# Microsoft Maia 300 chip to run most of its AI from 2027, cutting reliance on Nvidia

Source: Searchable (https://searchable.pk/news/ai/microsoft-maia-300-chip-to-run-most-of-its-ai-from-2027-cutting-reliance-on)
Type: News article
Author: Searchable Editorial
Published: 12 Oct 2026
Updated: 12 Oct 2026
Retrieved: 13 Oct 2026

> Microsoft's third-generation chip packs 96 to a rack and claims 40% more output per megawatt than GPUs. Nvidia still trains its models, and analysts put Maia at about a tenth of its AI chip spending.

**Microsoft will run most of its own AI products on its own chips rather than Nvidia's from 2027.** The company unveiled its third-generation AI accelerator, the **Maia 300**, on Monday 12 October 2026 and said that, starting with this chip, its homegrown silicon becomes the main engine for handling incoming requests to AI models, [CNBC reported](https://www.cnbc.com/2026/10/12/microsoft-will-mainly-use-its-maia-chips-for-ai-inference.html). Production starts in 2027.

That is a significant shift for the biggest buyer of AI computing on the planet. Nvidia's graphics processors fill Microsoft's existing data centres and train the models of OpenAI and Anthropic; Maia is Microsoft's attempt to bring the cost of running those models down.

## Maia 300 in numbers

| | Maia 300 |
| --- | --- |
| Chips per rack | **96** (double the Maia 200) |
| Power per rack | about **167 kW** |
| Output per rack | up to **7 times** the Maia 200 |
| Efficiency | **40% more tokens per megawatt** than GPU systems, by Microsoft's count |
| Manufacturing | TSMC **N2P** process, eight stacks of **HBM4** memory per chip |
| Production | from **2027** |

The rack, power and process figures come from Microsoft's announcement as reported by [The Next Web](https://thenextweb.com/news/microsoft-maia-300-ai-chip-2027) from Reuters. Microsoft did not name the GPU systems it compared against. For scale, Reuters noted that an Nvidia GB300 rack holds 72 GPUs and draws about 142 kW.

## Inference, not training

Maia is built for **inference**, the work of answering a prompt once a model is trained: long chats, reasoning and multi-step agents. Training stays with Nvidia and AMD, general manager **Alistair Speirs** told CNBC. "We need new architectures that will give us essentially better performance, better use of the power that we have," he said, adding that power translates "pretty directly into operational cost" for a data centre.

Executive Vice President **Scott Guthrie** wrote that Microsoft is aiming for **multi-gigawatt** deployments of the Maia family, though Speirs declined to give a timeline. Technical fellow **Jim Kleewein** said parts of **Microsoft 365 Copilot** should get faster on the new chip.

The current Maia 200, launched in January, already serves Copilot and runs OpenAI models alongside Microsoft's own MAI models; CEO **Satya Nadella** told analysts in July it "delivers 30% better performance per dollar" than the latest hardware in its fleet. By the end of 2026 it is due to support GitHub Copilot and Microsoft Foundry too.

## How much less Nvidia?

Not much, yet. **Evercore ISI** analysts estimate Maia will be about **10% of Microsoft's AI chip spending** by the first quarter of 2027, CNBC said. Unlike Google's TPUs or Amazon's Trainium, Maia cannot be rented by outside companies. "Nobody has tested them outside of Microsoft yet," said **Patrick Moorhead** of Moor Insights and Strategy. The Information reported in August that Microsoft was in talks with TSMC for more than 300,000 Maia 300 chips for 2027; Microsoft has not confirmed that number.

Microsoft says it will share more at its **Ignite** conference in San Francisco next month.

## What it means for users in Pakistan

Nothing changes in your Copilot or Office app this year. The aim is cheaper AI answers for Microsoft, which matters for what it charges for Copilot and Azure, priced in dollars for Pakistani freelancers and software houses; our [currency converter](/tools/finance/currency-converter) and the [dollar rate page](/data/usd-pkr) show what any dollar subscription costs in rupees today. TSMC, which makes the chip, reports its own quarterly results on Thursday: see [our preview](/news/technology/tsmc-results-thursday-analysts-expect-a-65percent-jump-to-a-record-dollar234).

## Frequently asked questions

**What is Microsoft Maia 300?**

Maia 300 is Microsoft's third-generation in-house AI accelerator, unveiled on 12 October 2026. It is built to run (not train) AI models, holds 96 chips per rack, and goes into production in 2027 as the main chip behind Microsoft's AI services.

**Will Microsoft stop using Nvidia chips?**

No. Microsoft still trains models on Nvidia and AMD hardware, and Evercore ISI estimates Maia will be only about 10% of its AI chip spending by early 2027.

## Sources

- Microsoft will mainly use its Maia chips for AI inference (12 Oct 2026) (CNBC): https://www.cnbc.com/2026/10/12/microsoft-will-mainly-use-its-maia-chips-for-ai-inference.html
- Microsoft says Maia 300 chip will run its AI services from 2027 (12 Oct 2026) (The Next Web): https://thenextweb.com/news/microsoft-maia-300-ai-chip-2027
- Microsoft is making its homegrown Maia 300 chip the primary engine for AI inference (12 Oct 2026) (Quartz): https://qz.com/microsoft-maia-300-ai-chip-inference-primary-101226

How to cite: "Microsoft Maia 300 chip to run most of its AI from 2027, cutting reliance on Nvidia", Searchable, 12 Oct 2026, https://searchable.pk/news/ai/microsoft-maia-300-chip-to-run-most-of-its-ai-from-2027-cutting-reliance-on