Anthropic’s $900.9B Compute Bet, Explained | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Anthropic’s $900.9B Compute Bet, Explained
Anthropic Direct could lower the cost of serving Claude, but its long-term leases and purchase commitments bring payment risks before all the capacity arrives.
Anthropic has disclosed $900.9 billion in long-term compute commitments through 2045, shifting from bundled cloud services to its Anthropic Direct strategy. This hybrid model involves separately securing data-center leases, TPU-based accelerator equipment, and specialized operational services. While this approach aims to lower unit costs and improve infrastructure control, it introduces significant financial risks. Anthropic must now synchronize complex, interdependent schedules for power, facility construction, and hardware delivery. If utilization rates fail to match these fixed, long-term payment obligations, the company faces substantial liquidity pressure before this capacity becomes productive.
Anthropic has disclosed approximately $900.9 billion in aggregate purchase commitments and lease agreements, substantially all tied to compute and infrastructure. The figure appears in its November 10, 2026 IPO materials and describes obligations stretching across years, with most of the later payments extending through 2045. It is not a bill payable immediately or a measure of data centers already built. (SEC-filed materials)
The strategy behind that number is consequential. Through Anthropic Direct, the Claude developer is moving beyond buying bundled cloud capacity toward separately securing data-center space, accelerator equipment and operating services. TPU-based capacity supplied through its relationships with Google and Broadcom forms an important part of that approach.
The economic bet is straightforward: commit early to large amounts of infrastructure, gain more control over the supply, and aim for better long-term costs. The financial challenge is harder. Payments, equipment deliveries, available electricity and paying customer demand must line up closely enough to make those commitments productive.
The $900.9 Billion Has Several Different Clocks
Anthropic’s filing gives three useful time horizons for understanding the headline total.
Period
Disclosed amount
How to read it
Twelve months after September 30, 2026
About $70.7 billion
The nearest payment window, running through September 2027
Through the end of 2030
About $451.6 billion
Roughly half the total, including the near-term amount above
Primarily 2031–2045
Most of the remaining approximately $449.3 billion
A long tail of obligations extending well beyond the initial buildout
The approximately $449.3 billion remainder is calculated by subtracting $451.6 billion from $900.9 billion. The first two rows overlap, so adding all three rows would double-count the near-term obligations. (SEC-filed materials)
That structure changes the interpretation of the headline. About 7.8% of the aggregate total falls within the first twelve-month window, while roughly half is expected through 2030. Neither percentage tells readers whether Anthropic can comfortably fund the payments. That requires cash balances, operating cash flow, financing availability and the detailed contractual schedule.
I’d treat the $70.7 billion near-term figure as the more immediate liquidity checkpoint. The $900.9 billion total captures the scale of the long-term bet, but the payment schedule determines when that bet needs funding.
The accounting perimeter matters, too. Mission Critical Research’s November 11 analysis of the amended filing distinguishes the broad commitments total, covering arrangements entered into as of and after September 30, from a contractual-obligations table totaling $320.0 billion at that date. It also identifies additional arrangements without fixed dollar commitments, with estimated payments of $150 billion to $175 billion as capacity is delivered. These figures describe different populations and timing; they should not be treated as interchangeable measures of debt or added together casually. (Mission Critical Research)
Anthropic Direct Separates the Building From the Chips
Anthropic Direct is a hybrid infrastructure model. Anthropic continues buying cloud compute while also arranging dedicated facilities and sourcing the accelerator infrastructure installed inside them.
The prospectus describes a separation between commitments for data-center capacity and commitments for the equipment that turns that capacity into usable compute. That distinction is central to understanding both the potential savings and the risks. (Anthropic prospectus)
The model has three practical layers:
Facilities and power: Partners can secure land, buildings and electricity, with Anthropic leasing capacity. Anthropic also says it has signed data-center leases itself.
Accelerator equipment: Anthropic obtains specialized chips and associated components through arrangements including master equipment leases.
Operations: Specialist companies develop and operate facilities. The filing identifies Fluidstack as one example of a partner providing colocation and services.
Mission Critical Research’s breakdown says the dedicated facilities are built to Anthropic’s specifications, and that Anthropic expects Direct to account for a growing share of its compute investment. (Mission Critical Research)
“Direct” therefore does not mean Anthropic owns every building, generator and chip outright. It means the company takes a more direct role in assembling those pieces instead of relying entirely on a cloud provider’s bundled service.
That can improve control over configuration, procurement and capacity planning. It also leaves Anthropic coordinating several interdependent schedules. A completed building cannot serve Claude without equipment and power; delivered accelerators cannot earn revenue while waiting for a usable facility.
My view is that Direct changes where the infrastructure risk sits. A cloud supplier normally packages much of that coordination into its service. Separating the components gives Anthropic more commercial choices, while exposing it more directly to mismatches between them.
TPU Capacity Changes the Supplier Relationship
TPUs, Google’s specialized AI accelerators, are a significant component of the Direct strategy. Mission Critical Research reports that the amended filing describes five-year master equipment leases with Broadcom, portions of which have been assigned to financing entities. Those leases provide TPU-based capacity, with some capacity online in 2026 and more expected from 2027. (Mission Critical Research)
The reported rent over those leases is approximately $161.2 billion before interest. Payments are quarterly and generally begin when the equipment can go into service. That is a different financial profile from paying only for short-term cloud usage, and the before-interest qualification matters when considering the eventual financing cost.
