Anthropic Akamai Cloud Deal Makes CPUs a Multibillion-Dollar AI Bet

Anthropic committed $11.6 billion to Akamai over seven years, turning the Anthropic Akamai cloud deal into a major bet on CPU-based AI infrastructure.

The agreement stands out because the current AI infrastructure race usually centers on GPUs, custom accelerators, and enormous training clusters. Anthropic instead wants dedicated CPU capacity and managed services across Akamai’s distributed cloud.

The arrangement also reverses the usual relationship between a cloud provider and its largest customer. Akamai issued Anthropic a warrant that can grow into an economic interest representing approximately 5% of Akamai’s outstanding common stock.

Amazon, Google, CoreWeave, and emerging infrastructure providers remain part of Anthropic’s wider compute strategy. Akamai is not replacing them. It is taking responsibility for a different and increasingly valuable part of Claude’s workload.

The Anthropic Akamai cloud deal is bigger than its headline number

Anthropic has committed to buy dedicated capacity, but Akamai must first build and operate the infrastructure needed to earn that revenue.

Akamai announced the agreement on September 24, 2026. Its material agreement filing says the companies signed two project plans on September 18.

Those plans sit under a master services agreement dated May 5, 2026. Akamai will provide dedicated cloud computing capacity and related managed support services.

The $11.6 billion commitment covers seven-year initial terms beginning on each project’s service start date. Payment remains subject to Akamai meeting specified delivery and service-availability requirements.

That distinction matters. The figure represents contracted commercial value, not revenue that Akamai received when the agreement was signed.

Akamai expects the initial infrastructure build to require approximately $5.5 billion in cumulative capital expenditures. It plans to spend about $1.7 billion during the fourth quarter of 2026, primarily to secure memory and other constrained components.

Another estimated $3.1 billion of capital spending follows in 2027. Akamai expects initial service to begin late in the second quarter, with revenue increasing during the second half.

The company then expects approximately $700 million of remaining capital expenditures in 2028. Its published capacity timeline targets the full contracted run rate by the end of that year.

After reaching steady state, Akamai forecasts about $1.7 billion in annual revenue through the remainder of the term. These figures are management estimates, not guaranteed outcomes.

The transaction will not affect Akamai’s 2026 revenue guidance because the new capacity is not ready to serve Anthropic. The near-term effect is cash leaving Akamai to fund equipment before corresponding service revenue arrives.

That sequence creates the article’s central tension. Anthropic secures long-term capacity while Akamai assumes significant construction, supply-chain, and execution risk.

The contract includes protections for both parties. Akamai can terminate the master agreement following an uncured breach by Anthropic.

Anthropic can terminate after an uncured material breach by Akamai. It also has termination rights following certain material outages or an Akamai change of control favoring one of Anthropic’s direct competitors.

Separate project plans can sometimes survive a breach affecting another plan. That structure reduces the chance that one disputed workload automatically ends the entire relationship.

Akamai has also arranged the supply chain behind its commitment. It signed a seven-year statement of work with Lenovo for hardware, software, and related services.

Akamai separately authorized Jabil to acquire approximately $1.7 billion in memory components. Jabil will hold unused components on consignment until they enter production.

The hardware commitments show that this is more than a marketing partnership. Akamai has already begun translating Anthropic’s demand forecast into servers, memory, and data-center capacity.

Why Anthropic needs CPUs when AI spending revolves around GPUs

The agreement separates the infrastructure required to create an AI model from the infrastructure required to operate a growing AI service.

GPUs and specialized accelerators perform the dense parallel calculations used in model training and many inference tasks. Inference is the process of running a trained model to generate an answer or take an action.

However, a production AI service contains far more than model inference. It must receive requests, authenticate users, retrieve data, run software tools, manage containers, filter outputs, and deliver responses.

Those supporting jobs often run efficiently on conventional server CPUs. They can also grow rapidly when an AI system becomes more agentic.

An AI agent does not simply produce one answer. It may plan a task, search stored information, call external tools, execute code, inspect results, and repeat the process.

Each loop generates orchestration work around the model. CPUs handle much of the scheduling, data movement, application logic, and tool execution surrounding the expensive accelerator calls.

Anthropic and Akamai have not published a detailed workload inventory. Readers should therefore avoid assuming that every server covered by the deal will directly generate Claude responses.

