SambaNova: the quiet chipmaker chasing Nvidia.

01 / Right NowWho they are, in May 2026

It's a Tuesday at 2200 Geng Road. The badge readers hum. A rack of SN50s blinks behind glass. A salesperson is on the phone with a sovereign-wealth fund that wants its own private Llama. None of this is hypothetical anymore.

SambaNova Systems is the rarest creature in AI - a hardware company that ships, with paying customers, that isn't named Nvidia. From a 430-person headquarters in Palo Alto, the company builds the chips, the systems, the cloud, and the software stack to run very large AI models very quickly. The pitch is not subtle. It is: stop renting GPUs.

For most of the last decade, this was an eccentric thing to say. Then ChatGPT arrived, GPU demand outran supply, and every CFO with a model in production discovered the joy of a six-figure cloud invoice. Suddenly an alternative was less eccentric. Suddenly a company called SambaNova was getting calls back.

02 / The ProblemWhat they noticed before the rest of us

The founding observation, back in 2017, sounded academic. Modern AI was choking on hardware built for graphics. GPUs were repurposed, not designed for the job. They were fast, sure, but rigid - you fed them instructions, they fetched, they computed, they wrote back. Data was the tourist. Compute was the city.

Kunle Olukotun, who had spent years at Stanford helping invent the multicore CPU, thought you could flip that around. What if the compute was the tourist? What if data flowed through a chip the way water flows downhill - finding the cheapest, fastest path? Christopher Ré, also at Stanford, was thinking similar thoughts on the software side. Rodrigo Liang, a former Oracle hardware executive, brought the third ingredient: someone who knew how to ship product to enterprises that bought things in racks.

Their thesis was that the bottleneck in AI was not raw FLOPs. It was memory bandwidth, model size, and the absurd dance required to move parameters in and out of accelerator memory. Their answer was the Reconfigurable Dataflow Unit, or RDU - a chip whose layout reconfigures itself to fit the model, not the other way around.

03 / The BetThree founders, one heretical idea

Hardware is famously the worst business model in technology. It takes longer than you think, costs more than you budgeted, and is unforgiving in ways software people find genuinely upsetting. To start a hardware company in 2017 was to mostly invite polite skepticism from venture capitalists who had spent the previous decade unlearning hardware.

SambaNova raised anyway. The early investors - Walden, GV, Intel Capital - were patient and technical enough to follow the argument. By 2021, the company had pulled in a $676M Series D from SoftBank, Temasek, GIC, and BlackRock at a roughly $5 billion valuation. Then the AI cycle turned, GPUs became a national obsession, and being the un-Nvidia stopped feeling like a punchline.

The team that has to build it

The headcount is small for the ambition. About 430 people, depending on the week. The job listings lean hard on systems work - compiler engineers, kernel developers, model performance leads, the kinds of titles you scroll past if you don't recognize them. There is, in fairness, no other way to build a competing AI compute stack from scratch.

04 / The ProductWhat you actually buy

The SambaNova catalog reads like four products with the same prefix, which is a deliberate choice and not, as it sometimes appears, a marketing accident.

If you squint, this is the AWS playbook applied to a chip company - sell the silicon to the customers who want it, host it for the ones who don't, and stay out of the model wars. SambaNova does not, conspicuously, train its own foundation models. It runs other people's.

05 / The ProofCustomers who don't make speculative bets

It is one thing for a private company to claim performance leadership. It is another for Argonne National Laboratory - which runs experiments that occasionally win Nobel prizes - to actually plug your hardware into the wall. SambaNova has those customers. Argonne, Lawrence Livermore, and a handful of European and Middle Eastern banks have all been named publicly. The unnamed list is reportedly more interesting.

The pattern is consistent. SambaNova wins where the customer cares about three things at once: speed, privacy, and not handing their model weights to a U.S. hyperscaler. Sovereign AI, in industry parlance. National labs, banks, telcos, governments - the kind of buyer who reads a procurement contract for fun.

The Intel partnership announced alongside the SN50 in February 2026 is the structural change worth watching. Intel needs a credible AI story; SambaNova needs more manufacturing reach. The mutual logic is obvious. The execution will, of course, be harder than the press release.

06 / The MissionWhat the company actually believes

The official mission - the one printed on the website - is to make state-of-the-art AI accessible to every enterprise. That is fine, and largely true, and a sentence you have read in some form on every AI company's homepage since 2018.

The unwritten mission is more interesting. SambaNova believes that AI compute should not be a near-monopoly held by a single chip vendor in Santa Clara. It believes large models should be runnable by the organizations that own them, on hardware those organizations control, without the toll booths of the hyperscaler era. It is, in a small way, a political argument dressed up as a product roadmap.

This is the part that energizes the team and irritates the competition. It is also the part that explains why customers like Argonne keep showing up.

07 / TomorrowWhy it matters past this quarter

Inference - the act of actually running a trained model in front of users - is on track to become the largest line item in enterprise AI budgets. By a wide margin. Training is glamorous; inference is recurring revenue. If you believe even half of the analyst forecasts, the market that SambaNova has spent eight years building toward is the one that is about to open.

None of which is a guarantee. Hardware competitors keep arriving - Cerebras, Groq, Graphcore (still around), and a dozen others. Nvidia is not asleep. The capital required to keep up is enormous, which is why the Series E and the Intel deal matter so much. SambaNova has to keep proving, every quarter, that its chip is faster than the chip people already own.

But the company is in the position every hardware startup dreams about. Real customers. Real revenue. Real silicon, in real racks, running real production models. The argument has moved from theoretical to empirical. That is a different kind of company than the one that pitched a Series A in 2018.

08 / Back to Geng RoadThe closing scene

Back at 2200 Geng Road, the racks are still blinking. The salesperson is off the phone. The fund is buying.

Eight years in, SambaNova is no longer the unusual hardware company in Palo Alto. It is the unusual hardware company that has receipts. A chip that ships. A cloud that serves. Customers who say its name out loud, sometimes in keynotes. A partner in Intel. A bank account refreshed by $350 million in fresh capital.

The Tuesday afternoon is calm. The Wednesday will not be.