Microsoft Maia 300: 7x Throughput for Enterprise AI Inference
Microsoft announced Maia 300 on October 12, 2026, its third-generation custom AI inference chip, and committed to scaling the Azure Maia family to multi-gigawatt deployments starting in 2027. The chip delivers 40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale token throughput of its predecessor, making it the clearest signal yet that the economics of enterprise AI inference are about to shift.
For teams deploying AI agents at scale, this matters now, even though Maia 300 does not go into production for another year. Token costs are the hidden tax on every agentic workflow, and the infrastructure that serves those tokens is being redesigned from the ground up.
What Microsoft Maia 300 Actually Delivers
Scott Guthrie, Microsoft’s executive vice president of cloud, published the announcement directly, signaling that this is a strategic infrastructure commitment, not a marketing preview.
The headline numbers from the official release:
- 40% more tokens per megawatt compared to leading GPU-based systems
- Up to 7x rack-scale token throughput compared to Maia 200
- 96 accelerators per rack within a 167kW power envelope
- 27.6TB of HBM capacity and 1.7 PB/s aggregate HBM bandwidth per rack
- Production deployment begins in 2027 at multi-gigawatt scale
The Verge confirmed that Maia 300 will become “the primary AI inference system for Microsoft products and services,” replacing the company’s current dependence on Nvidia and AMD GPUs for its own inference workloads.
| Metric | Maia 200 | Maia 300 (Projected) |
|---|---|---|
| Token throughput vs leading GPUs | 1.7x tokens/watt | 40% more tokens/megawatt |
| Rack-scale throughput | Baseline | 7x Maia 200 |
| Accelerators per rack | N/A (standard) | 96 |
| HBM capacity per rack | N/A | 27.6 TB |
| HBM bandwidth per rack | N/A | 1.7 PB/s |
| Production deployment | US, Europe, Australia (active) | 2027 |
| Primary use | Microsoft internal models | All Microsoft products and services |
Why Token Economics Define Enterprise AI Strategy
Most enterprise AI deployments fail their ROI projections not because models are too slow, but because token costs compound at every step of an agentic pipeline. A workflow that calls three models, generates a report, and validates outputs against a knowledge base can easily consume ten times the tokens of a single chat completion. At scale, that bill becomes the budget conversation that kills adoption.
Maia 300 is Microsoft’s answer to that problem, at least for workloads running inside Azure. The chip is designed specifically for three categories of AI workload that define enterprise deployments in 2026: long-context reasoning, latency-sensitive conversational experiences, and sustained multi-step agentic tasks. That framing, published in the official Microsoft technical announcement, maps directly to the cost centers that enterprise AI teams are fighting right now.
The efficiency improvement is not incremental. Going from 1.7x tokens per watt (Maia 200) to 7x rack-scale throughput at 40% better power efficiency represents a genuine step change in what Microsoft can profitably serve. That eventually translates to lower or more predictable pricing for enterprise Azure AI contracts.
The Nvidia Independence Play
Reuters reported that Microsoft plans to deploy potentially more than a million Maia 300 chips in 2027, targeting “multi-gigawatt deployments” as the company runs its own AI services at scale. Microsoft has said it expects to spend more than $175 billion in capital expenditures this calendar year, a significant portion directed at AI infrastructure.
This is not a statement that Microsoft is abandoning Nvidia for customer-facing GPU capacity. Azure will continue to offer Nvidia GB300, AMD MI300X, and other hardware to external workloads. The Maia strategy is narrower: Microsoft will stop paying Nvidia inference costs for its own products.
The comparison is instructive. A Maia 300 rack uses 167kW for 96 accelerators. Nvidia’s GB300 NVL72 rack uses roughly 142kW for 72 chips. Microsoft is packing more accelerators per rack, operating at higher efficiency, and targeting the specific inference latency profiles of its own product stack. That is a purpose-built system, not a general-purpose GPU alternative.
For enterprise buyers, this creates a useful signal: when a cloud provider stops relying on third-party chips for its own AI products, it is telling you something about the long-term cost curve of inference on that platform.
What Runs on Maia 300
Microsoft has identified the specific products that will transition to Maia 300 inference:
- Microsoft 365 Copilot (Teams, Word, Excel, Outlook)
- GitHub Copilot (code completion and agent workflows)
- Microsoft Foundry (the model deployment and routing layer)
- Dragon Copilot (healthcare-specific AI)
- Cowork experiences across the Microsoft 365 stack
Maia 200 is already running Microsoft’s in-house AI models, including the MAI family, in production across the United States, Europe, and Australia. Maia 300 will absorb all of that and scale it to the full enterprise surface.
For teams building on Microsoft Foundry and its Decision-1 routing layer, the Maia 300 transition means that the orchestration layer they use today will eventually run on substantially cheaper and faster underlying inference. The workflows do not change, but the economics behind them do.
Agentic AI Needs Inference-First Hardware
The Maia 300 design philosophy, “inference-first,” is worth taking seriously as a signal about where enterprise AI compute is heading. The chip is not optimized for training. It is optimized for serving tokens at scale, with low latency, high throughput, and high memory bandwidth per rack.
That matches exactly how enterprises are starting to use AI. Training a model is a one-time event. Serving it across thousands of simultaneous agent workflows, every hour of every day, is the persistent cost. Hardware designed for training economics (Nvidia’s historical strength) is a different product than hardware designed for inference economics at the scale Microsoft is targeting.
Enterprise multi-agent systems multiply inference costs nonlinearly. An orchestration layer that routes tasks to specialist agents, validates outputs, manages memory retrieval, and handles tool calls can easily run five to twenty inference steps per user request. At those multiples, a 40% efficiency gain per megawatt translates directly to whether a workflow is cost-effective to run continuously or only on demand.
The 2027 Infrastructure Gap
The gap between now and Maia 300’s 2027 production deployment creates a planning window for enterprise AI teams. The chip’s specs are public. Microsoft’s commitment to making it the primary inference substrate for its product stack is explicit. What remains open is pricing.
Enterprises negotiating Azure AI contracts over the next six to twelve months are doing so against a backdrop of substantially better infrastructure arriving in 2027. That is a negotiating data point. The efficiency gains built into Maia 300 will reduce Microsoft’s inference cost per token, and competitive pressure from Google’s TPU 8i and AMD Helion (announced the same day) will push at least some of those gains toward enterprise pricing.
The Microsoft MAI frontier diffusion strategy established the playbook: build proprietary models, deploy them on proprietary chips, and lock in enterprises at the application layer while continuously improving the cost structure underneath. Maia 300 is the next hardware layer of that strategy. Teams building agentic workflows on Azure are building on infrastructure that will get meaningfully cheaper and faster. The question is how much of that improvement flows through to customers, and on what timeline.
Watch for Microsoft’s Azure AI pricing updates in Q1 2027, when Maia 300 deployments begin. That is when the theoretical efficiency gains become observable in actual token costs.
Sources: Microsoft Official Blog, Microsoft Source EMEA, The Verge