With Maia 300, Microsoft Wants to Break Free From Nvidia
Trending Topics Jakob Steinschaden ● Covered by 2 sources
Microsoft’s new Maia 300 chip is built to run AI answers, not train models. It’s meant to cut Nvidia out of Microsoft’s own AI systems starting in 2027.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft is trying to stop paying Nvidia for every AI response it serves. The company says its Maia chips will become the main system for AI inference across its own products, with multi-gigawatt deployments starting in 2027. Maia 300 is the latest step in that plan, and Microsoft is pitching it as the chip that will finally let the company keep more of the AI money at home.
Maia 300 is built for inference from the start, not for training models. That matters because inference is the part that gets hit every time a user asks Copilot something, opens a conversational app, or kicks off an AI agent that has to chain together several steps. Microsoft says the chip is aimed at long-context reasoning and fast responses, which is basically the boring, expensive part of AI that runs all day.
The hardware numbers are doing a lot of the selling here. Microsoft says the compute chiplets are made on TSMC’s 2-nanometer N2P process, and each chip includes eight HBM4 stacks with 288 gigabytes of memory and 18 terabytes per second of bandwidth. A rack holds 96 accelerators, draws 167 kilowatts, and packs 27.6 terabytes of HBM memory. Microsoft says that setup should deliver up to seven times as many tokens as a Maia 200 rack, plus 40 percent more tokens per megawatt than “leading GPU-based systems.”
That last phrase is doing some careful evasive work. Microsoft never names Nvidia in that claim, even though Nvidia is the company everyone is really talking about here. The reason is obvious enough: Nvidia’s data center business recently brought in $89 billion in a single quarter, and Microsoft, like every other hyperscaler, has been helping fund that feast. If Microsoft can shift inference onto its own silicon, it keeps more margin for itself and frees up its Nvidia systems for paying Azure customers.
The catch is that this has been a slow program. Maia 100 showed up only in small numbers, and Maia 200 arrived late. It went live in January in Iowa, then expanded to Europe and Australia, and it already runs Copilot, supports Microsoft’s MAI team, and is lined up for GitHub Copilot and Foundry by the end of the year. Microsoft says Maia 300 will not go into operation until 2027, while Nvidia is already shipping Vera Rubin. The company says more details are coming at Ignite, which feels about right: the hardware race is always loudest right before the numbers get real.
My take — AI-written commentary, not fact-checked reporting
This is the only sensible move for a hyperscaler that’s serious about AI economics. Paying Nvidia for every inference request is a fine hobby for a while, not a business model. The catch, as always, is that custom chips are easy to announce and much harder to ship on schedule without turning the data center into a very expensive art installation.
Read more about this at: Trending Topics