OpenAI's Identity Crisis, Datacenter Wars, Market Up on Iran News, Mamdani's First Tax, Swalwell Out
Summary
A proposed annual NYC pied-à-terre levy—speculated at 3.9% on second homes above $5 million—would target the market’s most mobile buyers and could suppress new construction. Sacks calculates that interest and inflation could make a $10 million unit effectively require a $20 million breakeven after 10 or 11 years. Friedberg said the measure may also affect homes rented to people for whom New York is not a primary residence. The counterexample is Austin, where permissive building accompanied three straight years of falling rents and housing prices despite rising migration.
OpenAI’s strategic risk is not weak technology but an enterprise growth gap that could compound into an insurmountable Anthropic lead. Chamath rates Codex above Claude for complex, long-horizon coding, but Sacks puts OpenAI’s annual growth at 3–4X versus roughly 10X for Anthropic. Enterprise code tokens scale “like electricity,” while consumer monetization is constrained by perhaps 3–4% premium conversion and expectations of a $20 all-you-can-eat plan.
The frontier-model contest is becoming an infrastructure war because physical compute limits may arrive before demand limits. Colossus is described as expanding to 555,000 GPUs across three buildings with $18 billion invested, versus Meta’s planned 150,000-GPU Prometheus cluster in 2026. Chamath argued that efficiency and contribution profit will eventually outstrip subsidy: labs need usage revenue to fund capacity, not endless mega-rounds.
Compute scarcity looks investable, but power, permits, and local politics are becoming the binding variables. The panel cited roughly 100 contested data centers representing $162 billion, with about 40 potentially canceled, while one town allegedly replaced half its board after approving a $6 billion project. Jason’s darker framing is that the data center has become “the temple of the wealthy,” a physical symbol of gains consumers do not yet feel.
Allbirds’ 450% AI-pivot rally is both a ZIRP postmortem and a sign that markets will capitalize almost any credible compute narrative. After raising $350 million in its 2021 IPO, Jason described the company as Newbird AI and cited a $50 million convertible note alongside the joking claim that it bought eight H100s. The stock reached $14. Chamath explained the original bubble through COVID-era ZIRP and investors extrapolating one year of growth two or three years forward.
Eric Swalwell’s exit was framed less as a resolved misconduct case than as an example of political information being timed and weaponized. Friedberg said several sources described alleged conduct months earlier, which he initially dismissed as rumor because nobody had acted; all allegations remain unproven. Sacks speculated that Democratic insiders chose to “lance the boil” before California’s jungle primary could produce two Republicans in the runoff.
US equities are trading as though the Iran war is nearing resolution even while classic valuation measures flash caution. Sacks said the market recovered all war-related losses by Tuesday and made fresh highs Wednesday and Thursday; Friedberg called the stock market Trump’s “weather vane.” Yet the Shiller P/E and Buffett indicator were described as near records, leaving Chamath risk-off and eager for SpaceX and frontier-lab IPO liquidity.
AI’s model-layer economics are finally visible, but the panel split over whether enterprise application profits justify current valuations. Jason sees properly deployed AI making top employees 10–30X more productive; Chamath answered that he has not seen a “tsunami of more revenue and more profit” or scaled enterprise proof. Travis’s reconciliation: founder-led technology companies are shipping faster, but agents remain tasteless, easily lost, and dependent on humans—“AGI is not here.”
Deep dive
1. The pied-à-terre tax could destroy the marginal buyer it targets
Jason stressed that 3.9% was speculation, not a finalized rate, but the proposal would reportedly apply annually to second homes above $5 million. Sacks’s arithmetic was blunt: with interest and inflation, a $10 million unit could require something like a $20 million purchase-price breakeven after 10 or 11 years. “The math doesn’t work anymore.”
Sacks’s mechanism is elasticity: people choosing a second or third home can buy anywhere, so they are precisely the buyers most able to leave. Friedberg added that the measure could also affect a home rented to someone for whom New York is a second residence, even if the owner does not occupy it. Lower prices might superficially improve affordability, but the panel’s rebuttal was that fewer price-insensitive buyers mean fewer projects pencil—and a luxury penthouse does not magically become low-income housing.
Friedberg argued that absentee owners already pay property taxes while consuming few city services, making them unusually profitable residents. A “whale like Ken Griffin” overpaying per square foot for a top floor may subsidize an entire development; London’s high-end collapse and the migration of non-dom wealth toward Zurich, Lugano, and Milan were offered as warnings.
