Before Akamai books a dollar of revenue from its new Anthropic contract, it has authorized hardware maker Jabil to buy about $1.7 billion of memory. The components will sit with Jabil on consignment until Akamai uses them, while service revenue is expected to begin only in the second half of 2027. That timing turns a large cloud headline into a more revealing infrastructure wager: Akamai is spending now on the CPU and memory layer around AI systems, and Anthropic gets a path to own part of the supplier if it keeps buying. Akamai's SEC filing lays out the purchase and contract terms, while its investor presentation gives the delayed revenue schedule.
The headline number is $11.6 billion over seven years. The machinery underneath it is more useful to developers and infrastructure teams. Anthropic's models still need accelerators for inference, yet an agent also parses results, runs code, opens files, calls APIs and moves data. Those steps consume general-purpose compute. Akamai describes the contract as support for Anthropic's growing CPU workloads, without saying which products or tasks will run there.
The $11.6 billion is conditional
Akamai and Anthropic signed the two new project plans on September 18 under a master services agreement dating from May. The plans cover dedicated cloud capacity and managed support, and each has a seven-year term starting when its service begins. Anthropic's aggregate commitment is about $11.6 billion, provided Akamai meets delivery and service-availability requirements, according to the Form 8-K.
That distinction matters. A contracted commitment is firmer than a forecast, though it is not a prepaid cheque. Anthropic can terminate a project plan after a qualifying material outage. It can also end the master agreement after an uncured material breach by Akamai, or if Akamai changes control in favor of one of Anthropic's direct competitors. Akamai has its own termination right for an uncured breach by Anthropic. The full master agreement is due to be filed with Akamai's quarterly report for the period ending September 30, so some commercial details are still absent, the filing says.
The build is front-loaded. Akamai estimates about $5.5 billion in capital expenditure tied to the current commitment. Its schedule assigns about $1.7 billion to the fourth quarter of 2026 and roughly $3.1 billion to 2027, followed by about $700 million in 2028. Akamai expects no related revenue in 2026. It projects $150 million to $300 million in 2027, with service starting late in the second quarter and ramping through the rest of that year. The company expects to reach an annual revenue rate of about $1.7 billion by the end of 2028, then recognize revenue at roughly that yearly level through the remaining term. These are Akamai estimates and may change.
The memory order shows how tangible the commitment has already become. Under Akamai's filing, Jabil will purchase the components, hold unused stock for Akamai and repurchase it at cost as it is consumed. Akamai bears supplier invoices when Jabil receives the parts. In effect, Akamai is financing a supply chain two years ahead of the expected run rate.
Why a frontier AI lab is buying CPUs
Most large AI infrastructure announcements lead with accelerators. Anthropic has plenty of those arrangements. In April, the company said an expanded Amazon agreement would secure up to five gigawatts of capacity for training and serving Claude, including AWS Trainium chips. It previously announced plans to use up to one million Google Cloud TPUs. The Akamai disclosure points at a different part of the stack: distributed CPU infrastructure and its supporting software.
A model invocation is only one part of an agent run. When an agent writes a program, for example, it may inspect a repository, edit files, install packages, execute a build and send the result back to the model. IEEE Spectrum's account of rising CPU demand says model inference generally stays on a GPU or another accelerator, while tasks such as parsing output, invoking tools and making API calls run on CPUs. One cited AMD test put seven of eight stages in a realistic agent pipeline on the CPU. That result comes from AMD's testing rather than Anthropic's workloads, but it explains why general-purpose servers can become a bottleneck even when the model itself runs elsewhere.
Akamai has not identified the Anthropic services involved, so it would be premature to label this an agent-inference contract. The disclosed CPU focus still gives engineering teams a useful signal. Capacity planning for AI applications cannot stop at tokens per second or accelerator availability. Tool execution, web requests, sandboxes, storage and orchestration may set the end-to-end limit. A coding agent waiting on a package install gains nothing from a faster matrix multiplication.
Distribution may also matter. Akamai says its cloud spans core and edge locations, but the contract materials do not map Anthropic workloads to particular regions. The investor presentation says the new Anthropic deal and Akamai's earlier large cloud commitments together require about 95 to 105 megawatts. That combined figure should not be read as Anthropic's power allocation. It does show the physical scale Akamai expects to provision across this part of its cloud business.
Anthropic can earn a stake by spending more
The strangest part of the arrangement is the warrant. Akamai gave Anthropic the right to buy non-voting preferred shares equivalent to as many as 7.7 million common shares, or about 5% of Akamai's outstanding stock, at a common-share equivalent exercise price of $111.33. The first tranche represents 40% of the warrant, equal to about 2% of Akamai's outstanding shares, and vests after Anthropic or an affiliate makes the first payment under one of the new project plans. The filing says the warrant remains exercisable until the seventh anniversary of its issue.
The rest tracks new business. Each additional $3 billion of contractual value that Anthropic commits unlocks another tranche equivalent to roughly 1% of Akamai's shares. Three such steps would add $9 billion to the relationship, taking its possible value to about $20 billion and vesting the remaining 3%. Anthropic still has to pay cash to exercise the warrant, according to the filing. More spending does not hand it free stock.
This structure gives the customer a financial interest in the infrastructure company it is helping to expand. TechCrunch reported that this is Akamai's first cloud contract with a warrant and its largest agreement. The article also notes the direction of the incentive: instead of a supplier investing in the AI lab that buys its capacity, the supplier is giving the customer a route to a stake. Akamai's own presentation warns that the initial 3.1 million common-share equivalents will be dilutive under its accounting assumptions, with greater dilution possible if its share price rises and Anthropic signs more commitments.
The warrant also affects reported revenue. Akamai says the fair value of the first vested portion will reduce total revenue recognized from the $11.6 billion agreement. Contract value, cash received, recognized revenue and retained profit are different numbers. The build cost, service milestones and warrant accounting will come into view over several reporting periods.
What developers should watch
For people building with Claude, the disclosures contain no new API, region, price, workload map or latency promise. Their immediate value is as evidence about where a frontier lab expects pressure: in the ordinary compute surrounding models as well as the accelerators that run them.
The first meaningful check comes in 2027. Akamai says service should start late in the second quarter, followed by a revenue ramp in the second half. Watch whether that timetable holds, whether the company revises its $5.5 billion build estimate, and whether Anthropic commits another $3 billion and vests the next warrant tranche. Those events will show whether the $1.7 billion memory order was the opening purchase for a durable CPU business or an expensive reservation made before demand became visible.