This episode opens by focusing on the mythos-class model Fable 5 that Anthropic released Tuesday: it nearly sweeps the benchmarks, but its per-token price is double that of Opus 4.8 (in theory it consumes less overall because it's smarter). What truly ignited the developer community were three things — ① it forcibly retains 30 days of all prompts + outputs + context (with no exception even for enterprise customers who signed 'zero data retention,' who can only choose not to use the new model); ② it uses this data to 'profile and tier' users and decides what capabilities to unlock accordingly; ③ when the system judges you're doing frontier AI research (competing directions like machine learning / chip design), it **covertly downgrades** you to a weaker model while still charging you the premium-model price and even rewriting your prompt in the background. After the backlash it only 'walked back' one step: it now informs you of the downgrade, but still downgrades. People asking about mitochondria or the relationship between GLP-1 and cancer were downgraded, and even host Jason, asking on air about fertilizer regulations and about the components for building a nuclear bomb, was switched on the spot from Fable to Opus.

Chamath argues this exposes two big risks: one is censorship risk (to individuals) and the other is governance and business-continuity risk (to enterprises) — any downstream employee/scientist inadvertently triggering it could get key capabilities cut off, so enterprises must begin investing in 'control over AI, who gets to learn from your data, avoiding single points of failure,' needing more distributed, more governable solutions. Sacks adds: this is a betrayal of trust — a company that once opposed government surveillance is now doing mandatory surveillance + deciding at its 'discretion' whether you deserve frontier capabilities, creating AI's 'haves and have-nots.'

Freeberg speaks from first-hand experience: over the past two years his agricultural genomics company relied on these models for gene-construct design, RNA-guide design, and the like, with extremely high efficiency; but over the past two weeks it was restricted on grounds of 'bioweapon risk,' forcing a switch to locally deployed open-source models — and 'the best open-source models today are Chinese' (like the Ark Institute's genomics large model, already open-sourced and downloadable and fine-tunable by anyone). His core argument: you can't 'turn off AI,' and self-censorship will only push US companies toward Chinese models and hand over the advantage; regulation should target 'weaponization at the output end' (there are already product-liability laws, anti-bioweapon laws, and synthetic-nucleic-acid screening), rather than bluntly gating capabilities at the 'access end.'

Chamath reveals that two years ago he bought 2,000 acres in Arizona and got 2 GW of data-center capacity approved, originally intending to flip it to Blackstone/Google, but now feels he 'may have to build it himself,' because the vast majority of compute still flows to large models and the usable compute for open source is pitifully scarce; he even bid for and secured a third GW — but 'a 1 GW now costs $100 billion' (two years ago his project was only $4-5 billion, a 20x increase), so 3 GW is $300 billion, which an individual can't shoulder. He points out the fundamental difference between AI and internet economics: the internet's marginal-user cost ≈ 0 (which is why Google/Meta earned excess profits for decades), whereas **every marginal AI user has a real cost** — consuming GPU, power, and memory, and requiring enormous critical infrastructure just to get on the field. This is both the root of why token costs are hard to bring down and the strongest argument for 'why should the government hold equity in AI companies' (analogous to the federally built interstate highways — should the logistics companies running on them owe part of their profits to the government).

The second half turns to Bernie Sanders's proposal: in a New York Times op-ed he advocates a one-time levy of **50% equity** on the largest AI companies (OpenAI, Anthropic, xAI) to inject into a sovereign wealth fund, giving the public voting rights and board seats, on the grounds that 'AI is built on humanity's collective intelligence (books, songs, news, research, code), stolen by a handful of the ultra-rich.' All four oppose 'confiscation' (unconstitutional, a bad precedent), but all resonate with the sentiment behind it: since AI CEOs (Dario says entry-level white-collar jobs will halve in 1-5 years) keep telling the public 'we'll put half of people out of work,' ordinary people will naturally ask 'what's in it for me.' Sacks proposes: since OpenAI/Anthropic are 'public benefit corporations (PBCs)' already bearing a public mission, perhaps part of their profit flow should go toward paying down the national debt. Freeberg seizes the moment to restate his own proposal: convert the already-bankrupt Social Security trust fund (which currently holds only a single $4 trillion special Treasury bond) into an **individual-account sovereign fund** that can invest in stocks, investing in (rather than confiscating) quality American companies including AI, making everyone a shareholder. He also strongly rebuts 'AI causes unemployment': across the dozens of companies he's seen, AI is mainly used on the **revenue side** (one engineer can do 100x the work of before and build more products), not the cost-cutting side, so everyone is hiring like crazy — and May's nonfarm payrolls beat expectations.

The closing segment speeds through two non-investment topics: May CPI was +4.2% YoY (the highest since April 2023) and PPI 6.5%, mainly due to the Iran war pushing up energy + government overspending, with the odds of a rate hike this year rising to 49%; Chamath warns that if China's reserves run dry and it returns to the spot market to buy oil, oil prices could spike to $150-200; plus a heated debate over the Los Angeles mayoral primary's 'mail-in ballots / legalized fraud.'

(For AI investors: the 'regulatory capture + censorship' of closed frontier models is creating real demand for open-source/localization, benefiting the open-source model ecosystem and local inference compute; data-center power is an increasingly hard bottleneck, and the capital threshold of 1 GW = $100 billion is reshaping the landscape of players across the entire supply chain.)