The sharpest disagreement concerned the mayor’s video outside Griffin’s known property. One speaker argued that a widely marketed unit with a public owner was not meaningfully doxxed; Jason called it a dangerous “dog whistle,” particularly after a Molotov cocktail and bullet reportedly struck Sam Altman’s home, and urged applying the same standard if ideological roles were reversed.
2. OpenAI’s consumer franchise does not remove the enterprise imperative
OpenAI revenue chief Denise Dresser’s leaked memo disputed Anthropic’s roughly $30 billion run rate, alleging about $8 billion was attributable to revenue-sharing and channel accounting. She characterized Anthropic’s positioning as “fear, restriction, and the idea that a small group of elites should control AI,” while directing OpenAI toward business customers and the agent-platform layer.
Anonymous investors offered the opposite complaint: ChatGPT reportedly has 1 billion users and is growing 50–100% annually, so “what are you doing talking about enterprise and code?” The FT questioned an $850 billion valuation; Anthropic reportedly traded higher in secondary markets, no buyers were available at OpenAI’s latest price, and one investor said the round required a $1.2 trillion IPO to make sense.
Chamath’s product-level view complicates the identity-crisis narrative. His team finds Claude’s model ensemble more reliable for ordinary work, but Codex generally better for “very tricky,” complex, long-horizon coding. He sketches perhaps $3–4 trillion of eventual consumer value and another $2–3 trillion from enterprise, supporting a theoretical $7–9 trillion company “in the fullness of time, not tomorrow.”
His operating prescription is structural separation: let the consumer organization “double down and crush consumer,” isolate the enterprise team, and prevent context switching between them. Sacks agreed that OpenAI has been unfocused, but called the claim that it should avoid enterprise “totally misguided”—enterprise coding is exactly where revenue scales fastest.
3. Anthropic’s 10X trajectory matters more than accounting symmetry
Sacks accepted OpenAI’s apples-to-apples correction: Anthropic may be roughly 20% smaller when channel-partner revenue is excluded. But OpenAI has been growing about 3–4X annually while Anthropic has grown around 10X—roughly $1 billion to $10 billion of ARR last year, about $30 billion by the end of Q1, and potentially $80–100 billion by year-end on its current trajectory.
The causal difference is monetization. Businesses pay for coding tokens on a meter—the more useful work they consume, the more they spend—whereas consumers expect a $20 monthly all-you-can-eat subscription, with perhaps only 3–4% converting to premium. If one competitor takes a year to 10X and another takes two, Sacks argued, “it’s obvious which one’s going to win.”
Travis translated that gap into an Uber-style flywheel: customers generate tokens, revenue funds compute, and scale may improve reinforcement learning and the product. If a similarly sized rival is growing two, three, or five times faster, “I’d be worried,” because network effects around compute and customer volume can make today’s lead self-reinforcing.
Friedberg saw the same momentum operationally: his organization moved from heavy Cursor and Gemini use to roughly 90% Anthropic in six months, while Anthropic’s release cadence appeared “head and shoulders above everyone else.” Chamath warned that capital can temporarily buy scale, but a competitor funding expansion through revenue and contribution profit becomes “a very scary machine.”
4. Frontier labs now need owned compute, not rented abundance
Early-stage labs can rent from Amazon, Google, or Microsoft, but Chamath argued that hyperscaler dependence becomes a strategic mistake at frontier scale. With hyperscalers said to control 60% of compute, constraining independent labs can buy Google, Microsoft, or Meta time to catch up while forcing OpenAI and Anthropic into capital-intensive infrastructure ownership.
His analogy was Friendster: it was once “the cat’s meow,” but latency and availability created an opening for MySpace, then Facebook. Even a superior model can lose through the same “Friendster effect” if users cannot access enough inference capacity; product adoption then hits a wall for reasons unrelated to model quality.
The cluster race illustrates the stakes. Elon’s Colossus was described as expanding to 555,000 GPUs across three buildings with $18 billion invested, while Meta’s Prometheus was planned for 150,000 GPUs in 2026. Elon’s Cursor deal suggested another strategy: overbuild, privilege xAI’s models, and rent surplus capacity to outside model developers.
Sacks tied recent model behavior to scarcity: users reportedly complained that Claude was thinking about two-thirds less. A tweet then claimed that Opus 4.7 had replaced Opus 4.6 and restored the thinking, possibly at a higher price. Mythos may have cost 10–20X as much per token as Opus; withholding it could both preserve compute and create scarcity marketing, though Sacks preserved the altruistic explanation that code-base owners needed time to patch newly exposed vulnerabilities.
