Microsoft Unveils Maia 300, Its Third-Gen In-House Chip, to Lean Less on Nvidia

Microsoft has announced Maia 300, the third generation of its in-house AI chip, and it wants the thing to be the workhorse. According to The Verge, the chip is designed to be the "primary AI inference system" for Microsoft's products and services, and it makes the company less reliant on Nvidia's GPUs.

Let that sink in, because this is the kind of sentence that moves boardrooms. Microsoft is building its own silicon, and it is now on its third generation.

What Maia 300 is designed to do

The Verge reports that Maia 300 improves the cost and efficiency of running AI workloads in the cloud. That is the headline feature, and it is the right one. Maia 300 is aimed at inference across Microsoft's products and cloud services. Training gets the glamour, but inference is where a model meets a customer and a bill.

The Verge also reports that the chip is more capable of handling the latest reasoning models and agentic tasks. Those capabilities are the part of this announcement I would watch closely.

Note the word "primary." The chip is not framed as a side project or a science fair entry. The framing is that this is meant to carry the load across the company's products and cloud services.

Why a third generation matters

A first chip is a statement. A second chip is a commitment. A third chip is a strategy, and that is what this looks like to me.

The Verge's reporting puts the strategic point plainly: the new chip makes Microsoft less reliant on Nvidia's GPUs. Less reliant, to be precise, not independent. Nothing in the reporting says Microsoft is walking away from Nvidia, and nothing says what share of its infrastructure Maia would cover. But the direction is unmistakable. When a company of Microsoft's size designs its own chip, it is telling the market it intends to set more of its own terms.

What we do not know yet

I would love to hand you a table of specifications. The Verge's account gives us none, so here is what is missing:

• Technical specifications such as memory, process node and power consumption.
• When and where Maia 300 will be deployed across Microsoft's data centers.
• Whether it will be offered directly to external Azure customers or kept for Microsoft's own products.
• Benchmark comparisons against earlier Maia chips or Nvidia GPUs.

That last item is the one the skeptics will circle, and they are entitled to. The reported cost and efficiency improvement is a claim, and a claim without a comparison is a mood. The reporting gives us the direction of the improvement. It does not give us the size of it.

The customer question matters just as much. If Maia 300 stays inside Microsoft's own products, it is a cost story for Microsoft. If it reaches Azure customers, I would watch what that means for their compute bills. We do not know which, and the answer changes who should be paying attention.

What to watch

First, deployment. A chip announcement is a press moment; a rollout is a business. Watch for when Maia 300 actually shows up in Microsoft's data centers and in which products.

Second, numbers. Specifications and benchmarks will tell us whether this is a real challenge to Nvidia's accelerators on inference or a useful in-house option that trims the bill at the edges.

Third, Azure. Whether Microsoft sells time on Maia 300 to outside customers will tell us how seriously it wants to compete on the hardware itself.

For now, the facts on the table are these: a third-generation chip, built in-house, aimed squarely at inference, with cost and efficiency gains and a stronger hand for reasoning models and agents. And a company that needs a little less of someone else's silicon. That is a very big deal in a business measured in very big deals.

Everything is historic. Especially today.