> ## Content Index
> Fetch the complete content index at: https://carussignal.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# SambaNova's $11 Billion Bet: Why JPMorgan Chose On-Prem AI Chips Over the Cloud
- URL: https://carussignal.com/sambanova-11-billion-jpmorgan-on-prem-ai-inference/
- Published: 2026-07-12T20:00:09.000Z
- Updated: 2026-07-12T20:00:09.000Z
- Author: Carus Cha
- Tags: SambaNova, AI inference, Nvidia, JPMorgan, on-premise AI, data sovereignty, AI chips

# SambaNova's $11 Billion Bet: Why JPMorgan Chose On-Prem AI Chips Over the Cloud

> SambaNova just raised $1 billion at an $11 billion valuation and landed JPMorgan as an on-premises inference customer. The deal is a bet against the cloud-only AI thesis — here is what the numbers and the strategy actually say.

On July 8, 2026, at the RAISE summit in Paris, AI-chip startup SambaNova announced two things at once: the first close of a **$1 billion Series F** round that pushed its valuation to **$11 billion**, and a multi-year deal to power **on-premises AI inference for JPMorgan Chase**. Both signals point the same direction — money and a systemically important bank are betting that a meaningful slice of AI will run *inside* enterprises, not only in someone else's cloud. This post unpacks the funding, the chip claims, and why the JPMorgan choice matters more than the dollar figure.

## The raise: who's backing the Nvidia challengers

The Series F was led by **General Atlantic**, with participation from **T. Rowe Price, Capital Group, and Seligman Ventures**. That is a notable investor mix: T. Rowe Price and Capital Group are large public-market managers, the kind that typically show up late in a company's life. Their presence signals that the "Nvidia challenger" thesis has moved from venture speculation toward mainstream institutional conviction.

SambaNova's focus is narrow on purpose. It does not try to out-train Nvidia on giant training clusters; it builds **inference** chips — the semiconductors that *run* trained models quickly and cost-efficiently — sold as full server units for data centers and, increasingly, on-premise deployments.

![Graphic showing SambaNova's $1 billion Series F round and $11 billion valuation led by General Atlantic](https://carussignal.com/content/images/2026/07/sambanova-11-billion-jpmorgan-on-prem-ai-inference-1.webp)

\## The chip claim: SN50 vs. Nvidia's B200

SambaNova's pitch rests on its **SN50** accelerator. The company's benchmarks — and it's important that these are *vendor-reported*, not independent — claim the SN50 delivers roughly **3x the throughput** of Nvidia's B200 under real latency constraints. Here is the comparison as SambaNova presents it:

| Metric (SambaNova-reported)                                                                         | SambaNova SN50   | Nvidia B200      |
| --------------------------------------------------------------------------------------------------- | ---------------- | ---------------- |
| Llama 3.3 70B, FP8, 1K/1K tokens (per-user throughput)                                              | \~895 tokens/sec | \~184 tokens/sec |
| Average throughput advantage (latency-constrained, across Llama 70B / GPT-OSS 120B / DeepSeek 671B) | \~3x higher      | baseline         |
| Raw FP8 compute (dense / sparse TFLOPS)                                                             | 3,200            | 4,500 / 9,000    |

Notice the twist in the last row: **the B200 actually wins on raw FLOPS.** SambaNova's edge does not come from more compute — it comes from architecture. Its reconfigurable dataflow design and three-tier memory hierarchy keep data on-chip and cut the memory round-trips and scheduling overhead that eat GPU cycles. The claim, in plain terms: the SN50 wastes fewer cycles moving data around, so it converts less raw horsepower into more usable tokens per second.

The honest reading: these are impressive numbers, but they are SambaNova's own, on workloads it selected. Treat the 3x as a vendor claim worth watching, not a settled fact.

![Diagram contrasting GPU memory round-trips with SambaNova dataflow architecture keeping data on-chip](https://carussignal.com/content/images/2026/07/sambanova-11-billion-jpmorgan-on-prem-ai-inference-2.webp)

\## Why JPMorgan's choice is the real headline

The dollar figures grab attention, but the strategically interesting part is *who* signed up and *why*. JPMorgan selected SambaNova's SN40L and SN50 systems to run inference **on-premises** — inside the bank's own infrastructure — rather than shipping its most sensitive data to a third-party cloud.

