Anthropic compute deals now total roughly $517 billion in capacity commitments signed over 11 months, according to reporting by The Information, covering at least 14.8 gigawatts of new computing power. The latest is a seven-year, $11.6 billion agreement with Akamai announced on September 24, 2026 — notable because it is for CPUs, not GPUs. The company's annualized revenue passed $100 billion in mid-September.
Set those two numbers side by side. Anthropic has committed to more than five times its current annual revenue run rate in compute. Its own CEO has warned the company "could go bankrupt" if its estimates are off by even a small margin.
This article breaks down who the money is going to, why a company famous for buying GPUs just signed its biggest-ever contract with a CDN provider for ordinary processors, how the Akamai warrant works, and whether the math behind half a trillion dollars of commitments holds up.
Key Takeaways
- Anthropic has signed about $517 billion in compute capacity deals since October 2025, locking in at least 14.8 gigawatts.
- The largest named commitments are AWS ($100B), Broadcom (~$63B), Fluidstack ($50B), Nscale and Monarch WV ($45B), SpaceX ($45B) and Lambda ($35B).
- The Akamai deal is $11.6B over seven years for CPU capacity, expandable by $9B to about $20B.
- Akamai issued Anthropic a warrant for up to roughly 5% of its stock at a $111.33 strike, vesting as Anthropic spends.
- Anthropic's revenue run rate rose from $9B at the end of 2025 to $65B by the end of July and past $100B in September 2026.
How big are the Anthropic compute deals?
Anthropic has signed approximately $517 billion in compute capacity agreements in the 11 months since October 2025. Those deals lock in at least 14.8 gigawatts of computing power on top of the one to two gigawatts it already had. The figure is a ceiling on capacity purchases over many years, not cash that has been paid.
That distinction matters. As Data Center Dynamics reported, these are capacity leases stretching across the next decade, many of them contingent on the provider actually delivering working data centers on schedule.
The named commitments:
| Partner | Reported commitment | Notes |
|---|---|---|
| Amazon Web Services | $100 billion | Over a decade; AWS is the primary cloud partner |
| Broadcom | ~$63 billion | By 2027; custom silicon |
| Fluidstack | $50 billion | Data center capacity |
| Nscale and Monarch WV | $45 billion | Announced August 2026 |
| SpaceX | $45 billion | — |
| Lambda | $35 billion | Nvidia-backed cloud; reported August 31, 2026 |
| Akamai | $11.6 billion | Seven years, CPU capacity; announced September 24, 2026 |
| up to $40 billion | Cash and compute | |
| AMD | up to 2 gigawatts | MI450 GPUs; strategic partnership |
A few lines of arithmetic put the scale in context:
deals_bn = {
"AWS": 100, "Broadcom": 63, "Fluidstack": 50,
"Nscale + Monarch WV": 45, "SpaceX": 45, "Lambda": 35,
}
total_bn, gigawatts, run_rate_bn = 517, 14.8, 100
top_six = sum(deals_bn.values())
print(f"Top six deals: ${top_six}B ({top_six / total_bn:.0%} of total)")
print(f"Implied cost: ${total_bn / gigawatts:.1f}B per gigawatt")
print(f"Commitments: {total_bn / run_rate_bn:.1f}x current revenue run rate")
# Top six deals: $338B (65% of total)
# Implied cost: $34.9B per gigawatt
# Commitments: 5.2x current revenue run rate
Three things fall out of that.
Concentration. Six counterparties account for about two-thirds of the total. But the list is unusually diverse for a company of this kind: a hyperscaler, a chip designer, three neoclouds and a rocket company. Anthropic is deliberately not dependent on one supplier.
Cost per gigawatt. Roughly $35 billion per gigawatt is a blended figure across chips, buildings, power and multi-year operations, so it is not comparable to a construction cost. It does show the order of magnitude: a gigawatt of AI capacity, fully loaded over its contract life, is a tens-of-billions commitment.
The revenue multiple. 5.2 times run-rate revenue sounds alarming until you remember the commitments are spread over up to a decade. It still only works if revenue keeps climbing.
Why is Anthropic buying CPUs from Akamai?
Anthropic is buying CPU capacity from Akamai because AI agents consume a great deal of ordinary computing, not just GPU inference. The $11.6 billion, seven-year agreement uses Akamai Cloud's distributed infrastructure for CPU workloads. It is the largest contract in Akamai's history and can expand by up to $9 billion, to about $20 billion.
