Microsoft unveiled the third generation of its in-house AI chip on Monday at the OCP Global Summit in San Jose, and said it could deploy more than a million of them next year to run its own AI services. The Maia 300 is no longer a side research project — at that volume, it becomes core infrastructure, arriving as Microsoft says it expects to spend more than $175 billion in capital expenditures this calendar year, a meaningful share of which has been going to Nvidia.
That’s the pressure the chip is designed to relieve. Microsoft says Maia 300 systems deliver roughly 40% more useful computing per megawatt than what it calls “leading GPU-based systems,” though it didn’t name which systems it measured against, a gap worth noting before treating the figure as settled. The claim matters mainly for inference — the work of running an already-trained model to answer a question, summarize a meeting, or write code — rather than the heavier job of training new models from scratch, where Nvidia still holds the advantage.
The engineering path is a departure from Nvidia’s own. Where Nvidia builds its accelerators around one large chip ringed by high-bandwidth memory, Microsoft went with a “chiplet” design, bonding together smaller specialized components. The central computing processor uses 2-nanometer technology from Taiwan Semiconductor Manufacturing Co for key computing functions, while other chiplets use cheaper 3-nanometer tech for moving data around the system. At the rack level, that trade-off is visible: Maia 300 processors will go into racks holding 96 chips and drawing about 167 kilowatts of power, compared with 142 kilowatts for Nvidia’s GB300 racks, which hold 72 GPUs. More chips per rack, more power drawn — but, if Microsoft’s numbers hold, more work extracted from each watt.
There’s also a strategic choice buried in how Microsoft plans to use these chips. Unlike Amazon and Google, which rent custom silicon out to cloud customers, Microsoft will use Maia exclusively for its own AI services, including GitHub Copilot and the Copilot tools built into Microsoft 365. There’s no indication yet that outside Azure customers will get access to Maia capacity the way they can rent AWS’s Trainium chips or Google’s TPUs. For now, this is a cost play aimed inward, not a new product Microsoft is trying to sell.
Rani Borkar, president of Azure hardware systems and infrastructure, described the goal to Reuters in plain terms: “It’s basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system.” That framing — watts, dollars, memory — is really the whole pitch. Microsoft isn’t claiming Maia 300 beats Nvidia on raw performance. It’s claiming it costs less to run at the scale Microsoft now operates at, which is arguably the more important number for a company whose AI bill keeps climbing.
Microsoft is hardly alone in making this bet. The introduction of Maia 300 arrives amid a wave of custom silicon efforts from major cloud operators and AI developers, including OpenAI, aimed at reducing reliance on Nvidia. Google has been running TPUs for years, Amazon has Trainium, and OpenAI has reportedly been developing chips of its own. What’s notable about Microsoft’s version is less the technology than the timeline: mass production is expected to begin in 2027, with deployments described as “multi-gigawatt” in scale — language that, in this industry right now, basically means as big as the company can manage.
Nvidia isn’t at risk of losing its position anytime soon; it remains the default for training the largest models, and nothing in Monday’s announcement changes that. But the fact that Microsoft, Google, and Amazon are all now running their own silicon for inference — the layer of AI computing closest to actual users — says something about where each of these companies thinks the real cost pressure is going to sit over the next few years.