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

Summary

• OpenAI missed both its 1 billion weekly-active-user target and its 2025 ChatGPT revenue target per the Wall Street Journal, with $600B in compute commitments now roughly equal to its entire secondary-market value and CFO Sarah Friar reportedly in conflict—or under natural tension—with Altman over IPO readiness — Polymarket puts an OpenAI IPO by end-2026 at just 32%, down from 60% in December. Sacks’s contrarian read: the press week was terrible but the product week was won — ChatGPT 5.5 (on new base model “Spud”) is drawing rave developer reviews while Opus 4.7 “appears to be a bust,” so Sam “may end up being right but for the wrong reason”: the compute he overcommitted for a consumer miss is now helping him catch up in coding.
• Chamath’s core frame: every miss in AI right now is supply, not demand — “the limiting resource is power.” Less than half of announced gigawatts are actually under construction and he predicts 40% of announced projects get canceled, which hurts OpenAI and Anthropic most and hands leverage to the hyperscalers; the coming negotiation is “how much equity do I have to give up… to get access to the compute.” That opens a lane for Groq and SpaceX’s excess capacity — the Cursor deal “was the appetizer,” and “he and Dario should do a deal tomorrow.”
• Hyperscaler blowout quarter: $725B of 2026 CapEx guidance from four companies (Amazon $200B, Microsoft and Google $190B each, Meta $145B), Google Cloud +63% y/y, AWS +28%, while free cash flow collapses (Amazon -97%). Chamath’s call: the Mag 7 are converting from asset-light FCF machines into levered industrials with no obvious valuation case in five years — so “follow the dollars” and buy the underpriced suppliers receiving the trillion. Sacks rejects the Cisco-2000 analogy — “there’s no dark GPUs” — and declares AI, at 75% of GDP growth, now “synonymous with the American economy.”
• Cyber is the new front: GPT 5.5 Cyber matched Mythos on the AI Security Institute’s end-to-end attack simulation and, unlike compute-constrained Anthropic, looks commercially servable — possibly the first cyber model defenders actually get to use. Sacks says the panic is overdone: models “don’t create the vulnerabilities, they just discover them,” expect a one-time upgrade cycle then equilibrium; Chinese models arrive in ~6 months. Chamath goes further — all the world’s operational software gets rewritten in 5-6 years — and drops that the best cybersecurity company in the world (he indicates Nikesh Arora’s) has “penetrated and can essentially manipulate every model.”
• Musk v. Altman: Greg Brockman’s diary — “The true answer is that we want Elon out… this story will correctly be that we weren’t honest with him” — is, per Sacks, “not just journal maxing, it’s discovery maxing.” It’s a bench trial with the jury only advisory, yet Polymarket hasn’t budged from 42-43% Elon wins despite the discovery — one theory being it’s pricing a technical win that merely refunds his $40M. Sacks’s added fact: OpenAI apparently did offer Elon shares as a make-right and he refused on principle, wanting it to remain charitable.
• The vibe-coding cautionary tale: an agent running Opus 4.6 through Cursor deleted a founder’s production Railway volume — including backups — over a credential mismatch. Sacks: not AI scheming, just “old-fashioned bugs at an edge case,” but the systemic problem is real — “AI still doesn’t know what it doesn’t know,” so agents must be supervised, middle-to-middle (Balaji), and eliminating all software developers was “the peak of inflated expectations.”
• Retatrutide’s phase 3 is the peptide story: 37 lbs lost vs 6 on placebo in 40 weeks, liver fat down 80%, triglycerides -41%, A1C from 7.9% to 6%, plus anti-inflammatory and even de-aging chatter. FDA approval projected mid-2027, possibly sooner; Lilly’s portfolio logic is $50-Medicare tirzepatide as the entry product and Reta as the premium upgrade — “people will pay for this.”
• Sleeper legal risk repriced: the Bayer/Monsanto Supreme Court case pits EPA federal preemption against state failure-to-warn laws with $10B already paid, $10B reserved, and 90,000 cases outstanding. Friedberg walked in expecting 6-3 for Bayer and walked out calling it “a 50/50 coin flip, 5-4 either way” after opposing counsel turned the Chevron-doctrine overturn against the company — and if states win the right to ignore federal regulators, that logic extends to the FDA and USDA.

