As Anthropic moves toward an IPO, attention is turning to its AI infrastructure payment obligations, which total at least $518 billion over the next 10 years or so.
According to Reuters' September 29 report, which cites a confidential IPO filing, about 80% of that amount comes from contracts that cannot be canceled or that require fixed payments even if actual usage falls below the contracted amount.
While the IPO valuation has been reported at about $2 trillion, Anthropic is committing huge sums up front, in anticipation of future demand, to secure the computing resources it needs to develop and serve Claude.
Adding more sources of computing resources is an effective way to spread supply risk. But that does not necessarily mean spending can be flexibly reduced if demand growth slows.
About 80% of payment obligations remain even if usage declines
Looking at the major contracts Reuters reported through the lens of how far payments can be reduced if usage drops, four large contracts stand out.
| Counterparty | Reported payment amount | Terms if usage falls short, etc. |
|---|---|---|
| At least $111.1 billion | Anthropic pays the difference if it does not reach the contracted usage amount | |
| Amazon | At least $110 billion | As with Google, a fixed amount must be paid even if usage is insufficient |
| Microsoft | $31.4 billion | Generally non-cancelable, except where Microsoft's material breach goes uncured |
| Broadcom-related equipment leases | About $161.2 billion | Neither party can cancel, except in cases such as default |
Adding up the payment obligations to Google, Amazon, Microsoft, and Broadcom-related parties gives $413.7 billion, equal to about 79.9% of the reported $518 billion total.
The calculation is (111.1 + 110 + 31.4 + 161.2) ÷ 518 × 100, a rough estimate using the figures in the same Reuters report.
However, some figures are stated as minimums ("at least"), and contract terms differ. The $413.7 billion therefore does not represent the amount payable in any single year, nor current debt.
By contrast, the NVIDIA-based computing capacity Anthropic is procuring from xAI could reach up to $8.45 billion by 2029, and most of it can reportedly be canceled with 90 days' notice.
Even within the same $518 billion, the risk if demand forecasts miss differs greatly between contracts whose usage can be adjusted later and those that require fixed payments whether or not the capacity is used.
More compute arriving in 2026, with facilities coming online from 2027
In its April 20 announcement of the Amazon agreement, Anthropic said it would spend more than $100 billion on AWS-related technologies over the next 10 years and secure up to 5GW of computing capacity.
The plan was to have roughly 1GW of Trainium2 and Trainium3 capacity combined running by the end of 2026.
Anthropic also acknowledged at the time that a surge in users had affected service stability and performance during peak hours. Part of the background to these long-term contracts is that computing resources are currently short relative to demand.
There are already examples of added capacity flowing through to users.
In its May 6 announcement of the SpaceX partnership, the company said it would use Colossus 1's computing capacity to gain access to more than 300MW and over 220,000 NVIDIA GPUs within that month.
On the same day, it doubled the five-hour usage limits for Claude Code on eligible plans and removed the reduction of usage limits during peak hours for Claude Code on Pro and Max.
Anthropic said the new agreements and other infrastructure expansion made it possible to raise usage limits.
Meanwhile, many facilities will not actually go into operation until 2027 or later.
The April 6 announcement with Google and Broadcom laid out a plan to bring multiple gigawatts of next-generation TPU computing capacity online starting in 2027.
The partnership AMD announced on July 22 also plans to deploy up to 2GW of Instinct MI450 series GPUs, with the first 1GW to begin deployment in the first half of 2027.
In other words, the timelines differ between computing resources meant to handle the 2026 surge in demand and facilities secured in anticipation of further growth from 2027 onward.
The AMD collaboration also includes efforts to use Claude to optimize workloads on GPUs and accelerate the development of ROCm software.
Adding computing sources requires not only securing chips but also optimizing models to run efficiently on each type of hardware.
Anthropic says it will use Trainium, TPUs, NVIDIA GPUs, and others in different roles precisely so that it can choose the hardware best suited to each workload.
Note that GW is a measure of a facility's power scale and is not a figure that allows direct comparison of processing performance across different AI chips.
Having multiple platforms to draw on gives Anthropic more options if one supplier suffers outages or facility delays.
