Microsoft to Make Maia AI Chips Its Main Inference Engine by 2027
Microsoft will shift its AI inference workloads to its in-house Maia chips starting in 2027, deploying them at multi-gigawatt scale. The Maia 300 is expected to deliver 40 percent more tokens per megawatt than leading GPU systems, a move CEO Satya Nadella calls the company's biggest bet yet on its own silicon.
Microsoft will begin using its in-house Azure Maia chips as the primary system for AI inference across its products in 2027, deploying them at multi-gigawatt scale, the company has announced. The shift marks a significant expansion of Microsoft's custom silicon strategy and a direct challenge to the dominance of general-purpose graphics processing units in AI data centers.
The upcoming Maia 300 chip is expected to deliver 40 percent more tokens per megawatt than leading GPU-based systems, according to the company. That efficiency gain is central to Microsoft's plan to scale its AI infrastructure without a proportional increase in energy consumption. CEO Satya Nadella described the initiative as «our biggest bet yet on our own silicon,» underscoring the strategic importance Microsoft places on controlling its AI hardware stack.
Inference — the process of running trained AI models to generate responses — represents the vast majority of AI compute demand in production environments. By designing chips specifically for inference rather than training, Microsoft aims to optimize performance per watt, a critical metric as data centers face growing scrutiny over their power use. The multi-gigawatt deployment target suggests the company expects Maia to handle a substantial share of its global AI workload within the next few years.
The move reflects a broader trend among major technology firms to reduce reliance on external chip suppliers, particularly Nvidia, whose GPUs have been the industry standard for AI workloads. Custom silicon allows companies to tailor hardware to their specific software and services, potentially lowering costs and improving integration. Microsoft's Azure cloud platform, which hosts AI services for countless businesses, would be a primary beneficiary of the Maia rollout.
Microsoft has not disclosed which specific products will transition first or how the multi-gigawatt capacity will be distributed geographically. However, the 2027 timeline gives the company several years to refine the Maia architecture and prepare its data center footprint for the change. The Maia 300 follows earlier generations of the chip, which have been used in limited deployments as Microsoft tested and iterated on the design.
Energy efficiency is a key selling point. Tokens per megawatt measures how many units of AI-generated text or data a system can produce for a given amount of power. A 40 percent improvement over leading GPU systems would represent a meaningful reduction in operational costs and carbon footprint for large-scale AI services. As AI models grow larger and more capable, the demand for inference compute is expected to rise sharply, making efficiency gains increasingly valuable.
The announcement comes amid intensifying competition in the AI infrastructure space. Cloud providers and technology companies are racing to build out capacity while managing energy constraints and supply chain challenges. Microsoft's decision to prioritize its own chips at gigawatt scale signals confidence in its ability to compete on both performance and efficiency without depending solely on merchant silicon.
For Microsoft's customers, the transition is unlikely to be visible in the short term, as the company will continue to support a mix of hardware. But over time, the Maia chips could power a wide range of AI features across Microsoft's productivity software, cloud services, and consumer products. The company's ability to execute on the 2027 target will be closely watched as a bellwether for the custom AI chip movement.
5
