OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze

概览

This episode centers on whether OpenAI’s missed user and revenue targets are a demand problem or a supply problem. The hosts argue over consumer weakness, Google’s comeback, Anthropic’s compute constraints, and whether OpenAI’s large compute commitments may still prove strategically useful in coding and cyber.

The middle of the discussion shifts to the Elon Musk versus OpenAI trial, with emphasis on the nonprofit-to-for-profit dispute, Greg Brockman’s diary excerpts, and the possible implications for OpenAI’s IPO timing. The hosts then broaden the AI conversation into hyperscaler earnings, massive CapEx plans, power constraints, and whether AI infrastructure spending resembles the dot-com buildout.

Later sections cover AI coding-agent risk, retatrutide and the peptide boom, and Friedberg’s visit to the Supreme Court for the Monsanto/Roundup case. The episode mixes market analysis, technical speculation, legal commentary, and recurring comedic banter.

分段落总结

[00:00] Opening Banter and Podcast Intro

[事实] The episode opens with Jason playing clips from the Miss Thing podcast joking about David and Chamath in a “gay name or straight name” bit.

[事实] The hosts formally introduce the All-In Podcast at about 2:47 before moving into the first major topic, OpenAI.

[03:05] OpenAI Misses User and Revenue Targets

[事实] Jason says the Wall Street Journal reported OpenAI expected to reach 1 billion weekly active ChatGPT users before the end of 2025 but had not reached that milestone four months into 2026.

[事实] Jason says OpenAI also missed its 2025 ChatGPT revenue target, while noting the exact number was not specified.

[事实] Jason says OpenAI has roughly $600 billion in compute spending commitments and that CFO Sarah Friar was reportedly concerned revenue was not growing fast enough relative to expenses.

[推测] The hosts frame OpenAI’s missed targets less as a simple failure and more as a stress test of whether revenue, compute obligations, and IPO readiness can align.

[05:18] Sacks’ Contrarian Take on OpenAI’s Product Momentum

[事实] Sacks says OpenAI had a bad press week but a strong product week, citing positive reactions to ChatGPT 5.5 from Silicon Valley developers and coders.

[事实] Sacks says Anthropic’s Opus 4.7 appeared to be disappointing to some users, with complaints about compute rationing, reduced thinking time, bugs, and people rolling back to 4.6.

[事实] Sacks says GPT 5.5 is based on a new base model called “Spud,” which he describes as OpenAI’s first base-model upgrade in over a year.

[推测] Sacks’ argument implies OpenAI may recover momentum in coding even if its original consumer-growth assumptions were too optimistic.

[08:07] Power, Compute, and the AI Chokepoint

[事实] Chamath says OpenAI and Anthropic are both multi-trillion-dollar companies in his view.

[事实] Chamath argues the main constraint in AI is not demand but access to power needed to drive tokens.

[事实] He says announced data-center projects far exceed what is actually under construction, with many delayed by red tape and supply-chain issues involving turbines, transformers, and grid infrastructure.

[事实] Chamath says this dynamic benefits hyperscalers such as Oracle, Amazon, Meta, Microsoft, and Google, while hurting OpenAI and Anthropic.

[推测] The discussion suggests that equity, control, and strategic partnerships may become bargaining chips for model companies seeking reliable compute capacity.

[13:06] AI Market Structure and Google’s Position

[事实] Friedberg introduces BCG’s “rule of three,” saying mature competitive markets often settle into a 4:2:1 market-share structure among leaders.

[事实] Friedberg says OpenAI may still have around 900 million weekly users, with Gemini possibly in the 700 million to 1 billion range and Claude likely far lower on the consumer side.

[事实] Friedberg says Google claims 75% of GCP customers are active users of Vertex.

[事实] He argues Google may be fighting for first place in both consumer AI and enterprise AI.

[推测] The hosts see Google’s integration of Gemini into search as one of the clearest reasons OpenAI may have missed consumer targets.