But being able to move workloads to a different chip is a separate matter from being able to transfer payment obligations under an already-signed contract to another provider.
The more suppliers there are, the more profitability depends on how much of the capacity secured under each contract is actually used, and how much of that use translates into revenue.
Cloud companies are investors, sellers, and competitors at once
Amazon's April partnership explanation shows why partnerships with major cloud providers matter on the sales side as well.
The plan is to make Claude Platform available from existing AWS accounts, removing the need to manage additional credentials, contracts, or separate billing relationships.
Enterprises can adopt Claude while keeping the access management and monitoring systems they already use. Being able to tap existing sales channels, not just computing infrastructure, is an advantage that can speed adoption.
At the same time, dependence on these channels is growing.
According to a separate Reuters exclusive, the share of revenue coming through Amazon and Google rose from 11% in 2023 and 32% in 2024 to 47% in 2025.
Both companies are sales partners that bring Claude to customers and also suppliers that provide computing resources to Anthropic.
They also invest in Anthropic while developing their own AI models and AI services.
For Anthropic, partnerships with major cloud providers allow it to advance customer acquisition, fundraising, and compute procurement at the same time, but they also create a structure in which several critical roles depend on the same counterparties.
There is also a difference in contract flexibility between the revenue side and the spending side.
According to Reuters, 2025 usage-based revenue was about $3.8 billion, and subscription revenue was about $789 million.
Many large customers are reportedly not bound by strong long-term contracts and can reduce their usage or stop using the service.
If customers cut usage, Anthropic's revenue falls. But the minimum payments committed on the infrastructure side will not necessarily fall in the same proportion.
While demand is growing rapidly, a strategy of securing future computing resources in advance works well.
If growth slows, however, the gap between usage-linked revenue and payment obligations fixed over a long period widens.
Public inquiries have long pointed out that partnerships with cloud companies carry this kind of tension.
The U.S. Federal Trade Commission's market study published in January 2025 addressed arrangements in which AI developers pay a large portion of their funding back to the cloud providers that invest in them, as well as the technical and contractual burdens of switching to another cloud platform.
It also noted the possible competitive effects of sharing sensitive information such as pricing and customer usage.
The FTC examined information through September 2024 and public information through January 2025, and it did not rule that the Anthropic contracts now being reported are illegal.
Even so, it offers a reminder that a wider choice of AI chips does not make it easy to unwind relationships with cloud companies that span financing, sales, and compute supply.
The pattern in which major companies supporting Anthropic's growth also pursue growth in their own AI and cloud businesses is likely to continue.
What's in the $65 billion raise, and the payments still to come
In the Series H announced on May 28, Anthropic raised $65 billion at a post-money valuation of $965 billion.
The company said annualized revenue that month exceeded $47 billion.
That figure, however, assumes the most recent revenue pace continues for a full year and differs from revenue actually recognized over a year.
The roughly $2 trillion IPO valuation is likewise a target reported by Reuters; at the stage of the confidential filing announced on June 1, the number of shares to be offered and the public offering price had not been decided.
The breakdown of the $65 billion raise also deserves care.
Series H includes $15 billion previously committed by major cloud companies and others, of which $5 billion is Amazon's investment.
Adding the $5 billion Amazon investment announced in April on top of the $65 billion announced in May would therefore double count it.
The additional investment of up to $20 billion that Amazon has separately indicated is also a future investment contingent on meeting certain commercial conditions, and cannot be treated as funds already received in full.
The same Series H announcement also named Micron, Samsung, and SK hynix as strategic infrastructure partners.
Expanding AI computing power requires a large supply of memory and storage in addition to compute chips.
However, the announcement did not give the value or quantity of products to be purchased from each company. Being named as a partner cannot be taken as confirmed order value.
From 2027 onward, when many of the new facilities begin operating, it will be necessary to look at how much actual demand can fill the capacity secured under contract, alongside the payment schedules.
How much of the cost of training models, and of running inference in response to customer requests, can be covered by revenue from service fees will also matter.
Using up the secured computing resources is not enough.
Only if Anthropic can build a revenue structure that leaves a profit as usage grows will the huge fundraising and long-term infrastructure contracts function as investments that support its next stage of growth.