Microsoft is seeking manufacturing capacity for more than 300,000 Maia 300 artificial intelligence accelerators from Taiwan Semiconductor Manufacturing Co, signalling a sharp expansion of its custom-chip programme as competition for advanced semiconductor production intensifies. The company is preparing to unveil the Maia 300 as early as September, while discussions with TSMC cover production intended for delivery during 2027. Microsoft’s longer-term ambitions extend beyond the initial order under negotiation, with the company aiming eventually to secure manufacturing capacity exceeding one million chips as demand for computing power across Azure continues to rise.
The planned expansion would represent a major step for Microsoft’s internally designed AI processors. The company introduced the Maia family in 2023 and launched Maia 200 in January this year as an accelerator designed primarily for inference, the computational process involved when deployed AI models generate answers and other outputs.
Maia 200 is manufactured using TSMC’s 3-nanometre process and contains more than 140 billion transistors. It provides more than 10 petaFLOPS of FP4 performance and above 5 petaFLOPS at FP8 precision, while carrying 216GB of HBM3e memory with bandwidth of 7 terabytes per second. Microsoft has deployed the processor in its Iowa and Arizona data centres and says it delivers more than 30% better performance per dollar than the latest comparable silicon in its fleet.
Maia 300 is expected to deepen that strategy by giving Microsoft greater control over the economics and availability of the processors powering its expanding AI services. The company continues to buy large quantities of processors from outside suppliers, including Nvidia, and has emphasised that its infrastructure will remain heterogeneous rather than depend exclusively on its own silicon.
Securing enough production capacity may prove as important as designing the processor itself. TSMC remains at the centre of the global AI semiconductor supply chain, producing leading-edge chips for Nvidia, AMD and major technology groups while also supplying advanced packaging needed to combine AI processors with high-bandwidth memory.
The pressure is especially acute around Chip-on-Wafer-on-Substrate, or CoWoS, packaging. Demand for the technology has surged because high-performance accelerators require processors and multiple HBM stacks to be integrated into increasingly large packages. Industry estimates put 2026 CoWoS demand at close to one million wafer equivalents, despite aggressive expansion of TSMC’s packaging operations.
Nvidia is estimated to account for roughly 60% of available CoWoS allocation, equivalent to around 595,000 wafer equivalents under some supply-chain projections. Nvidia, Broadcom and AMD together are estimated to absorb more than 85% of capacity. Those figures remain industry estimates rather than disclosed customer allocations, and individual company volumes are subject to considerable uncertainty.
TSMC has been expanding CoWoS output rapidly as AI customers compete for production slots. The foundry previously described its packaging operations as fully loaded even while doubling capacity, illustrating why hyperscalers increasingly negotiate manufacturing arrangements well before products enter mass deployment. Its management continues to identify AI accelerators as a major structural growth driver and expects strong demand for leading-edge fabrication and advanced packaging technologies.
Microsoft’s push comes as cloud rivals increasingly use proprietary accelerators to control infrastructure costs. Google has developed multiple generations of Tensor Processing Units, while Amazon Web Services has expanded its Trainium family. The approach allows cloud operators to optimise processors, networking, software and data-centre architecture for their own workloads rather than rely exclusively on general-purpose accelerators.
Custom silicon has become more strategically important as inference accounts for a growing share of AI computing. Once models are trained, millions or billions of user requests can generate continuous demand for processing, making performance per watt and performance per dollar critical to cloud profitability.
Microsoft is also seeking broader external use of Maia hardware rather than limiting the chips to its own applications. Azure supports workloads from OpenAI and Anthropic alongside Microsoft models and enterprise customers, creating an opportunity to deploy internal processors across a wider range of demand when their economics are competitive.
Topics
Technology