Microsoft is preparing to publicly unveil its Maia 300 AI accelerator as early as September, according to a report by The Information. The company is simultaneously in discussions with Taiwan Semiconductor Manufacturing Company (TSMC) to secure more than 300,000 units by next year, signaling a significant escalation in its custom silicon ambitions.

TSMC Capacity Reserved for 2027 Delivery

Microsoft has contacted TSMC to reserve manufacturing capacity for the Maia 300 chips, which are slated for delivery in 2027. The order would far exceed the tens of thousands of Maia 200 units produced to date. Microsoft's longer-term target is to surpass one million units, though packaging negotiations and component supply constraints may limit the actual volume the company receives.

The preceding Maia 200 experienced delays after early testing failed to meet internal performance targets. The chip has since been deployed to only a limited number of data centers.

Microsoft CEO Satya Nadella informed investors during the company's fiscal Q4 2026 earnings call on July 29 that the chip delivers 30% better performance per dollar than existing hardware. He added that it is being scaled to support both OpenAI workloads and Microsoft's own MAI models.

When Cryptopolitan reported on the Maia 200 launch in January 2026, Microsoft described it as a second-generation inference chip built on TSMC's 3-nanometer process. Scott Guthrie, Microsoft's executive vice president for cloud and AI, called it "the most efficient inference system that Microsoft has ever built."

Strategic Push to Reduce Nvidia Dependence

Nadella has publicly stated that reducing Microsoft's reliance on Nvidia is a strategic objective. Nvidia currently dominates the AI accelerator market, and its GPUs command premium pricing that has driven up capital expenditures across the hyperscaler industry. The company says its custom chips can run internal and OpenAI workloads at lower cost, and it is expanding internal adoption through Azure AI Foundry and Copilot. Microsoft is also marketing the hardware to large external cloud customers.

Microsoft's effort follows similar initiatives by other major cloud providers. Google has deployed its Tensor Processing Units (TPUs) for internal and cloud workloads since 2015, while Amazon Web Services offers its Trainium and Inferentia chips to EC2 customers. Meta has also developed its MTIA inference accelerators.

Earlier this month, Anthropic confirmed it is assembling its own semiconductor team to design custom chips for its Claude models, positioning the company as both a potential Maia 300 buyer and a future competitor in the custom silicon space.

Industry projections from last year estimated that in-house chips from Google, Amazon, Meta, and OpenAI would account for 45% of the AI-chip market by 2028, up from 37% in 2024.

Microsoft reported fiscal Q4 2026 revenues of $90.0 billion, an 18% year-over-year increase, with Azure and other cloud services growing 43%.

TSMC's N3 node and CoWoS advanced packaging are expected to face supply constraints through 2027 — the same period during which Microsoft aims to scale up Maia 300 production. The timeline means Maia 300 chips would arrive well after the current wave of AI infrastructure buildout is already underway, potentially limiting their near-term impact on Microsoft's Nvidia procurement.