Anthropic has signed a six-year, $35 billion deal to purchase compute capacity from Lambda, a US AI cloud provider. The Wall Street Journal (WSJ), Reuters, and AFP reported the deal based on interviews with people familiar with the matter. Beyond the sheer size of the figure, what stands out is that NVIDIA—a company that sells semiconductors—is also becoming a tenant of the data center housing the compute capacity. A credit backstop supporting facility construction and GPU procurement has been woven into what would normally be a straightforward deal between an AI model company and a cloud provider.
However, the four companies have not jointly announced the $35 billion deal, and no contract has been made public. What can be confirmed is the roughly 350MW deal reported in the press, alongside a separate disclosure from Hut 8—the company building the facility—of two 352MW triple-net leases (long-term lease agreements in which the tenant bears taxes, insurance, and maintenance costs). Overlaying the two reveals why NVIDIA ended up as a tenant of the facility, and hints at who bears which risks.
From GPU Sales to Facility Tenant
According to the WSJ, Anthropic will pay $35 billion to Lambda, a company NVIDIA has invested in, to obtain compute capacity from a data center in Nueces County, Texas. The facility will be built by Hut 8, with NVIDIA serving as the tenant. AFP reported the contract term as six years, while Reuters confirmed the capacity in question at roughly 350MW. NVIDIA also participated in Lambda's $480 million funding round in February 2025.
Based on publicly available information, the roles can be reconstructed as follows:
| Company | Layer of Responsibility | Confirmed Details |
|---|---|---|
| Hut 8 | Land, power, building | Building Beacon Point and signing long-term leases of 352MW each |
| NVIDIA | Semiconductors, facility lease | Supplies GPUs and, per the WSJ, also leases the data center |
| Lambda | GPU procurement, cloud operations | Builds GPU infrastructure and provides compute capacity to Anthropic |
| Anthropic | Long-term demand | Purchasing six years of compute capacity for training and serving Claude |
Under this division of labor, Hut 8 handles construction, Lambda handles GPU deployment, and Anthropic represents the end demand. NVIDIA sits in between, offering both its semiconductor supply capacity and its corporate creditworthiness at the same time. This makes NVIDIA's obligations extend further than a transaction that simply ends with delivering GPUs.
At the same time, it remains undisclosed exactly which facility NVIDIA is leasing, how much capacity, or how much Lambda is paying. Without knowing NVIDIA's lease liabilities, Lambda's GPU purchase amount, or Anthropic's payment terms, it cannot be concluded that "NVIDIA created its own demand to circulate revenue." While the deal structure can be confirmed, where credit risk ultimately lands remains unclear.
350MW Is Not the Full 700MW
Hut 8 announced two phases for the same Beacon Point site in Nueces County: Phase 1 in May 2026 and Phase 2 in July 2026. Each phase provides 352MW of IT capacity, and both are 15-year, $9.8 billion long-term leases. Combined, that comes to 704MW and a base contract value of $19.6 billion. Hut 8 describes the tenant only as a "highly investment-grade company," without naming it.
| Disclosed Scope | IT Capacity | Term | Amount | Source Status |
|---|---|---|---|---|
| Anthropic–Lambda deal | ~350MW | 6 years | $35 billion | Based on reporting citing people familiar with the matter |
| Beacon Point Phase 1 | 352MW | 15 years | $9.8 billion | Official Hut 8 disclosure |
| Beacon Point Phase 2 | 352MW | 15 years | $9.8 billion | Official Hut 8 disclosure |
| Beacon Point total | 704MW | 15 years | $19.6 billion | Official Hut 8 disclosure |
The roughly 350MW figure is close to half of 704MW—roughly equivalent to a single 352MW phase. It would therefore be premature to conclude that Anthropic has secured an entire 700MW-class site. Public materials alone cannot determine whether the 350MW corresponds to Phase 1 or Phase 2. Furthermore, while Beacon Point has a total power receiving capacity of 1,000MW, only 704MW has been contracted for IT equipment—power receiving capacity and computing capacity are not the same figure.
Timelines also vary. Hut 8 expects the site's first power to come online in Q1 2027, with the first data hall delivery for Phase 2 expected in Q2 2028. Meanwhile, the start date for utilizing the $35 billion deal's capacity has not been reported. Even once the contract amount is finalized, the compute won't be available to run Claude until power, cooling infrastructure, and GPUs are all in place.