5. Data-center opposition is becoming a balance-sheet constraint
Chamath cited a town where half the board was voted out after approving a $6 billion data center, plus a Maine bill described as banning new builds. About 100 projects worth $162 billion were said to be contested; roughly 40 of every 100 might be canceled, and the transcript said that cited figure had already more than doubled from the prior year.
Sacks separated legitimate grid concerns from broader opposition. Some “wildcatter” developers sought permits without power plans, creating real fear that residential rates would rise; the administration’s ratepayer-protection pledge therefore required hyperscalers to bring their own generation, remain power-neutral to the grid, and potentially return energy outside peak use. The shorthand became “BYOE”—bring your own energy.
Sacks also argued that behind-the-meter power does not change a regulated utility’s incentive to expand its rate base. If a utility can earn roughly 10% on approved investment, a $10 billion line-burial project creates a powerful motive to keep spending and raising prices. The pledge helps individual projects, but not the monopoly’s underlying business model.
Jason argued that economics alone misses the populist symbolism: data centers are “the temple of the wealthy,” while consumers see limited direct benefit from AI. Sacks countered that construction supports tens of thousands of jobs and wages 25–30% higher for electricians, carpenters, concrete crews, and others; Jason replied that those benefits are less permanent than fab employment.
6. Doomer politics may have salted Anthropic’s own infrastructure path
Sacks described a second opposition channel: well-funded AI-doomer groups allegedly discovered that water-use and local-cost claims mobilized residents more effectively than Terminator scenarios. “We have to meet people where they are” was his paraphrase of that strategy; he called the water-use claim untrue and parts of the NIMBY campaign astroturfed.
His more pointed claim was that Anthropic politically allied with those groups while relying on third-party hyperscalers, perhaps believing data-center resistance would mostly hurt OpenAI and xAI. If Anthropic has now exhausted rentable capacity, that strategy has backfired: it must build, and Sacks expects “certain kinds of data centers” to become acceptable when they serve the effective-altruist mission.
Jason offered the consumer-resistance mechanism: the average person still does not see AI materially improving daily life. He added that the industry’s public story is dominated by Dario Amodei warning of hacks and job losses and by deeply negative characterizations of Sam Altman, rather than tangible promises around healthcare, housing, and education.
Chamath’s geographic answer was that data moves “at the speed of light,” so bans will redirect capacity toward Texas, Iceland, space, or other jurisdictions, with some latency cost. Sacks argued that allied, energy-rich Gulf states should also host American technology; he said Anthropic opposed those projects, while Iranian threats against them, in his framing, underscored their status as American-linked strategic assets.
7. Allbirds’ AI rally exposes both compute scarcity and ZIRP memory loss
Allbirds raised $350 million in its 2021 IPO. Jason later described the pivot as Newbird AI and cited a $50 million convertible note alongside the joking claim that it bought eight H100s. The stock reached $14 and rose about 450% in one week—an echo of companies adding “.com” in the late 1990s.
Sacks diagnosed the original mistake as treating physical businesses like software: investors ignored cost of goods and gross margins even though sneakers lack software’s near-zero incremental delivery cost. Chamath located it more specifically in COVID-era ZIRP, when investors extrapolated one year of rapid growth two or three years forward and paid the imagined future valuation immediately.
Chamath nevertheless saw signal inside the absurdity. A new neoscaler reportedly received a $1 billion Jane Street investment plus a $6 billion compute deal; Bloom Energy had surged because on-site natural-gas generation could obtain clean-air permits faster than grid connections. Scarcity now covers power, entitled land, and the physical data-center shell.
Bird supplied the regulatory warning. After it had, as Sacks’s example, reached roughly 80% share in some cities, officials sometimes divided a fixed scooter allocation equally among four operators, erasing marketplace network effects and unit economics. Jason and Sacks warned that cities could repeat this with autonomous vehicles—caps and operator quotas would recreate medallions, raise prices, and “delete the market.”
8. Swalwell’s fall exposed the value of politically timed information
Friedberg said several independent contacts told him in December and January about alleged inappropriate material involving Swalwell and employees. He dismissed it: if multiple people knew and nothing had surfaced, he assumed it was opposition rumor-mongering. When the information later emerged together, the striking fact was not merely its content but how long knowledgeable people had withheld it.
The panel repeatedly preserved the legal hedge: these remain allegations, nothing had been proven, and Swalwell deserves his day in court. Friedberg’s question was institutional—why neither intermediaries nor alleged victims acted earlier, and why multiple accounts appeared at the same moment in what looked to him like a deliberate, coordinated release.
