Microsoft Ramps Up Homegrown AI Chip Push

There's lots of readership and interest in this. The basics of the Maya 300 generation math, and they plan to deploy it at scale. So this generation inserts a system on a chip. It's designed for AI tasks. It's designed for AI inference, though we're not talking about training new models, it's about running existing ones efficiently. Uh, Amazon, actually, Microsoft thinks they can plug enough of these into their data center to make a meaningful difference on their costs, on performance. And ultimately, they say these things are going to be backing, uh, most of their AI workloads internally. My 300 promises 40% more tokens per megawatt than what Microsoft said was leading GPU systems, but because it uses, uh, HBM for I'm pretty sure they're testing against Nvidia's, um, they're benchmarking. It's Vera Rubin. Um, as the kind of compare to what what's the history and backstory of Microsoft's Chip program? Right. Like all of the hyperscalers have custom silicon programs. They different players put different emphasis on how much this is about diversifying compute away from Nvidia. Some of them put a focus on more on the economics or economics of throughput. And that's exactly where where Microsoft has landed. I know they got their start a little bit after their big hyperscale competitors and Google got start on their TPUs years earlier. Amazon, with its own imprint acquisition and its, uh, training and inferentia chip lines, uh, got going. So Microsoft was a bit of a late comer to this. Um, but I think what people were looking for in this third generation of Chip is whether, you know, it's at the scale where they're going to start maturing the thing and getting it out there in force. A more if they're going to go a different direction. And this signal multi gigawatt deployment, that's a that's a ton. They probably only got about 12GW in their in their data center today. That's a real indicator that they want to continue to continue pushing for it. Um not just because they think they can build a better better mousetrap than Nvidia when it comes to efficiency. It is interesting. You know, the stock's up more than a percentage point. But again near fresh record highs itself. Um I think we need to remind the audience Matt. The main point is that, um Microsoft is at hyperscale. It is a huge buyer of accelerators or GPUs. It is working on its own chips, trying to sort of explain that scale and what their commitment is to data centers for AI. So they're spending hundreds of billions of dollars like their big peers, uh, Amazon and Google and meta, um, building data centers and then most expensively fitting them out with Nvidia rigs, uh, to make the AI inside tech. Um, so one of the surprising things about Microsoft's announcement this morning is that they're not going to rent the computing power of their chips to third parties. I think there is enough internal demand for things like Copilot, for things like GitHub, um, that they can just use the chips to satisfy some of that work and then save those Nvidia precious chips for customers.