Microsoft Is Moving Copilot Onto Its Own Chips

Microsoft announced Maia 300 on October 12, the third generation of its in-house AI inference chip. The sentence to read twice came from Scott Guthrie, who runs Microsoft's cloud and AI group: the company will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.

What was announced, and how firm each part is

Claim	Status
Maia becomes the primary inference system for Microsoft products	Stated plan, from 2027
40% more tokens per megawatt than leading GPU systems	Microsoft expects; no baseline named
Up to 7x the rack-scale throughput of Maia 200	Microsoft figure, 96-accelerator config
Up to 1.7x more tokens per watt on Maia 200	Microsoft figure, shipping today
Maia 300 production deployments	Begin 2027

Maia 200 is the part already running. Microsoft says it is deploying across the United States, Europe and Australia and powering built-in AI features in Teams and Copilot. Maia 300's numbers are the company's own projections for silicon that ships next year, against an unnamed comparison, which is normal for a chip announcement and worth holding loosely.

The hardware detail that supports the efficiency story is memory. Microsoft describes eight HBM4 stacks delivering 18 terabytes per second of peak bandwidth inside a 950-watt envelope, with a rack of 96 accelerators fitting into 167 kilowatts. Memory bandwidth is what long-context and agent workloads actually wait on, which is the shift described in what cheaper compute means for business.

The unit tells you the constraint

Chip announcements usually lead on speed. Leading on output per megawatt is a different message, and it is consistent with what has been happening around data centre power all year: grid connections queued, large-load rules being written by regulators, and operators competing for electricity rather than for silicon.

For a provider with a fixed power envelope, efficiency per watt converts directly into how many customers can be served from the same site. That makes it a commercial metric rather than an engineering one, and it explains why a company would publish it prominently. The underlying squeeze is covered in the AI power bottleneck.

Efficiency gains at this layer tend to show up for customers as capacity and availability before they show up as lower prices. More tokens per megawatt means fewer waiting lists and fewer rate limits, which is worth more to most businesses than a percentage off a bill they struggle to forecast anyway.

What this changes for a business on Azure

Nothing this quarter. Inference hardware is not a customer-selectable option, the rollout begins in 2027, and the AI features in Teams and Copilot will behave the same way on the day the hardware underneath them changes.

Three things change gradually and belong in your notes rather than your calendar. The cost structure behind the AI you pay for becomes more of Microsoft's own making, which affects how much room it has to price. Model availability on Azure may increasingly reflect what runs well on Maia, so the newest third-party model and the Azure-available model can diverge. And the gap between a workload built on a raw model API and one built on provider-specific services gets wider in terms of what moving would cost.

Every major provider is doing a version of this, which is the relevant context. Custom inference silicon is now standard practice among the large clouds, so this is not Microsoft breaking ranks. It does mean the question of what your AI actually runs on has three different answers depending on who you buy from.

The portability question worth answering now

Custom silicon does not create lock-in on its own, because the layer you build against is the model and its API. The exposure comes from what gets wrapped around it: proprietary agent frameworks, provider-specific data connectors, tuning formats, and evaluation tooling that exists in one place.

A concrete test beats a worry. For each AI workload you run, ask how many engineering weeks a move to a different provider would take, and whether anyone has checked. Keeping one workload deliberately portable is the cheapest way to keep that number honest, which is the hedging approach in model hedging to avoid vendor lock-in.

For businesses buying Copilot seats rather than building, the exposure is commercial rather than technical, and the lever is the renewal. Current Canadian pricing is in Microsoft Copilot pricing Canada.

What stays the same

Your AI costs still move with your usage, and a more efficient chip in 2027 does not change the arithmetic of a workflow that makes eight model calls per run today. Forecasting that remains the same exercise described in cash flow forecasting when AI costs are variable.

Hardware announcements are also not a reason to delay. The capability available to you today is already ahead of what most businesses have deployed, and waiting for the next chip generation has been a losing strategy for three years running.

Frequently Asked Questions

What did Microsoft announce about Maia 300?

On October 12, 2026, Microsoft announced Maia 300, the third generation of its in-house AI inference chip, and said it will scale the Azure Maia family to multi-gigawatt deployments starting in 2027. Scott Guthrie, the executive vice president of Cloud and AI, said this would make Maia the primary inference system for Microsoft products and services. Production deployments of Maia 300 begin in 2027.

Are the performance numbers measured or projected?

Projected, and the wording matters. Microsoft says it expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and that a full 96-accelerator configuration delivers up to seven times the rack-scale token throughput of Maia 200. These are the company’s own figures for a chip shipping next year, with no named baseline system, so treat them as a direction of travel rather than a benchmark. The previous generation, Maia 200, is in production now and is credited by Microsoft with up to 1.7 times more tokens per watt for built-in AI features.

Why measure tokens per megawatt instead of speed?

Because power has become the thing in short supply. A data centre operator with a fixed grid connection cannot buy its way to more electricity in the short term, so the question stops being how fast a chip is and becomes how much useful output each megawatt produces. When a company optimises for that unit publicly, it is telling you which constraint it expects to be binding.

Does this affect my business if we use Copilot or Azure?

Not immediately, since the rollout starts in 2027 and inference hardware is not something a customer selects. Over time it affects three things worth tracking: the cost structure behind the AI features you pay for, which models are available and when on Azure versus elsewhere, and how portable a workload built on Azure AI services really is. None of that calls for action this quarter, and all of it belongs in your next renewal conversation.

Should we be worried about lock-in?

Worth examining rather than worried about. Custom silicon does not by itself lock you in, since the models and APIs are the layer you actually build against. The practical exposure comes from provider-specific services wrapped around the model, such as proprietary agent frameworks, data connectors and tuning formats. A reasonable test is to ask what a move would cost in engineering weeks, and to keep one workload deliberately portable so the answer stays known.