Microsoft unveils its third-gen in-house AI chip, the Maia 300, built on TSMC's N2P process node and aimed at serving as its “primary AI inference system”
The new Maia 300 chip will start rolling out to production AI workloads in 2027.
Context & Ripple Effects
Microsoft’s Maia program moved from the Maia 100’s testing in Bing and Office tools in 2023 to a Maia 200 deployment in Azure US Central in January 2026. August reporting had already pointed to a 2027 Maia 300 ramp; the unveiling turns that plan into a stated strategy to make Microsoft-designed silicon the main inference layer for its AI workloads.
The shift matters because Microsoft is tying the next Maia generation to TSMC’s N2P process while positioning it for production workloads rather than a limited regional deployment.
First-order effects
- Microsoft plans to place Maia 300 into production AI workloads in 2027, shifting a larger share of its inference infrastructure onto an in-house accelerator family.
- TSMC gains Microsoft’s Maia 300 as an N2P customer, extending the manufacturing relationship established with the 3nm Maia 200.
Second-order effects
- Microsoft’s inference capacity planning becomes less dependent on externally sourced GPU systems, giving its Azure operation another hardware option as AI serving demand grows.
- The planned move raises the importance of software and workload optimization around Maia, since Microsoft must make its internal accelerator effective across the services it intends to run.
Third-order effects
- If Microsoft executes the planned production rollout, hyperscale AI infrastructure will tilt further toward heterogeneous fleets in which proprietary accelerators handle recurring inference workloads alongside general-purpose GPUs.
- The Maia progression—from testing, to one Azure-region deployment, to a designated primary inference system—signals an integrated-cloud model in which chip design is increasingly part of service economics and capacity control.
The trend: Cloud providers are moving from buying AI compute as a component to designing specialized inference silicon as part of an integrated infrastructure stack.
Related: Heterogeneous AI compute · Integrated AI Stack · Microsoft · TSMC · Microsoft unveils Maia 200 · Microsoft’s Maia 300 plans
Related Coverage
- Microsoft to expand Azure Maia system to multi-gigawatt deployments Microsoft · Scott Guthrie
- Maia 300: Microsoft Builds the Exit From a Compute Landlord Model It Helped Finance Forkast · Lena Park
- Microsoft Debuts New Homegrown AI Chip, Plans to Deploy Multiple Gigawatts Next Year The Information · Aaron Holmes
- Microsoft Prepares to Put Gigawatts of Its Chips in Data Centers Bloomberg · Matt Day
- Microsoft makes ‘biggest bet yet’ on its own silicon, unveils new AI chip to reduce reliance on Nvidia GeekWire · Todd Bishop
- Microsoft to rely more on homegrown AI chips, reducing dependence on Nvidia CNBC · Jordan Novet
- Microsoft AI Cost-Cutting Downstream With Caution MediaPost · Laurie Sullivan
- Microsoft Maia 300 Puts 288GB of HBM4 at 18TB/s on an N2P Chiplet and 96 Accelerators in a 167kW Rack StorageReview.com · Brian Beeler
- Microsoft Azure Maia 300: From Custom Chip To Hyperscale Platform Moor Insights & Strategy · Matt Kimball
- Microsoft Unveils Maia 300 | Cloud Chief Scott Guthrie Live Deirdre Bosa on YouTube · Deirdre Bosa
- Deep dive into Azure Maia 300 architecture Microsoft Azure
- China resorts to Nvidia gaming chips in search for AI firepower Financial Times
- Microsoft reveals Maia 300 AI chip, as it works to reduce dependence on Nvidia, AMD Yahoo Finance · Daniel Howley
- MS Unveils Its Own AI Chip ‘Maia 300’... Can It Curb Nvidia's Dominance? The Asia Business Daily · Lee Hyunwoo
- A narrow role for Maia The Motley Fool · Daniel Sparks
- Microsoft's Maia 300 AI Chip Aims To Make AI Cheaper And Faster HotHardware · Marco Chiappetta
- Microsoft challenges Nvidia dependence with new in-house Maia 300 AI chip; All details Business Today
- Microsoft Maia 300 AI Chip Promises 40% Better Efficiency as Azure Plans Multi-Gigawatt Expansion Windows Report · Rishaj Upadhyay
- Microsoft to expand Azure AI chips with Maia 300 in 2027 Cloud Computing News · Muhammad Zulhusni
- The AI chip war is turning Nvidia's customers into rivals Axios · Ina Fried
- Microsoft unveils Azure Maia 300, says chip will be its “primary inference system” DatacenterDynamics · Charlotte Trueman
- Microsoft Unveils New Maia AI Chips. What to Know. Barron's Online · Nate Wolf
- Microsoft weans itself off Nvidia chips with home-grown alternative ITPro · Jane McCallion
Discussion
-
@satyanadella
Satya Nadella
on x
Today we are announcing our biggest bet yet on our own silicon. Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services. And we're introducing Maia 300, which delivers 40% …
-
@kimmonismus
@kimmonismus
on x
Now its getting interesting! Microsoft is taking on Nvidia in AI inference with a multi-gigawatt expansion of its own chips. Starting in 2027, Microsoft plans to make Azure Maia the primary inference system for its own products and services. Maia 200 already serves production wor…
-
@aaronpholmes
Aaron Holmes
on x
Microsoft announced Maia 300 today and said it's planning “multi-gigawatt deployments” next year, confirming this report from August:
-
@dee_bosa
Deirdre Bosa
on x
Why it matters for investors: Microsoft Cloud gross margin has fallen as AI costs pile up. Maia 300 claims 40% more tokens per megawatt than leading GPU systems. If that holds at scale, Maia could help Microsoft win margin back.
