Microsoft is moving Copilot onto its own Maia chips, with multi-gigawatt deployments from 2027
Microsoft plans to deploy its own Maia chips for AI features like Copilot starting 2027, aiming for multi-gigawatt power usage. Maia 300, the latest version, promises 40% more efficiency than leading GPU systems. Maia 200 is already running Copilot in some data centers, offering up to 1.7x the efficiency of GPUs. The shift focuses on power efficiency and memory improvements, but Microsoft has not disclosed full details on costs or availability.
How this was made
The 30-second read
Why it matters
For traders, the key is whether custom inference silicon meaningfully improves unit economics and capacity for Copilot and Teams, and whether that advantage can be monetized through Azure and Microsoft’s own AI features.
Market read
The disclosure is a strategic AI infrastructure roadmap that could influence expectations for MSFT’s AI cost structure and datacentre scaling, but it lacks several commercially critical details.
What to watch
The article does not specify the exact gigawatt scale, which GPU systems were benchmarked, or whether savings flow through to Copilot pricing or usage limits, all of which affect revenue and demand sensitivity.
Background
Microsoft has been deploying Maia inference chips already (Maia 200) and is now outlining a next-generation Maia 300 roadmap focused on inference efficiency and datacentre power constraints.
Ticker impact
Microsoft says it will move Copilot and other AI features to its Maia inference chips, with Maia 300 deployments starting in 2027.
Likely supportive for MSFT sentiment as investors price in longer-term margin and capacity benefits from custom inference hardware, though near-term details are limited.
The article provides a concrete roadmap (Maia 300 in 2027, multi-gigawatt datacentres) and efficiency framing (tokens per watt), but lacks verifiable third-party benchmarks, exact power figures, and commercial terms for Azure customers.
Market effects
Custom inference silicon could intensify competitive pressure on GPU-centric AI infrastructure providers by targeting the inference bottleneck (tokens per watt) rather than training compute.
Multi-gigawatt datacentre deployments from 2027 imply large incremental power demand, with named regions in the US, Europe, and Australia but none in Africa.
If replicated by other hyperscalers, the shift toward inference-optimized chips could reshape AI datacentre capex and supply-chain dynamics worldwide.
Counterpoint
Efficiency claims may not translate into economic advantage if performance comparisons are not standardized, if deployment costs are higher than expected, or if Azure customers cannot access Maia 300 directly.
Key entities
- companyMicrosoft
Announced a shift of Copilot and other AI features to its Maia inference chips, with Maia 300 deployments beginning in 2027.
- productMaia 300
Third-generation Microsoft-designed inference chip, with higher memory bandwidth/capacity and projected tokens-per-watt improvements.
- productCopilot
Microsoft AI feature expected to run on Maia chips for inference.
- productTeams
Microsoft collaboration product with built-in AI features expected to run on Maia chips.