Deep dive

1. OpenAI missed its numbers — Sacks argues it still won the product war

• The setup per the Journal: OpenAI expected 1 billion weekly active users before end of 2025, missed, and still hasn’t hit it four months into 2026; it also missed its 2025 ChatGPT revenue target, sitting at a $20-30B run rate with “a little bit of accounting nuance yet to be worked out.” Jason’s framing of the stakes: $600B in compute commitments roughly equals what OpenAI trades for on secondary markets — “the entire value of the OpenAI enterprise equals their spend commitments.”
• Inside the company, CFO Sarah Friar is reportedly worried revenue isn’t growing fast enough to keep up with expense and doesn’t think OpenAI is ready for public reporting standards; Jason framed this as possible conflict—or natural tension—with Altman, who wants to move faster toward an IPO later this year. Polymarket now prices a 32% chance OpenAI goes public by end of 2026, down from 60% in December.
• Sacks’s contrarian take: bad press week, great product week. ChatGPT 5.5 reviews from Silicon Valley developers have been “really strong,” while Opus 4.7 “appears to be a bust” — users rolling back to 4.6, complaints of compute rationing and reduced thinking time, bugs, and servers going down. 5.5 rides a new base model called “Spud,” the first base-model upgrade in over a year, paving the way for further gains.
• His deeper thesis: Sam “may end up being right but for the wrong reason.” The compute commitments were sized to consumer forecasts that missed — but coding became the all-important sector, and OpenAI now has more compute than token-constrained Anthropic, whose scarcity is limiting Mythos serving and forcing compute-gating on 4.7 (“a prudent business decision,” he stresses, not a knock on Dario). “Developer mojo is shifting”; Codex is taking share of coding tokens. Two weeks ago Anthropic growing 10X against OpenAI’s 3X looked like a runaway — the companies are now constantly one-upping each other.

2. Power, not demand, is the single choke point

• Chamath waves off the misses entirely: OpenAI and Anthropic are both “multi-trillion dollar” companies, and what’s happening is “a complete misunderstanding” — misses are “entirely one hundred percent due to the supply of the power necessary to generate the output token,” not demand. Anthropic just economically induced Amazon into direct capacity so users needn’t route through Bedrock, and is doing differentiated deals with economic participation on top of its Google capacity.
• The bottleneck is compounding: beyond power itself, backlogs are building in recips, nat gas turbines, and now transformers and tactical grid infrastructure. Of announced gigawatts, “less than half of it is actually being built” — stuck in red tape and supply chain delays — so “there’s no credible strategy to turn any of this stuff on.”
• His prediction: 40% of announced projects get canceled, partly because the industry “has done such a poor job of creating a good positive halo around AI” — matching the 40% cancellation rate of the last four years. This hurts Anthropic and OpenAI most and benefits Oracle, Amazon, Meta, Microsoft, and Google, setting up the trade: “How much equity do I have to give up, how much control… to get access to the compute, versus how badly will I miss my growth forecasts if I don’t?”
• The Elon angle: Groq and SpaceX sit on a ton of excess capacity, so “the Cursor deal was the appetizer” — and if the models catch up in quality, “he and Dario should do a deal tomorrow.” Not OpenAI, “because of the baggage.”

