Microsoft Maia 300: Specs, 2027 Rollout and How It Compares

Every Copilot answer you receive comes with a hardware and electricity bill. Microsoft Maia 300 is Microsoft’s third-generation custom AI accelerator, announced on October 12, 2026, at the OCP Global Summit. Microsoft says its Maia family will become the main system handling inference for its products and services from 2027.

Why build another chip? Microsoft wants more AI work from the power and money it spends, and Maia 300 pairs a TSMC N2P compute die with 288GB of HBM4 memory. Before we get into what that means for your Copilot requests, here are the main numbers.

Key fact	Microsoft Maia 300
Announcement	October 12, 2026, at the OCP Global Summit
Design and manufacturing	Microsoft-designed chip with a TSMC N2P compute die
Memory	Eight HBM4 stacks providing 288GB and 18TB/s
Peak compute	13.8 PFLOPS at FP4 and 6.9 PFLOPS at FP8
Power	950W chip TDP, with 96 accelerators in a 167kW rack
Intended workloads	Inference supporting Copilot, GitHub Copilot, Foundry, OpenAI and MAI models
Production plans	Production and deployments measured in multiple gigawatts beginning in 2027
Efficiency claim	Microsoft claims about 40% more useful computation per megawatt than GPU systems it has not identified

What Microsoft Maia 300 actually is

Maia 300 is the latest member of Microsoft’s own AI chip family, developed by its Azure Hardware Systems team under Rani Borkar. To understand why you should care, start with the work Microsoft built it to handle.

Let’s backtrack a little. Maia 100 arrived in November 2023, followed by Maia 200 in January 2026. Maia 300 is the third generation, with hardware designed around running large language models.

Training and inference do different jobs. Training teaches a model using data, while inference uses that trained model to process your request and produce an answer. When you ask Copilot a question, inference is the work needed to respond.

You see, Microsoft is paying particular attention to the stage where models generate their responses. Maia 300 targets tasks involving long conversations, reasoning, tool use and AI agents, where generating output can be demanding. That helps explain the emphasis on memory and communication between chips.

Does this mean Microsoft has finished buying other AI hardware? No, Microsoft’s Alistair Speirs told CNBC that Nvidia and AMD remain partners for model training. You should read Maia 300 as an inference investment, alongside those existing relationships.

Maia 300 specs at a glance

The numbers tell you where Microsoft spent its engineering effort. Put Maia 300 beside Maia 200, and the gains in memory and rack capacity become easier to see.

Specification	Maia 300	Maia 200
Compute manufacturing process	TSMC N2P	TSMC 3nm
Peak FP4 compute	13.8 PFLOPS	10.1 PFLOPS
Peak FP8 compute	6.9 PFLOPS	5.1 PFLOPS
Chip TDP	950W	750W
HBM memory	Eight HBM4 stacks, 288GB	Six HBM3E stacks, 216GB
Memory bandwidth per chip	18TB/s	7.1TB/s
Bidirectional scale-up bandwidth	3.2TB/s	2.8TB/s
Accelerators per rack	96	48
Rack power	167kW	64kW
HBM capacity per rack	27.6TB	10.4TB
Maximum linked accelerator count	12,288	Up to 6,144

Here’s the thing. Microsoft puts the memory bandwidth gain at about 2.5 times, while compute grows by up to roughly 36%. Its explanation is that moving data around large inference systems is becoming a tighter constraint than arithmetic, so the extra bandwidth addresses that problem.

Inside the chip and the rack

Maia 300’s design goes beyond the accelerator itself. You need to look at its silicon, cooling and network together, so let’s take those pieces one at a time.

Chiplets on 2nm and 3nm silicon

Microsoft separates the compute work from memory control and input-output functions across several dies. The compute die uses TSMC N2P, while Reuters describes the supporting data-transfer chiplets as using 3nm manufacturing. StorageReview lists two I/O dies and four memory dies alongside the compute die.