Akamai describes the deployment as infrastructure for Anthropic’s accelerating CPU workload demand. The company says its distributed architecture can support applications from centralized cloud regions to locations nearer users.

That network design is relevant because inference systems do not face one universal performance problem. Some tasks need enormous accelerator clusters, while others benefit from placing application services closer to customers.

Akamai operates infrastructure across a broad geographic footprint inherited from its content-delivery business. It later expanded into general cloud computing, including through its acquisition of Linode.

That history gives Akamai experience with distributed networking and application delivery. It does not automatically establish that Akamai can execute a build of this size on schedule.

The company’s existing cloud infrastructure business remains much smaller than the announced commitment. Akamai reported $99 million in cloud infrastructure services revenue during the second quarter of 2026.

That represented 39% year-over-year growth, according to its quarterly cloud results. The Anthropic contract is therefore both a validation and an enormous scaling test.

The agreement follows more than $2.8 billion in other multiyear cloud infrastructure commitments that Akamai announced during 2026. Yet the Anthropic deal is several times larger than that combined figure.

It also changes how investors must evaluate Akamai’s cloud ambitions. The business is no longer only a gradual extension of a content-delivery network.

Akamai is committing billions of dollars to dedicated capacity for one AI developer. Its future cloud results will become closely connected to Anthropic’s demand and Akamai’s delivery performance.

For Anthropic, adding Akamai creates another specialized infrastructure lane. The company can reserve accelerator capacity from other providers while sending suitable CPU workloads to a distributed platform.

That division may improve flexibility and reduce pressure on more expensive accelerator environments. The companies have not disclosed enough technical detail to quantify any savings.

The deal’s significance therefore lies in its workload split. GPUs remain central to frontier AI, but the systems around them now require their own industrial-scale procurement strategy.

Akamai is competing with a route, not one cloud company

The primary contest is between a diversified infrastructure portfolio and the idea that one hyperscale cloud should handle nearly everything.

Anthropic has not abandoned the largest cloud platforms. Its expanding relationship with Akamai instead shows how frontier AI developers are assembling capacity across providers, chips, and regions.

Amazon remains Anthropic’s primary cloud provider and training partner. Their expanded Amazon compute pact covers up to five gigawatts of capacity and includes multiple generations of Trainium accelerators.

Claude is also available through Google Cloud and Microsoft Azure. Anthropic uses Google’s tensor processing units and has pursued further capacity through several specialist infrastructure companies.

These relationships address different combinations of model training, inference, product distribution, geographic reach, and supply assurance. They should not be treated as interchangeable contracts.

Akamai’s deal is explicitly focused on CPUs. That puts it beside Anthropic’s accelerator portfolio rather than directly against every AWS Trainium or Google TPU deployment.

The competitive pressure falls on providers that want customers to keep all supporting workloads inside one cloud. Anthropic is showing that a major model developer can separate those jobs and procure them independently.

That choice can improve negotiating leverage. A company with multiple infrastructure paths is less exposed to one supplier’s capacity schedule, hardware roadmap, or commercial terms.

Diversification also creates operational costs. Anthropic must coordinate software, security, monitoring, data movement, and reliability across a growing collection of environments.

Workloads cannot always move freely between providers. Differences in processors, networking, storage, orchestration systems, and management tools can create migration work.

Data location introduces another complication. Distributing services can shorten the network path to customers, but moving information among clouds may increase latency, transfer costs, and governance demands.

Anthropic must therefore decide which work belongs on Akamai and which work should remain beside its accelerator clusters. That boundary will influence whether the arrangement produces lasting efficiency.

Akamai must also prove that its platform can function as more than supplemental capacity. Dedicated infrastructure has limited strategic value if deployment delays or outages force Anthropic to maintain equivalent backup capacity elsewhere.

This is why the Anthropic Akamai cloud deal is not a simple challenge to AWS, Google Cloud, or Microsoft Azure. It challenges the all-in-one procurement model those companies often encourage.

The approach also reflects the scale of Anthropic’s capacity requirements. The developer cannot assume that any single supplier will always have the right hardware available in the required location.

A broader supplier base protects against shortages, but it does not remove them. Akamai’s immediate memory purchases reveal how quickly an AI contract can become a supply-chain obligation.

The model also puts smaller infrastructure companies under pressure. They must decide whether to fund large customer-specific builds, even when their balance sheets cannot absorb hyperscaler-sized mistakes.

