Microsoft's Maia 300: 96 chips per rack, 40% more tokens per megawatt than GPUs
What happened
Microsoft's Maia 300 packs 96 chips into each data center rack, double the Maia 200's count, and Microsoft says a rack can process up to seven times more AI tokens than its predecessor. The company unveiled the third-generation in-house AI chip on Monday and said Maia will become its primary architecture for running AI models, with production starting in 2027. Microsoft also claims 40% more tokens per megawatt than GPU systems, but it did not say which GPUs it compared against, and outside customers cannot yet rent or test the chips.
What we know
- Microsoft says Maia 300 delivers 40% more tokens per megawatt than GPUs and up to seven times the per-rack output of Maia 200; it did not name the GPU systems it measured against, per The Next Web and CNBC.
- Each Maia 300 rack holds 96 chips and draws about 167 kilowatts, versus 72 GPUs and about 142 kilowatts for Nvidia's GB300 racks, according to Reuters as cited by The Next Web.
- The chip uses TSMC's N2P process for its compute chiplets and carries eight stacks of HBM4 memory, per The Next Web.
- Evercore ISI analysts estimate Maia will account for about 10% of Microsoft's AI chip spending by the first quarter of 2027, per CNBC.
- Microsoft's accelerators cannot be rented by the hour like Google's TPUs or Amazon's Trainium; Moor Insights & Strategy CEO Patrick Moorhead told CNBC that nobody has tested them outside Microsoft.
What we don't know yet
- Which GPU systems the 40% tokens-per-megawatt figure was measured against, and under what workloads.
- How many Maia 300 chips Microsoft will deploy in 2027: The Next Web says gigawatt-scale deployments could mean more than a million chips, while a Microsoft manager declined to give CNBC a timeline.
- Whether Microsoft will offer Maia to outside cloud customers; neither source reports such a plan.
- Whether the reported TSMC talks to make more than 300,000 Maia 300 chips for 2027 have produced a contract.
Why it matters
If Maia 300 performs as Microsoft claims, the company could shift a growing share of inference spending from Nvidia GPUs to its own silicon, which affects Nvidia's largest cloud buyers and Microsoft's own AI serving costs and margins. The unverified comparison baseline and lack of external testing leave the claim unproven until Microsoft shares more at its Ignite conference.
Claims
- ConfirmedMicrosoft unveiled the Maia 300, its third-generation custom AI chip, on Monday.
- ConfirmedMicrosoft says it will make Maia its primary architecture for running AI models.
- ConfirmedMicrosoft says Maia 300 produces 40% more tokens per megawatt than GPUs.
- ConfirmedMicrosoft did not specify which GPU systems the 40% tokens-per-megawatt comparison used.
- ConfirmedA Maia 300 rack holds 96 chips, double the Maia 200 rack.
- ConfirmedMicrosoft says a Maia 300 rack delivers up to seven times the output of a Maia 200 rack.
- ConfirmedMaia 300 production starts in 2027.
- ConfirmedEvercore ISI estimates Maia will be about 10% of Microsoft's AI chip spending by Q1 2027.
- ConfirmedMicrosoft's Maia accelerators cannot be rented by the hour by other companies.
- UnconfirmedMicrosoft is in talks with TSMC to make more than 300,000 Maia 300 chips for 2027.
Related coverage
Why you can trust this story
72%Source map · 2 outlets / 2 articles
- The Next Web
- Article 1thenextweb.comreport
- CNBC
- Article 1cnbc.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 · 10
- Claim: Microsoft unveiled the Maia 300, its third-generation custom AI chip, on Monday.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft says it will make Maia its primary architecture for running AI models.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft says Maia 300 produces 40% more tokens per megawatt than GPUs.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft did not specify which GPU systems the 40% tokens-per-megawatt comparison used.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: A Maia 300 rack holds 96 chips, double the Maia 200 rack.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft says a Maia 300 rack delivers up to seven times the output of a Maia 200 rack.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Maia 300 production starts in 2027.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Evercore ISI estimates Maia will be about 10% of Microsoft's AI chip spending by Q1 2027.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft's Maia accelerators cannot be rented by the hour by other companies.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Microsoft is in talks with TSMC to make more than 300,000 Maia 300 chips for 2027.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
Update log · 0
AI accelerates. Humans approve.