The same filing analysis describes a facility allowing Anthropic to issue up to $42 billion in convertible notes to Broadcom or another designated financial partner, specifically to fund quarterly rent under one lease arrangement. That provides a financing mechanism, not a reduction in the underlying cost.
A TPU-heavy buildout also changes dependency rather than eliminating it. Anthropic may reduce its reliance on a cloud provider’s allocation and pricing decisions while maintaining substantial dependencies on chip suppliers, equipment financing, facility operators and electricity providers.
Anthropic still uses multiple accelerator families: AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. Fortune’s October 26 reporting describes a proprietary orchestration layer intended to place workloads across different chips and cloud providers. Such software matters because access to several hardware families becomes more useful when workloads can actually be assigned across them. (Fortune)
The disclosed arrangements explain the direction of the TPU strategy. They do not supply an independent, like-for-like measurement of Direct’s cost per token against rented cloud capacity.
The Savings Depend on Keeping Capacity Useful
Anthropic has a clear reason to seek alternatives to expensive short-term supply. According to Mission Critical Research’s account of the filing, the company took delivery of $5 billion of short-term compute in the quarter ended September 30, 2026. Anthropic says that capacity currently costs substantially more than compute secured through its multiyear agreements. That comparison is the company’s assessment, not an independent benchmark. (Mission Critical Research)
Longer commitments can offer suppliers more predictable demand, supporting better commercial terms. Dedicated infrastructure can also give Anthropic greater control over how capacity is configured and used.
The tradeoff is utilization. A favorable contracted price is valuable when equipment is online and doing useful work. If demand arrives later than expected, or a facility cannot operate at its intended capacity, fixed payments are spread across less output.
For readers evaluating the strategy, I’d distinguish two questions:
Is committed capacity cheaper to obtain? Anthropic says its multiyear supply is cheaper than current short-term alternatives.
Will that capacity be cheaper per unit of useful work delivered? That also depends on deployment timing, utilization, operating costs and workload performance.
The second question is the more important economic test.
The contract durations add another planning challenge. The disclosed accelerator leases run for five years, while the aggregate obligations extend much further. That does not establish a contractual problem, but it means a long-lived infrastructure strategy must accommodate changing equipment generations and model requirements.
Conditional Commitments Still Require Careful Reading
Approximately $186.2 billion of the total sits in arrangements with exercisable termination rights or conditions that must be satisfied before they proceed, according to Mission Critical Research’s filing analysis. Examples include counterparties obtaining financing, facilities or regulatory approvals. (Mission Critical Research)
That is about one-fifth of the headline total, but it should not be described as money Anthropic can freely cancel. The relevant rights and conditions vary by contract. A project depending on a partner securing financing is different from an unrestricted option to walk away.
Conditional arrangements can limit exposure if a project fails to meet its requirements. They can also leave future capacity uncertain. If a planned facility falls through, Anthropic may avoid some obligations while still needing replacement compute.
Power introduces a similar two-sided risk. The filing describes separate utility arrangements and, in some cases, behind-the-meter generation, which supplies a facility directly instead of relying entirely on electricity delivered through the wider grid. Securing either route can require long lead times and substantial spending before a site opens. (Mission Critical Research)
Fortune’s reporting on the prospectus also identifies permitting, legal challenges and supply-chain disruption as potential obstacles to power connections and deployment. Anthropic warns of significant cash outflows before it receives the expected benefits. (Fortune)
What Investors Should Watch as Direct Expands
Anthropic Direct could strengthen the company’s negotiating position in the AI industry. A credible alternative source of dedicated capacity gives Anthropic more options when cloud supply is scarce or expensive. Large commitments can also give equipment suppliers and infrastructure partners a strong reason to support its expansion.
Cloud providers remain important, however. They supply flexibility, additional capacity and access to different hardware. Direct appears designed to change the balance of that relationship, with Anthropic assembling more of its own supply chain.
For the IPO, the most useful milestones will be operational and financial:
How much committed capacity enters service on schedule.
When lease payments begin relative to usable deployment.
How much expensive short-term compute remains necessary.
Whether customer demand keeps the new capacity productive.
Whether realized savings compensate for financing and infrastructure costs.
The strongest evidence for the strategy would be lower delivered compute costs alongside manageable cash requirements. A rising commitment total alone cannot show that. Anthropic’s central execution task is turning separately contracted buildings, power and equipment into productive capacity before the payment schedule gets too far ahead of the business.
Frequently Asked Questions
3 questions
1
Has Anthropic Already Spent $900.9 Billion?
No. The figure aggregates purchase commitments and lease agreements across many years. About $70.7 billion falls due in the twelve months after September 30, 2026.
2
Does Anthropic Direct Replace Its Cloud Providers?
No. Anthropic describes a hybrid approach combining cloud services with dedicated facilities and separately sourced accelerator equipment.
3
Can Anthropic Cancel All Its Conditional Commitments?
The disclosed $186.2 billion includes arrangements with termination rights or conditions. Whether an agreement can end depends on its specific terms; the entire amount is not an unrestricted cancellation option.