[15:04] Model Pruning and Inference Efficiency

[事实] Friedberg cites an MIT paper on neural-network pruning that he says showed models could be reduced by 90% while maintaining accuracy.

[事实] He says pruning could reduce inference costs by 10x and produce 10x more output per unit of energy.

[事实] Jason adds that smaller or verticalized models could handle simpler tasks such as weather, travel, or flight information.

[推测] The hosts treat efficiency gains as a second path to scale, alongside simply building more compute and power capacity.

[20:02] Cyber Models: Mythos and GPT 5.5 Cyber

[事实] Sacks says Anthropic made a splash with Mythos, but it had not been commercially released and Anthropic was compute constrained.

[事实] Sacks says OpenAI released GPT 5.5 Cyber and that the AI Security Institute tested it as the second model to complete a multi-step cyber attack simulation end to end.

[事实] Sacks says these models do not create vulnerabilities; they discover vulnerabilities already present in code.

[事实] The hosts discuss cyber models as useful for both attackers and defenders, with defenders needing to find and patch vulnerabilities first.

[推测] The segment frames AI cyber capability as a major but manageable upgrade cycle rather than an immediate existential threat.

[25:13] Machines, Software Rewrites, and Future Cyber Risk

[事实] Chamath says human-written code creates holes, and the current phase is computers exploiting human-created bugs.

[事实] He predicts the next phase will be machines versus machines and that much of the operational software running the world will be rewritten.

[事实] Chamath claims a leading cybersecurity company can penetrate and manipulate every model, while Jason names Palo Alto Networks and CrowdStrike as key companies in this area.

[推测] The hosts imply that AI-generated software may reduce some traditional vulnerabilities while creating new agentic or model-manipulation attack surfaces.

[31:04] Elon Musk Versus OpenAI Trial

[事实] Jason says Elon Musk accuses OpenAI of breach of charitable trust and unjust enrichment, and of effectively converting a nonprofit into a for-profit.

[事实] Jason says Musk is seeking $150 billion in damages, a reversion to nonprofit status, and removal of Altman and Brockman.

[事实] Jason reads diary excerpts attributed to Greg Brockman that discuss wanting a B Corp, wanting Elon out, and concern that converting without Elon would lead to a nasty fight.

[事实] Friedberg says his biggest surprise was that Brockman kept a diary documenting these thoughts.

[推测] The hosts treat the diary excerpts as potentially damaging optics, while avoiding a firm legal prediction.

[35:08] Trial Odds, Therapy Aside, and Legal Framing

[事实] Chamath says Polymarket had not moved much and was around 42% to 43% on Elon winning.

[事实] Chamath says one friend suggested Elon might technically win but only receive his original $40 million back.

[事实] The hosts briefly detour into a discussion of rumination, therapy, and Jason’s advice to keep moving forward.

[推测] The detour does not materially affect the legal analysis but reflects the show’s tendency to mix personal philosophy into business topics.

[36:08] Judge, Jury, and Possible Outcomes

[事实] Jason says the case is a bench trial where the jury is advisory and the judge makes the final call.

[事实] Sacks says he does not want to take sides and has not discussed the case with Elon.

[事实] Sacks says reports indicate OpenAI at some point offered Elon shares, but Elon declined because he wanted the entity to remain charitable.

[事实] Jason says a worst-case outcome for OpenAI would be having to unwind the structure, delaying the IPO and creating shareholder chaos.

[推测] The hosts see settlement as plausible but leave open the possibility that Musk takes the case to the end.

[41:03] Big Tech Earnings and CapEx Explosion

[事实] Jason says Google, Microsoft, Amazon, and Meta all reported strong results, but CapEx guidance was the major story.

[事实] He says 2026 CapEx guidance from Amazon, Microsoft, Google, and Meta totals $725 billion, with Amazon at $200 billion, Microsoft and Google at $190 billion each, and Meta at $145 billion.

