Microsoft Maia 300 to hit gigawatt scale from 2027
Microsoft says Maia 300 will carry its inference workloads from 2027, claiming 40 percent more tokens per megawatt than leading GPU racks.
In short
Microsoft plans to run its own Azure Maia chips at multi-gigawatt scale starting in 2027, with the upcoming Maia 300 becoming the main inference engine behind Copilot, GitHub Copilot and Teams.
At a glance
- Maia 300 ships in 2027 and is claimed to deliver 40 percent more tokens per megawatt than leading GPU systems.
- A 96-chip Maia 300 rack should produce up to seven times the tokens of a Maia 200 rack.
- Per chip, Microsoft lists eight HBM4 stacks, 288 GB of memory, 18 TB/s of bandwidth and a 950-watt power rating.
- Maia 200 already runs Copilot in the US, Europe and Australia; GitHub Copilot and Foundry follow by the end of 2026.
- Nadella calls it the company's biggest bet yet on its own silicon, with more detail due at Ignite 2026 in November.
Microsoft intends to operate its in-house Azure Maia accelerators at multi-gigawatt scale from 2027, and to make the upcoming Maia 300 the company's primary system for AI inference. Inference means serving finished models in production — not training new ones.
The numbers Microsoft is putting forward
Microsoft says Maia 300 delivers 40 percent more tokens per megawatt than leading GPU systems, and that a rack of 96 chips produces up to seven times the tokens of a Maia 200 rack. Each chip is specified with eight HBM4 memory stacks, 288 GB of memory, 18 TB/s of bandwidth and a 950-watt power rating.
Note what is being measured. The headline figure is throughput per unit of power, not raw performance per chip — and power is the constraint that currently caps data center buildout, together with the cooling and grid capacity that comes with it.
Maia 200 is already in production
The previous generation serves Microsoft Copilot in the US, Europe and Australia today, with GitHub Copilot and Foundry scheduled to follow by the end of 2026. In Teams, Microsoft puts Maia 200 at up to 1.7 times more tokens per watt than leading GPUs.
The silicon runs OpenAI models, Microsoft's own MAI models and open-source models. The workloads Microsoft names are long inputs, fast chat turns and agentic tasks.
Why own the inference layer
CEO Satya Nadella describes the move as the company's biggest bet yet on its own silicon. The logic is margin and fit: serving your own products on your own chips keeps the accelerator markup in house, and lets the chip, the rack and the cooling loop be tuned to a known set of models.
Inference is the sensible place to start. The load is predictable, it repeats at volume across the same handful of products every day, and it does not need to flex around constantly changing training runs.
What the sources do not establish
Every performance figure here is Microsoft's own; no independent measurements are available. Neither article identifies which GPU systems were used as the comparison baseline, and neither says whether Microsoft keeps buying Nvidia and AMD accelerators alongside Maia, or at what volume.
What “multiple gigawatts” translates to in chip counts or sites is also left open. Microsoft has said more will follow at Ignite 2026 in November.
FAQ
When does Maia 300 ship?
Microsoft says 2027, which is also when the multi-gigawatt deployments are meant to begin.
How much faster is Maia 300 than GPUs?
Microsoft claims 40 percent more tokens per megawatt than leading GPU systems. That figure has not been independently verified.
Does this mean Microsoft is dropping Nvidia?
Neither source addresses that. Maia is announced as Microsoft's main inference system for its own products, which is not the same as replacing all GPU purchases.
