Key Takeaways
- Akamai’s investment in its global AI infrastructure comes in the shape of thousands of new GPUs.
- Its senior vice president, Jon Alexander, said AI hardware is driving rack densities toward the 100kW era.
- And thanks to the sheer demand and scale of AI inference, the future of it may just move to the edge.
Cloud computing and security leader Akamai just obtained thousands of NVIDIA Blackwell GPUs to expand its infrastructure — all in the name of keeping up with AI inference.
“We’re seeing our partners re-architect their infrastructure specifically to support the coming explosion of AI agents and machine-to-machine interactions,” Jon Alexander, Akamai’s Senior Vice President of Cloud Technology, told HostingAdvice.
Akamai specifically bought NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, which are designed for enterprise data centers and include 96GB of GDDR7 memory.
Instead of running AI only in giant centralized data centers, Akamai wants to run AI closer to users and machines by distributing GPUs across its global network. It’s simple: At the edge, AI systems can operate faster and at a lower cost.
“We’ve engineered our core-to-edge platform to give customers the flexibility they need to place workloads and route data where it needs to be based on their specific governance, compliance, or regulatory requirements,” Alexander explained.
Akamai plans to deploy the GPUs across its global edge network, which includes more than 4,200 locations worldwide. Its GPU cloud regions already operate in major cities, including Chicago, Seattle, Frankfurt, Paris, Singapore, Osaka, and Mumbai.
The 100kW Rack Era
AI inference is fast-pushing rack power usage: Akamai says infrastructure that once ran 10kW-15kW racks are now hovering around 40kW. Alexander added there’s a “clear trajectory toward 100kW or more.”
Gartner estimates that by 2030, 75% of enterprise AI spending will be focused on inference. And that’s probably because CIOs expect AI to be involved in essentially every IT workflow by 2030.
Think about that for a moment: In just a few years, very little technical work will take place without some form of AI in the loop. It’s no wonder that the data centers we’ve built now need to be optimized to meet those demands.
But one of the biggest concerns surrounding the AI data center boom is the strain it could place on resources and the environment. Data centers already consume 1-3% of the world’s electricity, and that share is expected to grow over the next coming years.
Cooling is a major contributor: The International Energy Agency says cooling systems alone can account for 7% of energy use in efficient hyperscale facilities and more than 30% in less efficient enterprise data centers.
To deal with the heat generated by dense AI hardware, many operators are turning to liquid cooling. But those systems can require large amounts of water, pushing some data centers to rely on recycled wastewater, rainwater, or other non-potable sources instead of municipal drinking water.
Akamai hears this loud and clear — Alexander says it’s working hard to ensure that all of its AI inference data centers are built with sustainability in mind.
“We are focusing our investment into data center facilities with very low PUE ratings to ensure less energy is wasted,” he said. “Our hardware engineering teams are also actively working with our OEMs to optimize air flow and cooling architecture, including rear door heat exchangers.”
AI Is Moving to the Edge
AI happens in two stages: training and inference stages.
Training builds the AI model, and once the model is built, most of the work moves to “inferencing” — that is, running the model to do everything today’s modern user wants: answer questions, power their systems, and detect security issues as they come.
Because those responses need to happen quickly, companies like Cloudflare, AWS, Cisco, Google, and Akamai are pushing AI closer to users and devices rather than routing every request through distant cloud regions. Distributed networks like Akamai’s help deliver those responses faster and at a lower cost.
“The goal is no longer just providing rack space; it’s about providing the intelligent, high-speed distribution necessary to place workloads exactly where the use case demands,” said Alexander.
