2 minute read
Microsoft Just Built Its Own Ride
Microsoft rolled out the Maia 300 today, its third generation homemade AI chip, and Reuters reports the company could deploy more than 1 million of them next year. One million. This is a breakup letter written in silicon. The chips will run Microsoft’s own AI services, think Copilot and friends, and the goal is pretty obvious. Stop writing giant checks to Nvidia.
The Numbers Sound Great (Says Microsoft)
Microsoft claims Maia 300 squeezes out about 40% more tokens per megawatt than “leading GPU-based systems.” Which systems? They didn’t say. Funny how that works. A rack holds 96 chips and pulls roughly 167 kilowatts, and Microsoft says a full rack pushes up to seven times the output of the Maia 200 it launched in January. For comparison, Reuters notes Nvidia’s GB300 racks hold 72 GPUs at about 142 kilowatts. So more chips, more power, and a report card graded by the kid who took the test.
Why This One Matters
Microsoft calls itself Azure’s “customer zero,” and its own AI products are some of the hungriest mouths in its data centers. Business Insider reported Microsoft was already feeding its newest GPUs to its own AI products first while paying Azure customers waited in line. When the biggest eater in the building starts cooking for itself, the grocery store notices. Nvidia is the grocery store in this story.
Everybody Wants Their Own Chip (Sort Of)
Microsoft is late to this party. Google has been building its TPUs for over a decade and now sells them directly. OpenAI showed off its Jalapeño chip with Broadcom in June, with real volume expected in 2027. Apple has been working on a Broadcom-backed AI server chip codenamed Baltra, but it reportedly slipped and Apple is still renting Nvidia GPUs in the meantime. And SpaceX? It talked about building its own GPUs in its IPO filing, then Musk turned around in August and declared the company exclusive to Nvidia. So much for the great escape.
Don’t Plan Nvidia’s Retirement Party Yet
Production deployments don’t start until 2027, and Maia runs Microsoft’s own stuff, not Azure customers renting compute. Back in August, The Information reported Microsoft was haggling with TSMC for about 300,000 of these chips, with the million-unit dream depending on component supplies. Translation: Microsoft is still one of Nvidia’s best customers for now, and probably for a while.
What It Means for the Rest of Us
If you’re an Azure customer, here’s the quiet part. This chip is built to make Microsoft’s AI cheaper to run, not yours. Maybe those savings trickle down to you. Maybe Copilot just gets fatter margins. The AI arms race is officially a chip arms race now, and every big player wants to own the whole stack, from the silicon all the way to your subscription bill. Keep an eye on whether Microsoft ever rents Maia to outsiders. That’s when Nvidia should really start sweating.
Sources: Reuters, Microsoft, The Information, Business Insider, TechCrunch, HotHardware