Two sources familiar with the matter have disclosed that Microsoft intends to significantly ramp up production of its in-house next-generation artificial intelligence chips next year, aiming to attract major cloud service clients like Anthropic to adopt its silicon. Despite the current generation Maia 200 chip seeing limited market acceptance, the company is pushing forward with expansion plans. One insider stated that Microsoft plans to officially unveil the new Maia 300 chip this autumn, with a potential debut as early as next month.
Another source indicated that the cloud computing giant has engaged with chip foundry TSMC to secure production capacity for over 300,000 units of the Maia 300, with delivery scheduled for 2027. This order volume far exceeds the tens of thousands of Maia 200 chips Microsoft has produced to date. The source added that Microsoft's ultimate goal is to secure capacity for over one million Maia 300 chips, though constraints on chip component supply and ongoing capacity negotiations with TSMC could limit the ceiling on expansion.
Deploying the Maia series is a core strategy for Microsoft to reduce its reliance on NVIDIA chips and a top priority for CEO Satya Nadella. The self-development path has faced obstacles: last year, the Maia 200 chip was delayed due to early testing failing to meet internal benchmarks, and it has only been deployed in a few of Microsoft's data centers since its launch.
In a statement, Andrew Walker, General Manager of Microsoft Azure Maia, declined to disclose specific production plans but noted that Microsoft's long-term goal is to achieve gigawatt-scale computing capacity for Maia chip production. Walker did not provide exact numbers or timelines, while a multi-gigawatt data center typically requires millions of AI chips. "Custom in-house silicon is part of Microsoft's long-term AI infrastructure strategy. We are not disclosing specific output, but existing reports do not reflect the full scope of this project. We expect Azure Maia chip clusters to handle AI workloads at gigawatt-scale computing demand," Walker said.
In terms of rolling out self-developed AI chips, Microsoft lags behind competitors like Google and Amazon. Both have secured major clients, deploying Tensor Processing Units (TPUs) and Trainium training chips in their cloud businesses; Google even sells TPUs externally, allowing customers to use them outside its own data centers. In contrast, the Maia chips have so far been used only internally by Microsoft, to run OpenAI's large models and its own MAI foundation models, powering its Copilot assistant software. However, the majority of Microsoft's AI business still relies on NVIDIA chips.
Nevertheless, Microsoft is confident about attracting more customers to lease the upcoming Maia 300, with Anthropic being a potential client, primarily due to lower usage costs. Reports indicate that Anthropic has been in discussions with Microsoft for months about adopting the Maia series. In June, Nadella stated that two data centers were already using the Maia 200, with plans to expand to more sites, including overseas locations. But sources revealed that, as of late last month, the Maia 200 was still only operational in two data centers within the United States.
Even if the Maia 300 fails to secure external major clients, Microsoft has a backup plan: increasing internal AI business usage of Maia chips while continuing to rent out expensive NVIDIA chips to Azure cloud customers. Last month, Microsoft disclosed to investors that the operating cost of the Maia 200 is 30% to 40% lower than NVIDIA's flagship chips when running OpenAI and its own models; sources say the Maia 300, optimized for Microsoft's models, offers even better performance.
Even if Microsoft finalizes a large order for the Maia 300, the procurement volume will still be far below the annual chip orders of competitors like Google and Amazon. Morgan Stanley estimates that Google plans to produce over 3 million TPUs this year, with capacity increasing to 5 million units next year. Meanwhile, Microsoft's other self-developed chip line, the Cobalt central processing unit, has recently shown significant progress. Unlike the Maia AI-specific chips that compete with NVIDIA GPUs, CPUs are a traditional processor category, but market demand for them is also surging. Microsoft revealed last week that major clients like OpenAI and Adobe have deployed Cobalt processors in over 25 data centers globally.
In recent years, Nadella has privately criticized NVIDIA for monopolizing the AI chip market. In a 2022 email later disclosed as court evidence, Nadella lamented, "At this stage, we are just a thin layer built on top of NVIDIA, with all core intellectual property held by OpenAI," and mentioned that an unnamed business unit at Microsoft “will lose $4 billion next year.” It is currently unclear which specific business line incurred the loss, but sources say Nadella was referring to the high cost of operating OpenAI models on the Azure platform, stemming from Microsoft's inability to control the two core costs of chips and large models. The company's subsequent push into self-developed AI chips is aimed at addressing these two pain points.
Comments