The memory layout changes too. Maia 300 moves from uniform memory access, or UMA, to non-uniform memory access, or NUMA, across its chiplets. That distinction matters when you examine how software manages data across the separate parts.

For numerical formats, the chip supports enhanced MXFP4 and storage formats below three bits, with hardware handling type conversion. Its vector engine is optimized for softmax and low-precision casting. These are technical details, but they help explain how Microsoft designed the silicon for language-model workloads.

A 96-chip liquid-cooled rack

A complete Maia 300 rack brings substantial power and cooling requirements. Here is what sits behind the 96-chip total.

• Accelerator trays. The rack contains 24 trays, each holding four accelerators.
• Power supply. Rack power reaches 167kW, delivered through a 50V busbar with two power feeds.
• Cooling capacity. Liquid cooling carries away more than 80% of the rack’s heat.
• Network switches. Six switch ASICs sit within three liquid-cooled switch trays.
• Site compatibility. Cooling can use facility water or a heat exchanger at an air-cooled location.

Ethernet networking up to 12,288 chips

Each accelerator has an integrated AI network controller with 32 Ethernet ports rated at 400Gb/s each. Connections inside the rack are electrical, while optical switches connect racks into larger groups. A superpod can contain 128 racks, giving you a maximum of 12,288 accelerators.

Microsoft specifies communication rates of up to 750GB/s from each accelerator to any peer within that superpod. Those transfers do not require a host CPU or a separate network interface card. The takeaway is that Microsoft designed the connections between chips alongside the chips themselves.

How Microsoft will use Maia 300 for Copilot, Azure and OpenAI models

Microsoft intends Maia to handle most inference work across its own products and services as deployments grow from 2027. That plan builds on Maia 200’s existing work, so start with what is already running and what comes next.

Maia 200 currently serves Microsoft’s in-house models and OpenAI models supplied through Azure. It also runs open-source models and supports Copilot workloads. Microsoft is deploying that generation across the US, Europe and Australia, with Des Moines hosting the initial deployment and Phoenix next.

For your day-to-day use, the product plans break down like this.

• Copilot. Maia already supports Copilot features, and Microsoft’s Jim Kleewein expects parts of Microsoft 365 Copilot to run faster with Maia 300.
• GitHub Copilot. Microsoft plans to have it using Maia 200 by the end of 2026.
• Microsoft Foundry. This service is also scheduled to use Maia 200 before 2026 ends.
• Claude. Anthropic was reportedly negotiating to rent Microsoft’s Maia server chips in May 2026, according to The Information as relayed by Bloomberg. That report does not establish a completed deal.

Can you rent Maia by the hour through Azure? It is not currently offered that way, unlike Google TPUs or AWS Trainium. For now, you encounter Maia through the services it powers, with Copilot and Teams retaining the same appearance regardless of the underlying chip.

Maia 300 release date and deployment timeline

Maia 300 has been announced, but Microsoft plans to begin production and large deployments in 2027. To see where that leaves you today, separate completed events from reports and future plans.

Date	Event or planned step
November 2023	Microsoft introduced Maia 100
January 26, 2026	Maia 200 made its debut
August 2026	Microsoft reportedly sought TSMC capacity for more than 300,000 Maia 300 chips for 2027
October 12, 2026	Maia 300 was announced at the OCP Global Summit
November 17-20, 2026	Microsoft plans to share further details at Ignite
End of 2026	GitHub Copilot and Foundry are scheduled to use Maia 200
2027	Maia 300 production and deployments measured in multiple gigawatts are planned

How many chips will Microsoft deploy? Reuters says the total could exceed one million during 2027, but that figure is not a chip-count target announced by Microsoft. Treat it separately from the company’s stated deployment plans.

Microsoft Maia 300 vs Nvidia, Google TPU and AWS Trainium

You can compare memory, compute and system size, but the published figures use different measurements. Before declaring a winner, check whether each number describes one chip or an entire rack.

