Oxmiq raises $35m Series A to license CUDA‑compatible AI GPU architecture

Oxmiq is going straight at one of the biggest choke points in the AI stack: Nvidia’s effective monopoly on CUDA‑compatible compute. By offering OxCore as licensable GPU IP that can run CUDA workloads on non‑Nvidia hardware, the company is trying to decouple AI software ecosystems from a single vendor’s chips. The $35 million Series A, with backers like Samsung Catalyst Fund, MediaTek and Intel Capital, is a strong vote of confidence that chip and system makers are hungry for an ARM‑style model in AI accelerators.

If Oxmiq succeeds, it could meaningfully broaden who can field custom AI silicon. Instead of every hyperscaler and startup spending hundreds of millions on from‑scratch chip programs, they could license a proven GPU core and focus on differentiation higher up the stack. That’s good for resilience and pricing power across the industry; it also potentially accelerates the total amount of compute available for frontier training runs.

From an AGI‑race lens, more CUDA‑compatible architectures mean more actors can train large models at scale. That amplifies innovation but also raises proliferation and safety questions. The more drop‑in alternatives to Nvidia exist, the easier it becomes for both allied and adversarial players to spin up serious AI infrastructure without being as visible in Nvidia’s order books.