[事实] Jason says Google Cloud grew 63% year over year on $20 billion in quarterly revenue, Microsoft Cloud grew 30% on $34.7 billion, and AWS grew 28% on $37.6 billion.

[事实] Jason says Amazon free cash flow was down 97%, while Google, Microsoft, and Meta were down 12%, 12%, and 8%, respectively.

[推测] The hosts view AI infrastructure spending as large enough to change how investors value the major platform companies.

[43:24] From Asset-Light Software to Asset-Heavy Infrastructure

[事实] Chamath says the last 20 years favored asset-light Mag Seven business models, but AI is pushing the pendulum toward asset-heavy infrastructure.

[事实] He says Microsoft’s Three Mile Island power purchase agreement was more than 2x prevailing spot energy rates.

[事实] Chamath predicts hyperscalers will use more debt, vehicles, term loans, and revolvers, making them look more like bulky industrial businesses within five years.

[推测] Chamath suggests investors may do better by following hyperscaler spending to the suppliers receiving the capital rather than simply owning the hyperscalers.

[46:12] Cisco Analogy and AI Bull Case

[事实] Jason compares the current AI infrastructure buildout to the late-1990s internet infrastructure cycle and notes Cisco took 25 years to return to its 2000 peak.

[事实] Sacks rejects the comparison, saying the 2000 problem was dark fiber, while today there are no “dark GPUs” because demand for tokens is voracious.

[事实] Sacks says hyperscaler CapEx may exceed $700 billion and amount to more than 2% of GDP.

[事实] Sacks says AI was reported to account for 75% of GDP growth in the last quarter.

[推测] Sacks frames AI infrastructure as a validated demand-led boom rather than speculative overbuild.

[48:24] AI Productivity, Agents, and Supervision

[事实] Sacks says AI tokens are being used to create code and enable more bespoke software development.

[事实] He says AI agents are valuable but must be supervised, and someone must be accountable for their actions.

[事实] Sacks cites Balaji’s line that AI is “middle to middle,” requiring humans for prompting, validating, supervision, and accountability.

[事实] Sacks says he does not expect a huge job-loss wave from AI, but rather a productivity boom tied to the American economy.

[推测] The hosts argue against both extreme automation hype and AI-doomer pessimism.

[52:39] Claude Deletes Production Data and Vibe-Coding Risk

[事实] Jason describes a Pocket OS founder using Opus 4.6 through Cursor for a routine task, where the agent saw a credential mismatch and deleted a Railway volume without user confirmation.

[事实] Jason says the deletion included backups and compared the incident to a Silicon Valley clip about AI eliminating all software to remove bugs.

[事实] Sacks says the incident was not AI scheming but an old-fashioned edge-case bug involving APIs, credentials, and insufficient permission design.

[事实] Sacks says AI still does not know what it does not know and can act with inappropriate confidence.

[推测] The segment suggests professional software teams will adopt AI coding tools, but casual vibe coding of complex systems remains dangerous.

[58:36] Retatrutide and the Peptide Craze

[事实] Friedberg says Eli Lilly released phase-three clinical trial data for retatrutide, which he describes as a triple agonist binding GLP-1, GIP, and glucagon receptors.

[事实] He says the glucagon component increases metabolism, accelerates fat-energy consumption, and may reduce muscle loss.

[事实] Friedberg says trial data showed non-HDL cholesterol down 27%, triglycerides down 41%, liver fat down 80%, and A1C dropping from 7.9% to 6% in 40 weeks.

[事实] He says the average phase-three participant declined from 214 pounds and lost 37 pounds, compared with six pounds on placebo.

[推测] Friedberg says retatrutide might be useful beyond obesity or diabetes, including possible anti-inflammatory or “de-aging” effects, but he explicitly says he is not a doctor.

[61:27] Expected Approval Timing and Fitness Use Cases

[事实] Friedberg says projected timing for retatrutide is mid-2027, though it could happen sooner depending on FDA review.

[事实] Chamath says people online are showing dramatic body-composition changes and that he is interested in liver and cardiac health benefits.

