The host had a blunt objection. Whatever AI ends up producing, he asked, won't a lot of it "flow disproportionately into the billionaires, the most wealthy, that are building the robots?"
David Friedberg, billed by the episode as a top White House adviser, answered with a wager. "I will bet you anything that in the next 10 years, there will be a billionaire that will emerge that has absolutely zero net worth today, downloaded open-sourced AI, and built a company that made them a billionaire five years from now, seven years from now," he said. "And they're going to be someone that's going to come from nowhere." There would be a lot of those stories, he expected, the way there were with the internet.
The wager rests on one thing being true: that a person with no capital can now get a capable AI model without asking anyone's permission.
What "open source" means here
Asked what the term means "for Janet and Dave who might not know," Friedberg gave the short version: "open source just means the software is free for anyone to use."
The AI version of this differs a little from free software, and the difference matters. What is published is the model's weights — the trained parameters that define its behavior. Friedberg described them as unglamorous: "It's literally just a list of numbers. In a document, it's just, you know, trillions of numbers." Download the file, and you can run the model on hardware you control, or adapt it, instead of renting access through someone else's service.
Whether you may also sell what you build depends on the license attached. DeepSeek's January 2025 release of R1 sits at the permissive end: the company published code and model weights under an MIT license, allowed commercial use and distillation — training a smaller model on a larger one's outputs — and shipped six smaller distilled models alongside the full one.
Friedberg's other qualification is his own: the model is free, but running it is not. "The only cost is the cost of finding a computer to run that model," he said. "So you could have your own computer, you could rent one, but that cost is coming down like crazy now too." In his telling that is the ordinary behavior of technology left alone: "It always gets cheaper and then everyone benefits."
The Netscape story he tells
To explain why this changes who gets rich, Friedberg went back to the early web. Netscape sold the server software you needed to publish a website and the browser you needed to read one. "So charge, charge, gatekeeper, gatekeeper," he said. Then the Mozilla Foundation produced Firefox and the Apache Foundation produced a free web server, and anyone could put a site online or browse the web without paying a toll. "Open source then enabled all the entrepreneurs that built websites that made money to proliferate," he said. "They weren't gatekept."
Mozilla's own account of its history is tangled up with the company it is being contrasted against. The project began when Netscape released the source code of its browser suite in 1998; volunteers extended the work well beyond one company's browser; Mozilla 1.0 arrived in 2002, the independent nonprofit foundation formed in 2003, and Firefox 1.0 shipped in 2004, passing more than 100 million downloads in its first year by Mozilla's count. The free alternative grew out of the incumbent rather than simply arriving to displace it.
Applied to AI, Friedberg's version is that you no longer have to "go pay Anthropic or Claude or Gemini or ChatGPT for AI."
The argument inside the government
Earlier in the conversation he described what he considers the live policy question in Washington, one that splits factions within the same political party: whether American businesses should be allowed to use Chinese open-source models. Those models, he said, are "as good as, or in many cases, better than private American models," and because the weights are published you can download them and run them on whatever computer you like.
Companies want them because they are cheap. His illustration was a gap of roughly a hundred to one: about 50 cents for a million tokens of output — a token being roughly a word — against $50 from a private model such as Anthropic's. The comparison is his own rough measure rather than a quoted rate card; DeepSeek's launch announcement for R1, for instance, listed output at $2.19 per million tokens on its own hosted service, with separate prices for cached and uncached input. Friedberg's point is about magnitude, and about the option a downloadable model creates: skip the vendor's meter entirely and pay only for the machine.
So, he said, the question becomes whether to let those models become ubiquitous in industry, and what that does to American AI, to the capacity to compete, and to security.
On security he separated two complaints. The first is that a Chinese model might carry hidden backdoors. "That's not how open-weight models work, by the way," he said, arguing that "people that are technically deficient don't fully grok what the open-weight model really is" — a file of numbers, in his description, not a program phoning home.
The host raised the second: anyone who downloads a model can jailbreak it — strip away its refusals — and use it for whatever they like. Friedberg accepted the shape of the worry, listing cyber weapons, bioweapons and physical weapons as things people argue AI now puts within reach. His response was that society already lives with tools of that kind. Anyone can buy a DNA sequencer and a DNA printer and, in theory, make a virus; that exists today, and neither sequencing nor printing has been stopped. He reached for firearms as a further example, and the comparison was waved off in the room almost as soon as it was made.
He granted the general case: there is a good argument that any technology capable of causing harm should be banned or controlled by government. His flip side was that "if it's allowed to proliferate, there's far more benefit than risk. And the risk side needs to be mitigated. And that's where you build good defense systems and so on" — monitoring, cyber defense, biodefense. The remedy he prefers is spending on defense rather than restricting who may hold the technology.
Why he thinks cheap models mean more data centers, not fewer
Asked later what would let people on the ground share in any of this, Friedberg first rejected the premise that there is little to share. "Everyone thinks that AI is going to accumulate to two or three companies. I don't think that's how it's going to play out," he said. Every company using AI is a company with more money to pay employees, hire and train people, in his account. "I don't buy into the narrative that Dario, Elon, and Sam end up having all the value in AI. I think the Chinese blew that up."
The host pressed the obvious consequence: if the models are nearly free, doesn't the spending on data centers to run them look like a bubble? Friedberg said no, and gave the reason in the first person: "if the cost to run a Chinese model comes down by some number of fold over the cost to run a proprietary model, I'm going to run more models. I'm going to do it more. And as I do it more, I'm going to need more compute. I need more data centers." Cheaper inference, on this reading, enlarges the bill rather than shrinking it.
That leaves the infrastructure dependence intact — and with it the complaint that data centers push up everyone's electricity bills. Friedberg said he fully accepts that concern; the water objection he dismissed as "BS." His proposed fix for power is a construction standard: "every data center should be on its own grid. And every data center should make enough power not just to be on their own grid, but to pump energy back into the grid." Set that as the requirement for every new project in the US, he argued, and "you're dropping the price of electricity for everyone in America."
His last step is about who holds the deeds. Earlier he had named direct ownership as the mechanism he wanted: "401ks to own this stuff would be what we need to change from a public policy perspective," so that people could look at their phones and see themselves benefiting. Data centers, he said, fit that shape already. "They can all be in REITs and people can buy stock. And many of those are publicly traded. So they should be in your 401k and everyone can own a piece."