Microsoft says Maia 300 chip will run its AI services from 2027
Microsoft announced the Maia 300 chip, set to power its AI services starting 2027. The company plans gigawatt-scale deployments, with each megawatt generating 40% more tokens than top GPU systems. Maia 300 is designed for inference, targeting long conversations and multi-step tasks. Microsoft aims to optimize intelligence per watt, dollar, and byte of memory. The chip will support Microsoft Copilot, OpenAI, and GitHub Copilot.
How this was made
The 30-second read
Why it matters
The disclosed deployment scale (gigawatts starting next year, potentially over a million chips in 2027) and efficiency targets (more tokens per watt than top GPU systems) can influence investor expectations for Azure AI capacity planning and compute cost trajectory.
Market read
This is a concrete product and deployment roadmap for Microsoft’s in-house AI inference hardware, with quantified efficiency and scaling details that can affect AI infrastructure expectations.
What to watch
The article does not specify pricing, margins, or how Maia 300 compares on total cost of ownership versus competing GPU stacks, which are key for Azure unit economics.
Background
Microsoft is unveiling Maia 300, its third-generation custom AI chip, and positioning it as the primary inference system for its AI services beginning in 2027.
Ticker impact
Microsoft says its Maia 300 custom AI inference chip becomes the primary platform for running its AI services starting in 2027.
Likely supportive for MSFT sentiment as investors price greater AI infrastructure control and efficiency, though near-term impact may be limited by deployment ramp.
The article provides specific product and deployment timelines (2027 primary use, gigawatt deployments starting next year) plus performance-per-watt claims, which can affect expectations for Azure AI cost structure and capacity strategy.
Market effects
Reinforces the broader industry push to reduce reliance on Nvidia by moving more inference workloads to custom silicon, which could pressure GPU demand expectations at the margin.
Highlights continued dependence on TSMC advanced nodes (N2P) and HBM4 supply, keeping semiconductor supply-chain focus on Taiwan and memory ecosystem.
If widely adopted, custom inference chips could shift AI infrastructure capex and procurement patterns across global cloud providers and AI platforms.
Counterpoint
Performance-per-watt claims may not translate into broad cost advantages if software optimization, yield, and supply constraints limit real-world throughput.
Key entities
- companyMicrosoft
Plans to make Maia 300 the primary system for running its AI services starting in 2027, with gigawatt deployments beginning next year.
- productMaia 300
Third-generation Microsoft custom AI inference chip with chiplets, TSMC N2P process, and HBM4 memory.
- supplierTSMC
Manufacturing partner for Maia 300’s computing parts using N2P process.
- peerNvidia
Referenced for comparison via GB300 rack power and GPU count.