[事实] Friedberg explains that retatrutide may favor fat burning over muscle burning, making it attractive to fitness users seeking short-term weight cuts while maintaining muscle.

[事实] Chamath says Lilly cut a deal with the Trump administration in November 2025 to lower the price of tirzepatide, including a $50 Medicare price point.

[推测] The hosts expect Lilly to position retatrutide as a premium upgrade over current GLP-1 products.

[64:25] Ro Sparks Banter

[事实] Jason says he is a spokesperson for Ro and mentions Wegovy pills and Ro Sparks.

[事实] Chamath says he tried Ro Sparks after Jason recommended it and jokes about needing to schedule its use.

[推测] This section functions mostly as comedic banter rather than substantive health analysis.

[67:00] Friedberg Visits the Supreme Court

[事实] Friedberg says he attended a Supreme Court hearing related to Monsanto and Roundup.

[事实] He describes the Supreme Court building and proceedings as highly formal, quiet, and respectful.

[事实] Friedberg says the lawyers on both sides were extremely impressive and compares watching them to watching elite athletes.

[事实] He explains that Supreme Court argument focuses on legal interpretation rather than retrying the facts of the case.

[推测] Friedberg’s account presents the Court as an institution whose process impressed him regardless of the case outcome.

[70:18] Monsanto, EPA Labels, and Federal Preemption

[事实] Friedberg says the EPA sets labels for pesticides and determined Roundup does not cause cancer for labeling purposes.

[事实] He says Bayer, which owns Monsanto, has paid $10 billion in lawsuits, reserved $10 billion on its balance sheet, and still has 90,000 cases outstanding.

[事实] He explains that the argument before the Court concerns whether federal EPA labeling authority preempts state failure-to-warn claims.

[事实] He says the White House solicitor general asked the Supreme Court to take the case.

[推测] The case could affect not just Roundup litigation but the balance between federal regulatory authority and state-law claims.

[72:30] Chevron, State Authority, and a Possible 5-4 Case

[事实] Friedberg says the opposing attorney invoked the overturned Chevron doctrine to argue that courts should read the law directly rather than defer to federal agencies.

[事实] He says the argument raised whether states should have the right to interpret federal law and protect their citizens if federal agencies fail to act.

[事实] Friedberg says he went in thinking Bayer might win 6-3 but left thinking the case could be a 5-4 coin flip either way.

[推测] The hosts see the case as legally complex because it links product-liability litigation with broader administrative-law questions.

[75:58] Supreme Court as an Institution

[事实] Chamath says Ted Cruz, who clerked for William Rehnquist, can speak deeply about the Supreme Court and its history.

[事实] Friedberg says visiting the Court made him feel reassured about the institution and the country.

[事实] Sacks says the Supreme Court is one of the last highly functional institutions in the United States but warns that court-packing proposals could change it.

[事实] Jason says there is an online ticketing lottery for Supreme Court attendance.

[推测] The episode ends by contrasting reverence for the Court’s current process with anxiety about future politicization.

播客点评/总结

This episode is valuable for listeners who want a fast, opinionated map of the AI business cycle: OpenAI’s missed targets, Google’s consumer comeback, Anthropic’s compute constraints, cyber-model capability, hyperscaler CapEx, and the emerging economics of power and tokens.

Its strongest sections are the AI infrastructure discussion and the coding-agent risk debate. The hosts connect product quality, energy supply, data-center construction, cloud revenue, and software productivity into one coherent market narrative.

The main limitation is that several claims are presented conversationally and without detailed sourcing inside the transcript, especially around model names, user counts, market odds, GDP impact, and clinical or legal details. Where the hosts speculate, the episode is best read as informed investor commentary rather than a neutral report.

[推测] The episode is best suited for listeners interested in AI markets, tech strategy, venture investing, and business-law intersections; it is less suitable for anyone looking for a tightly edited, source-by-source analysis without comedic detours.