For most of the AI boom, the hyperscaler shopping list has looked the same: as many Nvidia GPUs as the supply chain allows, at whatever price the market sets. Microsoft just made it clear it wants a second option—and it’s willing to do the work itself.
Microsoft’s biggest silicon bet yet
On October 12, the company unveiled Maia 300, the third generation of its in-house AI accelerator, and said production deployments will begin in 2027 before scaling to multi-gigawatt capacity. Reuters reports that Microsoft could put more than a million of its own chips into service next year to run its AI products.
The hardware is a serious step up. According to Microsoft’s own architecture deep dive, the compute die is built on TSMC’s N2P process and paired with eight stacks of HBM4 memory, giving each accelerator 288 GB of capacity and 18 TB/s of bandwidth. Racks hold 96 accelerators, draw 167 kW and rely on liquid cooling for more than 80% of their thermal load, while an Ethernet-based fabric links up to 12,288 accelerators in a single superpod. Microsoft claims up to 7x the rack-level token throughput of Maia 200, which only launched in January. More details are promised at Ignite in November.
Inference is where the money goes
Maia 300 isn’t trying to be everything. Microsoft built it for inference—specifically the agentic, long-context, reasoning-heavy workloads that are fast becoming the bulk of AI demand. Training a frontier model is a very large bill, but a finite one. Serving AI to hundreds of millions of Microsoft 365 and Copilot users is a bill that never stops adding up.
That’s where custom silicon earns its keep. Microsoft says it expects Maia 300 to deliver around 40% more tokens per megawatt than leading GPU-based systems. It hasn’t said which systems, and nobody outside Microsoft has tested the figure, so treat it as a target rather than a benchmark. Still, the choice of metric says a lot. Power, not chips, is increasingly the hard ceiling on how much AI a data center can serve, and tokens per megawatt is how you squeeze more output from a fixed grid connection.
There’s a cost story, too. Every workload that moves off premium-priced GPUs and onto in-house chips reduces spend.
Not goodbye to Nvidia
Nobody should read this as Microsoft walking away from Nvidia. Maia runs Microsoft’s own services rather than the compute Azure customers rent, and the company remains one of Nvidia’s biggest buyers. What changes is leverage. A hyperscaler that can run a meaningful share of its own inference on its own chips negotiates very differently from one that can’t.
Microsoft is also playing catch-up. AWS has Trainium and Google has its TPUs, and both already rent their custom silicon to Cloud customers. Arguably the safest winner in all of this is TSMC, which gets paid whichever chip ends up in the rack.
What it means for the rest of the Cloud
For hosters, Cloud providers and data center operators outside the hyperscaler club, the lessons aren’t really about Maia. They’re about where the whole market is heading.
Efficiency is the new headline metric. If the biggest platforms are optimizing for tokens per megawatt, buyers will start asking everyone else the same question. Providers selling AI capacity will need a clear answer on power efficiency, not just a count of the GPUs they have racked.
Rack density keeps climbing. At 167 kW per rack, with liquid cooling carrying most of the load, Maia 300 is another reminder that facilities designed around traditional air-cooled densities can’t host the next wave of AI hardware without serious retrofits. Anyone planning a new build or a major upgrade should be designing for liquid cooling and high-density power from day one.
Lock-in gets more subtle. Custom silicon comes with its own software stack. As more AI services run on proprietary accelerators, moving workloads between Clouds gets harder to guarantee—which opens a door for providers that build on open standards and give customers a credible exit.
GPU supply could ease at the margins. If hyperscalers shift more internal inference onto their own chips, some of the pressure on GPU supply may lift over time. That’s speculative for now, but it’s worth watching for anyone who has been stuck at the back of the allocation queue.
The bottom line
Maia 300 won’t move Microsoft’s numbers before 2027, and the 40% efficiency claim still has to survive contact with real-life workloads, but the signal is hard to miss. The largest Cloud platforms are building vertically, from silicon to service, and power efficiency has become their main battleground. Everyone else in the ecosystem needs a plan for that world.
Those are exactly the conversations happening at CloudFest Global, where hosters, hyperscalers, chipmakers and data center operators come together every year to work out what’s next for the Cloud. Secure your spot at CloudFest Global 2027 and be part of it. Registration is now open.
