Akamai's potential $20B Anthropic deal scales AI beyond GPUs
The new agreement calls for CPU infrastructure that can fit within conventional co-location facilities, affirming that a distributed data center model is emerging for AI workloads.
A new partnership between Akamai and Anthropic demonstrates that AI infrastructure demand is extending beyond the massive data center campuses being built for GPU clusters.
Akamai said Sept. 24 that it had signed an $11.6 billion, seven-year commitment with Anthropic to provide distributed cloud infrastructure for the AI company’s growing CPU workload. The agreement can expand by another $9 billion, bringing the potential value of the relationship to approximately $20 billion.
For data center operators, this emerging opportunity could apply to facilities with conventional power and cooling infrastructure as AI companies add capacity for workloads surrounding model training, which lend themselves to more conventional CPU hardware.
"The critical difference is rack density," compared with GPU-based server clusters, said Dave McCarthy, group vice president, cloud and datacenters, enterprise infrastructure at IDC.
GPU clusters can draw from 40 kW to more than 100 kW per rack and require specialized liquid cooling, he said, while CPU environments typically operate at 10 kW to 20 kW per rack.
That lower density allows CPU workloads to fit into existing co-location facilities without the power and cooling requirements of large GPU deployments, McCarthy said.
Akamai did not specifically disclose the power requirement for the Anthropic contract. The company said the $14.4 billion portfolio of large, multiyear cloud contracts it has signed this year, including the Anthropic agreement, will require 95-105 MW of power.
Akamai will assemble capacity across multiple sites
Akamai said it will use its distributed cloud infrastructure and work with multiple co-location providers rather than build its own data centers for the contracts.
The company can secure roughly 10-30 MW at individual locations, said Akamai CFO Ed McGowan during an investor call Sept. 24. Some of that capacity is already secured.
"We have many, many providers we work with," he said. "We are able to get agreements for multi-years for say, 10-30 MW in certain locations."
That approach could create an opportunity for data center operators with available capacity that cannot accommodate the largest GPU deployments hyperscalers seek.
"Operators don't all need 500 MW gigawatt campuses with bespoke liquid loops to win AI business," McCarthy said.
Operators with 10-30 MW parcels in secondary metros can instead provide what McCarthy described as a distributed serving tier for AI application pipelines.
Those sites can use air-cooled, standard-power enterprise infrastructure rather than the specialized cooling systems required by the densest GPU deployments, he said.
AI workloads are moving beyond GPU clusters
The CPU requirement reflects the increased infrastructure diversity supporting AI applications as workloads move beyond model training, McCarthy said.
Training frontier models requires tightly packed GPU clusters connected by high-speed networks. Application serving, preprocessing, context caching and API termination have different requirements and can benefit from proximity to end users, data sources and client systems.
CPUs have a different dynamic … you can get a lot more CPU [capacity], therefore generally more revenue per dollar of megawatt of power.
Ed McGowanCFO, Akamai
CPU workloads can include agentic orchestration and tool calling, context and memory management, and pre- and post-processing and guardrails, McCarthy said.
Akamai's McGowan made a similar point during the investor call.
"There's an awful lot of CPU that's needed to run all this," McGowan said. "It's not all just GPU."
McGowan said CPU deployments can also generate more revenue per megawatt due to their power efficiency.
"CPUs have a different dynamic … you can get a lot more CPU [capacity], therefore generally more revenue per dollar of megawatt of power," he said.
Akamai is spending before the revenue arrives
The Anthropic agreement also illustrates the capital required to build the capacity before the associated revenue arrives.
Akamai expects to spend approximately $5.5 billion in capital expenditure over the next two years to support the $11.6 billion commitment.
About $1.7 billion of that spending is expected in the fourth quarter of 2026, primarily to secure and pre-purchase supply chain components, including memory.
Akamai expects to spend another $3.1 billion in 2027 and approximately $700 million in 2028.
The company expects $150 million to $300 million in revenue from the Anthropic contract during 2027, with the deployment beginning to generate revenue in the second half of the year. Akamai expects the contract to reach an annualized revenue run rate of approximately $1.7 billion by the end of 2028.
Akamai's broader $14.4 billion portfolio of large cloud contracts is expected to generate approximately $2.2 billion in annual recurring revenue once fully deployed and to require 95-105 MW of power.
The Anthropic relationship could add another $9 billion if the expansion option is exercised, with additional capital expenditures required.
For data center operators, this spending plan offers an opportunity to capture AI infrastructure spending without incurring the capital requirements of a massive GPU campus. McCarthy said operators with 10-30 MW parcels can provide the distributed capacity AI applications need using conventional power and cooling infrastructure.
Shane Snider is a senior news writer at TechTarget, covering AI infrastructure, hyperscale data centers, cloud platforms, and the power and energy systems driving modern compute expansion. You can reach Shane at [email protected] or on LinkedIn.
