Infrastructure for the frontier era: Scaling intelligence through open industry collaboration

Key Points

• Azure Maia 300 targets 40% more tokens per watt than leading GPU systems and 7x rack-scale throughput over Maia 200.
• Cobalt 200 cuts agent call latency by 33% and boosts agentic workload throughput 23% versus Cobalt 100.
• Microsoft contributes 800V DC power, zonal cooling, and open HSM specs to the Open Compute Project for industry-wide adoption.

What is changing

Microsoft unveiled its next infrastructure layer for AI at the OCP Global Summit 2026. The Azure Maia 300 accelerator is built for inference-heavy workloads like long-context reasoning and agentic tasks. In a 96-accelerator rack, it delivers up to seven times the token throughput of Maia 200 while improving tokens per watt by 40 percent versus leading GPU systems. Maia 200 is already running in U.S., European, and Australian datacenters, powering Teams, Copilot, and Cowork with 1.7x better efficiency.

The Cobalt 200 CPU adds more than 50 percent performance over its predecessor and reduces agent call latency by a third. Microsoft also detailed its 800-volt DC power architecture (Mt. Diablo), third-gen liquid cooling (Mt. Atlas), Multipath Reliable Connection networking for 100,000-plus GPU clusters, and the Tessera Integrated HSM open-source security module developed with Marvell and Oracle.

Why it matters

This matters most to cloud architects and AI platform teams planning large-scale inference deployments. They gain concrete efficiency metrics for Maia 300 and Cobalt 200 to model capacity and cost. The open specifications for power, cooling, networking, and hardware security mean vendors can build compatible gear, reducing lock-in risk for hybrid or multi-cloud strategies.

The impact is phased and collaborative. Maia 300 and Cobalt 200 are rolling into Microsoft’s fleet first; general availability timelines for customer-facing VMs were not specified. OCP contributions like 800V DC, ESUN networking, and Tessera HSM are at various spec stages, so broad ecosystem hardware will take time to appear.

What inference efficiency gains are you seeing with Maia 200 or Cobalt 200 in your workloads? Share your experience in the comments.

Read the original source.