Sacks’s theory, explicitly speculation, centered on California’s jungle primary: two Republicans were each polling around 14–15%, while Democrats were fragmented. Insiders may have wanted to “lance the boil” before a general-election matchup, he argued, comparing the pressure campaign to Biden’s abrupt withdrawal after Pelosi reportedly warned that matters could proceed “the hard way or the easy way.”
9. Congressional information advantages sharpen the valuation debate
The conversation pivoted from political control to political trading: Ro Khanna was said to have traded $600 million of stock, prompting jokes that he traded more frequently than Citadel Securities. Chamath’s serious point was that Reg FD does not apply to members of Congress; Jason added that information learned in committees or secure briefings could intersect uncomfortably with real-time trading.
Chamath framed Buffett’s history similarly but carefully: he called Buffett “the goat of goats,” while observing that his returns were distributed very differently before and after Reg FD. The claim was not that respect was unwarranted, but that broad disclosure rules materially reduced the informational edge available to investors subject to them.
Berkshire’s roughly $300 billion cash position then became a market signal. Chamath said both the Shiller valuation measure and Buffett indicator—total US equity value divided by GDP—were near all-time highs, while only eight or nine companies were making highs. That dispersion makes the index difficult to read, but Berkshire’s refusal to deploy cash suggests, in Chamath’s view, that opportunity remains scarce.
10. Markets are pricing an Iran resolution while valuation signals conflict
Sacks interpreted the post-Islamabad rally straightforwardly: investors believe the war is moving toward resolution after the president said military objectives were nearly complete. He never expected two countries with almost 50 years of hostile relations to settle everything in 24 hours, and expressly disclaimed any administration knowledge beyond public statements.
By his account, the market had recovered all losses since the war began by Tuesday, set a new high Wednesday, and was reaching fresh highs Thursday. Calling equities “the ultimate prediction market,” he concluded that they were pricing the conflict as an excursion approaching its end, even without a signed Islamabad deal.
Friedberg’s higher-altitude model is that the stock market is Trump’s “weather vane.” He said Trump moves in the policy space and does not let the S&P 500 fall too far, explaining a relatively tight index band despite elevated anxiety and VIX.
Chamath found data for either bias: historically, a roughly 5% gain in the first half of April corresponded to an average near-32% rise during the rest of the year, yet valuation measures remained extreme. He is therefore personally risk-off and wants SpaceX plus either Anthropic or OpenAI to IPO quickly so he can delever and “get some chips off the table.”
11. Model-layer ROI is real; enterprise transformation remains unproven
Jason’s bull case is a productivity discontinuity: Office might have improved efficiency 30% and the internet 50%, whereas top workers using AI correctly can become 10, 20, or 30 times more productive. Perhaps 10–20% already know how; in his portfolio, Micro1 was using AI to identify and recruit data contributors, while TaxGPT reportedly served 6–7% of accountants.
Chamath’s honest non-answer: “It has not translated into a tsunami of more revenue and more profit for me yet.” Edge startups can grow quickly, but he wants scaled, profitable deployments inside complex enterprises. “If you can’t prove that this works in the big-time, prime-time, big-league use cases, it’s a toy.”
Sacks sided somewhat with Jason: large-enterprise transformations often fail because change management is hard, but bottom-up coding activity has reached a new level. The missing model-layer ROI once used to justify calling AI a bubble is now visible in exponential coding revenue; the unresolved question has simply moved upward to application-layer profits.
Travis supplied the dividing line. Founder-led public technology companies report much faster feature development after making their cultures explicitly pro-AI, so “this stuff’s real and it’s not just hype.” Legacy enterprises face undocumented processes, middle managers, technocrats, bureaucrats, and human resistance—the “big boy” problem is change management, not model availability.
12. Agents multiply capable humans but still lack taste and judgment
Travis cautioned against equating faster software delivery with AGI. The best agents remain “not that smart yet”: they handle repetitive work, but lose the forest, lack taste, and require a human in the loop. Agentic productivity is real without supporting claims that autonomous general intelligence has arrived.
His investing side project supplied the episode’s best specimen. Even after extensive setup, the agents had to be taught that “if you want to make money investing, you can’t be on both sides of the same bet.” Sophisticated tooling did not prevent elementary strategic incoherence.
The panel’s closing synthesis was therefore conditional: organizations that constrain agents, supply context, and manage people can ship materially faster; unmanaged or poorly controlled token spending could instead produce “vibe-coded slop” and add 30–50% operating expense. “AGI is not here,” but neither is the technology merely a toy.