Dario Defends Himself, Datacenter Panic, AI Doomer Trap, Senate Toss-Up

Why does it matter? Because the greatest threat to American AI dominance isn't China — it's a San Francisco CEO named Dario Amodei.

Dario published a two-part essay defending Anthropic's regulatory push. The hosts read it as confirmation the machinery is already moving — and spent the episode mapping exactly where it leads. The argument isn't that Dario is lying. It's that the specific mechanism he's built, combined with forces already in motion, eliminates open-source competition, squeezes frontier labs out of capital markets, and hands the jurisdiction question to whoever offers the most freedom.

• Sacks maps a precise three-step sequence by which "fairness" language outlaws open-source AI without ever calling it a ban • Chamath identifies a feedback loop connecting AI doomerism, bipartisan data-center backlash, and rising yields that specifically threatens frontier labs — including Anthropic itself • Open-source models improve in production harnesses while closed-source models decay — meaning benchmark-driven AI valuations may be measuring the wrong thing entirely • 53% of self-identified conservatives under 40 support government-run grocery stores; the enemy of the data-center buildout is economic rage, not safety logic

Anthropic's regulatory push is a three-step mechanism to quietly outlaw open-source AI

The open-source ban is coming — Sacks just says they won't call it that. What they'll say is that the same standards must apply equally to open and closed models. The framing sounds neutral. The mechanism is not.

Here's the sequence Sacks maps step by step. First: push for a standard-setting body. Framed as "self-regulatory," branded as a FINRA for AI, though Sacks argues it isn't really self-regulatory at all. It does pre-release model testing and reports to government. He calls it a DMV for AI: "all these models are going to get lined up in a queue waiting to get their test done and then released and it's going to slow us down horribly."

Second: that body gets codified in law. Third — the kill shot — standards get applied equally to open and closed models in the name of fairness. But Dario testified in 2023 that open models are dangerous precisely because they can't be centrally monitored, controlled, or rolled back. That characteristic is immutable. "Daario basically says that open models are dangerous because of these characteristics they have. And those characteristics are immutable." Once the rules exist, "the open models cannot comply in the same way and gradually they will be shut off."

The tell, Sacks argues, is who funds the whole apparatus. "Daario and OpenAI, they're going to fund the whole thing. They're going to contribute all the compute." The people writing the safety standards are the same people whose closed-source business model benefits from eliminating open-source competition. Watch for any "fairness" argument that standards should apply equally — that framing is the mechanism for elimination, not a neutral safety principle.

AI doomerism has already triggered a capital doom loop that threatens the frontier labs that started it

Abbott in Texas. Shapiro in Pennsylvania. A GOP memo to AI executives warning they're about to lose the Ohio Senate race. Chamath connects these into something more alarming than any individual policy move.

"I was really shocked when I saw Governor Abbott of Texas put down this executive order around Texas data centers and limiting its growth. Josh Shapiro in Pennsylvania just did the same thing." These were two of the most data-center-friendly governors in the country. Now they're not.

Three vectors are compounding simultaneously. Frontier lab doomerism seeded public fear. That fear produced bipartisan political backlash. And then there are yields: the Treasury Department just doubled its bond buying. "Rising yields constrains the money supply in terms of how people want to invest in risk-seeking assets because the risk-free rate just keeps going up."

Trace the logic forward. Fewer data centers. Less available compute. Higher cost of capital. "Where does that leave frontier model companies? They're the most at risk. Why? Because they're not investment grade, they're in the worst position in terms of their balance sheets." As frontier labs grow, they need more compute. But the fear narrative they helped generate is now making it harder to fund that compute. Growth makes the problem worse, not better.

This is the loop: the labs that created the doomer narrative are the most capital-constrained players in the system the doomer narrative is now damaging. Chamath's verdict is blunt: "we are now in a very precarious situation that we could have frankly avoided."

Closed-source models decay when deployed in harnesses — and that gap is where frontier lab valuations are hiding

Three years ago, the benchmark was everything. Higher score, better model. Chamath says that era is over, and the data moving to replace it is coming "like a tsunami."

The key shift: harnesses. Codex, Claude Code, Cursor — wrappers around foundation models. And something unexpected happens when you wrap a closed-source model in its own harness. "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." The reverse holds for open-source: "when you take an open-source model and you use any other open source model, it improves in capability."

Chamath's read on what this means: "there are all kinds of ways of using open-source technologies that are meaningfully more performant and dramatically cheaper than the closed source alternative."

The investment implication is significant. Frontier closed-source labs are valued primarily on benchmark performance — lab results produced in conditions that don't survive contact with real deployment. If harness-wrapping systematically degrades closed-source performance while open-source compounds, the benchmark scores justifying those valuations are measuring the wrong environment.

Chamath's conclusion is direct: "Unless America stops us, us being consumers, business owners, from using the cheapest, fastest, best thing. Unless they stop us from doing that, we will win." The debate about which private American company wins or loses is secondary. The variable that determines the outcome is whether open-source stays accessible at all.

If recursive self-improvement is real, every regulatory body and data-center ban just picks the winning jurisdiction — it doesn't stop the outcome

An RSI factory needs three things. "The chips and the power and a connection. And if it's got the chips and the power and a communication connection, it can run." Friedberg's framing, and the logic that follows should unsettle anyone invested in the regulatory slowdown argument.

Recursive self-improvement — AI agents building better AI, which builds still better AI, without human architects in the loop — is what Friedberg puts under pressure. Anthropic has acknowledged it may happen. Enough people inside frontier labs believe it's real. And if it is real: "If RSI is a thing, you don't have a human process where you're stepping in to run this thing."