That is a direct challenge to the cloud-only AI narrative. The reasoning:

- **Data sovereignty.** Regulated institutions increasingly cannot send their most sensitive data to external servers. Tightening privacy and security rules push banks to keep data where they control it.
- **Predictable cost at scale.** For steady, high-volume inference, owning the hardware can beat metered cloud pricing over a multi-year horizon.
- **Control and latency.** On-prem means the bank controls the stack end to end and avoids network hops to a cloud region.

The takeaway is not "the cloud loses." It's that the market is splitting. Bursty, experimental, or variable workloads still favor the cloud's elasticity. But **steady, sensitive, high-volume inference in regulated industries** is exactly the segment the cloud-first story underweights — and it's the segment SambaNova is aiming at. A bank the size of JPMorgan validating that thesis is worth more than the $1 billion headline.

## Frequently Asked Questions

**How much did SambaNova raise and at what valuation?** $1 billion in a Series F first close, at an $11 billion post-money valuation, led by General Atlantic.

**Is the SN50 really faster than Nvidia's B200?** On SambaNova's own benchmarks, yes — about 3x higher throughput under latency constraints, and \~895 vs. \~184 tokens/sec per user on one Llama 3.3 70B test. But the B200 leads on raw FLOPS, and these figures are vendor-reported, not independently verified.

**What did JPMorgan actually agree to?** A multi-year deal to run secure, on-premises AI inference using SambaNova's SN40L and SN50 systems, keeping sensitive data inside the bank's infrastructure.

**Does this mean enterprises are abandoning the cloud for AI?** No. It means the market is segmenting. Elastic and experimental workloads still favor the cloud; steady, sensitive, high-volume inference in regulated sectors is where on-prem makes sense.

**Why inference and not training?** Training the largest models is a capital-and-scale fight where Nvidia dominates. Inference is a different battle — cost-efficiency and deployment flexibility matter more, giving specialized challengers a real opening.

## Key Takeaways

- SambaNova raised **$1B at an $11B valuation** (Series F, led by General Atlantic), with late-stage public-market investors joining.
- Its **SN50** claims \~**3x** the throughput of Nvidia's B200 under latency limits despite *lower* raw FLOPS — an architecture story, and a *vendor-reported* one.
- **JPMorgan** picking SN40L/SN50 for **on-premises** inference is the bigger signal: a challenge to the cloud-only thesis, driven by data sovereignty and cost.
- The real story is market segmentation, not cloud vs. on-prem as winner-take-all.

**How this was written** AI assisted with gathering sources and structuring a first draft — fact-checking and final edits were done by a person.

---

## References

- CNBC, "SambaNova hits $11 billion valuation as investors back Nvidia chip challengers": https://www.cnbc.com/2026/07/08/sambanova-ai-chip-funding-valuation.html
- TechCrunch, "AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round": https://techcrunch.com/2026/07/08/sambanova-draws-1b-at-11b-valuation-in-series-f-first-close/
- Tom's Hardware, "SambaNova claims SN50 chip is three times more efficient than Nvidia B200": https://www.tomshardware.com/tech-industry/artificial-intelligence/sambanova-introduces-new-ai-accelerator-partners-with-intel-to-deploy-xeon-cpus-for-inferencing-and-agentic-workloads-sambanova-claims-sn50-chip-is-three-times-more-efficient-than-nvidia-b200
- InvestorPlace, "JPMorgan Just Challenged the Cloud-Only AI Thesis": https://investorplace.com/hypergrowthinvesting/2026/07/jpmorgan-just-challenged-the-cloud-only-ai-thesis/
- Finextra, "JPMorgan Chase picks SambaNova for on-prem AI inference": https://www.finextra.com/newsarticle/48057/jpmorgan-chase-picks-sambanova-for-on-prem-ai-inference