Microsoft to make Maia chips the backbone of its inference and scale to gigawatts by 2027

From 2027, Microsoft will make its in-house Azure Maia accelerators the primary inference system across all of its products and deploy them at a scale of several gigawatts. The company's main hope is the Maia 300 chip: Microsoft claims it will deliver 40% more tokens per megawatt than leading GPU systems. CEO Satya Nadella called the move "our biggest bet on in-house silicon."

Why count tokens per megawatt

The metric is not an accident. For large data centers, the main constraint is increasingly not the number of chips but the available electrical power: grid connections, substations, cooling. If each megawatt yields more tokens, the same energy budget serves more requests. For high-volume inference services such as Copilot, this directly affects the cost of each response.

The claim comes with a caveat. The comparison against "leading GPU systems" does not name a specific configuration, and there are no independent Maia 300 benchmarks yet. The 40% figure remains a promise from the manufacturer, which will also be the one using the chip.

What we know about the plan

• Timeline: the transition begins in 2027.
• Role: Maia becomes the primary inference system in Microsoft products.
• Scale: multi-gigawatt deployment.
• Chip: Maia 300, the next generation of the Azure Maia line.
• Claimed advantage: 40% more tokens per megawatt than leading GPU systems.
• Leadership position: Nadella calls it the company's biggest bet on in-house silicon.

Note the wording around inference. The announcement is about serving model responses, not training. What will happen to GPU purchases for training does not follow from the statement.

Context: every major cloud is heading toward its own silicon

Microsoft showed the first generation of Maia in 2023, alongside the Cobalt processor. At the time it looked like an experiment against the backdrop of massive Nvidia purchases. Google has long developed TPUs, AWS has Trainium and Inferentia, and Meta has its own MTIA accelerators. The logic is the same for all of them: inference is a constant and growing expense, and paying the markup for a general-purpose GPU makes little sense when the workload is known in advance.

The scale of the new plan is different. Keeping in-house chips as a supplement to an Nvidia fleet is one thing; making them the foundation for all products is another. That requires a mature software stack, stable supply from the chip manufacturer, and sufficient memory capacity for the accelerators. The announcement gives no details on any of these points, and they will determine whether the claimed advantage holds up on real models.

Microsoft also has an extra lever: it controls which models run in its products, and in which modes. That lets it tune hardware to a specific workload, something a vendor of general-purpose GPUs cannot do.

What Azure customers and teams building on it should do

• Watch the price per token. If the energy savings are real, they will sooner or later show up in the pricing of Azure OpenAI and other managed services.
• Test latency and throughput on your own prompts, not on advertised averages. Changing the infrastructure under the hood can change how fast a service behaves.
• Do not tie your architecture to specific hardware. For those deploying their own models, Maia support in familiar frameworks will become a practical question as early as 2027.
• For investors and buyers: the signal is that part of inference demand is moving from Nvidia to the clouds' proprietary chips. How large that part is depends on whether the 40% is confirmed in practice.

The promise can be checked once Maia 300 starts serving real traffic. Until then, the figure remains a target for which Microsoft is prepared to rebuild its infrastructure.