Microsoft Unveils Maia 300 Chip to Cut Nvidia Dependence

Microsoft has unveiled its third in-house AI chip, Maia 300, built to run AI workloads inside its Azure cloud. Scott Guthrie, the executive vice president who runs Microsoft's cloud business, says the chip should deliver 40 percent more tokens (the text units a model processes) per megawatt than leading GPU-based systems. In a rack of 96 accelerators, throughput should be up to seven times higher than with the previous chip, Maia 200. The chip itself will not start handling real AI workloads in Microsoft's cloud services until 2027, not sooner.

This changes nothing about how smart AI models actually are. Maia 300 does not make a model smarter. It just runs already-trained models cheaper and faster inside Microsoft's own cloud. This is an infrastructure race, not a quality race. Here Microsoft is playing catch-up, building its own chip mainly to depend less on Nvidia. For now the numbers come from a press release, not a shipped product. Real use in production services will not begin until 2027.