What this covers

Gavin and Patrick work through a market drawdown that lacks a single named fear. Gavin observes that earlier selloffs (2022, DeepSeek, Liberation Day) had clear causes, which made them reassuring, and that the absence of one now is more worrying. The main negative data point Patrick raises is third-party data suggesting Anthropic's usage curve has moved off its trajectory. Gavin's answer to the open-source worry rests on Token Is a Token: a token takes the same flops, memory, and watts either way, so moving from frontier tokens at roughly 90 percent gross margin to open-source at about 30 percent shifts margin to infrastructure. Companies routing queries to cheaper models cut spend but, he argues, consume more GPU hours.

The middle of the conversation moves through the supply side. Gavin describes Installed Base Repricing, where contracted compute sits far below spot and lifts operating cash flow as contracts roll off. He also covers memory makers trading price upside for prepaid long-term agreements, Dwarkesh Patel's claim that a year's H100 rental could reach about $250,000, Nvidia's credit wrapper for GPU buyers, SRAM accelerators in disaggregated inference, and orbital compute.

Several turns are contested. Gavin says Chinese chipmaking cannot skip ahead, because Learning by Doing needs real production cycles, and calls US-China decoupling self-reinforcing. He compares a Debt-Fueled Buildout Unwind to the internet era. He argues data centers lower nearby electricity prices and sustain blue-collar jobs, and says a widely cited water-use figure was off by about 10,000 times. He also suggests that investors who all read news through Claude resemble a single trusted voice like Walter Cronkite.

Sharpest takeaway

Gavin argues that the July 2026 AI stock selloff is driven by narratives (open-source competition, Meta's compute rental, DUV news, spreads) while the underlying compute fundamentals are accelerating, and that credit financing is the one real risk to watch.

  • Token growth, GPU rental and spot prices, and hyperscaler operating cash flow all accelerated, so the selloff is not supported by quantitative data
  • Open-source token share gains shift margin away from frontier labs but do not reduce the compute needed per token
  • Credit and spread widening are the only undeniable negative, and they matter only if the buildout must be debt-financed

The argument · threads3 threads · 100 claims
0.71

Compute demand is a sustained shortage, not oversupply, despite one negative usage signal.

2 pointscentrality 4/5
  • No one Gavin asked reports having too many GPUs; demand for compute feels like a drug market, indicating a sustained shortage.

    “Not a single person.”

  • The main negative data point people cite is that third-party data suggests the Anthropic usage curve has started to move off its trajectory somewhat.

    “the third-party data suggests the Anthropic curve started to go off of its trajectory a little bit. That's the only thing.”

0.57

Founder-led firms' AI adoption signals they expect continued need for people.

2 pointscentrality 3/5
  • Founder-led and controlled companies, after adjusting for COVID-era overhiring, are not doing large layoffs even though they are most likely to adopt AI, which suggests they expect continued need for people.

    “nobody's really laying people off. These are the people that would probably be most quick to adopt AI”

  • Patrick suggests the distribution of pay inside a company, such as CEO-to-median ratios, could mirror the split between frontier and open-source tokens, and Gavin agrees.

    “The CEO makes X times more than the median person at a company, and maybe that's frontier tokens versus open source tokens.”

0.15

Factual errors persist long after they originate, as the spinach-iron example shows.

1 pointcentrality 1/5
  • The spinach-has-more-iron belief arose from a decimal-point error in an academic publication about 80 years ago that propagated widely, illustrating how factual errors persist, as Patrick explains.

    “In an academic book, they placed the decimal two things wrong.”