Microsoft has boosted its Azure Maia chip program, revealing a broad roll-out of its Maia 200 chips and significant expansion with Maia 300 once they come into production in 2027.

According to a blog from Scott Guthrie, EVP of Cloud and AI at Microsoft: “The multigenerational Azure Maia program has become a foundational element of Microsoft’s AI infrastructure strategy.”

Maia 200, which was announced in January and is already in production, is already serving production workloads in Microsoft Copilot, OpenAI models, and the company’s own Microsoft AI (MAI) models. GitHub Copilot and Microsoft foundry are to follow by the end of 2026, said Guthrie.

Maia 300, which will go into production in 2027, was described as “a complete intelligent system” that can support emerging AI workloads like long-context reasoning, latency-sensitive conversational experiences, and sustained, multi-step agentic tasks.

The company claimed that Maia 300 will offer significant improvements on token use efficiency, capable of delivering 40% more tokens per megawatt than “leading GPU-based systems”. It also will deliver up to 7x rack-scale throughput of Maia 200 when in a full 96-accelerator configuration.

Rani Borkar, president of Azure hardware systems and infrastructure, told Reuters: “It’s basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system.”

Nvidia’s customers become its competitors

While it’s not explicitly laid out in Guthrie’s blog, the shift to Maia 200 and Maia 300 means Microsoft is extricating itself from Nvidia’s ecosystem – at least for some of its AI systems.

Maia will be “the primary inference system for Microsoft products and services”, Guthrie said. This position had previously been held by Nvidia chips, notably Nvidia H100 and H200 ‘Hopper’ GPUs and Nvidia Blackwell GPUs.

Microsoft is not alone in taking an ‘in-house’ strategy, either.

Google has its Tensor Processing Units, which it has been developing and using for several years.

Amazon has its Tranium and Inferentia chip family, which is used both by itself and by generative AI pioneers Anthropic, with OpenAI signed on to use Tranium as part of an infrastructure agreement.

Speaking of OpenAI, it’s co-developing its own chips as well, called Jalapeño, with Broadcom. In August, the company said that when compared to “leading commercially available AI systems … (it) delivered a better combination of throughput, power efficiency, and latency”.

“Across all three public models (GPT-OSS, DeepSeek R1, Kimi K2.5), Jalapeño delivered a better combination of performance per watt and latency across the tested operating range, placing it on the Pareto frontier,” the company said.

Meta and Anthropic are also looking into building their own home-grown chips, though there have been no further announcements from either company.

All this doesn’t spell the end for Nvidia, however. Quoted by CNBC, Alistair Speirs, GM of global infrastructure, said the company partners with Nvidia and AMD for training large language models (LLMs).

The companies also have a close relationship when it comes to end-user hardware, such as the upcoming Surface Laptop Ultra.