SambaNova co-founder and CEO Rodrigo Liang joined CNBC's Squawk on the Street for an exclusive interview from the floor of the New York Stock Exchange on September 14, 2026, a session in which AI chip names sold off broadly as investors weighed whether frontier model training was starting to pace itself. Liang's answer to that question was to point at the other half of the market: inference, where the economics, not the benchmarks, decide what gets deployed.
He lays out the numbers behind that argument: a 6-month payback period, a 10x reduction in the cost of inference from deploying into data centers that already exist, and cost per token served as the metric that actually matters. He also discusses the Intel collaboration, the second close of SambaNova's Series F, AI safety, and IPO timing.
CNBC opened on the morning's question of whether frontier model training is slowing, and what that means for everyone building underneath it. Liang's read is that the attention on frontier models is welcome, but the harder problems are elsewhere.
"Inferencing is gonna be the center of the AI economy. And so here, we need to tackle the power issues, the cost issues. That's something that we can all agree on — that as we move into inference, the cost has to come down."
Asked what makes SambaNova faster than the field of NVIDIA challengers, Liang reframed the question. Speed is table stakes, and everyone will race toward it. What separates a demo from a business is what that speed costs to deliver.
"Ultimately, I think cost structure of that speed is gonna be incredibly important, because we want to democratize this. We want to make this available to everybody. Being able to drive good payback in order to allow companies to actually make money on inference is gonna be incredibly important to actually do a sustainable business in inference."
Pressed on what "cost" means in practice, Liang gave a number: 6 months to pay back the infrastructure investment on SambaNova, against an industry norm he puts at 12 to 24 months.
"For companies to be able to achieve sustainable growth, sustainable investments that allow you to make your money back, you gotta find a way to get your money paid back in a quick enough time because technology is moving really, really fast."
SambaNova completed the first close of $1 billion in Series F financing on July 8, 2026, at an $11 billion post-money valuation, in a round led by General Atlantic. Liang confirmed the second close is nearly done, with public-market investors participating in the private round.
"We're almost finished on that. You're seeing more and more of the public investors are coming into the private round, and so we're excited to actually share those shortly."
On Intel, and on the recurring reports that Intel might acquire SambaNova, Liang pointed to a relationship that predates Lip-Bu Tan's tenure as Intel CEO, and to a signed multi-year agreement to build and sell together.
"I've known Lip-Bu for 25 years. He was our chairman long before he was actually CEO of Intel. With Intel, we've actually signed a multi-year relationship where we're developing products together. We're actually going to market together and really driving this notion of lower cost of inference."
The clearest technical claim in the interview: Because SambaNova draws far less power, it can go into brownfield data centers that are already built, already powered, and already air-cooled, sidestepping the liquid cooling and new construction that competing systems require.
"One of the things that's incredibly important with SambaNova technology being a much, much lower power. We can actually use existing data centers. No more buildouts. If you do that, cost of inference is gonna drop 10x."
Asked what metric to measure that against, Liang named one, and defined what rolls into it.
"Cost per token served. If you think about every token that we issue, we want to actually see how much did it cost in terms of the cost of acquisition of chips, data center, power, cooling, all those things together. When you drive power down, when you drive speed up, when you drive concurrency up, you can actually then generate a much, much higher output for a fraction of the cost."
CNBC pushed on the implication: If inference runs in existing facilities, is the industry building more than it needs? Liang's answer was more specific than yes or no: The compute is needed, but it does not all require greenfield construction.
"If we can find a way to actually reuse existing data centers, brownfield data centers that are air cooled with the power already allocated to it, then you don't have to do that. That's part of the solution. It's not the entire solution."
On safety, Liang argued the moment calls for acceleration rather than a slowdown and pointed to what regulated customers are already doing: pulling AI back inside their own perimeter, a pattern he calls repatriation.
"I think it's less about the slowdown, but acceleration of the thinking around safety. Some of our customers have already taken steps towards it, where they bring the AI on-prem, bringing the AI inside their own infrastructure to protect the data of their customers. JPMorgan is doing exactly that."
"You had all this infrastructure going towards the cloud. Now people are starting to realize, for a class of use cases when it comes to AI, that data is so important to protect."
Asked, after 30 years as a semiconductor engineer, whether he shares the concerns of those at the frontier, Liang split the problem in two.
"The AI genie is out of the bottle. Now you have the models on one side and infrastructure on the other side. We have energy issues, energy crisis, data center crisis on the infrastructure side, and we need to be attacking both of those. Our solutions are really focused on the infrastructure, and making it much cheaper, much more sustainable, much more cost effective."
With OpenAI expected to go public next year, CNBC asked whether SambaNova might follow.
"We'll do it when it makes sense for the company. We just closed a billion dollar raise, and so we're in good position to continue to invest in the technology to drive this goal of ours, which is dropping the cost of serving."
On the use of proceeds, Liang pointed at supply chain lead times.
"We're buying supply. So that allows us to actually make sure that we have chips and infrastructure when the customers need them."