Microsoft Unveils Maia 300 AI Chip: Specifications, Purpose and 2027 Deployment Plan

Microsoft unveiled Maia 300, its third-generation custom AI accelerator, on 12 October 2026. The chip is built mainly for AI inference. Production deployments are planned to begin in 2027, so as of 14 October 2026 the unveiling is complete but the large-scale rollout has not started. The source article we used was published on 13 October 2026, one day after the event.

What happened and why it matters

Maia 300 is a custom chip designed in-house by Microsoft for its Azure cloud. Microsoft says it wants more than one million chips and multi-gigawatt deployment across Azure. For exam preparation, the story links three topics: AI hardware, semiconductor manufacturing and the power demands of data centres.

Event date and source date

• Event date: 12 October 2026 (unveiling, completed).
• Source publication date: 13 October 2026 (GKToday).
• Planned for 2027: start of production deployments and scaling of Azure Maia deployments to multiple gigawatts (scheduled, not yet happened).

What AI inference means

AI inference is the use of an already trained AI model to produce outputs from new inputs. Training builds the model, while inference is what happens when users actually query it. Maia 300 is designed for inference workloads such as long-context reasoning, latency-sensitive conversations and multi-step agent tasks.

Intended products and services

Microsoft plans to use the Maia family as an inference system for Microsoft Copilot, GitHub Copilot and OpenAI models hosted on Azure.

Maia 200 and the 40% claim

The predecessor, Maia 200, is already powering Microsoft Copilot and supporting Microsoft's MAI team. GitHub Copilot and Microsoft Foundry are scheduled to use Maia 200 by the end of 2026, which is still in the future as of 14 October 2026.

Microsoft anticipates that Maia 300 will deliver 40% more tokens per megawatt than leading GPU-based systems. This is an expected figure, not a measured result, and the comparison systems are not named. A token is a unit of text or other data processed by an AI model.

Key specifications at a glance

Chiplet design and HBM

In a multi-chiplet design, specialised computing components are combined in one processor package. Microsoft's design pairs compute chiplets with HBM4 inside a custom accelerator, rather than relying on a single large computing die. HBM stands for high-bandwidth memory, which is built to move data at high rates. This matters because AI computation depends on moving large amounts of data quickly.

Scale and power

The source says a deployment of about 10,400 Maia 300 racks would hold roughly one million accelerators and need about 1.7 gigawatts of power. Microsoft plans to scale Azure Maia deployments to multiple gigawatts from 2027.

Limits of the available evidence

• The account rests on a single secondary report (GKToday). No Microsoft statement or technical document was available for cross-checking.
• The 40% tokens-per-megawatt figure is Microsoft's expectation, and the GPU systems used for comparison are not identified.
• No pricing, availability regions or customer rollout schedule beyond 2027 was given.
• The evidence does not say that this topic has appeared or will appear in any exam. Treat it as general science and technology current affairs.

Frequently asked questions

What is Maia 300?

It is Microsoft's third-generation custom AI accelerator, designed mainly for AI inference on Azure.

Is Maia 300 already in production use?

No. Microsoft unveiled it on 12 October 2026, and production deployments are planned to begin in 2027. Its predecessor, Maia 200, is already powering Microsoft Copilot.

Practice MCQs

Question 1

On which date did Microsoft unveil the Maia 300 AI chip?

1. 12 October 2026
2. 13 October 2026
3. 1 January 2027
4. 31 December 2026

Answer: 12 October 2026 — the source states Microsoft unveiled Maia 300 on that date, while 13 October 2026 is the article's publication date.

Question 2

Which manufacturing process is used for the compute chiplets of Maia 300?

1. Intel 18A
2. TSMC N2P
3. Samsung 3GAP
4. TSMC N7

Answer: TSMC N2P — the source says the compute chiplets are built on TSMC's N2P, a 2-nanometre-class process.

Question 3

How many accelerators is each Maia 300 rack designed to hold, and within what power envelope?

1. 48 accelerators in 100 kilowatts
2. 96 accelerators in 167 kilowatts
3. 96 accelerators in 1.7 gigawatts
4. 10,400 accelerators in 167 kilowatts

Answer: 96 accelerators in 167 kilowatts — this is the stated rack design in the source.

Question 4

Which memory configuration does Maia 300 integrate?

1. Four HBM3 stacks with 144 GB
2. Eight HBM4 stacks with 288 GB and 18 TB per second bandwidth
3. Sixteen HBM4 stacks with 576 GB
4. Eight HBM4 stacks with 18 GB per second bandwidth

Answer: Eight HBM4 stacks with 288 GB and 18 TB per second bandwidth — these figures are given in the source.

Question 5

What is Maia 300 primarily designed for?

1. AI inference workloads
2. Quantum computing
3. Smartphone processing
4. Satellite communication

Answer: AI inference workloads — the source describes it as designed for inference, including long-context reasoning, conversations and multi-step agent tasks.