All-In: OpenAI's Identity Crisis, Datacenter Wars, Mamdani's Pied-à-Terre Tax, Swalwell Out

Mamdani's pied-à-terre tax[00:42]

NYC Mayor Mamdani's first big fiscal move: an annual tax on non-primary homes over $5M within 15 miles of Midtown — the besties cite an unconfirmed 3.9% figure.¹ Jason calls out that the announcement was staged on the sidewalk outside Ken Griffin's $238M penthouse⁴, which Chamath defends as fair game since the unit had been publicly marketed for years.

Chamath's frame is that the most elastic slice of the market is precisely the segment being taxed, and pied-à-terre buyers who can park capital anywhere will just go elsewhere — exactly what happened in London after non-dom status was gutted.² Sacks counters that a price-insensitive whale paying $3–4M/year in property tax while consuming zero city services is pure upside for the city, and that removing that demand "dries up" the developer math for new towers.

OpenAI's identity crisis[11:23]

OpenAI's chief revenue officer Denise Dresser sent a four-page Sunday memo — promptly leaked — that touted Amazon's $50B commitment and accused Microsoft of having "limited our ability" to serve customers.⁵ The same memo went after Anthropic, arguing its headline ~$30B run-rate is inflated by roughly $8B of channel-partner revenue.

The FT's simultaneous scoop was the more structural problem: secondary-market Anthropic stock is now trading above OpenAI for the first time, and investors argue OpenAI would need to IPO at $1.2T to make its last primary round pencil.⁶ At the Q2 benchmark both were near $30B ARR, but Anthropic is compounding at ~10× a year versus OpenAI at 3–4× — and in Sacks's framing that gap is already basically decisive.

Sacks's strongest point is that the critics have it backwards: OpenAI should do more enterprise, not less, because consumer is a structurally worse business — maybe 3–4% convert to premium and they want a $20/month all-you-can-eat bundle, while enterprise coding tokens meter like electricity. Chamath adds that Codex is genuinely better than Claude for long-horizon, tricky tasks — a view grudgingly shared by Friedberg despite his org now being "90% Anthropic."

Compute wars and the data-center backlash[27:28]

Chamath's central claim: hyperscalers control roughly 60% of all compute, which turns capacity allocation into a weapon. When AWS Bedrock loosened Anthropic's throttles, all the "Claude is thinking less" complaints evaporated — suggesting the model's scarcity narrative and its compute budget are uncomfortably correlated.

The comedic data point is Allbirds. The sneaker company rebranded to "NewBird AI" on an alleged eight-H100 purchase and the stock ripped 450% in a week off a ~$21M base.⁷ Chamath's serious reading: public-market froth for anything GPU-adjacent is a canary — the real signal is that power, permits, and shell space are now the binding constraint, not model architecture.

Sacks traces three reasons data centers keep getting voted down: residential rate-payer fear (addressed in part by the Trump administration's March 2026 Ratepayer Protection Pledge⁸), EA/doomer groups astroturfing "data centers use all our water" campaigns, and — most ironic — Anthropic's own longtime political alliance with those same groups now threatening the build-outs Anthropic itself will need. Maine's new statewide moratorium on 20MW+ sites is the early bellwether.⁹

Friedberg's wildcard thesis is cultural, not technical: the data center is the most attackable physical manifestation of the wealth the AI boom is creating. Most Americans don't feel value from AI yet — "some medical advice on ChatGPT" is the median experience — while the data center is visible, concrete, and rich-coded.

Jason's closing riff: the industry's two most visible spokespeople are Dario Amodei (preaching doom) and Sam Altman (the subject of Ronan Farrow's recent 70,000-word New Yorker investigation¹⁵). Until a Jensen, a Dell or an Elon takes the mic on healthcare, housing and education, "AI is less popular than ICE, the Democrats, and the government of Iran."

Swalwell out, and the Democratic machine[59:23]

Eric Swalwell dropped his California governor bid and resigned from Congress after a wave of sexual-misconduct allegations; he denies them, but lost all 21 Democratic House endorsements inside 24 hours.¹⁰ Friedberg's contribution is the anecdote: when he made discreet calls in December about a rumored run, multiple unconnected people already knew the allegations — and sat on them.

Sacks's reading is structural rather than moral: California's top-two jungle primary¹³ had produced a scenario where two Republicans (Hilton and Bianco) were polling into the general, so the Democratic establishment needed to winnow the field fast. "Powers that be made the decision to lance the boil" — and the parallel is explicitly Biden 2024, where Pelosi reportedly told the sitting president "we can do this the hard way or the easy way" days before his exit.¹¹

Added colour: Pelosi was Swalwell's original 20-years-ago political patron, and protected him during the 2020 Christine "Fang Fang" Fang Chinese-spy scandal.¹⁴ The bit then drifts, as All-In bits do, into a riff on congressional trading returns — Ro Khanna apparently beating Pelosi's legendary run this year — and a jab that Buffett's returns are "bimodally distributed pre- and post-Reg FD", while Reg FD famously doesn't apply to Congress.

All-time highs, Iran, and the Trump weather vane[1:10:31]

The S&P is hitting fresh all-time highs inside week seven of the Iran conflict, which feels wrong until you unpack it. Chamath flags both the Shiller CAPE and the Buffett Indicator at extremes¹², but also the weird dispersion signal — roughly eight names are at ATHs and the rest of the market is nowhere near it, which is why every macro narrative currently finds data to support itself.

Sacks reads the rally as a peace-dividend bet: the Islamabad meeting didn't produce a signed deal, but the price action says the market treats the conflict as "on its way to being resolved." Travis's blunter take: Trump's weather vane is the S&P, and he dials policy pressure up or down to keep it trading in a tight band — the "TACO trade" reframed as a deliberate instrument rather than a bug.

The tell on the AI side is that Chamath keeps pressing for a single real example of a scaled, profitable enterprise deployment and doesn't get one. Sacks concedes he's closer to Jason here than he'd like to admit: change management at big companies is real, most transformation projects are failing per McKinsey, and the ROI is only unambiguously visible at the model layer — not yet at the application layer. Travis closes on the honest note: "the agents are just not that smart yet" — investing agents on Kalshi and Polymarket have been his own quiet disillusionment.