Akamai has an established security and delivery business that can help finance expansion. Even so, its planned capital spending is large relative to the cloud operation it has built so far.

The result is a new competitive map. AI labs control valuable demand, while infrastructure companies compete to finance and deliver specialized pieces of the computing stack.

The stock warrant shifts risk back toward Akamai shareholders

Anthropic receives more than capacity because additional spending can increase its economic exposure to the company providing that capacity.

Akamai issued Anthropic a warrant for nonvoting convertible preferred stock. A warrant gives its holder the right to buy securities at a predetermined exercise price after specified conditions are met.

The instrument covers up to 387,051 preferred shares. Each preferred share initially converts into 20 common shares, producing a maximum equivalent of 7,741,020 Akamai shares.

That represents approximately 5% of Akamai’s outstanding common stock on an as-converted basis. The exercise price corresponds to $111.33 for each underlying common share.

The first tranche represents 40% of the warrant. It vests after Anthropic or an affiliated entity makes the first payment under the relevant project plan, subject to contractual conditions.

That tranche corresponds to approximately 2% of Akamai’s outstanding common stock. The remaining three tranches each represent about 1%.

Each later tranche vests when Anthropic commits another $3 billion in contractual value. Three additional commitments would add $9 billion and bring the potential relationship to approximately $20 billion.

The warrant therefore functions as a spending incentive. As Anthropic expands the commercial relationship, more of its potential stake becomes exercisable.

The structure also aligns Anthropic with Akamai’s performance. If the cloud expansion improves Akamai’s value, Anthropic can participate in some of that upside after exercising vested warrants.

However, alignment is only one interpretation. Existing shareholders must also consider dilution, which occurs when new shares reduce their proportional ownership.

The maximum 5% figure is meaningful, particularly because the award goes to the customer responsible for the contract. Akamai is effectively using potential equity participation to strengthen a major commercial relationship.

Akamai chief executive Tom Leighton reportedly described this as the first time the company had offered such a warrant within a cloud customer deal. That makes the structure an important precedent, not routine contract language.

The arrangement resembles a wider pattern in AI infrastructure. Providers increasingly combine capacity agreements with investments, warrants, financing, or other strategic ties.

Those links can accelerate construction by giving each party a reason to support the other’s growth. They can also make the economics harder to interpret.

A large headline commitment does not reveal the full cost of winning the customer. Analysts must account for equipment spending, financing needs, operating expenses, service credits, and potential dilution.

The deal announcement estimates strong cash conversion and operating margins once the deployment reaches scale. Independent operating results will be needed to test that claim.

Customer concentration presents another risk. The expected steady-state revenue would make Anthropic unusually important to Akamai’s cloud infrastructure segment.

A material outage could activate termination rights. A demand slowdown could also leave questions about how reusable the dedicated equipment would be for other customers.

The contract protects Akamai against certain breaches, but legal commitments do not eliminate deployment risk. Akamai must still secure components, construct capacity, and meet service levels.

Anthropic faces a different uncertainty. It is reserving infrastructure years before anyone can know exactly how AI model architectures or application workloads will evolve.

CPU demand may continue growing as agents perform more tools and application work. Software optimization could also change the amount or location of compute required for each task.

The unusual equity arrangement acknowledges that uncertainty indirectly. Both companies benefit more if the relationship expands, yet Akamai supplies much of the capital before that expansion becomes certain.

What the $11.6 billion commitment still does not prove

The contract validates demand for Akamai’s cloud, but it does not yet prove delivery, utilization, profitability, or technical advantage.

The most immediate uncertainty is execution. Akamai expects to spend billions before the deployment reaches its full revenue run rate.

Memory availability, server manufacturing, data-center construction, power delivery, and network integration must all proceed on schedule. A delay anywhere in that chain can postpone service and revenue.

The SEC filing identifies competition, capacity constraints, supply disruptions, security incidents, and cash-flow pressure among the transaction’s risks. These are specific concerns for a build of this size.

Akamai’s $1.7 billion memory commitment through Jabil is particularly notable. Buying components early can protect the schedule, but it also creates inventory exposure.

If requirements change, some hardware may need to be repurposed. The filing includes provisions covering unused inventory, carrying costs, and disposal.

The second uncertainty is workload visibility. Neither company has disclosed how many servers, CPU cores, data centers, or geographic regions the project requires.

They have not identified the processor suppliers in the public contract summary. They also have not explained how workloads will be divided between Akamai and Anthropic’s other clouds.

