Akamai pushes edge inference

Artificial intelligence inference is the part that answers a prompt, labels an image, or makes a decision after a model is trained. Akamai said it is now spreading that work across its edge network instead of sending every request back to a few central data centers. (edgeir.com) On April 14, Edge Infrastructure Review reported that Akamai’s system orchestrates inference across 4,400 edge locations using Nvidia graphics processing units. Akamai calls the service Akamai Inference Cloud and says it launched late in 2025 as a global-scale implementation of Nvidia AI Grid. (edgeir.com) The basic pitch is distance: the closer the graphics processing unit is to the user, device, or factory sensor, the less time a request spends crossing networks. Akamai said its software routes workloads across edge, regional, and core infrastructure to balance latency, cost, and performance. (finance.yahoo.com) Akamai is trying to turn a network built for content delivery into a distributed cloud for real-time artificial intelligence. Its own materials describe the platform as designed for “latency-sensitive inference at global scale,” with compute placed near users rather than concentrated in a few hyperscale campuses. (akamai.com) That matters for workloads that cannot wait for a round trip to a distant region, including fraud checks, industrial robots, and vehicle systems. Edge Infrastructure Review said Akamai has framed those as examples for inference placed closer to users, data, and control loops. (edgeir.com) Akamai had already been laying the hardware groundwork. Its cloud documentation says customers can run managed Kubernetes clusters on Nvidia RTX PRO 6000 Blackwell Server Edition and Nvidia RTX 4000 Ada graphics processing units, and can size systems from one to four Blackwell cards for inference-heavy jobs. (techdocs.akamai.com 1) (techdocs.akamai.com 2) The company has also been buying bigger clusters deeper in the network. Akamai said on March 5 that a four-year, $200 million service agreement with a major United States technology company would use a multi-thousand Nvidia Blackwell graphics processing unit cluster for high-performance artificial intelligence compute. (ir.akamai.com) Nvidia is pushing the same split between giant training clusters and widely distributed inference. A session posted for Nvidia’s 2026 GPU Technology Conference said “AI factories” suit model training, while production inference and agentic workflows increasingly need lower-latency infrastructure closer to where interactions happen. (nvidia.com) Akamai’s bet is that its edge footprint, built over decades for web traffic and security, can become a place to run the answering side of artificial intelligence. The next test is whether developers move enough live inference workloads onto that network to make 4,400 sites more than a coverage map. (edgeir.com)