aggregate AI compute commitments, NVIDIA-addressable infrastructure demand and NVIDIA's own capital-return program are three different quantities with three different clocks. Keeping them separate is more informative than treating the largest headline number as the options thesis.