Summary

Iran’s oil shock is large enough to lift inflation and cut growth, but the panel’s base case remains a short disruption. Brent swung from $84 to $119, back to $84, then toward $100; Goldman Sachs raised its PCE forecast from 2.1% to 2.9% and cut GDP growth by 30 basis points. Chamath’s strongest signal was oil collapsing from $120 to $90 after Trump said the war would end soon: a market “trial balloon” implying traders see no sustained conflict.
The preferred off-ramp is to declare Iran’s capabilities degraded and exit before escalation attacks the Gulf’s economic and human life-support systems. Sacks warned that retaliation against oil infrastructure could make reopening Hormuz irrelevant, while attacks on desalination could threaten roughly 100 million people and give Iran a “dead man’s switch over the economic fate of the Gulf states.” An extended war could also exhaust Israeli air defenses and raise the nuclear tail risk.
China may have more incentive than America to broker the settlement. The US produces and consumes roughly 20 million barrels a day, while Jason said Iran and Venezuela supply 20% of China’s oil consumption and its reserves cannot absorb five or six months of disruption. With young-male unemployment already cited at 25% and a three-day Trump-Xi summit approaching, Chamath’s call was categorical: “All roads lead to China.”
Brad argued that OpenAI and Anthropic have achieved unprecedented revenue velocity, with demand crossing from IT budgets into labor budgets. Jason cited Anthropic at a $14 billion run rate and $380 billion valuation, versus OpenAI at $20 billion and $840 billion; Brad separately described Anthropic producing a reported $6 billion February. Brad said Opus 4.6 and ChatGPT 5.4 marked the threshold where models and agents began “augmenting labor.”
The underwriting dispute is whether AI revenue is durable production spend or an enormous rolling experiment. Chamath sees board-driven AI checkboxes and regulated workflows still requiring humans. Friedberg said his team’s token consumption and spending are rising far faster than its economic output. Jason’s framing was: “They’re selling at a profit. I’m buying them and losing money.” Brad’s counter is that government, military, Palantir and Nvidia deployments are already production-grade—and experimentation that repeats indefinitely starts looking recurring.
AI’s communications failure is becoming a material infrastructure constraint. The panel contrasted doomer fundraising rhetoric with Sam Altman’s simpler utility model of metered intelligence; US optimism was cited in the 30s versus about 80% in China. Jason estimated that canceled or potentially canceled data centers could remove $120 billion of annual revenue across 2025–26, while open source already handles roughly 85% of one startup’s tokens without preventing frontier-lab growth.
State wealth taxes risk turning headline revenue into capital flight and larger budget holes. Washington approved an extra 9.9% tax above $1 million beginning in 2029, targeting 30,000 households and $4 billion, as Howard Schultz left Seattle for Surfside. Chamath cited California modeling in which 71% of 100,000 simulations produced negative NPV and an expected $25 billion hole; the broader political risk is a 5%-a-year federal wealth tax becoming Democratic “table stakes” by 2028.

Deep dive

Not yet available upstream; scheduled sync will retry.

Episode 176 - March 13, 2026 Dylan Patel — The single biggest bottleneck to scaling AI compute