Microsoft has unveiled the third generation of its custom artificial intelligence processor, the Maia 300, and signalled that it intends to run most of its AI services on silicon it designs itself rather than on graphics processors bought from Nvidia. The announcement, made on Monday, is the company's most aggressive move yet to loosen its dependence on the dominant chipmaker of the AI era.
Production-scale deployment of the Maia 300 is scheduled to begin in 2027, giving Microsoft roughly a year to ramp up manufacturing before the chips start carrying real customer workloads. The accelerator is being built to become "the primary AI inference system for Microsoft products and services", according to Scott Guthrie, the executive vice president who runs Microsoft's cloud and AI group.
Engineered for the inference era
Inference — the process of actually running a trained AI model to answer questions, write code, or carry out tasks — is where the economics of artificial intelligence are increasingly decided. Training large models makes headlines, but the day-to-day cost of serving millions of user requests is what determines whether AI products make money. That is the workload the Maia 300 is designed around.
Microsoft says the chip will deliver 40 per cent more tokens per megawatt than leading GPU-based systems, a claim aimed squarely at the metric that matters most to cloud operators: how much useful AI output you get for each unit of electricity consumed. A full rack of 96 Maia 300 accelerators is expected to deliver up to seven times the token throughput of the previous generation, the Maia 200.
The design targets reflect where AI usage is heading. The Maia 300 is aimed at newer workloads such as long-context reasoning, low-latency chat, and multi-step agentic tasks — systems that work through complex jobs in stages rather than answering a single prompt. Those are precisely the applications Microsoft is pushing into its flagship products, from Microsoft 365 Copilot to GitHub Copilot.
A different technical path from Nvidia
In designing the chip, Microsoft deliberately diverged from Nvidia's blueprint. Nvidia's AI accelerators are built around one enormous compute die surrounded by high-bandwidth memory chips. The Maia 300 instead uses a chiplet architecture: its central compute engine is made up of multiple specialised dies bonded together, rather than a single monolithic slab of silicon.
The most demanding compute chiplets are manufactured on TSMC's 2-nanometre process technology, while supporting chiplets that handle functions such as moving data off the processor and routing it within the system use the less expensive 3-nanometre node. By splitting compute, memory control, and input/output across separate pieces of silicon, Microsoft can devote more die area to raw compute where it counts.
Microsoft designs the processors itself and relies on Taiwan Semiconductor Manufacturing Co. to build them, rather than buying finished accelerators from Nvidia. Reuters reported that the company is negotiating with TSMC to secure capacity for more than 300,000 chips, with deliveries aimed at 2027 — and that Microsoft plans to deploy potentially more than a million of the chips next year to support its internal AI services.
At rack scale, the Maia 300 systems will pack 96 chips per rack drawing about 167 kilowatts of power. For comparison, Nvidia's latest GB300 AI server racks, which carry 72 graphics processing units, draw roughly 142 kilowatts.
Efficiency as a competitive weapon
Power has become one of the hardest constraints on AI expansion. Data centre operators around the world are scrambling to secure enough electricity to run their AI infrastructure, and every percentage point of efficiency translates directly into lower operating costs and smaller power requirements. Microsoft's framing of the Maia 300 is built around that reality: it is about "delivering more intelligence for every watt, every dollar, and every byte of memory", as Rani Borkar, president of Azure hardware systems and infrastructure, put it.
The company has not named the GPU-based systems it used as the basis for its 40-per-cent comparison, so independent verification will have to wait for production benchmarks. Even so, the direction of travel is clear. Azure customers could eventually see cheaper AI services as inference shifts onto Microsoft's own hardware, and tighter integration between the company's models and its silicon could accelerate feature rollouts for products like Copilot and Microsoft Foundry.
What it means for Nvidia
Microsoft is one of Nvidia's largest customers, and any shift in its purchasing behaviour reverberates through the chip industry. For years, the Redmond company — like every other hyperscaler — has had little choice but to queue up for Nvidia's latest accelerators, paying premium prices and waiting in long allocation lines. A credible in-house alternative weakens that dynamic, giving Microsoft negotiating leverage even if it never stops buying Nvidia hardware entirely.
That said, the transition will be gradual. The Maia 300 is an inference chip, not a training chip, and Microsoft expects Nvidia and AMD hardware to continue supporting model-training workloads for the foreseeable future. The new accelerators will also initially be dedicated to the company's internal operations rather than rented out to Azure customers, so the near-term impact on Nvidia's data centre revenues is likely to be modest.
The longer-term signal is harder to dismiss. Nvidia's dominance was built not just on fast silicon but on the CUDA software ecosystem that locks developers in. Every hyperscaler that builds its own inference stack — Google with its TPUs, Amazon with Trainium and Inferentia, Meta with its MTIA chips, and now Microsoft with three generations of Maia — chips away at that moat from a different angle.
The wider industry shift
The Maia 300 arrives amid a wave of custom silicon programmes from cloud computing rivals and AI developers — including OpenAI itself — all aimed at loosening Nvidia's hold on AI computing. The pattern is familiar from earlier technology transitions: whoever controls the hardware controls the margins, and the biggest buyers of AI compute have decided they would rather design their own.
Chief executive Satya Nadella described the announcement as "our biggest bet yet on our own silicon" in a post on X. It is a bet that builds on rapid momentum: Microsoft only announced the Maia 200 in January, and that chip is already running production workloads across the company's AI products, supporting both Microsoft's in-house models and OpenAI models delivered through the Azure cloud platform.
If the Maia 300 delivers on its promises when it reaches production in 2027, it would mark a structural shift in the economics of cloud AI — one in which the world's second-largest cloud provider buys far fewer GPUs from Nvidia and keeps far more of the AI value chain under its own roof. For an industry built on the assumption that Nvidia's accelerators are the only game in town, that is a prospect worth taking seriously.
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by The Verge. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