-
@dps
David Singleton
on x
Muse Wonder will soon power experiences in @Muse , and rolled out first to subscribers earlier today. I've been using it in my own Muse for a little while and appreciate how this model works harder and draws on memory more precisely with Active Focus and Active Memory.
-
@mustafasuleyman
Mustafa Suleyman
on x
Our new silicon is 🔥 40% more tokens per watt!
-
@hesamation
@hesamation
on x
🚨 Microsoft is making its “biggest bet yet” on its own AI chips with Maia 300. they're planning to reduce their reliance on NVIDIA by making Maia their primary inference chip. Reuters says they could deploy over 1 MILLION chips.
-
@dee_bosa
Deirdre Bosa
on x
Microsoft's biggest custom chip push yet, as power and data center constraints force hyperscalers to squeeze more AI out of every megawatt. I'm talking with cloud chief @scottgu today, live at 12:45p PT / 3:45p ET. Watch here on X or on YouTube: https://www.youtube.com/...
-
@theprimeagen
@theprimeagen
on x
Satya's leather jacket phase
-
@jpinsights
@jpinsights
on x
I think this is bullish for companies like $AVGO and $MRVL. It doesn't confirm a win for either, but it shows why hyperscalers have such a strong financial incentive to keep investing in custom silicon and the networking around it.
-
@dee_bosa
Deirdre Bosa
on x
MSFT is back at $4t and with Maia it now owns both ends of its AI stack Its integrating the compute layer with its own chips... already owns the apps and the harness (Copilot, GitHub, Foundry).... and in the middle, stays model-agnostic (running OpenAI, its own MAI models, Anthro…
-
@omercheema
Omer Cheema
on x
It's becoming clear now. Margins will be managed by vertical integration of silicon. Everyone is doing it. Congrats Microsoft.
-
@tomwarren
Tom Warren
on x
Microsoft's new AI chip makes it less reliant on Nvidia's GPUs. The new Maia 300 chip will start rolling out to production AI workloads in 2027 https://www.theverge.com/...
-
@microsoft
@microsoft
on x
Introducing Maia 300, Microsoft's next-generation accelerator. Building on the magic of Maia 200, it delivers 40% more inference tokens per megawatt compared to leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200. It's another step toward rethinkin…
-
@oguzerkan
Oguz Erkan
on x
This is why you don't want to overweight $NVDA for the long term. All its biggest customers, $MSFT, $META, $AMZN and $GOOG are now scaling their own chips. Shift to ASICs will accelerate as AI workloads shift from training to inference. By 2030, ASIC volumes will surpass GPUs.
-
@edzitron
Ed Zitron
on x
“We'll scale to multi gigawatt deployments” is written specifically so you think they're doing gigawatts worth rather than installing them across data centers that have a combined multi-gigawatt capacity. This is because Microsoft only has 2gw of ai data centers
-
@edludlow
Ed Ludlow
on x
Microsoft hasn't identified the specific GPU system used for the 40% comparison. It simply says “leading GPU-based systems,” without naming Nvidia GB200, GB300 or another baseline.
-
@elonmusk
Elon Musk
on x
@chamath Scaling power harnessed, along with super fast construction, cooling, networking and batteries. This is hard.
