Microsoft (NASDAQ: MSFT) has unveiled the Azure Maia 300 accelerator at the OCP Global Summit, marking a direct challenge to the chipmaker dominance it has long helped finance.

Scott Guthrie, Microsoft’s EVP of Cloud+AI, announced the chip on October 12, 2026, positioning it as a serious competitor to the industry’s leading GPU-based systems rather than a symbolic in-house project.

According to vendor-reported figures, the Maia 300 delivers 6.9 PFLOPS of FP8 performance and 13.8 PFLOPS of FP4, putting it ahead of NVIDIA’s (NASDAQ: NVDA) Blackwell B200 on dense FP8 compute.

The chip carries 288GB of HBM4 memory across eight 36GB stacks and delivers a reported 18 TB/s of bandwidth, representing a 2.25x bandwidth advantage over both the B200 and B300.

The Maia 300 operates at a thermal design power of 950 watts, significantly lower than the B300’s 1,400 watts, giving Microsoft a notable efficiency advantage on the metrics that govern inference throughput per dollar of power.

Built on TSMC’s (NYSE: TSM) N2P 2nm-class process with a multi-chiplet design and I/O chiplets on a separate 3nm process, Microsoft is packing 96 accelerators per rack with scale-up clusters targeting roughly 12,000 accelerators.

Microsoft claims the Maia 300 will deliver 40% more tokens per megawatt than leading GPU-based systems and up to seven times the rack-scale token throughput of the previous Maia 200 generation.

Reuters reports that Microsoft is negotiating with TSMC for more than 300,000 units to meet a 2027 production window, with multi-gigawatt deployment targets underpinning the chip’s commercial ambitions.

Evercore ISI estimates that Maia could represent approximately 10% of Microsoft’s total AI chip spending by the first quarter of 2027, signaling a meaningful structural hedge rather than an outright replacement for NVIDIA silicon.

The announcement carries particular financial significance given that Anthropic’s S-1 disclosed $31.4 billion in compute commitments to Microsoft for Azure, placing Microsoft deep inside the supplier dependency loops it is now working to reduce.

Broadcom’s (NASDAQ: AVGO) $42 billion convertible note to Anthropic and reported financing discussions with OpenAI spanning $50 billion or more represent the clearest expressions of a compute landlord model that ties customers to suppliers through layered financial obligations.

Microsoft is pitching Maia 300 to external customers including Anthropic, though no confirmed deployments have been announced and no external parties have tested the chip at this stage.

If Anthropic were to adopt Maia alongside its existing NVIDIA and Google TPU infrastructure, it would gain a third compute option with different incentive alignment, potentially loosening the financial grip of existing chip suppliers.

Unlike the financing structures Broadcom has used to bind Anthropic to its compute commitments, Microsoft is funding and deploying Maia through its own cloud, concentrating performance risk on its own balance sheet rather than distributing it through external tenants.

Markets responded immediately to the announcement, with Microsoft shares rising 1.07% intraday while NVIDIA fell 1.22%, according to market data confirmed via Yahoo Finance on October 12, 2026.

The Maia 300 does not match the B300 on sparse FP8 performance, where NVIDIA’s tensor core architecture delivers 14.0 PFLOPS against Maia’s 6.9, meaning the NVIDIA ecosystem retains a meaningful technical lead in certain workloads.

For enterprise teams that have already optimized their infrastructure around NVIDIA hardware, the central question is not whether Maia 300 wins on a spec sheet but whether Microsoft can deliver these performance claims at production scale starting in 2027.

All technical specifications cited are vendor-reported figures, and no independent benchmarks of the Maia 300 exist at this time, making execution and real-world delivery the only metrics that will ultimately determine whether this chip reshapes the competitive landscape.