Microsoft Officially Launches Maia 300: 7x Rack Throughput, 40% Better Tokens-Per-Megawatt, Multi-Gigawatt Production in 2027
Microsoft announced Azure Maia 300 on October 12, 2026 in a blog post from Scott Guthrie, executive vice president of Cloud + AI. The announcement resolves months of speculation with a full specification sheet and a production timeline: deployments begin in 2027, and Maia will become the primary inference system across Microsoft products and services at multi-gigawatt scale.
Specification Breakdown
| Metric | Maia 300 |
|---|---|
| Rack config | 96 accelerators |
| Rack power envelope | 167 kW |
| HBM stacks | 8 x HBM4 |
| Peak HBM bandwidth | 18.0 TB/s |
| HBM capacity (rack) | 27.6 TB |
| Aggregate HBM BW (rack) | 1.7 PB/s |
| TDP per accelerator | 950 W |
| Rack token throughput vs Maia 200 | 7x |
| Tokens per megawatt vs GPU systems | +40% |
Maia 300 replaces HBM3e (used in Maia 200) with HBM4, a move that drives the memory bandwidth increase. The 950W TDP per accelerator is higher than Maia 200’s 750W envelope, but the 40% efficiency claim against GPU-based systems means the per-token power cost is lower when the rack is fully utilized.
Architecture: Dataflow as a First-Class Primitive
Microsoft describes Maia 300 as an “inference-first” design with a software-defined dataflow architecture. Rather than treating networking as a separate communication layer, Maia 300 integrates dataflow operations across compute, memory, and network resources as a single system. The company says this reduces data movement — increasingly the bottleneck in large inference systems — and improves utilization across all three subsystems.
On-chip SRAM operates as a software-managed scratchpad to reduce pressure on HBM. The scale-up network uses a custom AI transport over Ethernet-based infrastructure, scaling from a single rack to datacenter-level deployments without proprietary interconnect requirements.
The Business Calculus
Maia 200, announced January 26, 2026, is already deployed across Microsoft Copilot and the MAI research team. GitHub Copilot and Microsoft Foundry are on schedule to follow by the end of 2026. Maia 200 runs OpenAI models, Microsoft’s own MAI models, and leading open-source families.
Maia 300 takes that foundation and targets the long-context reasoning and multi-step agentic workloads that dominate frontier AI usage in 2026. Microsoft’s August 2026 planning articles described Maia 300 in general terms; this announcement replaces projection with specification.
The multi-gigawatt deployment commitment is significant. Azure currently operates on a mix of NVIDIA H100/H200 and Maia 200 infrastructure. By making Maia the “primary inference system” from 2027, Microsoft is committing to a trajectory where its own silicon handles the majority of its AI inference load. At the efficiency numbers claimed, the economics favor in-house silicon at the scales Microsoft operates.
Competitive Context
The efficiency claim — 40% more tokens per megawatt versus leading GPU systems — places Maia 300 in direct competition with NVIDIA’s GB300 NVL72 rack configurations and Google’s TPU v7 deployments. Neither NVIDIA nor Google has published a direct response to this specific metric.
For enterprise Azure customers, the practical effect is lower inference latency at comparable cost as Maia 300 scales out, and broader availability of the long-context and agentic capabilities that Maia 300 is designed to serve. Pricing for Maia 300-backed services has not been announced.