Anthropic's $11.6B Akamai Deal: Why CPUs, Not GPUs

By Jaspal Singh September 25, 2026 Updated: September 25, 2026

Anthropic has committed $11.6 billion over seven years to Akamai for CPU capacity — not GPUs. Announced on 24 September 2026, the deal attaches a warrant for up to roughly 5% of Akamai's stock and can grow to about $20 billion. It is a bet that agentic AI runs on ordinary processors as much as on accelerators.

What Happened

According to Akamai's 24 September 2026 press release, the company signed an agreement under which Anthropic commits approximately $11.6 billion of contractual value over an initial seven-year term. Anthropic will leverage “Akamai Cloud's distributed infrastructure and software to support CPU workload growth at scale.” The release describes a “significantly expanded relationship” rather than a first contract, and provides for expansion by up to an additional $9 billion — a total potential commitment of roughly $20 billion.

Dr. Tom Leighton, Akamai's co-founder and chief executive, said Anthropic “chose Akamai's capabilities for building and operating AI infrastructure at scale,” and has described it as the largest contract in the company's history.

The unusual part sits in the equity. Akamai issued Anthropic a warrant for up to approximately 5% of its common stock outstanding — about 7.7 million shares on an as-converted basis, at an exercise price of $111.33 per common share. Roughly 2% is expected to vest in connection with the announced $11.6 billion commitment; the remaining 3% vests only if the relationship expands, at approximately 1% per additional $3 billion committed.

Akamai's Form 8-K filed with the Securities and Exchange Commission puts precise mechanics behind that summary. The warrant was entered on 18 September 2026, six days before the announcement, and covers 387,051 shares of Series B Non-Voting Convertible Preferred Stock at an exercise price of $2,226.60 per warrant share — a figure the filing describes as a 30-trading-day volume-weighted average price multiplied by 20. Vesting runs in four tranches: 40% on the first payment under what the filing calls Project Plan 3, then three tranches of 20% each as each additional $3.0 billion is committed. The vested portion is exercisable until 18 September 2033.

Akamai also guided to the cost, and it is front-loaded. Capital expenditure tied to the initial commitment is estimated at approximately $5.5 billion: roughly $1.7 billion in the fourth quarter of 2026 for supply-chain components including memory, about $3.1 billion in 2027, and around $700 million in 2028. Revenue is not expected until the second half of 2027, reaching an annualised run rate of roughly $1.7 billion by the end of 2028. Akamai said there is no impact to its 2026 revenue guidance. Shares rose more than 20% in extended trading.

This expands a $1.8 billion, seven-year Akamai–Anthropic agreement announced in May 2026, which Leighton also called the largest deal in the company's history at the time. The new commitment is roughly six times larger.

Why It Matters

Almost every large AI infrastructure announcement of the past two years has been a GPU announcement. This one is not, and that is the signal worth reading. Anthropic is not short of accelerator capacity: as TechCrunch catalogued in August, it has signed a $45 billion, six-year deal with Nscale for Nvidia Vera Rubin capacity, alongside roughly $10 billion with Volta, $5 billion with AMD, arrangements with SpaceX, five more gigawatts from Amazon, and expanded Google and Broadcom TPU capacity. Adding $11.6 billion of general-purpose processors on top says the bottleneck in serving agentic models is no longer only the accelerator.

For Akamai, the strategic re-rating is larger than the contract. A company built on content delivery is now underwriting $5.5 billion of capital expenditure against a single customer's seven-year promise. That concentration is the risk: the spending lands in 2026 and 2027, and revenue does not begin until the second half of 2027. If Anthropic's consumption ramps more slowly than modelled, Akamai has already bought the hardware. Similar demand-timing scepticism has dogged other large compute programmes, including the EU's €20 billion AI computing plan, which drew lawmaker criticism over exactly this uncertainty.

The warrant is the other structural story. Supplier-to-customer equity has become a defining feature of the 2026 AI build-out: the vendor hands the buyer upside in its own shares, and the buyer's committed spend is what vests it. That aligns both sides, and it also means the customer's cost of compute and the supplier's reported contract value are not independent numbers. The same entanglement runs in the other direction elsewhere, with Google committing up to $40 billion to Anthropic as both investor and compute supplier. Spreading commitments across Akamai, Nscale, AMD, Amazon and Google is also a hedge against single-supplier dependence — the logic behind Anthropic's early talks with UK chip startup Fractile.

One detail deserves attention from anyone buying hardware next year: Akamai is spending about $1.7 billion in a single quarter partly to pre-purchase memory. That is a large buyer pulling DRAM forward in an already strained market, compounding the pressure behind the 40–50% surge in RAM and SSD prices.

How It Works

The technical premise is that agentic AI changes the shape of an inference server. A single-turn chat completion is dominated by matrix multiplication on an accelerator. An agent is not: it calls tools, queries APIs, reads and writes files, executes code in sandboxes, retrieves documents and orchestrates subtasks before answering. That work is sequential, branch-heavy, I/O-bound control logic — exactly what general-purpose cores are for, and what a GPU is poor at.

Industry disclosures through 2026 make that shift explicit. As Futuriom summarised from first-quarter earnings calls, CPUs moved to the centre of the AI infrastructure story. AMD chief executive Lisa Su told investors that “inferencing and agentic AI are increasing the need for server CPU compute.” Amazon's Andy Jassy put it more bluntly: “AI is commonly seen as a GPU story, but the rise of agentic workloads…is driving massive CPU demand as well.” Arm told shareholders that data centres are expected to require “more than 4x current CPU capacity per gigawatt as agentic AI scales,” and AMD raised its server CPU addressable-market projection to over $120 billion by 2030.

