inference hardware

Microsoft’s Maia 200 is the latest, bold step in a multi-year pivot by hyperscalers to own the silicon that runs generative AI — a purpose-built, inference-first accelerator that promises significantly lower token costs, higher utilization for large models, and a path away from sole reliance on...

Microsoft’s Maia 200 is the clearest signal yet that hyperscalers are moving from buying commodity GPUs to building inference-optimized silicon and systems — a tightly integrated hardware + software play aimed at driving down the marginal cost of serving large language models and other reasoning...

Microsoft’s Maia 200 announcement marks a decisive escalation in the hyperscaler silicon arms race: an inference‑first accelerator built on TSMC’s 3 nm process that Microsoft says is already in Azure racks and is explicitly tuned to lower the per‑token cost of running large language models like...

Microsoft has announced Maia 200, a purpose-built AI inference accelerator that the company says will give Azure a material cost and performance edge for running large language models and other production inference workloads, promising multi-petaFLOPS low-precision throughput, a high-bandwidth...

Microsoft’s Maia 200 is not a subtle step — it’s a direct, public escalation in the hyperscaler silicon arms race: an inference‑first AI accelerator Microsoft says is built on TSMC’s 3 nm process, packed with massive on‑package HBM3e memory, and deployed in Azure with the explicit aim of...

Microsoft’s Maia 200 is a purpose-built AI inference accelerator that promises to reshape how Azure runs large language models and other high‑throughput generative AI workloads, claiming dramatic gains in token-generation efficiency, a major new memory and interconnect design, and an...

Microsoft has quietly begun deploying its second‑generation in‑house AI accelerator, the Maia 200, a TSMC‑built chip Microsoft says is designed to cut the company’s reliance on external GPU vendors and deliver a step change in inference cost, power efficiency, and scale for Azure‑hosted AI...

Microsoft’s Maia 200 is not a modest chip announcement — it’s a systems-level gambit that stitches custom silicon, huge on‑package memory, an Ethernet‑based scale‑up fabric and a developer SDK into a single inference‑first platform Microsoft says will materially lower per‑token costs for Azure...

Microsoft is rolling Copilot Vision into Windows — a permissioned, session‑based capability that lets the Copilot app “see” one or two app windows or a shared desktop region and provide contextual, step‑by‑step help, highlights that point to UI elements, and multimodal responses (voice or typed)...