TL;DR: on September 24, 2026, Anthropic signed an $11.6bn, seven-year contract with Akamai for CPU compute capacity, expandable to roughly $20bn. The deal confirms a fact that gets little airtime: AI agents spend most of their execution time off the GPU, on plain, ordinary computing.
Everyone is watching GPUs and mega data-center deals. Meanwhile, Anthropic just committed $11.6bn to the kind of processor that has powered web servers for thirty years. Here is why that choice is news in itself, and what a business owner can take away from it.
- On September 24, 2026, Anthropic signed an $11.6bn, seven-year contract with Akamai covering CPU workloads, with an option to expand by another $9bn.
- According to Jon Alexander, SVP of Product for Akamai's cloud group, up to 90% of an agentic task's execution time happens outside the GPU.
- Akamai granted Anthropic a warrant for roughly 5% of its share capital; Akamai stock jumped about 20% after the announcement.
- For an SME, the takeaway is direct: the running cost of an AI agent mostly sits on commodity, competitive compute, not on scarce hardware.
The news: a giant contract, for ordinary processors
Akamai Technologies, historically known for its content delivery network, announced on September 24, 2026 an $11.6bn, seven-year contractual commitment from Anthropic, the maker of Claude. The agreement includes a possible expansion of another $9bn, for a total potential commitment of roughly $20bn, according to Akamai's official press release.
The purpose of the contract is explicit: meeting Anthropic's fast-growing demand for CPU workloads, served by Akamai Cloud's distributed infrastructure, which spans thousands of points of presence worldwide. Akamai expects around $5.5bn in capital expenditure to fulfill the initial commitment, including $1.7bn as early as 2026, per Yahoo Finance. The market cheered immediately: Akamai shares gained about 20% in after-hours trading on September 24, 2026.
"Anthropic is advancing the AI revolution and we are thrilled they chose Akamai's capabilities for building and operating AI infrastructure at scale," said Tom Leighton, Akamai's co-founder and CEO, in the September 24, 2026 announcement.
Why CPUs rather than GPUs?
Because an AI agent only spends a fraction of its time generating text. Jon Alexander, SVP of Product for Akamai's cloud technology group, explained it back in July 2026 in an interview with TFiR: "for agentic applications, often what we're seeing is up to 90% of the overall task execution isn't on the GPU." The GPU (graphics processing unit) handles inference, meaning the model's actual text generation. Everything else, calling an API, querying a database, running code in a sandbox, chaining the steps of a task, is classic computing that runs on CPUs.
The table below sums up what actually runs where inside a working AI agent.
| Agent step | Hardware used | Concrete example |
|---|---|---|
| Text generation (inference) | GPU | Writing an answer, analyzing a document |
| Tool and API calls | CPU | Reading a CRM, sending an email, searching the web |
| Sandboxed code execution | CPU | Computing, transforming a file, producing a report |
| Orchestration and logic | CPU | Loops, decisions, retries, error handling |
Where that compute sits matters too. Jon Alexander offers a telling order of magnitude: add 100 milliseconds of latency to each step of an agent that loops 100 times, and you have added 10 seconds to the final result. Hence Anthropic's interest in a distributed infrastructure that brings execution closer to tools and data, rather than concentrating everything in a handful of giant data centers.
What the financial structure reveals
The contract carries a second, less visible signal: Akamai granted Anthropic a warrant over 7.7 million shares, roughly 5% of its capital, at an exercise price of $111.33. About 2% vests with the initial $11.6bn commitment; the rest unlocks at roughly 1% per additional $3bn in purchases.
In other words, the supplier is partly paying its customer to keep the business. This kind of structure, already seen in the big AI compute deals of 2025 and 2026, shows how badly infrastructure providers want a durable seat in the AI labs' value chain. It also shows the buyers' bargaining power: Anthropic, which already holds massive agreements with Amazon and Google, now treats compute as a raw material to be negotiated across several suppliers. Akamai, for its part, had already announced more than $2.8bn in multi-year cloud commitments earlier in 2026; this contract changes the scale of its infrastructure business.
What does this change for a small or mid-sized business?
Three concrete things, none of which require you to buy any hardware.
The cost of AI agents should keep falling. If most of an agent's execution time sits on CPU compute, then most of its running cost sits on a commodity, competitive market with well-known prices. That is a very different dynamic from the scarcity of high-end GPUs. No pricing was announced with this deal, but the direction is clear: several providers are now fighting for these workloads.
An agent's value is in the software, not the hardware. What this contract confirms is that agentic AI is mostly classic software engineering: wiring the agent into your real tools (CRM, email, invoicing), defining what it is allowed to do, handling errors. When you evaluate a vendor or a solution, that is the part to scrutinize; the underlying model is interchangeable.
Where the compute runs is a parameter again. An infrastructure distributed across thousands of points of presence means execution can move closer to your data. For a European company subject to GDPR (General Data Protection Regulation), the application layer of an agent, the part that touches your customer data, can be hosted in France or elsewhere in Europe without prohibitive extra cost. It is an architecture choice, not an insurmountable technical constraint.
How this connects to my day-to-day work
I see this CPU-GPU split every day in my own tools. My Bloomberg Dashboard, a daily financial brief generated with Claude Haiku 4.5 on my personal portfolio, illustrates the proportions well: fetching prices, computing variations and formatting the report is all classic code that would run on the most modest server. The model only steps in at the end, for analysis and writing. The same goes for the agents I build for SMEs: reliability comes from the wiring into data and the orchestration, not from raw power. A useful agent does not need a supercomputer; it needs clean access to your tools.
Next time a billion-dollar AI contract makes headlines, look at what is actually being bought. When the maker of Claude commits $11.6bn to ordinary processors, the message is simple: production AI is becoming software like any other. And software, unlike GPUs, is not something you endure; it is something you design.
Frequently asked questions
Why is Anthropic buying CPUs when AI runs on GPUs?
Because AI agents spend most of their time off the GPU. According to Jon Alexander, SVP of Product for Akamai's cloud group, up to 90% of an agentic task's execution time happens outside the GPU: API calls, code execution, orchestration. Those workloads run on ordinary processors, which is what the $11.6bn deal signed on September 24, 2026 covers.
Will this deal make AI agents cheaper?
That is the likely direction. CPU compute is a commodity market with plenty of competition, unlike high-end GPUs. When most of an agent's running cost sits on ordinary servers, competition between infrastructure providers pushes prices down. No specific price cut has been announced at this stage.
Does a small business need to buy servers to use AI agents?
No. For an SME, the question is not hardware but software: how well the agent is wired into your actual tools (CRM, email, ERP) determines its usefulness. A well-designed agent runs on standard cloud infrastructure billed at a few dozen euros per month, and the application layer can be hosted in Europe.