The open-source ban is coming: stand up the body, write it into law, then close open source in the name of fairness
Sacks lays out the three steps by which an open-source ban actually lands: first stand up the body, then write it into law, then shut open source down in the name of standards applying equally. Chamath argues the real political fuel is voters' revulsion at two men about to become trillionaires, and has nothing to do with AI safety or with water and power.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
What stung Dario was not the leak but the regulatory-capture charge
Gavin Baker said last week that some people inside Anthropic believe they will end up as the only private company in the world; Anthropic's Sholto Douglas flatly denied it, but Dario's two long posts never touched it at all — they answered only the charge that regulation equals regulatory capture equals concentration of power. Sacks's read is that what actually forced him out into the open was not that leak but being called a regulatory capturer over and over, because that is what damages the purity of his motives. Sacks's counterpunch deliberately steps around motive: he says Dario may well be sincere and not in it for the money, but Anthropic pushing its own preferred regulatory framework extremely aggressively at both the state and the federal level is itself regulatory capture — taking over the machinery of the state on behalf of your own political agenda, and there is not much to argue about on that point.
— David SacksThree trend lines converge exactly on the frontier model companies
Chamath puts three lines together. First, the doomer narrative has taken root and is spreading. Second, the mainstream political backlash is coming from both parties — Abbott in Texas and Shapiro in Pennsylvania, the two governors originally friendliest to data centers, have each moved to restrict them, and Axios reported that the Republican side sent memos to AI company executives telling them to stop rage baiting the public, because the key Ohio Senate seat is going to be lost mainly over this issue. Third, yields are rising, and the Treasury just doubled the size of its bond buybacks. A rising risk-free rate squeezes money out of risk assets, data centers themselves are being halted, and the intersection of the three lines falls exactly on the frontier model companies: they are not investment grade, they have the worst balance sheets, and the more they grow the more compute they need — compute they can neither buy nor finance. His characterization: a completely avoidable unforced error.
— ChamathIf you truly believe AI is dangerous, unmask the thinking tokens first
Chamath gives the doomer position a concrete, verifiable condition: if you really believe this thing is dangerous, the first thing to do is stop masking the model's thinking tokens, so that third parties can verify what you say. Claude masks them, and OpenAI's models mask them too; with open models you can watch in real time what the model is thinking and watch misalignment happen. From this he draws an observation: the open-source camp, Nvidia included, has produced no comparable doomerism. As things stand, outsiders can only accept the closed companies' unilateral account of tokens they cannot see. He also states the argument on the other side: if competitors could see every trace and span, that would amount to handing over the recipe — Anthropic will not even show thinking tokens to paying API customers, so there is no chance it shows them to OpenAI in a room.
— ChamathSelf-regulation is not the danger; building a DMV for AI is
Freeberg presses him: Dario, Sam and Elon all signed on to the industry self-regulation proposal Demis raised, and Sacks supported it at the time, so why oppose it now. Sacks's distinction is that Demis wrote at a very high level, and there are many versions of an SRO, some acceptable and some terrible. What he can accept is the MPAA kind — no reporting to government at all, the industry writing its own standards under government pressure and using them to hold off cruder intervention. What he opposes is the FINRA version — a genuine rule-making body that does pre-release model testing, reports to the government and gets written into law, where the word self is only a fig leaf. He names the thing a DMV for AI: every model lined up waiting for its test. He says he has talked to Elon, and that Elon supports the MPAA version, not the FINRA version. He also points out that Dario's recent blog post treats the FAA as the right model for AI, which in his view is enough to disqualify the entire proposal.
— David SacksClosed models get weaker once wrapped in their own harness
What Chamath offers is a technical observation rather than an ideological argument. Three years ago there were only models and benchmark scores, and whoever scored higher was better; over the past few years a layer has appeared in between — the harness — and Codex, Claude Code, Hermes and openclaw are all examples of it. He says closed models post very high benchmark scores, but their capability degrades once packed into their own harness, while open models actually gain capability when swapped into any other open combination. He expects data like this to become a flood over the coming year: many ways of combining the open-source stack are clearly more usable, and far cheaper. So his answer to Freeberg's question about what happens if China wins is this: as long as America does not use some severe measure to stop consumers and businesses from using the cheapest, fastest, best thing, America wins. What is actually being fought over is which American private company wins or loses, and ordinary Americans do not care about that — the direction of policy already reflects it.
— ChamathThe self-regulatory body is a Trojan horse; the endpoint is banning open source
Sacks lays out the implementation roadmap directly. Step one is to stand the regulatory body up: what Dario wants is an FDA for AI, but the political support is not there, so a FINRA for AI serves as the Trojan horse — self-regulatory in name, on the ground first. Step two, that body acquires the power to test models before release, and the pressure that follows pushes it into law. Step three, in the name of fairness, it is declared that the standards must apply equally to open and closed models. And open models technically cannot comply in the same way — he cites what Dario said in his 2023 Senate testimony: sufficiently advanced open models are dangerous because they cannot be centrally monitored, rolled back and withdrawn the way closed hosted services can, and those properties are technically immutable. The ones putting up the money, the compute and the liaison with government officials will be these same few companies, because the government itself has no idea how to set AI safety standards and can only turn around and ask them.
