Can Rivos Win Where Nvidia Is “Overkill”?

Rivos is asking for a lot of money before it has sold a single chip. The Santa Clara startup is reported to be seeking up to $400–500 million in new funding: a raise that would push its lifetime capital close to $900 million and value the business at north of $2 billion. That simple fact explains why this is a story worth watching: investors are being asked to back a product that exists today mostly on paper and in prototype form.

There’s logic to the ask. Rivos designs inference hardware, the kind of compute used to run trained AI models in production, and the company argues there’s room for alternatives to Nvidia’s expensive top-tier cards. CEO Puneet Kumar has said Rivos will aim for “smaller installations where Nvidia might seem like an overkill from a cost perspective.” As model deployment spreads, some customers will prioritize cost, power efficiency and simpler deployments over absolute peak performance.

An opening in Nvidia’s shadow

Nvidia’s dominance rests not only on silicon but on software and habit. CUDA, the company’s developer ecosystem, is deeply embedded in training and deployment pipelines; that’s the practical barrier any rival has to clear. Rivos is trying to lower that barrier. It’s building compatibility tools that aim to let existing CUDA-based models run with less rework, and it’s pairing that software pitch with a hardware strategy based on the open RISC-V architecture.

Those two choices, a migration-friendly software story and a RISC-V cost/licensing angle, are sensible if the goal is to win customers who want “good enough” inference at a far better total cost of ownership. They also mirror moves by other challengers in the market. Companies such as Cerebras, Groq and Positron have all staked different claims: wafer-scale performance, low-latency inference, or rapid, air-cooled deployments. Each has real customers and proven deployments; Rivos does not yet. The difference matters.