SambaNova, the Palo Alto AI-chip company, completed the first close of a $1 billion Series F on July 8, lifting its post-money valuation to $11 billion — roughly five months after its last mega-round. General Atlantic led, and the participant list reads like an institutional roll call: BlackRock, Intel Capital, T. Rowe Price, the Qatar Investment Authority and Vista Equity Partners among them. But the number that matters isn’t the valuation. It’s the customer the company put forward the same week.
JPMorgan Chase has selected SambaNova as its AI inference-infrastructure partner in a multi-year deal. The bank will deploy SambaNova’s Reconfigurable Dataflow Unit chips — the SN40 and SN50 lines — to run models on its own premises, inside its own walls, rather than shipping prompts and customer data to someone else’s servers. JPMorgan joins a customer list that already includes SoftBank, Saudi Aramco and Intel. For a chip startup that markets itself as a Nvidia challenger, a systemically important bank is the reference customer money can’t quite buy.
Read the two events together and the strategy comes into focus. The AI-infrastructure trade has been sold as cloud: rent GPUs from a hyperscaler, pay by the token, never touch a rack. SambaNova is selling the opposite — inference you own, in hardware you control. For most startups that’s a niche. For a bank bound by data-residency rules, model-risk governance and regulators who ask precisely where the computation runs, “on our own machines” isn’t a cost optimization; it’s a compliance answer. That’s the wedge — and it’s why, as with Anthropic’s own chip ambitions, the customer list, not the cap table, is the real story.
Our take: The valuation is the headline; the location is the news. Nearly every dollar of the AI build-out has assumed the workload lives in a hyperscaler’s cloud. JPMorgan just voted with its data center. SambaNova’s pitch — run inference on hardware you own, keep the data behind your own firewall — turns out to be exactly what the most regulated, most target-rich enterprises want to hear. It won’t dislodge the cloud for the internet at large. But it carves out a durable second market — banks, defense, healthcare, sovereign AI — where “where does the compute run” is a board-level question. A $1 billion round says investors think that market is real. A JPMorgan deployment says at least one very demanding customer already does.
What to watch
- Scale, not pilot. A test cluster and a fleet are different stories. Watch whether JPMorgan says how many workloads move on-prem — and whether it displaces existing cloud inference or merely adds to it.
- The second close. CEO Rodrigo Liang has signaled more investors are joining and a second tranche is coming. The final round size — and valuation — will show whether $11 billion was a floor or a ceiling.
- Nvidia’s counter. On-prem inference is the one flank where a challenger can win on data control rather than raw speed. If more banks follow, expect Nvidia to sharpen its own enterprise on-prem story — the same instinct pushing rivals to build inference silicon of their own.
- Copycat customers. Aramco and SoftBank were early adopters. The tell is whether a second U.S. money-center bank or a large insurer signs in the next two quarters.
The market spent 2026 arguing over which model is smartest. The quieter fight is over where the models actually run — and who owns the machines underneath them. Hyperscalers are racing to design their own silicon; governments are rationing imported chips; and now a Wall Street bank is planting inference in its own basement. SambaNova just raised a billion dollars betting that “in-house” is a category, not an exception.