Microsoft AI executive says new Maia 300 chip is designed to reduce reliance on Nvidia, arguing “competition is good”

Microsoft’s effort to build more of its own AI compute infrastructure is getting a fresh spotlight, with cloud and AI executive Scott Guthrie arguing that reducing reliance on any single chip supplier is both feasible and beneficial for the industry. In a recent interview, Guthrie discussed Microsoft’s Maia 300 AI chip and how it fits into Microsoft’s broader approach to powering AI workloads in the cloud, including weather-related uses that require large-scale inference.

Guthrie said Maia 300 is engineered to improve efficiency, describing it as capable of delivering about 40% more tokens per unit of power, expressed as “per megawatt.” In AI systems, “tokens” are the basic units of text that models generate or process, and tokens per megawatt is a way to measure how much model output a data center can produce relative to its energy use. The metric indicates that Microsoft is framing the chip not just as raw performance, but as an attempt to manage the costs and power constraints that come with large AI deployments.

The interview also centered on supply-chain and competitive dynamics in AI hardware. Guthrie said Maia 300 is part of Microsoft’s strategy to cut “Nvidia dependence,” making clear that the company is willing to develop and deploy alternatives to mainstream GPU offerings. He also said, in direct terms, that competition is good, an indication that Microsoft views multi-vendor AI compute as more than a procurement strategy, and instead as a market structure it wants to encourage.

A key implication of the discussion is that Microsoft is treating custom silicon as a tool to influence both cost and throughput across its cloud AI stack. If a chip can increase tokens per megawatt, it can potentially reduce the amount of electricity and supporting infrastructure required for a given level of AI output. While Microsoft did not provide the interview’s full comparative methodology or specify what exact “baseline” it was compared against in the portion available, Guthrie’s framing was clearly about improving efficiency at scale, which is a recurring challenge for AI operators.

Industry context matters here. Microsoft is one of the biggest buyers and operators of cloud AI capacity, and it sits at the intersection of three pressures: rising demand for AI compute, energy and infrastructure constraints in data centers, and the concentration of advanced AI accelerators among a small number of suppliers. In that setting, custom chips like Maia are a way to spread risk and potentially negotiate better long-term control over performance and economics.

At the same time, the interview did not lay out specific customer adoption details for Maia 300, nor did it provide a timeline for how quickly Microsoft plans to expand deployments or how workloads are allocated between Maia and other hardware. It also did not describe whether Maia 300 is limited to particular model types or inference pathways, or how it will integrate with Microsoft’s full software stack in production environments. Those gaps leave unanswered questions about the deployment footprint and the pace of hardware substitution.

What to watch next is whether Microsoft follows up with more concrete disclosures, such as performance comparisons that describe what tasks and models were used, whether Maia 300 is being rolled out broadly across Azure or targeted to specific workload classes, and how Microsoft’s stated efficiency goals translate into measurable capacity planning outcomes. Any additional detail on how Maia 300 affects end-to-end costs, scheduling, and throughput would help clarify whether the “tokens per megawatt” advantage will materially change the economics of running AI applications at Microsoft.