Microsoft Unveils Maia 300 AI Chip for Azure Data Centres
Microsoft unveiled Maia 300 on 12 October 2026, the third generation of its in-house AI accelerator family. The chip is built on TSMC's N2P process and is meant to become the primary system for running AI inference across Microsoft's products. Production deployments start in 2027, with a scale-up to multi-gigawatt capacity planned from that point. Microsoft says the chip should deliver about 40% more tokens per megawatt than leading GPU-based systems, but it has not named the systems it compared against and has not published independent test results. For businesses that automate work, the near-term effect is indirect: Maia 300 will run Microsoft's own AI services, and Azure customers cannot choose it for their own workloads yet.
In short
- Microsoft unveiled Maia 300 on 12 October 2026 as its primary inference chip for its own AI services.
- Production deployments begin in 2027, with a multi-gigawatt scale-up planned from there.
- The 40% tokens-per-megawatt gain is Microsoft's own estimate and has not been independently tested.
- Azure customers cannot currently rent or select Maia 300 for their own workloads.
- For now, the practical effect reaches businesses through Copilot, GitHub Copilot and Foundry.
Microsoft has announced Maia 300, the newest chip in its Azure Maia line of AI accelerators. According to the company's Azure team, the chip is designed for inference, which is the stage where a trained AI model answers questions, summarises documents or carries out a multi-step task. The announcement came on 12 October 2026 at the OCP Global Summit in San Jose, California. Scott Guthrie, executive vice president of Cloud and AI, presented the chip as the primary inference system for Microsoft products and services.
What was announced
The main compute die is built on TSMC's N2P process, a 2 nanometre node. Memory and input-output functions sit on separate chiplets, so each part can be made with a process suited to its job. Each accelerator carries eight HBM4 memory stacks with 288 GB of combined capacity and 18 TB/s of bandwidth, inside a 950-watt power budget.
Microsoft groups 96 accelerators into a rack that draws about 167 kilowatts. A single scale-up domain can reach 12,288 accelerators in what Microsoft calls a superpod. Compared with the Maia 200 generation, Microsoft says a full 96-accelerator rack delivers up to seven times the token throughput.
| Item | Microsoft's stated figure (October 2026) |
|---|---|
| Compute process | TSMC N2P (2 nm) |
| Memory per accelerator | 288 GB HBM4, 18 TB/s bandwidth |
| Efficiency claim | About 40% more tokens per megawatt than leading GPU-based systems |
| Rack configuration | 96 accelerators, about 167 kW |
| Production start | 2027 |
Why it matters for businesses that automate work
Most of the AI work in a business today, such as drafting, summarising, classifying invoices or answering routine queries, is inference. The cost of that work depends heavily on how much power and hardware the cloud provider spends per answer. Microsoft's argument is that more output per megawatt lets it add AI capacity inside the same data centre footprint.
Microsoft has already moved part of its workload onto the earlier Maia 200. According to the Moor Insights and Strategy analysis of the launch, Maia 200 is deploying across the United States, Europe and Australia. Maia 300 is the next step in that shift away from buying most inference capacity from outside suppliers.
What changes in practice
For most businesses, nothing changes on the day of the announcement. Microsoft has not announced Maia instances that Azure customers can rent, and it has not said that customers will be able to select Maia for their own workloads. Reuters, as cited by Tikr, noted that Microsoft will use Maia exclusively for its own AI services. Amazon and Google, by contrast, already rent their own custom chips to cloud customers.
The practical route to Maia will run through services a company already uses. Microsoft names Copilot, GitHub Copilot and Microsoft Foundry as the first places where Maia-based capacity is expected to appear. A firm that uses Microsoft 365 Copilot or builds agents on Foundry may see changes in capacity, speed or price over time, but no one has yet published how much those changes will be.
A business owner planning automation work should treat this as background, not a reason to change plans. The steps that matter now are the ones that do not depend on the chip:
- List the routine tasks that run on AI tools today, such as data entry, document handling and follow-up messages, and record how many hours each one takes each week.
- Ask your software vendor which AI model and cloud region it uses, and whether it has a plan for capacity from 2027 onward.
- Keep your automation logic in tools you control, so that a change of model or cloud does not force you to rebuild every workflow.
The limits of the claims
The 40% efficiency figure is Microsoft's own estimate. Tikr notes that the figure has no named benchmark behind it, and it has not been tested outside Microsoft. WindowsForum's analysis of the launch makes the same point and adds that the cost to customers is not yet known. Until Maia 300 ships and independent tests appear, the figure should be read as a design target. Maia 300 also does not replace the need for people who check outputs, approve payments and handle exceptions. An AI system that runs faster does not decide whether an invoice is correct.
What to watch next
Three things will show whether the announcement changes the economics for ordinary businesses. The first is the start of production in 2027, and whether it arrives on schedule. The second is whether Microsoft makes Maia capacity selectable for Azure customers, which it has not said it will do. The third is whether any independent benchmark confirms the tokens-per-megawatt gain. Reuters reporting cited in coverage also says Microsoft could deploy more than one million of the chips next year, but that figure is a reported plan, not a Microsoft commitment.
How AiStaffo would automate this
AiStaffo would not change anything about the chip itself. The useful work is in the layer above it: connecting the tools a business already uses, such as its accounting package, inbox, document store and order system, so that routine data entry, follow-ups, reconciliations and reports run automatically. Where a business uses Microsoft 365 Copilot or Foundry, the workflows would be built so that the model or cloud behind them can be changed later without rebuilding each process. The owner still approves payments, decides on exceptions and reviews the first weeks of output. Book a free automation audit to map which of your routine tasks are ready to run automatically today.
Questions people ask
What is the Microsoft Maia 300 chip?
When will Maia 300 be available?
Can my business rent Maia 300 on Azure?
Is the 40% efficiency gain proven?
Does Maia 300 make AI automation cheaper for small businesses?
Book a free automation audit
Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.