Long-Term Demand Draws Financing for Buildings and GPUs
Data centers require massive upfront construction costs before any revenue arrives. Hut 8 signed 15-year leases in which the tenant bears taxes, insurance, and maintenance, using future rental income as backing for construction financing. If Anthropic's demand continues for six years, Lambda in turn finds it easier to borrow money to purchase GPUs. A long-term contract transforms buildings and semiconductors into assets that financial institutions can evaluate.
On August 12, 2026, Lambda announced a $926 million secured loan. The funds are earmarked for acquiring and installing GPU servers to serve investment-grade, long-term contract customers, with the GPUs themselves and contract revenue serving as collateral. Since the customer's name was not disclosed, it cannot be confirmed whether this is the same deal as the Anthropic contract discussed here. Even so, Lambda's business model of procuring GPUs based on customer contracts can be verified through primary sources.
At this point, one should not subtract Hut 8's $19.6 billion from the $35 billion to estimate profits for Lambda or NVIDIA. The $19.6 billion represents 15 years of facility lease payments for the full 704MW. The $35 billion, according to reporting, represents six years of cloud compute fees for roughly 350MW. The latter may include GPUs, networking, and operations, but the breakdown has not been disclosed. The term, capacity, and scope of services all differ.
The advantage for NVIDIA in becoming a facility tenant is that it can quickly link compute demand to power and construction. Conversely, if Anthropic's demand falls short of projections or construction is delayed, the semiconductor maker could also be affected by the facility contract. Until the extent of the guarantees involved is disclosed, the scale of the benefits and burdens cannot be measured.
Anthropic Has Not Consolidated Around NVIDIA
Anthropic has not concentrated its computing infrastructure with a single company. In May 2026, the company explicitly stated its policy of using a mix of AWS Trainium, Google TPUs, and NVIDIA GPUs. Amazon is a key training and cloud partner, with Anthropic committing over $100 billion over ten years to AWS technology and securing a contract for up to 5GW of capacity.
Under its agreements with Google and Broadcom, Anthropic will add 5GW of next-generation TPU capacity starting in 2027. It has also secured a contract with SpaceX for access to more than 300MW and over 220,000 NVIDIA GPUs within one month. Additionally, its partnership with Microsoft and NVIDIA includes $30 billion in Azure capacity, and its agreement with Fluidstack involves a $50 billion investment in domestic US AI infrastructure. Since the overlap between these contract amounts and capacities has not been disclosed, they cannot simply be summed into a single "total investment figure."
The Lambda deal in question adds roughly 350MW of NVIDIA-based compute capacity to this diversified strategy. While it increases Anthropic's reliance on NVIDIA, it does not represent an abandonment of Trainium or TPUs. Rather, Anthropic is stacking large-scale contracts across different chip types and cloud providers, each becoming available at different times. This reflects the fact that, beyond model performance, the ability to secure power and semiconductors when needed determines the ceiling on service capacity.
That said, the generation and quantity of GPUs that will make up the roughly 350MW remain unknown. It has also not been disclosed how the capacity will be allocated among training, inference, and Claude Code. There is no basis for directly comparing this to the "over 300MW, over 220,000 units" figure from the SpaceX deal in terms of performance.
The $35 Billion's Breakdown and Timing Will Determine Its Real Impact
The headline figure of $35 billion alone cannot tell us when this deal will resolve Anthropic's compute shortage. First, it must be confirmed whether Hut 8 can achieve first power in Q1 2027. After that, which data halls will receive which generation of GPUs, and when Lambda will deliver capacity to Anthropic—both the construction and deployment timelines need to be tracked.
Contractual responsibilities also warrant scrutiny. Anthropic's minimum purchase commitment, Lambda's GPU procurement obligations, and the rent and guarantees NVIDIA bears, along with how delays would be handled—none of these have been disclosed. If Hut 8's future filings or announcements from the four companies clarify the tenant and scope of the contract, it will become possible to determine whether the $35 billion represents long-term demand backed by physical infrastructure, or a ceiling figure subject to multiple conditions.
Only once Beacon Point comes online as scheduled, GPUs are installed across the roughly 350MW, and Anthropic continues to use the capacity, will the chain of four companies linked through NVIDIA's creditworthiness be properly evaluated—not by contract value, but by actual Claude computing capacity.