Akamai's $11.6 Billion Anthropic Bet: Why the Biggest AI Cloud Deal of 2026 Is for CPUs, Not GPUs

Introduction

On the morning of September 25, 2026, Akamai Technologies shares jumped more than 20 percent in after-hours trading, touching as high as $133.71 in premarket activity. The trigger was an 8-K filing and a press release announcing that Anthropic had committed $11.6 billion over a seven-year initial term to run CPU workloads on Akamai Cloud. It is the largest customer contract in the nearly 30-year history of a company that built its reputation on delivering cached web pages, and by most measures it is the largest AI cloud deal signed anywhere in 2026.

The dollar figure drew the headlines. The architecture buried in the filing is the story. Almost every large AI infrastructure deal of the past three years has been a GPU deal — NVIDIA accelerators arranged in dense clusters, joined by InfiniBand, commissioned by a frontier lab or a neocloud. This one is explicitly not that. Anthropic is buying CPU capacity, and it is buying it from a company with 4,400-plus edge locations in more than 100 countries rather than from a handful of hyperscale regions. TechTimes documented the deal in detail, and the reasoning it lays out explains why the infrastructure market is starting to bifurcate along workload lines rather than geography.

Main Content

Why Inference and Training Are Splitting Apart

Training a frontier model is irreducibly centralized. Every step requires all-reduce communication between tens of thousands of accelerators at once, generating terabytes per second of inter-chip traffic that cannot cross a geographically dispersed network. There is no distributed version of that problem. The compute has to sit inside a purpose-built building.

Inference is a different shape of problem. Each query from a Claude user is an independent computation: read the weights, generate tokens one at a time, return the response. Academic work on large-model serving has established that the binding constraint here is memory bandwidth — the rate at which a system can read model weights from memory — rather than floating-point throughput. A 70-billion-parameter model quantized to 4-bit precision needs roughly 38 gigabytes of resident weights fetched per output token. Hardware with a sufficiently deep and fast memory pool can serve interactive traffic for models in that size class without a GPU at all.

At Anthropic's scale, a large fraction of production traffic sits squarely in that category: request routing, preprocessing, orchestration between model calls, and the infrastructure management of agentic workflows that chain multiple sequential Claude invocations together. These are CPU-intensive tasks. At a company whose revenue reached a $30 billion run rate in 2026, up from roughly $9 billion at the end of 2025, they aggregate into a compute bill large enough to justify a nine-figure annual contract with a non-hyperscaler.

What Akamai Actually Sold

Akamai's answer is a three-tier stack called AI Grid, launched in March 2026 and described by the company as the first global-scale implementation of NVIDIA's AI Grid reference design. The edge tier, spanning 4,400-plus points of presence, handles ultra-low-latency requests and acts as the point of contact with the user. A semantic caching layer running on WebAssembly-based serverless compute can return frequently requested outputs without a round trip to an accelerator at all. The core tier hosts multi-thousand GPU clusters built on NVIDIA RTX PRO 6000 Blackwell Server Edition parts for heavy multimodal and continuous post-training workloads.

The whole thing is governed by an orchestration layer that evaluates tokenomics — cost per token, time to first token, and total throughput — and routes each request to whichever resource is optimal at that moment. For latency-sensitive enterprise deployments where a 200-millisecond round trip to a Virginia data center degrades an agentic workflow's user experience, edge-proximate CPU capacity offers something centralized architectures structurally cannot: compute at the point of contact. This is the same conclusion our coverage of the compute continuum in cloud strategy has been tracking for months, now confirmed with a contract number attached.

The Financial Engineering

The deal's structure is at least as unusual as its technical shape. Akamai issued Anthropic a Series B warrant covering up to 387,051 shares of newly created Series B non-voting convertible preferred stock, each share convertible into 20 common shares — a total common-stock equivalent of 7,741,020 shares, or about 5 percent of Akamai's shares outstanding. The exercise price is $111.33 per common-share equivalent, reflecting $2,226.60 per warrant share at the 20-to-1 ratio, set at Akamai's 30-day volume-weighted average price as of September 18. VentureBurn reported the warrant terms alongside the after-hours share move.

Vesting is tied directly to how much Anthropic actually commits. Roughly 40 percent of the warrant — about 2 percent of Akamai's stock — vests on Anthropic's first payment under the plan. Each additional $3 billion committed triggers another 1 percent. The full 5 percent only materializes if the relationship expands to approximately $20 billion, which the agreement permits as an option. This is the first time Akamai has attached an equity warrant to a cloud customer deal, and it converts a supplier relationship into something closer to a strategic alliance with shared downside.

