JPMorgan Chooses SambaNova for On-Prem Inference, Valuing Startup at $11 Billion
Startup closes first tranche of a $1 billion Series F led by General Atlantic, five months after a $350 million Series E round, and anchors sale with a client that rarely appears in vendor press releases.
SambaNova Systems announced on Wednesday the partial closure of a $1 billion Series F funding round, valuing the company at $11 billion post-money. The tranche, led by General Atlantic, included participation from Seligman Ventures, T. Rowe Price Associates, and Capital Group. This round comes five months after a $350 million Series E round, closed in February alongside the announcement of the SN50 chip and its partnership with Intel.
The most valuable asset of the announcement is not in the captable. Alongside the press release, SambaNova revealed that JPMorgan Chase has selected the company as its inference infrastructure partner, deploying the SN40 and SN50 systems to run artificial intelligence workloads within the bank's data centers. This is the type of reference that a Silicon Valley startup almost never manages to showcase publicly: one of the largest banks in the world opting for dedicated hardware rather than renting GPU capacity via hyperscalers.
Why JPMorgan's Decision Matters
A regulated global bank adopting a dedicated inference stack, with an auditable trail and data that does not leave its perimeter, is an expensive signal to emit. Technology leadership in other financial institutions often quickly mimics this pattern, as it addresses three pressures simultaneously: the rising cost of tokens in frontier models, the requirement for data residency from regulators in multiple jurisdictions, and the demand for low latency in market products that operate in milliseconds.
CEO Rodrigo Liang has indicated to analysts that SambaNova's strategy is not to compete with Nvidia in training but to dominate the phase that generates recurring revenue for corporate clients: production inference. The February round positioned Intel as an investor and industrial partner, which grants SambaNova access to an alternative foundry at a time when TSMC's wafer queue has become the industry's structural bottleneck.
What Changes for the AI Infrastructure Market
The $11 billion valuation places SambaNova at a level comparable to Groq and Cerebras, although it still lags behind the valuations held by Nvidia and Anthropic. The point of interest for corporate buyers is another: the dedicated inference silicon market finally has a reliable second and third supplier, which mitigates the risk of vendor lock-in that CIOs have been signaling in supplier surveys since last year.
JPMorgan Chase is one of the few institutions that publishes technology capex with enough granularity for analysts to model returns. Part of this budget is now shifting to proprietary silicon within the bank's data center, altering the nature of the account: infrastructure capex gains weight over public cloud opex. For a treasury used to amortizing servers, this model saves margin over a three to five-year horizon, provided that the inference volume justifies the purchase.
Insights for Banks and Consultancies Outside the U.S.
SambaNova’s move is not merely American. European banks like Deutsche Bank and Barclays are under pressure from DORA and the EU AI Act review to prove that high-risk client data does not circulate through suppliers outside the bloc. Dedicated silicon within their own data center is a direct technical response to this regulatory framework. In Japan, Mizuho and MUFG announced plans in the first half of the year to enhance internal computational capacity for AI workloads in compliance and fraud prevention areas.
In Brazil, Itaú and Bradesco have been expanding their own computational capacity with a focus on generative AI in customer service and credit. The SambaNova model, if replicable on a smaller scale, enters the conversation as an alternative to direct GPU purchases. For consultancies serving these accounts, including the Big 4 and Brazilian firms like Falconi and CI&T, the project designs change: part of the work shifts from integration with hyperscaler APIs to MLOps over proprietary hardware.
The open question is whether SambaNova can deliver the volume required by a sales pipeline for 20 or 30 global banks. The company still does not publish the number of wafers shipped per quarter. Liang hinted at considering an IPO in 2027 in the United States, a timeline that gives the Series F some breathing room to prove delivery before going public.