Measure	Maia 300	Nvidia GB300 NVL72	Google Ironwood TPU	AWS Trainium3
HBM memory	288GB HBM4 per chip	20TB HBM3E per rack	192GB per chip	144GB HBM3e per chip
Memory bandwidth	18TB/s per chip	Up to 576TB/s per rack	7.37TB/s per chip	4.9TB/s per chip
FP8 compute	6.9 PFLOPS per chip	720 PFLOPS per rack with sparsity	4,614 TFLOPS per chip	362 MXFP8 PFLOPS per 144-chip UltraServer
System size	96 chips per rack, 12,288 per superpod	72 GPUs per rack	Up to 9,216 chips per pod	Up to 144 chips per UltraServer
Rack power	167kW	About 142kW, according to Reuters	Not disclosed	Not disclosed
Rental access	No hourly rental offering	Available through cloud providers	Available through Google Cloud	Available through AWS

Here’s the thing. Nvidia’s compute figure includes sparsity, and several entries describe whole systems rather than individual chips. Those differences prevent you from treating this table as a direct speed ranking.

Microsoft previously claimed Maia 200 delivered three times Trainium3’s FP4 performance and exceeded Google’s seventh-generation TPU at FP8. Those were Microsoft’s comparisons, rather than independent test results. Keep that distinction in mind when assessing Maia 300.

The same caution applies to Microsoft’s claim of about 40% more useful computation per megawatt than leading GPU systems. Microsoft has not identified the competing configurations, and analyst Patrick Moorhead told CNBC that nobody outside Microsoft had tested Maia. You have a company efficiency claim, with outside verification still missing.

What Maia 300 means for you and AI costs

The chip change happens behind the services you use. What matters for your experience is whether Microsoft turns that hardware investment into faster responses or different costs, so here is what you can expect so far.

• Familiar apps. Copilot and Teams retain the same appearance regardless of the hardware processing your requests.
• Possible speed gains. Microsoft’s Jim Kleewein expects some parts of Microsoft 365 Copilot to become faster with Maia 300. That is an expectation, rather than a published guarantee for every request.
• No promised price reduction. Microsoft has not said whether savings will change Copilot prices or usage limits.
• Electricity matters. Microsoft’s efficiency claims connect AI output to power consumption, making energy use part of the cost argument.
• A large spending commitment. Reuters reports that Microsoft expects capital expenditure above $175 billion during 2026.
• Capacity remains relevant. In July, Microsoft said it expected capacity shortages to continue at least through 2026.
• No direct hourly access. You cannot currently rent Maia through Azure in the same way you can rent Google TPUs or AWS Trainium.

Questions Microsoft still has not answered

The announcement gives you plenty of hardware numbers, while leaving several practical questions unresolved. Keep these questions beside the headline claims.

• GPU comparison. Which competing hardware and configuration produced the baseline for the 40% efficiency claim?
• Copilot usage. What proportion of Copilot requests will run on Maia?
• Deployment total. How many Maia 300 chips will Microsoft actually install?
• Customer access. Will Azure customers eventually get direct rental access?
• Regional locations. Which sites in Europe and Australia will host Maia deployments?
• Program spending. What has Microsoft spent developing and deploying the Maia family?

What to watch next

The next useful updates should tell you more about deployment and service performance. Start with the scheduled events, then check whether the reported plans become confirmed commitments.

• Ignite details. Microsoft has promised more information during its November 17-20 event in San Francisco and online.
• Year-end services. Watch for GitHub Copilot and Foundry to begin using Maia 200 as planned.
• Anthropic negotiations. Anthropic reportedly discussed renting Maia chips to run Claude. A confirmed agreement would give you firmer evidence than those reported talks.
• Production progress. Follow the 2027 manufacturing ramp and TSMC N2P capacity.
• Spending estimates. Evercore ISI estimates that Maia could represent about 10% of Microsoft’s AI chip spending by the first quarter of 2027. That remains an analyst estimate.

FAQs