Microsoft just told Nvidia something worth paying attention to: we’re building our own. The company revealed the Azure Maia 300 at the OCP Global Summit 2026 in San Jose, positioning it as the future primary AI inference system across Microsoft’s products and services. That’s not a side project. That’s a strategic shift.
The numbers Microsoft is citing are aggressive. The Maia 300 delivers 40% more tokens per megawatt than what Microsoft calls “leading GPU-based systems,” which is clearly aimed at Nvidia. A 96-chip rack also offers seven times the rack-scale token throughput of the previous Maia 200. For context, tokens are the basic unit AI systems use to process text, images, and video. More tokens per megawatt means the system does more work for every unit of power consumed. In data center economics, that matters enormously.
Microsoft says Maia 300 deployments will begin next year at multigigawatt scale, alongside the existing Maia 200 and Maia 100. The Maia 200 already runs production workloads including Microsoft 365 Copilot, OpenAI models, and Microsoft’s own MAI models. It reportedly delivers 1.7 times more tokens per watt than comparable GPU-based systems across Teams, Copilot, and Cowork.
So why does this matter beyond Microsoft’s internal efficiency gains? Because it’s part of a much larger pattern. Amazon, Google, and Meta are all either shipping or actively developing custom silicon. OpenAI and Anthropic are working on chips too. The hyperscaler crowd is clearly trying to reduce how much they depend on Nvidia, which still earns more than 50% of its quarterly revenue from this exact group of customers. That’s a real concentration risk, and these companies know it.
The threat to Nvidia is real but not immediate. Amazon and Google are already renting their custom chips to outside customers, which competes directly with Nvidia’s cloud GPU business. Amazon is also exploring selling chips for on-premise data center installations. Still, custom chips tend to be optimized for specific workloads, and enterprises running diverse AI tasks will keep buying Nvidia hardware for the foreseeable future.
The Maia 300 matters for developers and infrastructure teams evaluating Azure because it signals where Microsoft is putting its efficiency bets. If the token-per-watt claims hold up in production, Azure workloads running on Maia silicon could become meaningfully cheaper to operate than those running on third-party GPUs. That’s a real differentiator in a market where inference costs are a constant pressure point for anyone building AI products at scale.
Nvidia isn’t going anywhere. But the cushion it once had with hyperscalers is getting thinner every year, and the Maia 300 is another reason why.