At that point, pre-release testing bodies, FINRA for AI, data-center restrictions — none of these prevent the outcome. They only determine which jurisdiction achieves it first. A group smart enough can "go spin up an RSI factory anywhere in the world or anywhere in space." Banning data centers means the labs set up in Iceland, in Kazakhstan, wherever sovereignty and cheap power coexist. "Your attempts to ban the data centers or the attempts to ban or to create a regulatory body. That's why I call it a fool's errand."

Friedberg's pragmatic conclusion: "Wouldn't we be better off having systems where we actually have those labs based in the US, where we have jurisdiction over them, where we have capacity to monitor and control and police them?" The only coherent US strategy, if RSI is even plausible, is attraction — not restriction.

The anti-data-center backlash has nothing to do with AI safety — it's 63% of Americans living paycheck to paycheck looking for a target

53% of self-identified conservatives under 40 support government-run grocery stores. Not Democrats. Conservatives.

That's the opening line of a Wall Street Journal article Jason drops mid-episode, and it reframes everything. A Fox News poll from July 2026: 61% of Republicans view capitalism favorably, down from 72% in 2019. Strongly favorable views fell from 54% to 41%. The share who say the system is rigged toward the wealthy climbed from 30% to 42% among Republicans since 2018.

Jason's diagnosis is blunt: "Nobody, no regular rank and file American cares about AI safety. That's not the discussion for them. The discussion for them is why am I not getting rich and everybody else is?" Abbott and Shapiro aren't doing this because they read an alignment paper. They're doing it because constituents watch Waymo rides and Zipline deliveries siphon money upward while wages stagnate. "What two things can they do? They can vote for socialists and they can actually stop data centers."

Chamath sharpens the point by one degree: it's not purely economic logic, it's vibes. The tech class isn't just winning — it's winning in a way that's "pretty displeasing" and "not aspirational." When the next generation of founders looks even more remote than the last crop, slowing them down becomes the only available lever.

No message about national security, AI jobs, or safety reform reaches this. The underlying driver isn't policy disagreement. It's cultural animosity — and the data center has become the temple where that animosity concentrates.

Anthropic's safety case rests on reasoning no outside party can see — which makes it impossible to challenge democratically

Start with the most basic epistemological problem. Anthropic builds its case for AI doom on internal model behavior. But the mechanism by which that behavior is observed is structurally unverifiable from outside: "You cannot obfiscate the thinking tokens of a model... you give it a prompt, you get back something, but all the stuff in the middle is obfiscated."

Chamath's argument is clean. Open-source models let you watch the reasoning in real time. You can see misalignment as it develops. "I haven't seen the same doomerism from any open-source model company, including Nvidia. And the thing with the open source models is you can actually see it thinking." Closed models require trusting the lab's interpretation of internal processes no one else can access. "Instead we have to basically agree to their interpretation of tokens that we can't see that only they can see that they can interpret how they want."

This is the credibility problem none of Dario's essays actually resolve. The blackmail study — prompted over 200 times until they got the headline-grabbing result, ultimately criticized by the UK AI Safety Institute as produced under "highly pressurized conditions" — is the clearest example of what happens when safety claims aren't externally contestable. Sacks connects it directly: Anthropic engineered the study, amplified the result, put executives on 60 Minutes to breathlessly promote it. A lot of people have since looked at it and concluded it was contrived.

Chamath's proposed fix: "Make the bloody thing available to other people. Let somebody else with some amount of judgment as well sit in the room and steel man it with you." Interpretability as a precondition for policy — not a nice-to-have.

The DSA platform would cost $71–212 trillion — the entire Forbes 400 wealth is $6.6 trillion, less than 10% of the low end

The Cato Institute ran the numbers. New federal spending under the DSA policy platform: $71 trillion on the low end, $212 trillion on the high end.

Sacks has the comparisons ready. Corporate profits: $35 trillion. The Forbes 400's total accumulated wealth: $6.6 trillion — "less than 10% of the low end of DSA spending." Take 10% of every dollar held by everyone worth over $50 million in America, and it covers three months of today's fiscal spending. Not the socialist expansion on top of it. Three months.

The conclusion is unavoidable: "This is not about mathematical realism. This is about anger towards a class of people that are increasingly odious." Sacks draws the historical pattern — every socialist movement requires a scapegoat class, some identity assignment that absorbs blame for unaffordability. In 2026, the assignment is tech billionaires, and the data center is, as Chamath puts it, "the temple where they go to make their sacrifices."

Which means no policy concession satisfies the underlying dynamic. Raise the minimum wage, build affordable housing, fund trade schools — the hosts largely agree these are right — but they don't address the animosity. The math makes it explicit: the socialist wave isn't a redistributive calculation. It's rage that has found a scapegoat and will not be appeased by better messaging.

The doom loop is self-constructed — and the only exit is the one the doomer narrative is actively blocking

AI doomerism didn't just shape public opinion. It built the specific political conditions — bipartisan data-center backlash, rising yields, frontier capital squeeze — now threatening the labs that generated it. A loop of their own construction.

If RSI is real, regulation just hands the breakthrough to whoever offers the most freedom. If it isn't, open-source already wins in production and the benchmark valuations don't survive deployment. Either way, the policy debate being fought right now is the wrong one. The actual question is whether the US attracts or repels the labs that will determine which country gets there first — and the doomer narrative is actively answering that question in the wrong direction.

The scapegoat has already been selected.

---

Topics: AI regulation, Anthropic, Dario Amodei, regulatory capture, open-source AI, data centers, recursive self-improvement, RSI, AGI, midterm elections, democratic socialism, inflation, economic inequality, Silicon Valley, FINRA, open-source models, political backlash, a16z