Microsoft is betting big on in-house chip development as it seeks to reduce its heavy reliance on Nvidia.
According to a Monday report from The Information, two sources with knowledge of the matter revealed that Microsoft plans to launch its next-generation self-developed AI chip, the Maia 300, this fall, with a potential debut as early as next month. The company is in discussions with Taiwan Semiconductor Manufacturing regarding a production contract for over 300,000 chips by 2027—a massive leap from the current Maia 200's output of just tens of thousands. Simultaneously, Microsoft is actively negotiating with large cloud clients, including Anthropic, about using the Maia chip.
The core driver behind this expansion is cost efficiency. Microsoft disclosed to investors last month that the Maia 200 chip reduces operating costs by 30% to 40% compared to Nvidia's cutting-edge chips when running OpenAI and Microsoft's own models. According to sources, the Maia 300, optimized specifically for Microsoft's models, performs even better. Even without securing major external customers, Microsoft's fallback plan involves shifting more internal AI workloads to the Maia chip, while continuing to lease Nvidia's expensive chips to Azure cloud clients.
However, Microsoft still lags significantly behind Google and Amazon in the self-developed chip race. Google's TPUs and Amazon's Trainium chips have already been adopted by several large clients, with Google even selling TPUs to external customers. In contrast, the Maia 200 is currently used only by Microsoft itself, and as of the end of last month, it was operational in just two data centers within the United States.
The rollout of the Maia 200 has been slower than expected, limiting production expansion ambitions. The chip was delayed last year due to early testing failing to meet internal targets, and has since been deployed in only a few Microsoft data centers.
In June, Microsoft CEO Satya Nadella mentioned that two data centers were already operational, with plans to expand to more facilities, including overseas deployments. However, according to a source, as of the end of last month, the situation remained unchanged, with only two data centers in the U.S. showing no progress.
In comparison, the scale advantage of competitors is stark. Morgan Stanley estimates that Google plans to produce over 3 million TPU chips this year, with that number rising to 5 million next year. The Maia 300 order of over 300,000 chips that Microsoft is negotiating represents a significant gap.
Despite this, Microsoft has not scaled back its ambitions. The report indicates that Microsoft ultimately hopes to secure production capacity for over one million Maia 300 chips, but the actual scale remains uncertain due to component supply constraints and ongoing capacity negotiations with Taiwan Semiconductor Manufacturing.
In a statement, Andrew Wall, General Manager of Azure Maia at Microsoft, declined to comment on specific production plans but said Microsoft ultimately aims to produce Maia chips in gigawatt-scale quantities. " Microsoft continues to invest in custom chips as part of our long-term AI infrastructure strategy," he said. "We expect Azure Maia deployments to support AI workload demands measured in gigawatts." Typically, a data center requiring several gigawatts of power can house millions of AI chips.
On the client front, The Information previously reported that Anthropic has been in talks with Microsoft for months about potentially using the Maia chip. Microsoft's main appeal lies in cost efficiency: internal tests show that the Maia 300 offers particularly strong value for running Microsoft's own models.
The self-developed chip strategy is driven by Nadella's long-standing strategic concerns. In a 2022 email that was later made public through legal proceedings, Nadella wrote, "We are just a thin layer on top of Nvidia, with all the intellectual property in the hands of OpenAI," and mentioned that a Microsoft business unit "will lose $4 billion next year." According to a source, Nadella was referring to the high cost of running OpenAI models on Azure. At that time, Microsoft could control neither the chip costs nor the AI model itself, which became the starting point for a dual-track strategy involving self-developed chips and proprietary models.
Currently, Microsoft still relies heavily on Nvidia chips to run most of its AI software, including the MAI models that power Copilot. The commercialization path for the Maia chip involves gradually migrating internal workloads to the self-developed chip and eventually convincing external clients to follow suit with more competitive costs.
Interestingly, compared to the Maia's rocky progress, Microsoft's other self-developed chip line—the Cobalt central processing unit—has shown stronger market traction recently. Microsoft announced last week that large clients, including OpenAI and Adobe, have adopted the Cobalt chip in over 25 data centers worldwide.
The Cobalt chip falls under the traditional CPU category, differing from Nvidia's GPUs and AI-specific chips like the Maia. However, amid the current boom in AI infrastructure demand, CPU requirements are also surging. The interim success of the Cobalt chip provides some validation for Microsoft's overall self-developed silicon strategy and helps build client trust for the Maia's future rollout.
Comments