Summary: On March 3, 2026, Akamai said it bought thousands of NVIDIA Blackwell GPUs and would deploy them across its global distributed cloud. The article’s stack includes that cloud together with an edge network of more than 4,400 locations. It reports claimed gains of up to 2.5x latency improvement and 86% lower cost, while noting that baselines and traffic patterns are unpublished.
On March 3, 2026, Akamai (NASDAQ: AKAM) announced that it has bought thousands of NVIDIA Blackwell GPUs and will deploy them across its global distributed cloud, with the stated goal of creating one of the world’s most widely distributed AI platforms. Training is not the story here. Inference is.
The company that began life as a CDN is wagering on a particular shift: AI’s center of gravity is moving out of the handful of mega data centers where training is concentrated, toward inference that runs near the people using it. Akamai points to MIT Technology Review, which found that 56% of organizations see latency as the number-one obstacle to deploying AI at scale. Data Center Dynamics followed a day later with a footnote of its own: the exact GPU count was not disclosed.
The Case for an Inference-First Era
Akamai’s argument pulls no punches. AI’s first wave packed training into a few hyperscale regions, and the industry has since reached the point where inference matters as much as training. For models to graduate from the lab into autonomous delivery, smart grids, surgical robotics, and real-time fraud prevention, decisions have to happen at real-world speed — and round-tripping a distant hyperscaler region, with the latency and egress costs that entails, is now the bottleneck.
Adam Karon, Akamai’s COO and general manager of its Cloud Technology Group, frames the split this way: hyperscalers will keep advancing the frontier of training, but Akamai is focused on the “inference era.” Centralized AI factories remain necessary; what scaling AI into applications calls for is a “decentralized nervous system” that places inference compute at “the street corner and the hospital bed: where the work is happening, where the data lives, and where the ROI is realized.”
What Goes Into a “Decentralized Nervous System”
On paper, the stack reads as follows: NVIDIA RTX PRO servers running RTX PRO 6000 Blackwell Server Edition GPUs, paired with BlueField-3 DPUs, layered on Akamai’s own distributed cloud and an edge network of more than 4,400 locations worldwide. Three jobs define the platform: predictable, high-performance inference on dedicated GPU clusters; fine-tuning LLMs locally so data stays in-region for privacy and compliance; and post-training optimization against private datasets.
This is not a case of racking GPUs in a warehouse. It is an orchestration system that treats model placement as a variable and intelligently routes inference workloads to the location nearest the user.
The Claims: Up to 2.5x Better Latency, Up to 86% Lower Cost
The headline numbers: versus traditional hyperscaler infrastructure, running inference on NVIDIA AI infrastructure improves latency by up to 2.5x and cuts costs by as much as 86%. Treat both as marketing until you benchmark your own workload — the baselines and traffic patterns are not published. But the direction is credible: inference rewards proximity and minimal egress, and moving bits closer to users is the business a CDN has run for two decades. The only new part is moving the GPUs along with them.
The demand signal matters too. Akamai says its initial RTX PRO 6000 Blackwell deployments have seen “strong demand,” and it will keep adding GPU capacity. Edge inference is no longer a proof of concept; it is a product with a queue.
The Road From CDN and Linode to the Inference Cloud
This buildup took years. Akamai acquired cloud provider Linode in 2022 for roughly $900 million to gain IaaS capabilities, then unveiled Akamai Inference Cloud in October 2025 to pull inference workloads closer to users. The large-scale Blackwell buy extends that same line: CDN-grade global connectivity and caching, stacked with GPU clusters and DPUs, yields an alternative architecture that competes with hyperscalers for inference demand head-on.
What Developers and Platform Teams Should Take Away
Three observations. First, inference is acquiring a vendor spectrum of its own: training remains a fight between hyperscalers and sovereign clusters, while inference is territory that distributed-network players like Akamai can carve off — especially latency-sensitive, data-residency-sensitive workloads such as agentic AI, physical AI, and compliance-driven local fine-tuning. Second, architects gain a new option in the model-placement era: routing an LLM will increasingly mean choosing not just a model but an inference location. Third, cost structures shift — the 86% saving will not land wholesale on your bill, but egress fees and latency penalties are two of the most invisible lines in current AI bills, and both are worth testing yourself before committing.