SambaNova Systems closes $1B round at $11B valuation, targets enterprise on-prem AI inference

SambaNova Systems closes $1B round at $11B valuation

SambaNova Systems, founded in 2017 and focused on AI inference hardware and software for on-premises enterprise deployments, has closed a $1B first-close round at an $11B valuation. CEO and co-founder Rodrigo Liang frames the timing as deliberate: inference has become the dominant workload, and enterprises are now moving beyond cloud experimentation toward private, controlled AI deployments.

“We did a big funding announcement today with a billion dollar round that we did a first close on at an 11,000,000,000 valuation. What's happening now is people starting to realize, okay, there is a whole another class of workloads — whole another class of applications that you've got to go and do private. With JPMorgan, they selected us as the ones that can come in and go into an environment like JPMorgan, do full on prem secure private.”

The enterprise on-prem thesis

The core pitch is data sovereignty. Enterprises in regulated industries, financial services, pharma, healthcare, have workloads they will not send to frontier model APIs or hyperscale clouds. Liang argues those organizations need a secure, on-prem inference stack where they can run open-source models, fine-tune on proprietary data, and retain full control regardless of how the regulatory environment evolves.

JPMorgan is the anchor customer named publicly, selected SambaNova for secure on-prem AI inference. Last month, SambaNova also announced a partnership with Vista Equity, which has 92 portfolio companies building AI applications. Liang describes these relationships as co-development arrangements, not just hardware sales, with customers that are sophisticated enough to shape how enterprise AI infrastructure actually gets deployed.

Hardware positioning

SambaNova's rack runs at 10 kilowatts, air-cooled, versus the roughly 140 kilowatts per rack for traditional GPU clusters. That power profile means enterprises can deploy into existing data centers without new builds, new power contracts, or liquid cooling infrastructure. Liang says the company specializes in running very large models, trillion-parameter scale, faster than competitors, which is where it argues the power-per-token economics are most competitive.

Nine-year overnight success

SambaNova was doing recommender models and image classification work before ChatGPT reframed the market around language. Liang credits OpenAI with focusing the entire industry on a single use case, which let hardware, software, and model companies align around a common target. Voice and video are now following as additional inference workloads, but language gave the market the production traction to sequence everything else.

On the chip design cycle, Liang is direct about the structural tension: a three-to-four year design-to-deployment timeline through TSMC means roadmap bets are made years before customer demand becomes visible. The Vista and JPMorgan partnerships are partly an answer to that, providing demand signal, or in his word, offtakes, that justify the long hardware investment cycles.

The raise positions SambaNova at a moment when enterprise AI spend, historically large but slow to move, is finally pulling through to the infrastructure layer.