— David SacksForce companies onto worse models and capital simply goes offshore
Chamath's rebuttal to an open-source ban runs along the line of capital flows: suppose you tell Coca-Cola or Pfizer that inside the United States they have to run slower, more expensive, worse models. They will comply, then shrink their US business and move investment and growth overseas. The result is foreign direct investment falling off a cliff while direct investment into other countries rises. He thinks the closest precedent is China: after Jack Ma's speech on the eve of the Ant Financial IPO, Xi Jinping tightened across the board, requiring that every company of any value have a government representative on its board and a golden vote, after which FDI and the economy fell off a cliff. He says if America actually did this, you would see the result in the stock market and in corporate capital allocation almost immediately. Sacks agrees with all of these consequences, but asks one thing in return: bad consequences do not mean it will not happen — banning data centers has consequences that are just as bad, and people are voting for it anyway.
— ChamathHalf of young conservatives now back government-run grocery stores
The numbers Freeberg cites are the hardest political signal in the episode: among self-identified conservatives under 40, 53% support the government running grocery stores — not Democrats, conservatives. In the Fox News poll from July 2026, Republicans with a favorable view of capitalism fell from 72% in 2019 to 61%, and those with a strongly favorable view fell from 54% to 41%; Republicans who think the system is tilted toward the rich rose from 30% in 2018 to 42%. His conclusion is that this is no longer a principled partisan dividing line but a horseshoe: the part of America that lives on a paycheck will vote for whatever looks like it solves affordability, regardless of which party it comes from. On that basis he predicts that someone like AOC will be president in 2028. Chamath explicitly rejects the premise: he says he does not believe any polling data, that pollsters lie just as the media do, and that the two work together to find the data that gets clicks.
— David FriedbergIn their own words · checked verbatim
Now, look, I think that Daario may be sincere in his beliefs. I'm not saying he's doing this for pecuniary reasons. Nonetheless, he is seeking to capture the machinery of the state on behalf of his political agenda. It's just there's not much to debate there.
David Sacks3:02
When you take a closed source model, it does really well on a benchmark. When you wrap it in its own harness, it decays in capability. when you take an open-source model and you use any other open source model, it improves in capability. What is that trying to tell you?
Chamath28:15
An open source ban is coming. They're not going to call it that. You know what they're going to say? They're going to say that we simply have to apply the same standards to open models that we apply to closed.
David Sacks29:15
Data centers are going to be 3% of GDP. It's basically creating all the gains in the stock market. But they're not voting against that. They're voting against two people that they deeply dislike becoming trillionaires. That's what they're voting against.
Chamath34:20
The data center is the temple where the billionaires and trillionaires pray. It's where they go to make their sacrifices. It is their gathering spot. So I think the data center is this like, you know, temple on the mountain.
David Friedberg1:26:56
Figures
| Data centers as a share of US GDP (expected) | 3% | 34:20 |
| 30-year Treasury yield | 5.3% (this week, a 20-year high) | 1:07:42 |
| Average bias across 3000+ polls over four election cycles | 3.7 percentage points toward Democrats | 1:01:37 |
| Self-identified conservatives under 40 who support government-run grocery stores | 53% | 1:06:41 |
| Republicans with a favorable view of capitalism | 61%, versus 72% in 2019 | 1:06:41 |
| CATO's estimate of new federal spending under the DSA platform | $71 trillion–$211 trillion | 1:22:54 |
| Total net worth of people worth more than $50 million | $23 trillion (10% of it covers only 3 months of federal spending) | 1:23:55 |
Glossary
- harness
- The layer wrapped around a model that handles tool calls and the loop, such as Claude Code or Codex.
- SRO (self-regulatory organization)
- A body in which an industry writes and enforces its own standards; FINRA, NFA and MPAA are three different shapes of it.
- RSI (recursive self-improvement)
- A model automatically designing and training a stronger next generation of itself, with no human architect involved.
- interlocking directorates
- Section 8 of the 1914 Clayton Act bars the same party from sitting on the boards of competing companies at the same time.
- DSA (Democratic Socialists of America)
- The furthest-left organization inside the Democratic Party, used in the episode as the reference point for the cost of 2028 policy.
How to listen
Engineering leads currently choosing between open and closed models; investors who need to assess compute financing and the credit risk of the frontier labs; policy and compliance people who need to anticipate how AI regulation actually lands.
The a16z antitrust investigation segment starting at 56:33 — all four of them say themselves that there is no real information in it.