In practice this shows up as a ratio. Conventional AI servers have run roughly one CPU for every four to eight accelerators. In agentic deployments that tightens materially — Intel has told investors it could approach 1:1 as workloads move from training to inference. Multiply that across a frontier lab's serving fleet and the CPU line item stops being a rounding error.

Why Akamai? Its infrastructure is not shaped like a hyperscaler's: thousands of distributed points of presence close to end users rather than a handful of gigawatt campuses. That is the wrong architecture for training a frontier model and a defensible one for low-latency serving, orchestration and the retrieval traffic surrounding an agent. Leighton attributed the earlier win to Akamai's ability to manage and scale complex distributed systems, plus low latency. The complement to that serving layer is efficient scheduling of the accelerators themselves, which is why control-plane software has become an acquisition target in its own right — see IREN's $625 million purchase of Mirantis to unlock GPU utilisation.

The warrant arithmetic is worth working through, because the two sets of published figures look inconsistent until you do. The filing's 387,051 preferred shares each convert into 20 common shares, which is 7,741,020 — the “approximately 7.7 million” in the press release. Likewise, the $2,226.60 preferred exercise price divided by 20 is $111.33. The preferred is non-voting and converts automatically on transfer to any holder other than Anthropic and its wholly owned subsidiaries. So Anthropic gets economic exposure without a vote, and Akamai avoids handing a customer governance rights.

Finally, the revenue model explains the cash-flow shape. Akamai's earlier Anthropic contract was consumption-based, with revenue recognised as capacity is actually used. Capacity has to be built, powered and accepted first, which is why $5.5 billion of capital expenditure precedes the second-half-2027 revenue start.

What's Still Unknown

Which CPUs. Neither the press release nor the 8-K names a silicon vendor or architecture. Arm-based server parts, AMD EPYC and Intel Xeon all plausibly fit “CPU workloads at scale,” and the choice has real consequences for cost per token and for which chipmaker books the revenue.

Where the capacity lands, and its power profile. No sites, site count or megawatt figures have been disclosed. For a distributed operator it may be spread across many facilities rather than one campus, but that is inference, not disclosure.

What Project Plans 2 and 3 cover. The filing names them and ties the first vesting tranche to the first payment under Project Plan 3, without describing their scope. The commercial structure underneath the headline number is therefore opaque.

Whether the $11.6 billion is a firm floor. The documents describe contractual commitment, and the prior agreement was consumption-based. How much is take-or-pay versus usage-dependent is not public, and it determines who carries the risk if agent traffic disappoints. Three of the warrant's five percentage points likewise vest only against a further $9 billion.

Margin impact. Akamai has guided capital expenditure and revenue timing but not this business's gross margin relative to its delivery and security segments. Chief financial officer Ed McGowan has said the warrant's financial value is small relative to overall contract value; dilution and margin are separate questions, largely unanswered.

Frequently Asked Questions

Why would an AI company buy CPUs instead of GPUs?

Because agentic workloads are not just matrix maths. Agents call tools, hit APIs, run code, read files and orchestrate subtasks — sequential, I/O-heavy logic that general-purpose cores handle far better than accelerators. AMD and Amazon both told investors in 2026 that agentic inference is driving material server CPU demand alongside GPU demand, rather than instead of it.

How much is the Akamai–Anthropic deal actually worth?

Approximately $11.6 billion of contractual commitment over an initial seven-year term, with an option to expand by up to a further $9 billion for a total potential of roughly $20 billion. Only the $11.6 billion is committed today, and it expands a $1.8 billion agreement announced in May 2026.

What does Anthropic get besides compute?

A warrant for up to about 5% of Akamai's common stock — roughly 7.7 million shares on an as-converted basis at $111.33 each. About 2% is expected to vest against the announced commitment, with the rest vesting at roughly 1% per additional $3 billion committed. The underlying preferred stock carries no voting rights.

When does Akamai start earning money from this?

Not immediately. Akamai expects revenue to begin in the second half of 2027, then reach an annualised run rate of about $1.7 billion by the end of 2028. The company said the agreement does not change its 2026 revenue guidance, so the near-term effect is capital spending rather than sales.

What does it cost Akamai to serve the contract?

Roughly $5.5 billion of capital expenditure: about $1.7 billion in the fourth quarter of 2026 for supply-chain components including memory, around $3.1 billion in 2027, and some $700 million in 2028. That spending largely precedes the revenue, which is the main financial risk in the structure.

Is this a sign that GPU demand is slowing?

No, and nothing in the disclosures suggests it. Anthropic continues to commit heavily to accelerators, including a $45 billion Nscale agreement for Nvidia Vera Rubin capacity. The CPU commitment is additive: as inference shifts toward multi-step agents, the ratio of general-purpose cores to accelerators tightens, so both lines grow together.

Why is a content delivery network selling AI infrastructure?

Akamai's distributed footprint of points of presence close to users suits low-latency serving, orchestration and retrieval traffic rather than frontier-model training. Leighton attributed the earlier Anthropic win to Akamai's ability to manage complex distributed systems at low latency. It is competing on architecture and locality, not on gigawatt-scale training campuses.