3. Rule of three: Google may be in first place in both markets

• Friedberg’s structural lens is BCG’s rule of three — mature markets settle at a 4:2:1 share ratio. Consumer AI is trending toward a ChatGPT/Google fight for the top two: OpenAI at ~900M weekly users even after the miss, Gemini at 700-750M and “pretty neck and neck,” Claude “probably sub a hundred million” — with Elon possibly emerging, enabled by compute.
• Enterprise may already belong to Google: it claims 75% of GCP customers are active Vertex users, “probably why Google stock has absolutely ripped.” Sacks adds that if there’s a single reason OpenAI missed consumer targets, it’s that Google took meaningful share — Sergey came out of retirement, Gemini went to the top of search, and search revenue surged anyway: “they’ve done it brilliantly.”
• Friedberg’s efficiency kicker: an MIT paper on pruning shows you can cut network size 90% with no accuracy loss, reducing inference cost 10X per energy unit. Like search, a few queries dominate volume — weather, scores, stock prices — so a “macro model” dynamically calling many small models can serve them cheaply. And these submodels won’t map to human categories: “we don’t know why these models work the way they do when they get broken down.” Early innings of efficiency — another opening for Elon to reinvent the stack.

4. Cyber goes frontier: GPT 5.5 Cyber matches Mythos, and the panic is overdone

• OpenAI released GPT 5.5 Cyber, which per AI Security Institute testing became the second model to complete a multi-step cyber attack simulation end to end — Mythos-level capability, but commercially ready because OpenAI has the compute Anthropic lacks. Sacks: “5.5 might be the first cyber model that cyber defenders actually get to use.” All frontier models reach Mythos-level cyber within ~6 months, Chinese ones included — though he pegs DeepSeek 4 at “80, 85%” of the American frontier, not at it.
• Sacks’s anti-doom case: Mythos “doesn’t create the vulnerabilities. It just discovers them. The bugs were already in the code.” There are only a few thousand truly elite hackers, and a model scales their craft — so arm the white hats first, expect a one-time upgrade cycle that flushes out dormant bugs, then a new equilibrium between AI offense and AI defense “as long as everyone does what they’re supposed to do.”
• Chamath’s longer arc: we’ve gone from humans exploiting human-written holes to computers exploiting them; next is machines versus machines. Within five or six years “all the operational software that runs the world will get rewritten” — machine-written, “more impregnable” — after which the threat morphs into machine-on-machine injection into agentic loops. Then his bomb, hedged as “I’m not even sure if I’m allowed to say this”: probably the best cybersecurity company in the world — he indicates Nikesh Arora’s — has “penetrated and can essentially manipulate every model.”
• The demand signal: George Kurtz told Jason there’s “a line out the door” for this. Jason’s arithmetic — ~5M security experts on Earth; at roughly $100 per security expert you mint another 5M, or 50M, agents that never sleep — and his analogy: like the murder rate, attack vectors shrink until the weak link is human factors, “the secretary who puts her Post-It note with the password.”

5. Musk v. Altman: the diary is the smoking gun, but Polymarket hasn’t moved

• Elon is suing for breach of charitable trust and unjust enrichment — seeking $150B in damages, reversion to a nonprofit, and removal of Altman and Brockman — framing it as bigger than OpenAI: “If we make it okay to loot a charity, the entire foundation of charitable giving in America will be destroyed.”
• The discovery centerpiece is Greg Brockman’s diary: “Conclusion: We truly want the B corp. The true answer is that we want Elon out… In the end, it’s still about wanting a for-profit, just without him.” Friedberg’s incredulity — “let me write about it, let me record it, and by the way, let me never delete it… I don’t get it” — and Sacks’s line: “It’s not just journal maxing, it’s discovery maxing. It’s smoking gun maxing.” Cue The Wire: “Is you taking notes on a criminal conspiracy?”
• Mechanics matter: it’s a bench trial with the jury advisory only — Judge Rogers, a 61-year-old Obama appointee who oversaw Epic v. Apple (ruling for Apple with caveats), makes the final call and sets damages. Chamath finds it “weird” Polymarket sits unmoved at 42-43% Elon wins despite the published discovery; a friend’s theory is the market is front-running a technical win where Elon is simply credited back the $40 million.
• Sacks won’t take sides — last time he weighed in on Elon litigation he got deposed for six hours — but surfaces the make-right that wasn’t: OpenAI apparently did offer Elon shares, and he turned them down because he didn’t want the charity converted at all, “a principled view” per the reports. Jason handicaps a settlement as most likely — worst case for OpenAI is unraveling the structure and delaying the IPO; if Elon’s first $40-50M check holds up, he’s “due 10, 20, 30% of the company after dilution.”