The Buildout Bill

Akamai has committed approximately $5.5 billion in capital expenditure against this contract, and it started moving on supply chain before the announcement. On September 23 it signed a hardware supply agreement with Lenovo Global Technologies Ireland International Limited covering the hardware, software, and services needed for the buildout, with a three-year initial master term and a statement of work running seven years to match Anthropic's commitment. On September 24 it authorized Jabil to purchase roughly $1.7 billion in memory components under an existing master agreement, with Jabil holding the parts in consignment as bailee until Akamai consumes them. That pre-purchase is a direct response to memory supply tightening driven by AI demand, where data center operators now compete with consumer electronics manufacturers for the same high-bandwidth memory.

The balance sheet absorbed the strain. Akamai amended its credit agreement to raise its maximum consolidated leverage ratio and paused its share buyback. Revenue from the commitment does not begin until the second half of 2027, with an annualized run rate of roughly $1.7 billion targeted by the end of 2028. Management guided to no 2026 revenue impact, and the company ended 2026 with $4.6 billion in cash and $1 billion in credit lines. Morgan Stanley, per VentureBurn, estimates operating margins on the deal around 30 percent.

The Relationship Was Not New

Akamai and Anthropic signed a master services agreement on May 5, 2026, disclosed at first-quarter earnings when CEO Tom Leighton described a $1.8 billion, seven-year "landmark" commitment from an unnamed "leading frontier model company." Bloomberg later identified the customer. The September 18 signing of Project Plans 2 and 3 under that existing agreement expanded the arrangement more than sixfold. Add it to more than $2.8 billion in other multi-year cloud infrastructure commitments announced across 2026, including a $600 million four-year robotics cloud contract from August, and Akamai's year-to-date signed contract value reached approximately $14.4 billion.

Two disclosures deserve more attention than they got. First, Anthropic can terminate the master services agreement if Akamai undergoes a change of control in favor of a direct competitor of Anthropic — a real constraint on Akamai's merger options, written by a customer that now has both leverage and equity. Second, Cloud Infrastructure Services, the segment built on the Guardicore and Linode acquisitions, was already up 40 percent year-over-year in the first quarter and 39 percent in the second while the legacy content delivery business declined. This contract monetizes a three-year pivot that was already visible in the segment numbers.

The Portfolio Anthropic Is Building

The question any infrastructure team asks about a new supplier is what it adds that existing ones do not. Anthropic's other arrangements are extraordinary in scale: Amazon is investing up to $25 billion and committing up to 5 gigawatts of capacity, primarily Trainium2 and Trainium3; Google and Broadcom committed multiple gigawatts of next-generation TPU capacity arriving from 2027; SpaceX's Colossus 1 facility provides access to more than 300 megawatts and over 220,000 NVIDIA GPUs; CoreWeave supplies additional GPU inference capacity.

Every one of those is a high-performance accelerator designed for parallel matrix operations, physically concentrated in a small number of very large buildings. Akamai's network is the inverse: low-density compute spread thin across thousands of sites. That topology fits latency-bound, bandwidth-bound, CPU-heavy work rather than training or dense-batch GPU serving, which is exactly why the deal is framed as complementary rather than substitutional.

Conclusion

The durable signal is architectural, not financial. Akamai has spent more than a decade arguing that its distributed topology would matter again whenever the next compute wave arrived. That argument held in the CDN era, held through the Linode acquisition, and now has a $11.6 billion signature on it.

For enterprise teams planning AI infrastructure, the practical takeaway is that the compute market has structurally broadened past three hyperscalers. The question facing a frontier lab at Anthropic's scale is no longer simply which cloud to run on. It is now also which workloads belong at the edge, and who has already built a network capable of serving them. Akamai's answer to that question just became worth $11.6 billion, and the next test arrives in the second half of 2027, when the first revenue shows up.

Images

A rack-dense cluster room of the kind that centralized training workloads demand. Illustrative image, not an Akamai facility.

Hot-swappable drive bays and rack-mounted chassis. Edge capacity is deployed as many small dense racks rather than a few large halls.

Edge points of presence look like this at the wiring level: standard 19-inch cabinets patched locally, repeated across thousands of sites.

References

• Akamai Technologies, press release: "Akamai Announces $11.6 Billion Multi-year Agreement with Anthropic to Support Growing Demand" (akamai.com)
• TechTimes, Earl Bensen, "Akamai Lands $11.6 Billion Anthropic Deal as CPU Inference Defies GPU Consensus," September 28, 2026
• VentureBurn, Ekemini, "Akamai Signs $11.6B Seven-Year Deal With Anthropic to Power CPU Workloads on Akamai Cloud," September 25, 2026
• Data Center Dynamics, "Akamai signs $11.6bn compute deal with Anthropic"