The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Anthropic’s infrastructure sprint secures access to large amounts of future computing capacity across several chip platforms. That may help the company keep pace with reported demand and match workloads to different hardware. But the public announcements do not show whether the capacity is being used efficiently, what it costs Anthropic per unit of compute, or whether the commitments will produce attractive returns.
Why Anthropic is committing to so much compute
Anthropic is making a capacity-and-demand bet: secure infrastructure ahead of expected customer growth and future model-development needs, rather than wait until demand is already pressing against available capacity. The company reported annualized run-rate revenue above $30 billion in April 2026, up from approximately $9 billion at the end of 2025. It also said more than 1,000 customers were spending over $1 million annualized, more than twice the count it reported in February 2026. These are company-reported figures, not audited annual revenue or profit.
As an Amazon Associate I earn from qualifying purchases.
Earlier demand indicators point in the same direction. Anthropic said in 2025 that it had more than 300,000 business customers and that the number of large accounts with over $100,000 in run-rate revenue had grown nearly sevenfold year over year. Those figures help explain the urgency, but they do not establish how much of the announced capacity is needed, when it will be used, or what it will earn.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →What the announced agreements provide
The announcements describe different kinds of commitments: a spending agreement, an investment, planned chip capacity, and a separate long-term lease forecast. They should not be added together as if they were the same measure of cost or value.
#1 Best Overall
| Arrangement | Publicly announced amount or capacity | Timing and qualification | What the figure measures |
|---|---|---|---|
| AWS agreement | More than $100 billion of AWS technology spending over ten years; capacity of up to five gigawatts | Announced in April 2026; the announcement describes a commitment and capacity ceiling, not five gigawatts already online | Anthropic’s stated spending commitment and planned capacity access |
| Amazon investment | $5 billion invested, with up to another $20 billion tied to commercial milestones | Announced in April 2026; the additional amount is conditional on milestones | Investment in Anthropic, distinct from Anthropic’s AWS spending commitment |
| Google TPU expansion announced in 2025 | Up to one million Google TPUs; an expansion worth “tens of billions of dollars” and expected to bring well over one gigawatt online in 2026 | Anthropic’s 2025 announcement described expected capacity; it is not evidence that all of it was delivered or in service | Planned TPU capacity and an approximate value description, not a disclosed Anthropic cost breakdown |
| Google-Broadcom agreement | Multiple gigawatts of next-generation TPU capacity | Announced in April 2026; expected to come online starting in 2027, with most of the capacity expected to be in the United States | Future compute capacity, not a public spending total |
| TeraWulf Kentucky campus lease | About 401 MW of critical IT load; TeraWulf forecast about $19 billion in contracted lease revenue over an initial 20-year term | The SEC-filed announcement expected phased service to start in the second half of 2027 and full capacity by early 2028 | A landlord’s forecast of contracted lease revenue. The announcement does not establish that this lease is an Anthropic commitment. |
The AWS spending figure and Amazon’s investment are separate parts of the same commercial relationship: one is a commitment by Anthropic to buy AWS technology, while the other is capital Amazon invested in Anthropic, with further investment linked to milestones. The TPU announcements describe capacity at different points in time; the 2025 expansion and the 2026 Google-Broadcom agreement should not be treated as interchangeable or as one quantified spending total.
Why use several chip platforms?
Anthropic says it spreads workloads across Google TPUs, Amazon Trainium and NVIDIA GPUs, and has described Amazon as its primary cloud and training partner. Its stated rationale is to match workloads to hardware and improve resilience. In principle, more than one platform can give a company options if a particular type of compute is constrained or better suited to a specific task.
That flexibility is not automatic. Different hardware and cloud environments can require engineering work to adapt and operate workloads, and long-term capacity commitments can make it harder to shift spending elsewhere. The company’s public statements describe its strategy, but do not disclose how workloads are divided among platforms or quantify any savings from that approach.
Rank #2
Where the potential economics come from
Capacity before demand arrives
Large commitments can reserve room for growth and reduce the risk that infrastructure becomes a bottleneck. If customer and model demand grows as Anthropic expects, having capacity planned in advance could support more usage and revenue. The trade-off is that capacity arriving ahead of demand may sit underused, while a delay in customer growth can leave the company with obligations before it has corresponding revenue.
Hardware and infrastructure choices
Anthropic’s mix of cloud services and specialized chips gives it potential options in how it obtains compute. Amazon CEO Andy Jassy characterized Trainium as offering “high performance at significantly lower cost for customers.” That is Amazon’s claim, not an independent comparative benchmark, and it does not show Anthropic’s negotiated prices or realized cost per task.
Long-term contracting
Long-term arrangements can help a provider plan and finance infrastructure, while giving the customer a path to capacity over time. But the headline figures do not reveal the contract details that determine the customer’s risk: for example, how much capacity must be taken or paid for, what happens if delivery is delayed, or how prices change. The public AWS and TPU announcements do not provide enough information to calculate those effects.
What public numbers can—and cannot—prove
The announced amounts show the scale and timing of selected arrangements, not Anthropic’s realized economics. None of the listed figures is a direct measure of profit, and the public announcements do not disclose Anthropic’s realized cost per unit of compute, utilization, operating expenses, margins, or return on invested capital for these commitments.
- Capacity is not the same as delivered compute. Gigawatts and megawatts describe power or capacity measures, not how much useful computing work has been completed.
- Spending is not cost per task. The AWS commitment does not reveal the cost of running a particular model or serving a particular customer request.
- Run-rate revenue is not profit. An annualized revenue pace does not show the expenses required to generate it or how much of that revenue is recurring.
- A forecast is not realized revenue. TeraWulf’s approximately $19 billion figure is its forecast of contracted lease revenue over the initial term, not revenue already collected and not a disclosed Anthropic result.
For the same reason, the public figures cannot be combined into a single total called “Anthropic’s data-center spend.” They describe different counterparties, terms, measures and stages of delivery.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Dependence and competition risks
Securing capacity from major cloud and chip providers can give Anthropic access to infrastructure that would be difficult to assemble quickly on its own. It can also increase dependence on those providers. The Federal Trade Commission’s January 2025 report identified issues to watch in cloud and AI partnerships, including access to compute and engineering talent, switching costs, and partners’ access to sensitive technical or business information. The report reflects information available to FTC staff through September 2024 and public information through January 2025, so it is a framework for considering risk—not an evaluation of Anthropic’s agreements announced in 2026.
Anthropic CFO Krishna Rao described the Google-Broadcom deal as a “disciplined approach to scaling infrastructure,” intended to serve customer growth while supporting AI development. That states the company’s rationale; it does not resolve how much flexibility the contracts provide or what their eventual returns will be.
How to judge whether the sprint pays off
The announcements establish that Anthropic is arranging substantial future compute capacity and that the company reports rapid growth in revenue run rate and large customers. They do not yet establish that the capacity will be used at high utilization or that the economics are attractive. A clearer assessment would require disclosures showing delivered capacity, utilization, contract costs, operating expenses and financial returns over time.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
