JPMorgan Just Chose to Run AI On Its Own Hardware. SambaNova's $11B Valuation Is the Receipt.
The dominant AI storyline this year is that intelligence is getting radically cheaper. Blended inference costs are down 67% year over year. JPMorgan Chase just made a decision that runs the opposite direction: it's buying its own chips.
SambaNova completed the first close of a $1 billion Series F round on July 8, valuing the AI chip maker at $11 billion post-money. General Atlantic led the round, joined by T. Rowe Price, Capital Group, funds managed by BlackRock, the Qatar Investment Authority, Vista Equity Partners, and Intel Capital, among others. CEO Rodrigo Liang says a second close is coming in the weeks ahead. The round lands five months after a $350 million Series E and the unveiling of SambaNova's SN50 chip.
The detail that matters sits inside the financing announcement, not the headline number: JPMorgan Chase will deploy SambaNova's SN40 and SN50 systems to run its own AI workloads on its own premises, rather than through a third-party cloud provider. A systemically important bank is choosing to own its inference infrastructure at the exact moment renting it has never been cheaper.
The cost collapse is real and well documented. An analysis of 2.4 billion enterprise API calls found blended inference costs falling from $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026. Gartner forecasts a 90% reduction in frontier inference costs by 2030. If price were the only variable in this decision, renting should win every time.
For a regulated bank, it isn't the only variable. The marginal cost of a token is not the binding constraint — data residency, model auditability, and the ability to point a regulator at hardware the bank actually owns are. Choosing on-premise inference is a sovereignty decision dressed up as a procurement decision, and it explains why chip and infrastructure companies keep raising at expanding valuations even as the API layer above them commoditizes.
This mirrors a pattern showing up elsewhere in finance: as agentic AI moves into higher-stakes, regulated decision-making — underwriting, claims, compliance — the institutions best positioned to pay a premium for controlled infrastructure are precisely the ones that can least afford a third party holding their model weights or their customers' data.
If the pattern holds, the durable margins in AI infrastructure may not sit with the chips serving the cheapest tokens, but with the ones serving institutions willing to pay extra never to have to ask a cloud provider's permission.
| Metric | Value |
|---|---|
| SambaNova Series F first close | $1.0B at $11B valuation |
| SambaNova prior round (Feb 2026, Series E) | $350M |
| Blended inference cost, Q1 2025 | $18.40 / million tokens |
| Blended inference cost, Q1 2026 | $6.07 / million tokens (-67% YoY) |
| Gartner-forecast inference cost reduction by 2030 | 90% |
Frequently asked questions
How much did SambaNova raise and at what valuation?
SambaNova completed the first close of a $1 billion Series F financing on July 8, 2026, valuing the AI chip maker at $11 billion, led by General Atlantic with participation from T. Rowe Price, Capital Group, BlackRock-managed funds, and the Qatar Investment Authority.
Why is JPMorgan running AI on its own hardware instead of the cloud?
JPMorgan Chase is deploying SambaNova's SN40 and SN50 systems on its own premises to keep AI workloads inside infrastructure it controls directly, rather than relying on a third-party cloud provider — a priority for regulated institutions managing data residency and model auditability.
Are AI inference costs really falling?
Yes — blended enterprise AI inference costs dropped about 67% year over year, from $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026, and Gartner forecasts a 90% reduction in frontier inference costs by 2030.