Microsoft's Maia 300: 40% more tokens per megawatt, but not until 2027
What happened
Microsoft said Maia 300, its third-generation in-house AI chip, will deliver 40% more tokens per megawatt than leading GPU-based systems, per cloud chief Scott Guthrie. Production rollout begins in 2027, and Microsoft plans multi-gigawatt Azure Maia deployments from that year, with Maia as the primary inference system for its products. The efficiency figure is a company projection, not a measured result, and no independent benchmarks have been published.
What we know
- Guthrie said Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200 in a 96-accelerator configuration, per The Verge.
- Microsoft says Maia 300 will begin rolling out to production AI workloads in 2027, per The Verge and GeekWire.
- Maia 300 is built on TSMC's N2P process and uses a chiplet design that separates compute, memory controller and I/O functions, per The Verge.
- Maia 200, announced in January 2026, already runs production workloads including Microsoft Copilot, GitHub Copilot and Microsoft Foundry, per The Verge.
- Guthrie said Maia will become the primary AI inference system for Microsoft products and services, which The Verge reads as reducing reliance on Nvidia and AMD chips.
What we don't know yet
- Which 'leading GPU-based systems' the 40% tokens-per-megawatt comparison uses, and under what workloads and model sizes.
- Whether the 2027 schedule holds, given Maia 300 depends on TSMC's N2P node.
- How much Nvidia and AMD hardware Microsoft will still buy, since the sources describe Maia as 'primary' for inference and say nothing about training.
- Maia 300's absolute performance, memory and pricing; the sources give only relative figures.
Why it matters
Inference is the recurring cost of running Copilot and other AI services at scale. If Maia 300 meets its efficiency target, Microsoft would lower its power bill per token and depend less on Nvidia and AMD supply. Nvidia would then have one fewer guaranteed buyer among the largest cloud providers.
Claims
- UnconfirmedMicrosoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems.
- UnconfirmedIn a 96-accelerator configuration, Maia 300 delivers up to 7x the rack-scale token throughput of Maia 200, per Microsoft.
- ConfirmedMaia 300 will begin rolling out to production AI workloads in 2027.
- ConfirmedMaia 300 is built on TSMC's N2P process node.
- ConfirmedMaia 200 already runs production workloads including Microsoft Copilot and GitHub Copilot.
- ConfirmedMicrosoft plans multi-gigawatt Azure Maia deployments starting in 2027, making Maia its primary AI inference system.
Related coverage
Why you can trust this story
76%Source map · 2 outlets / 2 articles
- The Verge
- Article 1theverge.comreport
- GeekWire
- Article 1geekwire.comreport
How this credibility score is calculated
- Source reliability
- Corroboration
- Primary source
- Atom-verified claims
- No contradiction
- Settled
- Claim attribution
- AI disclosure
Atom-verified claims matched cited source text. Weights are fixed and explainable.
Human accountability
- 2 independent origins / 2 sources
- Drafted by the Newsmesis judgment agent (agent-cli); human approval required before publishing.
Verification ledger · 6
- Claim: Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
- Claim: In a 96-accelerator configuration, Maia 300 delivers up to 7x the rack-scale token throughput of Maia 200, per Microsoft.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
- Claim: Maia 300 will begin rolling out to production AI workloads in 2027.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Maia 300 is built on TSMC's N2P process node.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Maia 200 already runs production workloads including Microsoft Copilot and GitHub Copilot.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft plans multi-gigawatt Azure Maia deployments starting in 2027, making Maia its primary AI inference system.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
Update log · 0
AI accelerates. Humans approve.