This is the angle most coverage treats as a footnote, and it is the most informative part of the story.
Every other deal on the list above is about accelerators — GPUs or custom AI chips. Akamai is a content delivery and edge cloud company. It does not sell frontier AI silicon. So why is a model lab writing it an eleven-figure contract?
Neither company has published a detailed workload breakdown, so what follows is inference from how these products work. When a model answers a chat message, almost all the compute is GPU. When an agent does a task, the model is only part of the system:
- It runs code in a sandbox — a container or virtual machine.
- It executes tool calls: builds, test suites, file operations, web requests.
- It drives browsers and desktops for computer-use tasks.
- It holds sessions open for hours while work proceeds.
None of that runs on a GPU. It runs on general-purpose processors, memory and network. As agents go from minutes to hours per task, the CPU footprint per user grows with them. An agent that runs a 680,000-line code migration in a day, as one early Claude Opus 5.5 tester reported, spends that day compiling and running tests.
Akamai's network is also distributed, with capacity close to users in many regions. For sandboxes that need low latency to a user's browser or that must keep data in a particular geography, that is a better fit than a handful of giant GPU campuses.
The takeaway: the bottleneck for agentic AI is shifting from "enough GPUs" to "enough of everything." We looked at the power side of that constraint in the AI data center power crunch, and at the chip side in how custom AI chips are breaking Nvidia's grip.
How does the Akamai warrant work?
As part of the deal, Akamai issued Anthropic a warrant for nonvoting preferred stock convertible into about 7.7 million common shares — up to roughly 5% of Akamai — at a strike price of $111.33 per share. About 2% vests with the initial commitment, and each additional $3 billion of spending unlocks roughly another 1%.
According to TechCrunch, the structure looks like this:
| Term | Detail |
|---|---|
| Base commitment | $11.6 billion over seven years |
| Expansion option | Up to $9 billion more (about $20 billion total) |
| Warrant size | ~7.7 million shares, up to ~5% of Akamai |
| Strike price | $111.33 per share |
| Initial vesting | ~2% tied to the announced commitment |
| Further vesting | ~1% per additional $3 billion of spending |
| Akamai revenue, 2027 | $150–300 million, starting in the second half |
| Akamai run rate, end of 2028 | ~$1.7 billion annually |
| Akamai capex for the build-out | ~$5.5 billion, plus $1.7 billion in 2026 for components |
In effect, the customer gets equity in the supplier for being a customer. This has become the standard shape of large AI infrastructure deals, and the incentives are worth spelling out.
For Anthropic, the warrant is a discount paid in upside. If the contract makes Akamai more valuable, Anthropic captures part of that gain. The more it spends, the more it vests.
For Akamai, it is a price worth paying to land a customer that transforms its growth profile. The market agreed: Akamai shares rose as much as 17% in after-hours trading on the announcement.
The risk sits in the conditions. Akamai must spend about $5.5 billion in capital to build capacity, plus $1.7 billion on components this year, against revenue that only starts in the second half of 2027. The agreement is conditional on Akamai meeting delivery and service-availability requirements, and either party can exit under certain conditions. Akamai is taking on real construction risk before it sees meaningful revenue.
Can Anthropic afford $517 billion?
Only if its revenue keeps growing at close to the current pace. Anthropic's annualized revenue run rate went from $9 billion at the end of 2025 to $65 billion by the end of July 2026, and passed $100 billion in mid-September. The commitments are sized for a company several times larger than it is today.
The trajectory is steep by any standard:
| Date | Annualized revenue run rate |
|---|---|
| End of 2025 | $9 billion |
| End of July 2026 | $65 billion |
| Mid-September 2026 | More than $100 billion |
Axios reported the $100 billion milestone on September 18, with the company preparing for a stock market debut as soon as November. We covered the earlier leg of this run in Anthropic overtakes OpenAI on revenue and valuation.
The bull case is simple. Revenue grew more than elevenfold in under nine months. Demand is constrained by capacity, not by customers. Signing capacity is therefore the same thing as signing revenue.
The bear case is just as simple, and it comes from inside the industry. The Decoder noted that Anthropic CEO Dario Amodei warned the company "could go bankrupt if its estimates were off by even a small margin." Earlier in 2026 he said competitors "don't really understand the risks they're taking." OpenAI's Sam Altman has recently urged caution about what he called "unsustainable silliness" in compute build-outs, arguing that technical progress could make today's expensive infrastructure obsolete.
