Microsoft plans to lean harder on its own AI chips and reduce its dependence on Nvidia's GPUs, CNBC reports — a shift that touches the most expensive line item in the company's datacenter buildout.

The report gives no capacity figures or dollar amounts, but the direction is clear: Microsoft wants a larger share of its AI training and inference workloads running on silicon it designs itself, rather than on Nvidia accelerators it must buy at premium prices in a supply-constrained market.

Why does Microsoft want its own silicon?

Nvidia's GPUs remain the de facto standard for large AI model training, and Microsoft has been one of the company's largest customers. That dependence carries two costs. The first is financial: Nvidia's datacenter GPUs carry high margins, and every accelerator Microsoft buys transfers margin to a supplier rather than keeping it in-house.

The second is strategic. Nvidia allocates scarce flagship parts across hyperscalers — Microsoft, Google, Amazon, Meta and a wave of AI startups all compete for the same finite supply. A customer with its own accelerator roadmap gains negotiating leverage on price and delivery schedules, and hedges against allocation shortfalls.

Microsoft's homegrown effort is not new. The company has fielded its Maia accelerator family for AI workloads and Cobalt CPUs for general-purpose compute as part of its in-house silicon program. CNBC's report signals that Redmond intends to deepen that commitment rather than treat custom silicon as a side project.

How does this fit the hyperscaler pattern?

Microsoft is following a route its closest rivals already took. Google designed the TPU family more than a decade ago and now runs major portions of its AI training and serving on it. Amazon built its Trainium and Inferentia lines through Annapurna Labs. For both, custom accelerators shifted a meaningful slice of AI compute away from merchant GPU purchases.

Microsoft's position differs in one respect: its Azure AI services and its OpenAI partnership have made it one of Nvidia's most visible customers. Any material shift in Microsoft's accelerator mix therefore lands directly on Nvidia's hyperscaler revenue base — a segment that now anchors the GPU maker's datacenter business.

What are the limits of the shift?

Designing a competitive AI accelerator is easier than displacing Nvidia at scale. Nvidia's advantage rests not only on hardware throughput but on CUDA, the software ecosystem most AI researchers and frameworks target. Custom chips only pay off where the operator controls the full software stack and can port workloads — which is true for hyperscalers' internal services, less so for everything their cloud customers run.

That suggests a mixed fleet for years to come: Microsoft's own chips take a growing share of first-party AI workloads, while Nvidia GPUs remain essential for the broad Azure portfolio. Google and Amazon still buy Nvidia parts alongside their in-house lines, and Microsoft's pattern will likely resemble theirs.

CNBC's report does not specify volumes, timelines or which chip generations the shift involves, so any near-term revenue impact on Nvidia remains an analyst estimate rather than a disclosed figure. The competitive dynamics, though, are already moving: with Maia in production use and the commitment deepening, Microsoft joins Google and Amazon in treating internal silicon as core AI infrastructure rather than an experiment. Expect Redmond to detail more of that roadmap at its Build and Ignite conferences.