Anthropic’s $895.4 billion commitment to compute infrastructure expansion

Original author: Su Yang

Original editor: Xu Qingyang

Source: Tencent Tech

AI companies are becoming major buyers of infrastructure.

Anthropic’s expansion path has extended from training models and competing for enterprise customers to building data centers, procuring chips, and coordinating power resources. To secure sufficient computing power, the company has begun directly engaging in tasks that were previously primarily handled by cloud service providers and infrastructure operators.

On November 2 in the United States, Anthropic filed an updated prospectus with the U.S. Securities and Exchange Commission. The document added financial data for the third quarter of 2026 and disclosed long-term procurement and leasing arrangements.

Documents show that computing power has become the most significant cost item in Anthropic’s expansion. To meet the demands of training the Claude model and serving enterprise clients, the company has not only entered into long-term agreements with cloud providers such as Amazon, Google, and Microsoft, but has also begun directly purchasing chips, participating in data center construction, and sourcing power and facility resources from former cryptocurrency mining companies and new cloud service providers.

The updated prospectus shows that Anthropic's revenue for the third quarter was $16.96 billion, higher than the $11.56 billion in the second quarter and significantly above the approximately $1.38 billion in the same period of 2025.

In the first three quarters of this year, Anthropic's cumulative revenue reached $33.242 billion, compared to $2.545 billion in the same period last year, representing an approximately 12-fold increase.

However, revenue growth did not lead to a corresponding improvement in operating profit.

In the third quarter, Anthropic's operating expenses rose to $17.5 billion, up from $10.6 billion in the second quarter. The company recorded an operating loss of $567 million for the quarter, compared to an operating profit of $974 million in the second quarter. Computing and infrastructure costs reached approximately $9.9 billion, nearly doubling from about $5.6 billion in the second quarter.

Anthropic attributed its shift to a negative operating profit primarily to short-term compute procurement. Due to customer demand growth exceeding previously locked-in compute capacity, Anthropic had to purchase additional computing resources on the market at temporarily higher prices, which were significantly above the rates under multi-year agreements.

In the third quarter, Anthropic's short-term compute procurement amounted to approximately $5 billion. Additionally, a non-cash expense of approximately $2 billion related to an employee donation matching program also impacted the quarter's operating results.

After excluding equity-based compensation and the aforementioned donation matching expenses, Anthropic's adjusted operating margin for the third quarter was approximately 15%, down from 19% in the second quarter. The decline in adjusted margin reflects ongoing pressure on profit margins from increased spending on compute resources and other operating costs amid revenue expansion.

There have also been changes on the pricing side. Anthropic disclosed that revenue growth from increased customer usage from Q2 to Q3 was partially offset by price reductions on new model versions and contract discounts, impacting approximately $600 million.

Anthropic's net loss in the third quarter was $23.29 billion, compared to $66.3 billion in the second quarter. The cumulative net loss for the first three quarters reached $104.039 billion, of which approximately $94.3 billion was related to fair value changes of financial instruments such as convertible notes and derivative liabilities, primarily driven by the company's rising valuation.

The prospectus discloses that, as of the end of September, Anthropic's aggregate spending commitments arising from purchase obligations and lease agreements amounted to $895.4 billion, the vast majority of which relates to computing resources and infrastructure.

Of this amount, approximately $447.5 billion is expected to be settled by December 31, 2030, with the remainder extending through 2031 to 2045. Commitments maturing within the next 12 months amount to approximately $69.8 billion.

This amount does not equal accrued interest-bearing debt, nor does it represent Anthropic’s obligation to pay the full amount immediately. It includes various types of purchase agreements, equipment leases, and data center-related contracts, with payments spread over multiple future years. However, some contracts are highly binding, and the company may still be obligated to make payments even if actual usage falls below expectations.

From the perspective of contract composition, unstarted lease commitments amount to approximately $321.4 billion, and fixed minimum purchase commitments amount to approximately $265.6 billion. These long-term arrangements lock in future costs for computing power procurement and infrastructure usage, while also obligating the company to make payments over an extended period.

As of the end of September, Anthropic held a total of $131.845 billion in cash, cash equivalents, and short-term investments. When including available credit facilities, the company’s total liquidity is approximately $146.6 billion.

The scale of long-term commitments is significantly larger than the book value of funds. $895.4 billion is approximately 6.8 times Anthropic’s cash and short-term investments; the $69.8 billion in commitments maturing over the next 12 months represents about 53% of its current cash and short-term investments.

The prospectus also clearly highlights the operational risks involved. Anthropic's revenue primarily comes from usage-based consumer services, with most revenue not backed by long-term contracts, while compute and infrastructure expenses typically require multi-year commitments. The company may enter into long-term agreements based on expected revenue, but actual future revenue may not meet those expectations.