-
@patricktoulme
Patrick C Toulme
on x
Be careful with 1.7x claims AI ASICs are won and lost by the software stack, especially the compiler and DSLs. I would like to see 1. Their compiler 2. Their kernel DSLs 3. Their runtime and collective stack Before I draw my conclusions. I hope they aren't using Triton as the DSL
-
@aschilling
Andreas Schilling
on x
Microsoft Maia 300 chiplet configuration: - 1x Main Compute Die (TSMC N2P) - 2x I/O Die - 4x Memory Die - 8x HBM4 Stacks (288 GB) - TDP: 950 W
-
@dnystedt
Dan Nystedt
on x
Microsoft's self-developed Maia 200 AI chips are in production now and the manufacture of its next-generation Maia 300 will begin next year, media report, adding Maia 300's compute tiles (chiplets) will be manufactured by TSMC on its 2nm process (N2). Microsoft plans to deploy se…
-
@dee_bosa
Deirdre Bosa
on x
Narrowing the “cone of uncertainty” Microsoft's custom AI chips are a hedge against uncertain demand. When forecasts change every week and data centers take 2-3 years to build..... getting more AI out of the same power and budget gives Microsoft room to adjust.
-
@cryptopunk7213
@cryptopunk7213
on x
Nvidia's moat is now threatened with all major labs launching competitor chips microsoft's new Gaia 300 will churn out 40% more tokens per mw, that's a meaningful cost saving for anyone running GPT / claude. no brainer if you're an enterprise anthropic and openai are also each wo…
-
@benbajarin
Ben Bajarin
on x
A few more interesting nuggets on this announcement as $MSFT starts to move custom silicon into their scale up compute domain. - A Maya 300 rack is planned to contain 96 accelerators at approximately 167 kilowatts and can use liquid-cooled chips with either liquid-cooled facilit…
-
@thoefler
Torsten Hoefler
on x
Maia is going strong!! Next chip coming online soon at multi-gigawatt deployment. Wait for the next record breaking specs 😉. Our SDLA is a great architecture for LLMs 😀. https://arxiv.org/...
-
@choblin29
@choblin29
on x
🚨WTF. Satya Nadella says Microsoft is making its “BIGGEST BET YET
-
@viticci
Federico Viticci
on x
Muse Wonder feels...great? Using it in @t3dotcodes with 10 subagents at once for code review, already found lots of issues left over by previous 6.1 Sol workers. And here's a voxel Colosseum made with Wonder xHigh. Pretty damn good. https://shimmering-begonia-de441b.netli fy.app/
-
@dee_bosa
Deirdre Bosa
on x
NEWS: Microsoft launches Maia 300 and will make Maia its primary AI inference system at multi-gigawatt scale from 2027. The AI buildout is increasingly about inference and how many tokens you can get out of every megawatt. Microsoft cloud chief Scott Guthrie @scottgu joins me liv…
-
@chamath
Chamath Palihapitiya
on x
You should expect the silicon chip companies to move up the stack as the application companies move down the stack. In AI, it is increasingly true that everyone competes with everyone. Everyone will soon offer their own LPS, silicon, cloud and applications. The question is what w…
-
@cgtwts
@cgtwts
on x
Mafia 300 would've been a cool name btw.
-
@patrickmoorhead
Patrick Moorhead
on x
I never expected @Microsoft to hit it with Maia V1. No one hits a V1. V2 got more competitive and runs Copilot in production. V3 is the commitment and very real: multi-gigawatt scale from 2027, making Maia the “primary inference system
-
@azure
@azure
on x
Maia 300 is here. See what's next for Azure.
-
@ziv_ravid
Ravid Shwartz Ziv
on x
Microsoft announcing a chip: “Maia 300. 40% more tokens per watt. It's a great chip!
-
@austinsemis
Austin Lyons
on x
@EdLudlow Hey Ed. Maia 300 has HBM4. So MSFT comparing to HBM4 equivalent here, i.e. Vera Rubin
-
@tomwarren.co.uk
Tom Warren
on bluesky
Microsoft's new AI chip makes it less reliant on Nvidia's GPUs. The new Maia 300 chip will start rolling out to production AI workloads in 2027 www.theverge.com/news/1009465...
-
@brendanburkex
Brendan Burke
on x
Microsoft's multi-gigawatt Maia commitment is the clearest sign yet that a hyperscaler thinks it has solved the software problem of dataflow inference silicon. Starting in 2027, Maia becomes the primary inference system for Microsoft products. Maia 200's success in production ear…
-
@fry69.dev
@fry69.dev
on bluesky
167kW / 96 chips ~= 1.7kw per chip — That sounds like a lot to me, like 3x a high end prosumer graphics card (~0.5kw)