The list of startups vowing to loosen Nvidia’s stranglehold on artificial intelligence is long, expensive, and littered with companies that never quite lived up to their own hype. This week, SambaNova made the case that it belongs near the front of that line, though for how long remains to be seen.
On July 8, the Palo Alto chip maker announced it had raised $1 billion at an $11 billion valuation in the first close of its Series F, led by General Atlantic and backed by a veritable sh*tload of institutional money, including BlackRock, T. Rowe Price, Capital Group, Intel Capital, and Qatar’s sovereign wealth fund (and when Gulf oil money and Wall Street’s largest asset managers are all elbowing into the same AI chip round, it is worth remembering how these crowded, everyone-in trades have tended to end. Spoiler warning: no muy bueno).
What differentiates this raise from the others on the 2026 “let’s throw all the money we possibly can at AI” leaderboard is where SambaNova is aiming its chips: inference rather than training.
Training a frontier model is the glamorous, headline-grabbing half of the business that often ends in a few viral videos showing we’re not quite as close to AGI levels or, ahem “sentience” (Anthropic needs to calm down on the sensationalist marketing tbh).
Instead, SambaNova is focusing on inference, the far less sexy work of actually running that finished model billions of times a day, where the recurring revenue and the durable long-term market accumulate. SambaNova is wagering that the future belongs to whoever runs models fastest and cheapest, and it just secured serious validation for that thesis by signing JPMorgan Chase as an ‘inference partner.’ (infer from that what you will). The round also lands a mere five months after a $350 million Series E, a fundraising cadence that suggests either the technology is genuinely working or the capital simply has nowhere more urgent to be.
Why inference is the smarter place to fight
There is a genuinely clever strategic logic buried under all that cash. Training happens once per model and is dominated utterly by Nvidia, whose real moat is not the silicon but CUDA, the software ecosystem every AI engineer already knows. Inference, by contrast, happens forever, scales with actual usage, and is far less wedded to Nvidia’s software stack. SambaNova builds what it calls reconfigurable dataflow chips, purpose-built to run models efficiently, and its pitch to a bank like JPMorgan is straightforward: you are going to run these models a billion times a day, so let us make each of those times cheaper. That is a much better place to pick a fight than trying to out-Nvidia Nvidia at the thing Nvidia is best at.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
The catch, and there is always a catch
The graveyard of “Nvidia killers” is deep, well-populated, and littered with companies that had brilliant chips, glowing benchmarks, and absolutely no way to pry developers off CUDA. SambaNova has a real product and a marquee customer now, which puts it ahead of most, but an $11 billion valuation prices in an awful lot of success it has not yet delivered. A second close is expected within weeks, which will only push that number higher.
The optimistic read is that the AI market is finally splitting into training and inference as separate businesses, and that a specialist can win the second one. The cynical read is that we have simply reached the stage of the cycle where a billion dollars is the new Series A and everyone with a chip and a slide deck is getting funded. As ever with this stuff, both things can be true at once.
Sources
Not investment advice, just a market watcher’s notes.