Anthropic Committed $11.6 Billion to Akamai for CPUs. The Warrant Struck at $111.33 and the Stock Closed at $134.64.

Akamai announced on Thursday, September 24, 2026 that Anthropic has committed $11.6 billion over seven years to Akamai Cloud, with an option to extend by another $9 billion for a ceiling around $20.6 billion. The tape picked it up Friday. Akamai stock ran 17 to 22 percent after hours depending on when you clipped it, with one print at $134.64 against a $110.41 close.

Every writeup I read led with the dollar figure. The dollar figure is the least interesting thing in the release.

Two things here are genuinely new, and both of them are buried. The first is a word that does not appear: GPU. The second is the direction the equity travels.

The Word That Is Missing Is GPU

Akamai's own language is that the agreement supports "Anthropic's accelerating CPU workload demands" across a platform spanning thousands of points of presence with diversified hardware. No accelerator count. No gigawatt figure. No training cluster. For a company that has spent 2026 announcing capacity in gigawatts, Anthropic just signed its most linguistically unusual compute deal of the year, and the unusual part is that it reads like a hosting contract from 2014.

This is not Akamai being coy about accelerators it does not have. Akamai launched AI Grid in March 2026, orchestrating distributed inference across 4,400 plus edge locations on NVIDIA RTX PRO 6000 Blackwell servers with BlueField DPUs, and it sells an Inference Cloud product on top. If Anthropic wanted edge GPU, the SKU was sitting right there. The release says CPU.

Here is what I think is actually happening, and I want to flag it as a read rather than a disclosure, because neither company has confirmed it. Agents do not only burn accelerator time. An agent trajectory is a token generation step followed by a pile of work that no GPU touches: spawning a sandbox, executing the tool call, running the code the model wrote, parsing the result, hitting a retrieval index, holding session state, and doing it again a few hundred times. That middle layer is ordinary compute, it is spiky, it is latency sensitive, and it is the fastest growing line item nobody put on a slide.

Anthropic ships Claude Code, a managed agent platform, and by its own September disclosure runs on the order of 30,000 concurrent internal agents on its busiest platform. Every one of those trajectories needs somewhere cheap to execute. Accelerators are the wrong instrument for that and they are the scarcest thing Anthropic owns. Moving the non accelerator half of the agent loop onto a distributed CPU fleet is the most boring possible infrastructure decision and probably the correct one.

The Terms, In One Table

The strike is the detail worth sitting with. $111.33 against a $110.41 close the day before the announcement. Anthropic's warrant was struck essentially at the money on a stock that the announcement itself was always going to move.

Run the arithmetic on the print. At $134.64, the full 7.7 million shares carry about $179 million of intrinsic value. Only the first tranche vests now, roughly two fifths of the total, so call the day one paper gain about $72 million. Anthropic committed $11.6 billion and walked out of the announcement holding roughly $72 million of the counterparty, or about 0.6 percent of the contract value, handed back before a single vCPU was provisioned.

The Equity Is Flowing the Wrong Way

Every large compute arrangement this year has had the same shape. The supplier takes a position in the AI lab. NVIDIA into OpenAI. AMD writing OpenAI a warrant against purchase volume. Hyperscalers putting capital into labs that then spend it back on their own cloud. The chip and cloud vendors have been financing their own demand, and the whole sector has been arguing for a year about whether that is vendor financing with extra steps.

This one inverts. The supplier is issuing equity to the buyer. Akamai is paying Anthropic, in stock, for the privilege of being spent at.

That only makes sense if Anthropic's signature is the product. Akamai is not selling scarce silicon here, it is selling CPU capacity in a market where CPU capacity is a commodity with three enormous incumbents. What Akamai actually bought with 5 percent of itself is a seven year anchor tenant, a capex plan its board can defend, and a story that repriced the equity 20 percent in an evening. On those terms the warrant looks less like a concession and more like the cheapest marketing spend in the company's history.

It also tells you something about Anthropic's negotiating position that I think is underrated. A buyer who can extract equity from a public supplier is a buyer who had other bids.

What $1.7 Billion a Year Does to a $99 Million Segment

Now the part that should make an Akamai shareholder read the risk factors twice.

That last row needs its caveat stated plainly, so here it is: Akamai will be a larger company in 2028 and the real concentration percentage will be lower than 39. But the shape does not change. A CDN and security business growing revenue at 6 percent a year is strapping a $1.7 billion annual line to a single counterparty and funding it with about $5.5 billion of capex it has to spend first.

We covered this exact structure four days ago when Nscale went to market with 85 percent of backlog sitting in two customers. Akamai is a far sturdier business than Nscale and the comparison is not an equivalence. It is a pattern. The AI buildout is converting diversified infrastructure companies into single tenant landlords, and the tenant is always one of four labs.

One more line from the release that nobody picked up: Akamai is pre purchasing critical components, memory included. Memory is the bottleneck of the moment, DRAM contract pricing has been on a tear all year, and a company committing $5.5 billion of build is locking supply before the deal is even revenue generating. That is a supply chain judgment with real money behind it, and it is worth more as a signal than most analyst notes on the memory cycle.

Three Counterreads

The CPU framing may just be accounting. Akamai said CPU workloads and gave no specifics on intended usage. It is possible this is a general capacity agreement that happens to start with CPU tranches and adds accelerators later, in which case my agent tax thesis is reading tea leaves. I would take that bet against me, but it is a live possibility until either party says more.

The warrant is not free money. Anthropic has to spend $3 billion for each additional percent, and the shares are non voting convertible preferred. A 5 percent stake that requires $20.6 billion of committed spend to fully vest is an incentive structure, not a windfall, and it works in Akamai's favor every time it triggers.

Concentration cuts both ways. Anthropic is days away from an S-1 process with a founder supervoting proposal, a Sherman Act class action, and a Pentagon supply chain designation an appeals court just let stand. Akamai has tied a $5.5 billion capex plan to a counterparty with an unusually eventful legal calendar. Anthropic is not a credit risk in any ordinary sense. It is a headline risk with a seven year term.

Our Take

The number that matters is $99 million. That is the entire Akamai cloud infrastructure business for a quarter, growing at a genuinely good 39 percent, and it is about to be dwarfed by four times over by one customer who is not paying for the thing Akamai spent three years building. They built a distributed GPU inference grid across 4,400 locations, and the biggest contract in company history showed up asking for CPUs.

I find that funny and I also find it instructive. The infrastructure industry spent 2025 and 2026 convinced the scarce resource was accelerator time, and it was, for training. The agent era is quietly re pricing the other half of the stack: sandboxes, tool execution, orchestration, retrieval, session state. None of that is glamorous. All of it scales with trajectory count rather than parameter count, and trajectory count is the number that is actually compounding right now.

The honest uncertainty: none of the CPU reasoning is confirmed. Akamai gave a workload category and nothing else, and I have built an argument on one phrase in a press release. Treat it accordingly.

Three signposts for the next 60 days. Whether Akamai's Q3 call breaks out what Anthropic is actually running, because a CPU deal described in vCPU hours and a CPU deal described in racks are different businesses with different margins. Whether Anthropic's S-1, when it goes public, lists Akamai in the compute commitments table alongside the accelerator vendors or in a separate category, because that is the company telling you in its own words whether this is compute or something else. And whether a second lab signs a CPU heavy capacity deal with an edge provider before year end, because one deal is a procurement decision and two is a shift in where the agent stack actually runs. You can track the commitments as they land on our Anthropic provider page.