If customer revenue growth fails to cover related payments, or if the company is unable to secure further financing on acceptable terms, Anthropic may be unable to fulfill certain contractual obligations. Breaches of certain commercial agreements could also trigger cross-default provisions in credit agreements and convertible notes, leading to accelerated debt maturity or even termination of relationships with suppliers.

The cost increases have already been reflected in the financial data. For the first three quarters, Anthropic's computing and infrastructure expenses increased by 329% year-over-year, adding approximately $14.4 billion. In the third quarter, related expenses amounted to about $9.9 billion, accounting for roughly 58% of the quarter's revenue.

According to foreign media reports, by the end of last year, Anthropic's management had recognized significant uncertainty in computing power demand forecasts and subsequently signed dozens of intensive agreements for computing power and infrastructure, extending partnerships from major tech companies to data center startups, chip manufacturers, and former cryptocurrency mining enterprises.

This round of procurement is no longer limited to purchasing cloud services. Anthropic has begun hiring experts in data center construction and supply chains to directly purchase server equipment and cooling systems, helping partners accelerate data center development. It is also engaging with data center developers in multiple states to study grid connectivity, project permitting, and the potential to repurpose old coal-fired power plants and industrial facilities into data centers.

New cloud service providers and former cryptocurrency mining companies have joined Anthropic’s supply chain. External media, consolidating publicly available information, report that companies such as Fluidstack and Lambda provide computing resources, while firms like TeraWulf and Hut 8 are involved in data center construction.

These partners control land, power access resources, or existing data center facilities. For Anthropic, leveraging existing sites and power resources can reduce the time required to build large campuses from scratch. For some former mining companies, transitioning to AI data centers also offers new opportunities for business growth.

However, available space does not equate to immediate power delivery. Power supply, equipment installation, and construction personnel may all delay the project timeline.

Anthropic initially sought data center projects of at least around 200MW, later began accepting projects of approximately 50MW, and supplemented power supply through battery storage and natural gas generation. The company also considers combining multiple smaller projects to gradually increase computing capacity.

Anthropic is also directly involved in the design and procurement of infrastructure projects. The company purchases equipment such as chips and cooling systems, then hands them over to partner operators for installation, reducing the time from project construction to compute capacity going live.

This model expands the range of resources Anthropic can access, but also increases the complexity of project management. Construction timelines, equipment specifications, power conditions, and delivery schedules vary across different data centers. Even after contracts are signed, actual computing power must undergo construction, equipment installation, grid connection, and commissioning before it can be put into use.

John Tough, managing partner at energy investment firm Energize Capital, believes that Anthropic does not control a complete compute supply chain and must instead source power and infrastructure resources individually across the globe, piecing together projects of varying scales.

Anthropic refers to this dedicated infrastructure as Anthropic Direct in its prospectus. The company aims to reduce its reliance on more expensive short-term compute purchases by locking in dedicated capacity over the long term.

Foreign media, citing informed sources, revealed that Anthropic expects to have over 3.5 GW of computing capacity by 2026 and plans to double its scale by 2027.

Anthropic's long-term compute contracts are concentrated among a small number of large technology companies and infrastructure providers.

According to the prospectus and related disclosures, Anthropic’s contracts with Google amount to approximately $121.3 billion, with Amazon approximately $110 billion, and with Microsoft approximately $44.9 billion; its agreement with SpaceX is valued at up to $120.5 billion and extends through 2036. Additionally, the company has related agreements with Fluidstack and Lambda.

The commitment for Broadcom-related equipment leasing amounts to approximately $161.2 billion and is also a significant part of this compute expenditure structure.

In addition, Anthropic plans to procure an additional $125 billion to $150 billion in Google chips and expects to spend over $20 billion on AMD chips. These arrangements do not have fixed contract amounts and are therefore not included in the aforementioned $895.4 billion commitment total.

The multi-chip architecture is another arrangement by Anthropic to mitigate supply risks. According to its prospectus, the company sources its computing power from AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, and centrally orchestrates these diverse resources, allocating them based on the price and performance requirements of different tasks. This helps reduce dependence on a single chip supplier, but also makes Anthropic’s relationships with its supply chain partners more complex.

Amazon and Google are both major investors in Anthropic and also provide cloud services and chip resources. Microsoft similarly provides cloud infrastructure. Broadcom is involved in chip and equipment supply. Anthropic pays these companies for computing power while also receiving investment or funding support from some of them.

Evercore ISI analyst Mark Mahaney previously noted that Anthropic’s relationship with major cloud providers has multiple dimensions: as a customer, competitor, sales channel, and investment target.