Microsoft’s Maia 300 is the company’s third-generation custom AI chip, announced on October 12, 2026. Microsoft says it plans to deploy potentially more than a million of the chips next year to power its own AI services, with multi-gigawatt rollouts starting in 2027. The chip uses TSMC’s 2-nanometer technology for its core computing chiplets and, according to Microsoft, delivers about 40% more useful computing per megawatt than leading GPU-based systems.
What Microsoft Announced
On Monday, October 12, 2026, Microsoft unveiled the Maia 300, the next generation of its custom silicon for AI data crunching. The announcement came with an unusually specific deployment target: more than a million Maia 300 chips going into production next year, backing Microsoft’s internal AI services, including its own AI models and models from OpenAI that run on Azure.
Rani Borkar, president of Azure hardware systems and infrastructure, framed the chip as a full-system efficiency play. In an interview with Reuters, she said the Maia 300 is about delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system rather than just the silicon.
The timing matters. Microsoft has said it expects to spend more than $175 billion in capital expenditures this calendar year, most of it on AI infrastructure. Building its own chips is how the company plans to keep that spending from flowing entirely to one supplier.
Maia 300 Specs: What We Know So Far
Microsoft has not published a full datasheet, but the details disclosed paint a clear picture of a chip designed for inference at massive scale:
- Architecture: chiplet-based design, a different technical path from Nvidia’s approach of one massive computing die surrounded by high-bandwidth memory
- Process node: 2-nanometer technology from TSMC for the central computing chiplets; cheaper 3-nanometer chiplets for data movement and I/O functions
- Rack configuration: 96 Maia 300 chips per rack
- Power: about 167 kilowatts per rack, compared with roughly 142 kilowatts for Nvidia’s GB300 racks (72 GPUs)
- Efficiency claim: about 40% more useful computing per megawatt than leading GPU-based systems, though Microsoft did not name the systems it compared against
- Deployment: multi-gigawatt rollouts starting next year
The 2-nanometer choice is significant. It puts Microsoft among the first customers shipping volume product on TSMC’s most advanced node, the same foundry that builds Apple’s and Nvidia’s flagship silicon.
How Maia 300 Differs From Nvidia’s GPUs
Microsoft is deliberately not trying to build a general-purpose GPU. The Maia family is tuned for inference, the phase of AI that runs trained models to generate answers, code, and actions for real users. Nvidia’s GB300 and upcoming racks are built for both training and inference and are sold to everyone. Maia 300 is built for one customer with one problem: making Microsoft’s own AI services cheaper to run at cloud scale.
That shows in the design. Nvidia packs a single enormous compute die with high-bandwidth memory stacked around it. Microsoft splits the Maia 300 into chiplets, putting the expensive 2-nanometer silicon only where it does the actual computing and using cheaper 3-nanometer parts to shuffle data around. The result is a chip that trades raw peak performance for better tokens per watt and tokens per dollar on the workloads Microsoft actually runs.
For more on how Nvidia’s hardware line is evolving, see our RTX 6080 breakdown and the NVIDIA RTX Spark launch coverage.
Why Microsoft Is Building Its Own Chips
The announcement is the latest move in a multi-year strategy to bring more of Microsoft’s AI stack under its own roof. Three forces are driving it:
Cost control. Every ChatGPT query, Copilot suggestion, and AI agent action that runs on an Nvidia GPU rents margin to Nvidia. Microsoft’s chief technology officer Kevin Scott has suggested the long-term goal is to run most of the company’s AI workloads on its own silicon.
Supply security. Nvidia controls a large share of TSMC’s advanced packaging capacity, and hyperscalers must lock in massive chip orders years ahead to guarantee access. By betting on its own accelerator, Microsoft carves out dedicated manufacturing capacity that cannot be crowded out by rivals’ GPU demand.
