SambaNova just raised over $1 billion at an $11 billion valuation — five months after its last round. That’s the headline everyone will write.
But here’s what actually matters: JPMorgan Chase is using SambaNova’s hardware in production. Not a pilot. Not a proof-of-concept. The largest bank in America, by assets, has deployed the company’s RDU-based systems for “enterprise AI workloads that demand performance, control, and reliability,” per its CIO.
That’s not a press release talking point. It’s a reference customer that changes the conversation.
Most AI chip startups sell a story about what they could do. SambaNova just sold a story about what it is doing — inside one of the most technologically conservative institutions on the planet.
Why JPMorgan Matters More Than the Valuation
Banks don’t gamble on infrastructure. They don’t deploy unproven hardware into production environments where a latency spike or a security gap costs millions. When JPMorgan picks a vendor, it’s not a bet — it’s a validation.
The bank has been public about its AI ambitions. It has over 2,000 data and AI professionals on staff. It’s running hundreds of use cases across fraud detection, trading optimization, customer service automation, and risk modeling. But until now, most of that horsepower came from the usual suspects: Nvidia GPUs in the cloud, with all the cost and data-movement tradeoffs that implies.
SambaNova offers something different: on-premise inference at cloud-like scale.
For JPMorgan, the calculation is straightforward. Financial data can’t leave the building — not fully, not without layers of compliance overhead that make cloud AI expensive in ways that don’t show up on a per-token price sheet. Running inference locally eliminates the data exfiltration risk. It cuts the egress fees. And if SambaNova’s claims about inference throughput hold up, it does all of that without sacrificing performance.
I think this is the real story: enterprises are starting to realize that the cloud isn't the only answer for AI.
The On-Premise Pivot That No One Saw Coming
For the last three years, the AI narrative has been cloud-first, cloud-always. Train in the cloud, deploy in the cloud, pay as you go. The hyperscalers built their AI businesses on this assumption, and for a while, it made sense — the cloud offered flexibility, scale, and access to the best chips.
But inference changes the math.
Training is a sprint — expensive, intense, but temporary. Inference is a marathon — it runs forever, and every millisecond and every penny compounds. When you're running a customer service bot that handles 10 million queries a day, or a fraud detection model that scores every transaction in real time, the cost structure flips. The cloud's convenience premium starts to look like a tax.
SambaNova is betting that enterprises will eventually do the same thing they did with databases and ERP systems twenty years ago: bring the most critical workloads back in-house. Not everything — but the crown jewels. The workloads that touch customer data, financial records, or proprietary IP.
JPMorgan is the proof point. And Saudi Aramco, SambaNova's other named enterprise customer, reinforces the pattern. These are not startups playing with AI on the margin. These are institutions that move slowly and bet big.
What This Means for the Nvidia Narrative
Let’s be clear: SambaNova isn't replacing Nvidia at JPMorgan. The bank runs thousands of GPUs and will keep buying more. But Nvidia no longer has a monopoly on the inference conversation.
That’s the shift.
The market has treated Nvidia as the only real option for AI compute for so long that anything else felt like a footnote. But inference is where the challengers have their best shot. Training requires the full CUDA stack, the ecosystem, the developer mindshare — all the things Nvidia has spent fifteen years building. Inference is more forgiving. It’s about throughput, latency, and total cost of ownership. Those are measurable, comparable, and beatable.
SambaNova claims its SN50 chip delivers 5-10x better inference performance per dollar than GPUs. I haven't verified that independently, and neither has anyone else outside the company. But JPMorgan's willingness to deploy the hardware suggests the gap is real enough to matter.
The takeaway isn’t that Nvidia is doomed — it’s that the inference market is now open for competition. And that competition will push prices down, performance up, and options expand. That’s good for everyone buying AI hardware, and bad for Nvidia's margins over time.
The Intel Angle No One is Talking About
There's another layer to this story that deserves attention.
Intel CEO Lip-Bu Tan sits on SambaNova's board. Intel Capital is an investor. The two companies have a co-development agreement to build enterprise inference solutions using Intel's Xeon processors alongside SambaNova's RDU accelerators.
That partnership has an obvious logic: Intel needs a credible AI accelerator story to offer customers who don't want to buy Nvidia. SambaNova needs Intel's enterprise relationships and distribution channels. It’s a classic "frenemy" arrangement — but it’s also a signal that Intel sees on-premise inference as its best path back into the AI conversation.
The acquisition rumors from late 2025 — that Intel reportedly offered $1.6 billion for SambaNova — now look like a missed opportunity. At $11 billion, SambaNova is out of Intel's comfortable acquisition range. The company is talking about a 2027 IPO instead.
SambaNova has essentially outgrown the one strategic buyer that made the most sense. That's either a sign of strength or a vulnerability — it depends on whether the company can keep winning customers like JPMorgan without Intel's safety net.
What I'm Watching Next
SambaNova now has the funding, the customers, and the narrative. But it also has a long runway of execution risk ahead.
Three things to track over the next 12–18 months:
Customer concentration. JPMorgan and Saudi Aramco are marquee names, but they're two customers. The next phase requires SambaNova to expand into the mid-market and show that its solution is repeatable, not bespoke.
The Nvidia counterattack. Nvidia isn't standing still. Its next-generation Blackwell platform includes inference-specific optimizations, and the company has the R&D budget to match whatever SambaNova builds. The question is whether Nvidia can compete on cost, not just performance.
The Intel relationship. If Intel eventually acquires a different AI chip startup or builds its own competitive accelerator, SambaNova loses its most important strategic partner. The IPO timeline suggests the company is preparing for independence, not a sale — but that independence comes with higher stakes.
For now, though, SambaNova has done something few AI chip startups have accomplished: it convinced a Fortune 100 bank to actually run production workloads on its hardware. That's worth more than the billion-dollar round, and it changes the conversation about what's possible in enterprise AI.
P.S. If you're an enterprise CIO reading this and wondering whether to call SambaNova, your answer is probably "yes" — but only if you're willing to spend the next six months wrestling with integration, training, and operational overhead. The hardware works. The rest is up to you.
