Nvidia Anthropic $180 Billion Contracted Value Explained

In late September 2026, Nvidia put a number on how much computing one customer has lined up. In a September 2026 investor presentation, the company said that Anthropic's total reported contracted value across multiple cloud providers and neoclouds now exceeds $180 billion, and that Anthropic has contracted 2.6 gigawatts (GW) of Nvidia AI infrastructure to be shipped through 2028. The Nvidia Anthropic $180 billion figure has been repeated widely, often as if it were a single contract signed with Nvidia. The source slide says something narrower, and the difference shapes how the number should be read.

$180B+Anthropic contracted valueacross multiple CSPs and neoclouds, per Nvidia

2.6 GWNvidia AI infrastructurecontracted to date, shipped through 2028

$40BNvidia revenue opportunity per GWRubin reference design

$150BBuyback increaselifts authorization to $235B

NVIDIA

Anthropic

Nvidia disclosed the figures; Anthropic is the customer whose contracts they describe.

What Nvidia actually disclosed

The disclosure appears in Nvidia's September 2026 Non-Deal Roadshow deck, in a list of developments since its second-quarter earnings announcement. Two items sit next to each other: Anthropic "contracts 2.6 GW to date in NVIDIA AI infrastructure to be shipped through 2028," and Anthropic's "total reported contracted value across multiple CSPs/neoclouds now exceeds $180B." The same list mentions Nvidia's neocloud partners in Australia delivering up to 2 GW by 2027.

Three details are easy to miss:

It is a presentation, not a contract. Coverage from Yahoo Finance notes the figure came from the roadshow presentation, not a formal SEC filing or an earnings call transcript. Nvidia did not publish a contract document.
The $180 billion is not Nvidia revenue. It is Anthropic's total contracted value with the cloud providers and neoclouds that, in turn, buy Nvidia systems. Several analyses stress that the headline is distinct from revenue flowing directly to Nvidia.
The word "reported." The slide describes contracted value as "reported," meaning it aggregates deals already disclosed by Anthropic and its providers rather than a new audited total.

One discrepancy is worth flagging. Several secondary outlets, including TradingView's Seeking Alpha feed, report the capacity as 2.5 GW. Nvidia's own slide says 2.6 GW, and this article uses the primary figure.

What sits inside the $180 billion

The number is a stack of separate agreements. Metir's earlier coverage of individual deals gives a partial picture, and a few of them can be added up from public reporting:

A roughly $35 billion, six-year agreement with Lambda for about 350 megawatts in Texas, covered in Anthropic's $35 billion Lambda deal.
A $45 billion, six-year agreement with Nscale for roughly 460 megawatts of Nvidia Vera Rubin capacity in West Virginia, covered in the Nscale deal analysis.
A $10 billion agreement with Volta, covered in Reading the Volta deal.
A reported $30 billion Microsoft Azure commitment from November 2025, per remio.

The Lambda and Nscale deals alone come to about $80 billion and about 810 megawatts, which is under a third of the 2.6 GW Nvidia cited. That is simple arithmetic on reported figures, not something Nvidia stated, and it shows the remainder of the capacity sits with other providers and other contracts. Not every contract in the $180 billion necessarily involves Nvidia hardware either: Anthropic also buys capacity on other chip families, and the 2.6 GW figure is specifically the Nvidia portion.

Nvidia's Endeavor building in Santa Clara, California. The photo is illustrative of the company that made the disclosure, not of any Anthropic capacity. Photo: Coolcaesar, CC BY-SA 4.0.

Why gigawatts are the more telling unit

Dollar figures in AI infrastructure are hard to compare, because they mix lease terms, equipment, power and financing. Gigawatts of contracted capacity are closer to a physical quantity. Nvidia's deck also introduces a conversion: its revenue opportunity per gigawatt, based on its reference design, has risen from $18 billion in the Hopper generation to $25 billion with Blackwell and $40 billion with Rubin.

Nvidia revenue opportunity per gigawatt, by generation

Billions of dollars per gigawatt of AI capacity, based on Nvidia's reference design. Source: Nvidia Non-Deal Roadshow presentation, September 2026.

The deck attributes the increase to advances in the GPU and to a broader portfolio that now includes standalone CPU racks, LPU racks and networking. On that basis, 2.6 GW multiplied by the Rubin figure would be $104 billion, but that is an upper-bound illustration only. The 2.6 GW will be delivered across generations through 2028, the per-gigawatt number is a reference-design estimate, and Nvidia has not said how much of Anthropic's capacity is Rubin or what share comes with full-stack systems. Tech Times makes a related point: customers that pair Rubin GPUs with third-party networking would generate less Nvidia revenue per gigawatt.

The headline is a stack of agreements across many providers, not one purchase order signed with Nvidia.

The concentration question

The disclosure arrived alongside Nvidia raising its buyback authorization by $150 billion to $235 billion, as covered in Nvidia's $235 billion buyback. The two are separate decisions, though both speak to how much cash the AI buildout is generating and committing.

Analysts have raised a structural question. Nvidia has also been an investor in Anthropic, with a reported commitment of up to $10 billion announced in November 2025. When a supplier invests in a customer, and that customer's contracts with intermediaries then feed the supplier's order book, the relationships become interlinked. This is sometimes called circular financing. It does not by itself indicate a problem, but it means utilization shortfalls at one point could travel through the chain: Anthropic's demand supports neocloud leases, which support neocloud financing, which supports Nvidia's shipments.

The commitments also assume sustained usage. Contracted value is a promise to pay over years, typically six in the Lambda and Nscale cases, so what Anthropic's revenue trajectory looks like through 2028 matters as much as the signed total.

What to watch

Which providers fill the remaining capacity. Only part of the 2.6 GW is accounted for by the deals disclosed so far.
Delivery timing. Contracted capacity is not energized capacity. Power, permitting and rack supply determine when each tranche becomes usable.
Neocloud listings. Nvidia's deck notes that Nscale and Firmus have filed for IPOs, which would put more of this contract structure under public-market disclosure.
Primary documents. A future Nvidia filing or Anthropic's own disclosures would replace the secondary reporting around this figure.
Generation mix. The ratio of Hopper, Blackwell and Rubin systems in the 2.6 GW sets how the per-gigawatt economics play out.

For teams building on frontier models, the practical lesson is about supply. Capacity on this scale is being secured years ahead, and a single lab's access to compute can change quickly. Using a model-agnostic workspace such as Metir, which offers models from several labs in one place, keeps your work portable if availability or pricing shifts on any one provider.

Sources:

Image credits

Nvidia Endeavor building, Santa Clara: "NVIDIA Headquarters.jpg" by Coolcaesar, Wikimedia Commons, licensed CC BY-SA 4.0.