So both rival CEOs have publicly warned about exactly the behaviour both companies are engaged in. Neither currently generates enough revenue to cover its compute commitments on its own.
Three specific risks deserve attention:
- Efficiency cuts both ways. Claude Opus 5.5 does typical tasks for about 40% less than Opus 5. Good for customers — but if every model generation needs fewer tokens per task, revenue per unit of work falls unless usage grows faster.
- Price competition. In the same week as the Akamai deal, OpenAI cut its mainstream model prices in half. Capacity contracts are fixed; token prices are not.
- Delivery risk. At least 14.8 gigawatts has to be physically built, powered and cooled by counterparties of very different sizes. A delayed data center is capacity Anthropic planned revenue around.
There is one structural protection. These are capacity ceilings with delivery conditions, signed with many counterparties. That is more flexible than owning the buildings — though Anthropic is also reported to be planning its own data centers.
How does this compare with OpenAI?
OpenAI is targeting 30 gigawatts of capacity by 2030, roughly double Anthropic's 14.8 gigawatts of new commitments, though the two are hard to compare directly because Anthropic's contracts extend beyond 2030. On revenue, Anthropic's run rate of more than $100 billion is now ahead of OpenAI's, which is reported at nearly $70 billion.
The strategies differ in a way that matters for developers. OpenAI's build-out has been concentrated in a small number of very large projects. Anthropic's is spread across clouds, neoclouds, chipmakers and now an edge network.
For anyone building on these platforms, diversification has a practical consequence: Claude is available on AWS, Google Cloud and Microsoft Azure as well as Anthropic's own platform, and its capacity does not depend on any single provider shipping on time.
For the capital markets view of where this money comes from, see AI venture funding concentration. For the memory supply chain underneath all of it, see the Nvidia–SK Hynix HBM4 deal.
Frequently asked questions
How much has Anthropic committed to compute deals? Anthropic has signed about $517 billion in compute capacity agreements in the 11 months since October 2025, according to The Information. The deals cover at least 14.8 gigawatts and run over roughly the next decade; the figure is a ceiling on purchases, not cash already spent.
What is the Anthropic and Akamai deal? It is a seven-year, $11.6 billion agreement announced on September 24, 2026 for Akamai to supply CPU-based cloud capacity to Anthropic. It can expand by up to $9 billion, to about $20 billion, and is the largest contract in Akamai's history.
Why does Anthropic need CPUs instead of GPUs? AI agents run code, tests, browsers and tools in sandboxes, and that work executes on general-purpose processors rather than GPUs. As agent tasks get longer, the CPU capacity needed per user grows alongside the GPU capacity needed for the model itself.
What is Anthropic's revenue in 2026? Anthropic's annualized revenue run rate passed $100 billion in mid-September 2026. It stood at $9 billion at the end of 2025 and $65 billion at the end of July 2026.
Does Anthropic own part of Akamai? Not yet, but it can. Akamai issued Anthropic a warrant for up to roughly 5% of its stock, about 7.7 million shares at $111.33 each, which vests in stages as Anthropic's spending with Akamai increases.
Who are Anthropic's biggest compute partners? The largest reported commitments are with Amazon Web Services ($100 billion), Broadcom (about $63 billion), Fluidstack ($50 billion), Nscale and Monarch WV ($45 billion), SpaceX ($45 billion) and Lambda ($35 billion).
The verdict
Anthropic compute deals are a bet that demand for AI work is limited only by supply, and so far the revenue line supports it: $9 billion to more than $100 billion in under nine months is the fastest scaling any software business has recorded. If that continues, $517 billion of capacity looks prudent rather than reckless.
The Akamai contract is the more interesting signal. A model lab spending $11.6 billion on CPUs tells you where the product is going. The future being purchased is not a smarter chatbot. It is millions of long-running agents, each needing a sandbox, a network connection and hours of ordinary compute.
The risk is that the commitments are fixed while prices and efficiency keep moving against revenue per task. When the CEOs of both leading labs warn that this spending could end badly — while signing more of it — believe the warning and the signature at the same time.
If you are planning infrastructure around these models, read our piece on the AI data center power crunch next.
Half a trillion dollars says agents are the product. The invoices arrive whether or not the agents do.
