Rodrigo Liang says the money spent on inference infrastructure should come back in six months, against the 12, 18 or 24 months he says most companies are running at today.
The build-out case for AI assumes new data centers with liquid cooling and far more power per rack. Liang's pitch is the opposite: put the workload into buildings that already exist, already have their power allocated and are still air cooled.
"Inferencing is going to be the center of the AI economy."
Liang co-founded SambaNova and runs it, has spent 30 years as a semiconductor engineer, and has just raised what the interviewer described as a $1 billion Series F at an $11 billion post-money valuation.
The full segment is covered here so you can skip it.
Here are the 8 insights that matter.
👤 Guest: Rodrigo Liang, co-founder and CEO of SambaNova, a chip company competing with Nvidia on inference
🎙️ Hosts: Carl Quintanilla, David Faber and Sara Eisen, the anchors of CNBC's Squawk on the Street
👥 Also on: Bessemer Venture Partners' Byron Deeter, in a separate segment of the same hour
📰 Published: 14 September 2026 on the Squawk on the Street podcast feed (CNBC)
🟣 Apple Podcasts | 🔗 Segment page | ⏱️ length not available
Key Takeaways
A slowdown in frontier training is not a slowdown in inference, which is where his business sits
Six months is the payback period he is selling against an industry norm of 12 to 24 months
The metric he wants the industry priced on is cost per token served, all-in
Chips, data center, power and cooling together, divided by the tokens issued
Lower power means old air-cooled buildings can be used, so the build-out is not the only route
He says that alone takes the cost of inference down by a factor of ten
Banks are bringing AI inside their own firewalls, and JPMorgan is doing it now
He calls it a repatriation of infrastructure, after a decade of moving to the cloud
Intel is a multi-year product and go-to-market partner, and its chairman was SambaNova's chairman first
1. Inference Is the Economy
The interviewer opened on the market's problem of the morning: investors trying to price a world in which frontier-model training is paced, with some arguing that is simply good for inference.
Liang did not dispute the framing. "Inferencing is going to be the center of the AI economy."
He treated the safety conversation as welcome and the infrastructure conversation as the other half of the same problem. "And so here we need to tackle the power issues, the cost issues."
The one thing he thinks is not in dispute: "That's something that we can all agree on that as we move into inference, the cost has to come down."
2. Speed Needs a Payback
Asked what he could cite about the product that would make anyone believe it is the fastest on these workloads — with the interviewer noting there is no shortage of companies claiming to challenge Nvidia — Liang moved the question off speed.
"Look I think speed is going to be something that everybody's going to be racing towards. But ultimately, like I said, I think cost structure of that speed is going to be incredibly important because we want to democratize this." The goal, he said, is to make it available to everybody.
"And so being able to drive good payback in order to allow companies to actually make money on an inference is going to be incredibly important to actually do a sustainable business in inference."
3. Six Months, Not 24
Pressed on what he means by cost if it is not just speed, Liang gave the number the rest of the argument rests on.
"We're talking about six months. Six months." That, he said, is how long it takes to pay back the infrastructure invested in.
"Today, most companies are taking 12, 18, 24 months before they can actually make their money back."
His reason for caring about the gap is obsolescence: "So for companies to be able to achieve sustainable growth, sustainable investments that allow you to make your money back, you've got to find a way to get your money paid back in a quick enough time because technology is moving really, really fast."
4. The Series F and After
The interviewer put the funding record to him — a Series F on 8 July, described on air as "a series F, $1 billion, and 11 billion post-money valuation" — and asked whether the second close he had talked about had happened.
"Well almost finished on that. Really excited to see a number of really excited investors to come in."
The change he flagged in the register is who is showing up. "As you're seeing more and more of the public investors are coming into the private round." He said the names would be shared shortly.
"But look, it's going to be incredibly important for us to continue to bring in great investors that bring the company into a broader and broader visibility."
5. The Intel Relationship
Asked about Intel — the interviewer noted that Lip-Bu Tan is its executive chairman and that there are rumors Intel might buy SambaNova — Liang answered on the partnership and let the takeover question go by.
The relationship predates Intel: "One of the great things about Lip-Bu is you have known Lip-Bu for 25 years. He was our chairman long before he was CEO of Intel." Over the years, he said, Tan has been supportive of the company.
