Edge AI: hybrid compute push
Industry reporting says edge AI inference is set for rapid growth as workloads shift from cloud‑centric training toward heterogeneous inference at the network edge, and Nokia and Blaize have expanded a hybrid compute stack in Singapore. (digitimes.com) The framing emphasized hybrid, heterogeneous architectures rather than a single accelerator class. (digitimes.com)
Artificial intelligence is moving closer to where data is created, and Nokia and Blaize are building that shift around a hybrid stack in Singapore rather than a single chip. (digitimes.com) In plain terms, training is the part where a model learns from huge datasets in centralized data centers, while inference is the part where that trained model answers a prompt, flags a defect, or recognizes an object in real time. Blaize and Nokia said the next phase of adoption will be driven by inference at scale, with edge devices handling some work locally and cloud graphics processing units handling heavier jobs. (ir.blaize.com) The two companies first signed a memorandum of understanding on January 27, 2026, through Nokia Solutions and Networks Singapore Pte. Ltd., to develop hybrid inference systems across Asia-Pacific. That agreement covered telecom, industrial, and smart infrastructure use cases and paired Blaize’s inference platform with Nokia networking, automation, and cloud infrastructure. (prnewswire.com) On March 31, Blaize said the partnership had moved into joint validation at the Nokia Network Innovation Lab in Singapore, with a reference architecture that links edge and data center environments. The companies then took that work to GITEX Asia 2026 in Singapore on April 9 and April 10. (prnewswire.com; gitexasia.com) That architecture is aimed at a practical problem: companies often have to stitch together chips, software frameworks, networking, security, and management tools from different vendors before an artificial intelligence system can go live. Nokia’s Asia-Pacific artificial intelligence and cloud head Dion Leung told Light Reading the goal is a “pre-integrated validated stack” that combines inference hardware, networking, and automation. (lightreading.com) The pitch is also about cost and power. Leung said many edge workloads cannot justify the economics or electricity demands of deploying graphics processing units everywhere, so the companies are trying to place workloads across “far edge” and “near edge” locations to lower capital and operating costs. (lightreading.com) Blaize said customers are already shifting budgets toward deployment. Joseph Sulistyo, Blaize’s senior vice president of global marketing, told Light Reading that cloud providers are reporting 15% to 50% growth in inference-related workloads as customers move from experiments to production systems. (lightreading.com) DigiTimes went further, reporting on April 14 that edge inference could grow tenfold as generative artificial intelligence demand moves away from cloud-centric training and toward distributed execution. The same report said the Nokia-Blaize buildout in Singapore is framed around heterogeneous computing, meaning different kinds of processors share the job instead of one accelerator doing everything. (digitimes.com) Singapore is the test bed, but the target market is broader. Blaize said the January agreement was aimed at Asia-Pacific, and Light Reading reported the companies see both mature markets such as Singapore, Japan, and Australia and faster-growing markets such as India, Indonesia, and Vietnam as part of the rollout path. (prnewswire.com; lightreading.com) For now, the story is less about a new model than about the plumbing around it. Nokia and Blaize are betting that the next artificial intelligence buildout will be won by systems that decide, task by task, what runs on the device, what runs nearby, and what still belongs in the cloud. (ir.blaize.com)