Without those details, claims about cost advantages or performance remain speculative. The contract confirms demand, not a benchmark victory.

The third uncertainty is revenue timing. Initial service is not expected until late in the second quarter of 2027, while the full run rate arrives at the end of 2028.

Investors will see substantial capital spending before they see the complete revenue effect. Management’s current projections can change as delivery milestones become clearer.

The fourth uncertainty concerns Anthropic’s total infrastructure exposure. The company has announced or reportedly pursued commitments across Amazon, Google, Akamai, and specialist providers.

Those announcements do not share one reporting standard. Some describe firm spending, while others include options, capacity ceilings, contingent tranches, or reported negotiations.

Adding every headline figure together would create a misleading total. Contract terms, deployment dates, and overlapping workloads matter more than a raw sum.

The Akamai agreement offers more transparency than many reported deals because a public company filed material terms with regulators. Even here, the full master services agreement will not appear until a later quarterly filing.

The fifth uncertainty is customer demand. Anthropic is reserving infrastructure based on expectations for Claude usage and future AI agents.

If adoption exceeds expectations, Akamai’s optional expansion path becomes valuable. If adoption weakens, long-duration commitments can pressure Anthropic’s economics.

This risk is common across the sector. Frontier AI companies must secure capacity before they know precisely how customers will use models several years later.

Waiting carries its own cost because desirable power, memory, and server supply can disappear. Committing early transfers forecasting risk into the contract.

Anthropic’s choice suggests that management considers insufficient CPU capacity a greater threat than reserving too much. The agreement alone cannot establish whether that judgment is correct.

For developers and enterprise buyers, the practical issue is reliability. More infrastructure diversity should create additional room for Claude workloads, but capacity announcements do not immediately remove limits or outages.

Customers should watch actual product availability, latency, regional coverage, and service stability. Those outcomes will show whether the infrastructure program reaches users.

For knowledge workers, the same distinction applies. Larger compute commitments matter only when they produce more dependable tools, longer-running agents, or fewer interruptions in daily workflows.

Teams adopting AI should continue building a searchable record of outputs, decisions, and source material. A structured AI knowledge base reduces dependence on any single model session or provider.

Three signals will reveal whether the CPU strategy works

The next evidence should come from deployment milestones, financial conversion, and further warrant vesting, in that order.

The first signal is Akamai’s initial service launch, scheduled for late in the second quarter of 2027. That milestone tests whether the company can turn supply commitments into functioning capacity.

A timely launch would strengthen the case that Akamai can execute infrastructure projects far larger than its existing cloud business. A delay would weaken confidence in the planned revenue schedule.

The launch should also produce more technical information. Processor choices, deployment regions, workload categories, and service architecture would clarify why Anthropic selected Akamai.

The second signal is Akamai’s reported cloud infrastructure revenue and capital efficiency. Revenue should begin rising during the second half of 2027 before approaching the planned run rate in 2028.

Investors should compare that growth with capital spending, depreciation, operating expenses, and cash generation. Revenue alone will not show whether the contract creates attractive returns.

Service reliability belongs in the same test. Anthropic’s outage termination rights make uptime a commercial issue rather than a general product metric.

The third signal is any additional $3 billion commitment from Anthropic. Each such commitment would vest another warrant tranche representing approximately 1% of Akamai’s outstanding common stock.

An added tranche would indicate that Anthropic needs more capacity and remains comfortable expanding the relationship. It would push the agreement closer to its potential $20 billion scale.

No additional commitment would not automatically mean failure. Anthropic could simply have enough capacity from the original project or from other suppliers.

The timing and stated purpose of any expansion will matter. More CPU demand would support the thesis that production AI requires an enormous supporting layer beyond accelerators.

The Anthropic Akamai cloud deal ultimately tests a less visible part of the AI boom. Training chips attract attention, but reliable AI products also require conventional computing, networking, memory, and operations.

Akamai now has an opportunity to turn that supporting layer into a much larger business. It also carries the burden of financing and delivering capacity for a customer whose needs continue to change.

Anthropic gains another path for scaling Claude and a potential stake in its supplier. Whether that structure becomes a model for future cloud deals depends on what reaches production.

Watch the service start, then the revenue, and finally the warrant tranches. Those three signals will show whether this was disciplined infrastructure planning or an expensive wager on demand that arrived too early.