Microsoft는 자체 설계한 Maia 칩을 실제 운영에 투입하고 있으며, 이를 통해 AI 투자 비용에서 Nvidia에 지급하는 비중을 줄이는 것을 목표로 하고 있습니다.

Microsoft's Maia 200 chip is already handling AI workloads in production, not just lab demonstrations, and the company says its custom silicon is beginning to change the economics of Azure inference.

Satya Nadella's company introduced the Maia 200 on January 26, positioning it less as a GPU killer than as a way to lower costs. Microsoft now says Maia 200 is serving production workloads, including Microsoft Copilot, and running OpenAI models, Microsoft's own MAI models, and open-source model families within Azure. That is the key point: this is not a slide at an investor day but silicon doing real work inside Azure.

The specifications are concrete enough to verify. Built on TSMC's 3-nanometer process, the Maia 200 contains more than 140 billion transistors, delivers over 10 petaFLOPS at 4-bit precision, and carries 216GB of HBM3e memory with 7 terabytes per second of bandwidth. Microsoft says this yields more than 30% better total cost of ownership than the latest-generation hardware in its fleet. It later told investors that the Maia 200 offers 30% better performance per dollar and 40% better performance per watt when running MAI models.

The performance-per-dollar claim is the core of the story. Nvidia's dominance has never rested solely on having the fastest chip. It rests on hyperscalers having few alternatives. Every point Microsoft cuts from its infrastructure bill is revenue Nvidia does not collect.

Nvidia's margins are a major reason companies are racing to build their own silicon. Nvidia's H100 is estimated to cost $3,320 to manufacture and sells for $25,000 to $40,000. The newer B200 costs an estimated $5,700 to $7,300 to build and sells for $30,000 to $50,000. These figures put Nvidia's overall gross margin at roughly 73% to 75%. Microsoft, Google, and Amazon are all trying to engineer their way around those margins, and they have considerable leverage while doing so. As of Nvidia's fiscal 2026 third quarter, four direct customers each accounted for more than 10% of its revenue, together making up about 61%, up from roughly 36% a year earlier. Nvidia did not identify those customers. When its largest customers are also among the companies best equipped to fund their own chips, Nvidia's pricing power becomes fragile.

Microsoft is not abandoning Nvidia. Maia is intended to complement GPUs for specific workloads rather than replace them, and Microsoft continues to buy substantial amounts of Nvidia hardware for jobs Maia is not designed for. Still, the direction is clear. On October 12, 2026, Microsoft announced the Maia 300 and said it will scale Azure Maia to multi-gigawatt deployments starting in 2027, making the Maia 300 the primary AI inference system for Microsoft products and services.

Microsoft is also late to this race. Google has been shipping Tensor Processing Units since 2015 and is now on its seventh generation. Amazon's Trainium line is on its third iteration. Microsoft's Maia program did not produce its first real chip until late 2023, which helps explain why Microsoft emphasizes the Maia 200's cost metrics rather than claiming outright performance leadership.

According to Microsoft's own published specifications, the Maia 200 delivers roughly three times the FP4 throughput of Amazon's third-generation Trainium and slightly outperforms Google's seventh-generation TPU on FP8 workloads. These are the comparisons Microsoft wants readers to focus on.

The broader trend is the more important one. Shipments of custom AI accelerators are projected to grow 44.6% in 2026, compared with 16.1% for merchant GPUs. Some analysts project that Nvidia's share of hyperscalers' internal inference compute could fall to 20% to 30% by 2028. Training workloads, where Nvidia's software ecosystem and CUDA lock-in are strongest, will likely remain stickier. Inference, the less expensive and more repetitive work of running a trained model for users, is precisely where a focused chip like the Maia 200 can compete without matching Nvidia feature for feature.

For founders and investors tracking AI infrastructure costs, the signal is less about Microsoft defeating Nvidia and more about where power is shifting. Compute has been the largest line item in most AI startups' budgets, and its price has been set almost entirely by one company's margins. If Azure, Google Cloud, and AWS each move even partway toward self-sufficiency in inference silicon, downward pressure on GPU rental prices will not stay confined to the hyperscalers' balance sheets. Eventually it will reach everyone renting compute from them.