"With Intel, we've actually signed a multiyear relationship. We're developing products together. We're actually going to market together and really driving this, this notion of lower cost of inference, driving that cost structure down so that everybody can be part of the AI economy."
6. Brownfield Data Centers
Asked how much cheaper inference actually gets, Liang gave a mechanism rather than a projection, and the mechanism is a building.
"Well, you've got to look at where things are today. It's a combination of speed, power, data center."
Because the technology draws much less power, he said, existing data centers can be used — no new build-outs, brownfield sites with the power and cooling already in place. "If you do that, cost of inference is going to drop ten x."
The interviewer put the implication to him directly: does that mean the world needs less compute than it thought? Liang agreed with the reading. "Yeah, exactly. And so there are lots of data centers in the world today."
"The problem is that most of the new technologies need liquid cooling with much higher power outputs, much larger infrastructure in order to actually deploy these large existing clusters."
"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. And that's where SambaNova fits."
He was careful about the size of the claim. "That's part of the solution. It's not the entire solution."
7. Cost Per Token Served
The interviewer asked whether the right yardstick is revenue per gigawatt, and what metrics actually matter. Liang named one.
"Cost per token served."
The point is that it is an all-in number: "And so 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 chips, data center power cooling, all those things together." What was the total cost, he asked.
"And so when you drive power down, when you drive speed up, when you drive concurrence of many users at the same time, you can actually then generate a much, much higher output for a fraction of the cost."
8. Bringing AI On Prem
Asked whether he has safety concerns of his own, Liang redirected from model safety to data safety, and described customers moving infrastructure back inside their own walls.
"Look, I mean, when it comes to models, I think incredibly important for us to be talking about safety. I think it's great that the world's coming together and thinking about that."
His reframing of the week's debate: "I think it's less about the slowdown, but acceleration of the thinking around safety, what do we need to do?"
"And some of the banks, for example, 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." The logic, in his words: "Let me bring it safe, make my customers data safe within my own firewalls."
The interviewer added that this is now the big discussion among law firms as well as banks, after a model solved a math problem using other mathematicians' data.
Liang's name for the trend is a reversal of the last decade: "We're seeing this repatriation of infrastructure where you had all this infrastructure going towards the cloud."
"And now people are starting to realize for a class of use cases, when it comes to AI, that data is so important to protect."
"And so why don't I bring it back repatriate into my own firewalls, something that I've tested for 25, 30 years, protecting the way I've always protected my data and still get the benefits of AI in a way that allows you to actually safely make that transition."
Bonus Insights
Asked whether he is personally worried about the pace of the technology, the way some of the people at the frontier are, Liang started from the assumption that it is too late for that question: "The AI genie is out of the bottle." He split the problem in two — "Now you have the models on one side and infrastructure on the other side" — and said both need attention: "And where there are safety issues that we need to be very, very thoughtful about saying we have energy issues, energy crisis, data center crisis on the infrastructure side, and we need to be attacking both of those." His own lane, he said, is the second one. "And our solutions are really focused on the infrastructure and making it much cheaper, much more sustainable, much more cost effective."
On whether SambaNova would follow OpenAI to the public market next year, Liang would not commit. "Look, I think we'll see." The company will list when it makes sense for the company, he said, and the recent raise means it does not need to. "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 and actually get infrastructure much, much cheaper." "And so if that makes sense for the market, it makes sense for the company."
Asked where the new capital is going, he named supply rather than headcount: building infrastructure and buying supply, because supply chains carry long lead times.
The interview also produced the only light moment in the hour. Told he had been a semiconductor engineer for about 20 years, Liang corrected it to "30." The interviewer's response was that he looks far younger than the years would suggest.
Liang's bottom line is that the cost of inference, not the speed of it, decides whether AI is a sustainable business — and that the cheapest route to a lower cost is to stop building data centers and start reusing the air-cooled ones already standing.
Products, Companies & Tools Mentioned
SambaNova (Liang's company: a six-month payback on inference infrastructure, deployed into existing air-cooled data centers, priced on cost per token served)
Intel (A multi-year partnership to develop products and go to market together; its executive chairman Lip-Bu Tan was SambaNova's chairman for years first, and the interviewer raised takeover rumors Liang did not address)
Nvidia (The incumbent the interviewer said SambaNova is one of any number of challengers to)
JPMorgan (His named example of a bank bringing AI inside its own firewalls to protect customer data)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: