Microsoft's new AI chip cuts reliance on Nvidia's GPUs
chips inference microsoft nvidia
| Source: The Verge | Original article
Microsoft unveiled its third‑generation Maia AI chip, set as the primary inference engine for its services, aiming to cut costs and boost efficiency while reducing reliance on Nvidia GPUs.
Microsoft unveiled its third‑generation in‑house AI processor, the Maia 300, positioning it as the primary inference engine for the company’s cloud‑based products and services. The chip, announced today, is billed as more cost‑efficient and better suited to the latest reasoning and agentic AI models than the graphics processing units (GPUs) Microsoft has relied on from Nvidia. Production‑grade workloads are slated to begin in 2027, with the company targeting a deployment of potentially more than a million Maia chips by that time.
The move marks a strategic shift for the software giant, which has spent years building its AI stack on Nvidia hardware. By moving inference workloads onto its own silicon, Microsoft hopes to curb operating expenses and lessen its exposure to supply‑chain constraints tied to a single external supplier. Analysts note that the scale of the planned rollout could translate into multi‑billion‑dollar savings and give Microsoft greater control over performance optimisation for its AI services.
The announcement also signals a broader industry trend toward custom accelerators as firms seek to differentiate their AI offerings and manage costs. For Microsoft, the Maia line complements its ongoing investments in AI safety and governance, echoing recent internal calls for tighter oversight of powerful models.
What to watch next: how quickly Microsoft can transition existing workloads from Nvidia GPUs to Maia chips, and whether the performance claims hold up in real‑world cloud deployments. Observers will also track any reaction from Nvidia, which could adjust its pricing or partnership strategy in response to a major customer building its own inference hardware. Further details on the chip’s architecture and pricing are expected as the 2027 rollout approaches.
Sources
Back to AIPULSEN