Microsoft Plans Potential Million-Plus, Multi-GW Maia 300 Deployment, Reshaping Azure Inference Costs and GPU Procurement

A. Key facts

Microsoft unveiled Maia 300 on October 12, 2026 and said potential deployments could exceed one million chips and reach multi-GW scale beginning in 2027 for internal AI services and Azure workloads.

A 96-accelerator rack draws about 167 kW. Core compute chiplets use TSMC 2 nm and some data-movement/I/O chiplets use 3 nm technology. Microsoft claims roughly 40% more useful compute or tokens per MW than leading GPU systems and up to 7x Maia 200 rack throughput, but disclosed neither the comparator nor independent benchmarks.

B. Business significance · InveDeck interpretation

At a potential million-plus and multi-GW scale, Maia would become primary inference infrastructure rather than a supplemental chip. A larger custom-silicon mix could reduce NVIDIA/AMD procurement cost and supply constraints while improving Azure unit inference economics, gross margin and external GPU negotiating leverage.

Execution problems in 2 nm yields, advanced packaging, memory, networking or software compatibility could instead reduce utilization and lengthen payback periods on Microsoft's expanding data-center and power commitments.

C. Quantifiable impact

One million chips imply about 10,417 racks at 96 chips each. At 167 kW per rack, this equals roughly 1.74 GW of IT load, assuming identical configuration and simultaneous operation and excluding cooling, conversion and networking overhead.

Chip ASPs, rack cost and utilization are undisclosed, so profit, capex and FCF savings cannot be quantified. FY2026 Q4 capex was $41 billion, up more than 70% YoY, and FCF was $19.6 billion, down 23%. FY2027 Q1 capex was guided to about $50 billion and calendar-2026 capex to $175 billion. Maia changes investment mix and potential efficiency rather than raising guidance by itself.

D. Link to earnings and expectations

FY2026 Q4 revenue was $90 billion, up 18% YoY, and Azure grew 43%. Cloud contracted backlog reached $678 billion, up $51 billion from $627 billion in the prior quarter.

The lower calendar-2026 capex outlook reflected lease-accounting changes while underlying investment plans were unchanged. An August report discussed 2027 capacity above 300,000 units and a longer-term target above one million; the new company announcement officially confirms the product, rack design, efficiency claims and multi-GW intent.

E. Industry and supply chain · InveDeck interpretation

The plan is positive for TSMC 2 nm/N2-class and 3 nm chiplets and advanced packaging. If the million-chip scale is reached, server racks, switching, optics, storage, power and cooling demand would also be substantial.

It is negative for NVIDIA and AMD share and pricing leverage in internal Azure inference. Yet Microsoft's total AI capacity is expanding to multi-GW scale, so Maia is likely to coexist with GPUs for training and general workloads rather than eliminate them. Better efficiency is more likely to increase tokens produced from constrained power than reduce overall data-center investment.

F. What to watch next

• Actual 2027 production and installed units and quarterly ramp
• TSMC 2 nm yields and packaging, memory and networking constraints
• Comparator, workload and independent benchmark for the claimed 40% per-MW gain
• Adoption by external Azure customers, OpenAI and other model developers
• Realized changes in Azure margin, capex, depreciation, FCF and NVIDIA/AMD procurement