Inference economics. Analysts note that while training makes headlines, inference is the phase that will permeate enterprise workloads, since every application and interaction will involve some kind of inference. A chip optimized for tokens per watt and tokens per dollar matters more for Microsoft’s margins than one optimized for benchmark records.
Microsoft is not alone. Google has its TPUs, Amazon has Trainium and Inferentia, and OpenAI is developing its own silicon too. The AI chip race is now a foundry-level contest, and the Maia 300 is Microsoft’s biggest bid yet.
The Maia Family: From Maia 100 to Maia 300
Maia 300 did not appear out of nowhere. Microsoft’s custom silicon program has been building for three years:
- November 2023: Microsoft announces the Maia 100 at Ignite, alongside the Arm-based Azure Cobalt 100 CPU, framing both as a silicon-to-service approach. The Maia 100 is built on TSMC’s 5-nanometer process with roughly 105 billion transistors and 64GB of HBM2e memory.
- Late 2024: Maia 100 goes into production for Azure OpenAI Services, powering Copilot and Azure AI Studio workloads.
- January 2026: Maia 200 is announced, an inference-optimized chip on TSMC’s 3-nanometer process with 216GB of HBM3e memory. It deploys in Microsoft’s Iowa and Arizona regions.
- October 2026: Maia 300 is announced, jumping to 2-nanometer chiplets with deployment targets an order of magnitude larger.
Each generation has moved the program closer to Scott’s goal: Microsoft AI workloads running primarily on Microsoft silicon, with Nvidia GPUs reserved for peak demand and training jobs where they still win.
What the Maia 300 Means for Azure and the Cloud Market
For customers, the Maia 300 matters in two ways. First, cheaper inference should eventually show up as lower-priced Azure AI services, since the cost per token falls when the underlying hardware is more efficient. Second, Microsoft’s bet on its own silicon reduces the risk that Nvidia supply constraints or pricing power become Azure capacity constraints.
For the market, the message is competitive. Amazon is spending $1 billion on data center programs, Google is pushing its TPU roadmap, and OpenAI is designing its own chips. The cloud giants are all concluding the same thing: at the scale of hundreds of billions in AI capex, renting someone else’s chips forever is the most expensive option.
There is a check on the hype. Microsoft’s 40% efficiency claim is a vendor number with no published methodology, and the comparison systems were not named. Independent benchmarks, if they ever come, will tell the full story. What is not in doubt is the scale of the bet: more than a million chips, multi-gigawatt deployments, and $175 billion in capex behind them.
FAQ
What is the Microsoft Maia 300?
The Maia 300 is Microsoft’s third-generation custom AI chip, announced October 12, 2026. It is a chiplet-based accelerator built on TSMC’s 2-nanometer process, designed for AI inference workloads such as coding and agentic business tasks on Microsoft’s internal services.
When will the Maia 300 be available?
Microsoft says it plans to deploy potentially more than a million Maia 300 chips next year, with multi-gigawatt rollouts starting in 2027. The chip is for Microsoft’s own data centers, not for sale to other companies.
Is the Maia 300 faster than Nvidia’s GPUs?
Microsoft claims about 40% more useful computing per megawatt than leading GPU-based systems, but it did not name the comparison systems or publish a methodology. The Maia 300 is optimized for inference efficiency on Microsoft’s workloads rather than peak benchmark performance.
What workloads will the Maia 300 run?
According to Microsoft, the chip handles AI data crunching such as coding and carrying out complex business tasks on a user’s behalf. Its predecessor already runs Microsoft’s in-house AI models and OpenAI models delivered on Azure, and the Maia 300 continues that role at larger scale.
Why is Microsoft making its own AI chips instead of buying Nvidia?
Three reasons: cost control on AI inference at massive scale, supply security for TSMC manufacturing capacity, and efficiency gains tuned to Microsoft’s own workloads. CTO Kevin Scott has said the long-term goal is to run most of Microsoft’s AI workloads on its own silicon. For more on the agent workloads driving this, see our OpenAI Dots explainer.