There comes a point in every big tech company's life when it looks at its Nvidia invoice, sighs deeply, and says, "We could just make our own." Microsoft hit that point years ago. On October 12 it showed off the third act: Maia 300, a custom AI chip built to run the models behind its products more cheaply than the GPUs it currently rents its future from.

Built for Talking, Not Just Thinking

Maia 300 is an inference-first chip, meaning it's tuned for answering requests rather than training new models. It's built from chiplets, with compute components on TSMC's N2P process and eight stacks of HBM4 memory per chip. A rack holds 96 of them and draws roughly 167 kilowatts.

Microsoft's claims are bold: up to seven times the token throughput of a Maia 200 rack, and 40% more tokens per megawatt than leading GPU-based systems. The company says deployments start in 2027 at gigawatt scale, projecting more than one million chips deployed that year, with Maia becoming the primary system for AI inference across its products. Maia 200 already supports Copilot and OpenAI models today, and more details are promised at Microsoft Ignite.

The Real Product Is Tokens Per Watt

Notice what Microsoft is optimizing for: long conversations, reasoning, and multi-step agent tasks. Those are the workloads that chew through memory and electricity, and they're exactly where AI's cost curve hurts most. The company's pitch, in its own words, is "more intelligence for every watt, every dollar, and every byte of memory." In 2026, power is the scarcest ingredient in AI, and whoever squeezes the most tokens out of a megawatt wins the margin war.

The angle people miss: this doesn't mean Nvidia is getting kicked out. Training still leans heavily on GPUs, and Google, Amazon, and OpenAI are all chasing their own custom chips for the same reason Microsoft is. What's changing is that inference, the part every Copilot query and chatbot reply depends on, is becoming a vertically integrated business. That's good news for AI pricing over time, and a quiet signal that the per-request cost of AI features is a moving target, not a fixed tax.

Translation for normal humans: the AI features you're paying for today should get cheaper to run tomorrow, and the hyperscalers are spending billions to make sure that margin ends up in their pocket first.

Weighing whether AI features in your site or business tools are worth the running costs? Our team helps businesses scope AI integrations that make financial sense now and keep getting smarter as the price per token drops.

Source: The Next Web