The video argues the AI trade is colliding with a compute reset: Leopold Aschenbrenner’s long-only leveraged fund is reportedly down 30% to 50% in four weeks as semis sell off, while the same scarcity logic is used to justify higher GPU prices, bigger cloud spend, and a more concentrated winner set. The core counterpoint is that frontier labs and policy voices are now openly discussing slowing AI: SSI says it needs 10x more compute, Anthropic wants to slow development without banning open models, and a petition signed by 1,300 researchers plus Ilya Sutskever calls for deliberate pacing. The market framing stays bullish on infrastructure names benefiting from demand and long-term agreements, but bearish on smaller labs and anyone trying to compete without huge capital. Microsoft’s 43% cloud growth and $500 billion-style spending ambition are presented as proof the spend cycle is still accelerating, even as Apple, Intel, AMD, Samsung, and open-source models sit in the mix.
Leopold Aschenbrenner’s fund is reported down 30% to 50% over the last four weeks as semiconductor stocks draw down aggressively, and for the first time since inception it appears to be starting to sell positions it had accumulated over time.
The fund is described as a long-only leveraged strategy built to maximize upside returns, and the current drawdown is framed as its first real stress test after working incredibly well before now.
A rumor says Situation has sold half of its Anthropic stake, but it is explicitly framed as unconfirmed.
SSI is described as radio silent with no leaks, then announcing a long-term strategic partnership with NVDA that will let it 10x compute in the next 12 months.
SSI’s announcement says its research is now worth scaling; the lab is framed as trying to safely align AI with human intention, and the segment describes the announcement as effectively raising $5 million with a high hit-rate profile for smaller AI labs.
A rumor says SSI’s model could release in mid to late August and be superior to Fable, while Anthropic is expected by end of August to ship Fable 6 and "crush everything else."
An open petition signed by 1,300 researchers across frontier AI labs and the AI industry calls for the US government to slow AI development; Ilya Sutskever is said to have signed it that morning, placing him at roughly 1,301.
The petition is summarized as three claims: AI may soon recursively improve itself by building better AI, no single company or country can do this alone so the US should lead international coordination, and tools plus agreements may be needed to slow development if necessary.
Anthropic says it has never stated it wants to ban open models, only slow development so AGI-type intelligences do not escape control.
Dario’s three control options are listed as: stop selling powerful chips or chip-making equipment to China, crack down on distillation, and require capable models to pass through a safety-processing framework before release whether closed or open.
Distillation is defined as taking a released smart model’s prompts and outputs to train a different model, likened to stealing the blueprint.
Dwarkesh’s GPU thesis is that everything goes up — GPUs, usage, utilization — because the intelligence value per token rises while compute supply stays constrained, so pricing can climb with demand.
The segment says one AI chip running Claude Fable 6 could replace a software engineer worth $250K a year, implying the chip should be priced at that level and that smaller labs like SSI get priced out.
MSFT is said to have reported $90 billion in revenue, EPS of $4.75 versus $4.25 expected, and cloud revenue up 43% year over year; the stock is said to be up 15% this morning, with Azure as the number-one highlight.
SMCI is described as doing extremely well and signing many LTAs that lock supply through about 2028, leading to the claim that 2027 memory supply may already be accounted for.
INTC and AMD are described as having strong quarters, with AMD called a banging quarter and Intel framed as infrastructure for agents through CPUs.
A contrast is drawn with AAPL as the one major company not doing AI right now, despite being the world’s most valuable company; Siri is said to be incredible and Apple is expected to push more local inference on-device.
Kimiko 3 is said to have released open-source weights, but running the home open-source setup is claimed to cost about $35,000 a month, then $1 million a year, then roughly $500,000 one time for the weights, so the model is framed as expensive but open.
The Ledger Agent Stack sponsor pitch says agents propose, humans approve, and Ledger signers enforce, with open-source tools for code X, Claude code, and Cursor available now at developers.ledger.com.