Mistral AI eyes in-house chips to lower token costs soon

Mistral AI is evaluating in-house chip designs to cut token costs while it continues using Nvidia gear, CEO Arthur Mensch told CNBC in his first remark.

The disclosure marks Mensch’s first acknowledgment of semiconductor ambitions at the French startup. It highlights a push to control more infrastructure as the company competes with U.S. rivals OpenAI and Anthropic.

Chip talk tied to economics, not hype

Mensch said custom silicon can materially reduce the cost of deploying tokens, the units of data processed by AI models. Lowering per-token outlays is core to serving enterprise workloads at scale.

He indicated ownership of chips could come later, but that the company is not committing to a timeline. The firm is testing options while keeping near-term operations steady.

Nvidia remains the present-tense supplier

For now, Mistral relies on Nvidia hardware and described the relationship as strong. The company is building data centers with Nvidia chips as it trains and serves its models.

The approach balances ambition with capacity already in market. Any in-house effort would run alongside, not replace, existing deployments in the current phase.

Europe’s contender with enterprise focus

Paris-headquartered and valued at nearly 12 billion euros, Mistral develops AI models and sells into corporate accounts. The firm is often framed as Europe’s answer to OpenAI and Anthropic.

Mistral’s enterprise emphasis is reflected in its customer roster, which includes chip equipment leader ASML. The buyer mix points to practical use cases and service-level needs over consumer experiments.

The comments to CNBC clarify where the company is drawing the line today: efficiency gains through design exploration, with procurement anchored to Nvidia in the short run. Owning more of the stack is on the table, but only when it supports cost and reliability targets.

Next steps hinge on whether internal chip work can prove meaningful savings across deployments without sacrificing time to market. Until then, expect Mistral to press its enterprise strategy with existing infrastructure while iterating on potential custom silicon.