6. $725B of CapEx and the end of the free-cash-flow machines

• The earnings numbers: $725B in 2026 CapEx guidance from just four companies — Amazon $200B, Microsoft $190B, Google $190B, Meta $145B — and adding Grok, OpenAI, and whatever the new Apple CEO commits, Jason sees “a trillion dollars in build out over the next year.” Cloud growth validated the spend: Google Cloud +63% y/y on $20B in quarterly revenue, Microsoft Cloud +30% on $34.7B, AWS +28% on $37.6B.
• The cost: free cash flow is being sacrificed — Amazon’s FCF down 97%, Google and Microsoft down 12% each, Meta down 8% — along with buybacks and dividends.
• Chamath calls it a structural shift in capital markets: twenty years of asset-light Mag 7 absorbing all investment dollars, and now “the pendulum is swinging violently in the other direction.” His example: Microsoft’s Three Mile Island forward purchase agreement was more than 2x prevailing spot for energy — and that’s a small slice of its needs. These companies will get levered — term loans, revolvers, financial engineering — and “look like this big, bulky industrial business in five years. I’m not sure that there’s a good valuation case to be made at that point.” The trade: “just follow the dollars… buy those companies because those companies are already underpriced.”
• Jason raises the Cisco chart — 25 years to reclaim the 2000 peak. Sacks rejects the analogy outright: 2000 was dark fiber, infrastructure built and unused, whereas today “there’s no dark GPUs” (crediting Brad Gerstner) — voracious token demand is pulling the investment forward, and “the bull thesis for AI just got validated in a single afternoon.”

7. Sacks: AI is now synonymous with the American economy

• The macro case: hyperscaler CapEx going from $350B last year to over $700B — more than 2% of GDP — and AI contributing 75% of GDP growth last quarter. “And this is just the CapEx part… the building of the factories,” before counting the economic impact of the tokens themselves.
• Downstream, he sees an explosion of bespoke software: “every business that now wants to get code will be able to get code for the first time” — they couldn’t even hire the engineers before. His challenge to the doomers: “what do they want the American economy to do, just stop?… when you talk about stopping AI progress, what you’re really doing is stopping the American economy” — no growth means no money for social programs, debt, or defense. AI may poll badly, but a strong economy doesn’t, “and I believe that those things are now synonymous.”
• Credit where he assigns it: “President Trump set the table on this” — with Jason suggesting it was arranged with some good advice, maybe. Worth keeping Friedberg’s correction to the highway analogy: the interstate system was federally funded with no competition — inflation-adjusted, “the most expensive project in US history.”

8. An agent deleted production in nine seconds — supervision is the job

• The cautionary tale: the founder of Pocket OS, vibe coding with Opus 4.6 through Cursor’s top tier, watched an agent hit a credential mismatch on a routine task and “fix” it by deleting a Railway volume without user confirmation — including the backups. The pod runs the Silicon Valley “Son of Anton” clip: the most efficient way to get rid of all the bugs was to get rid of all the software — “technically and statistically correct.”
• Sacks pushes back on the “AI scheming” reading: this wasn’t paperclip-style misalignment but “old-fashioned bugs occurring at an edge case” — an API not designed for permissioned use, a credential left lying around. The systemic issue is different: “AI still doesn’t know what it doesn’t know.” A human pauses before deleting a production database; the model doesn’t, and “the longer the time horizon for a task, the more likely it is to go off the rails.”
• Hence his labor take: agents are valuable but must be supervised — AI is “not end to end, it’s middle to middle” (Balaji) — someone has to prompt, validate, and be accountable, and it won’t be a CEO with thousands of agent reports. Eliminating all software developers “was the peak of inflated expectations.” Aaron Levie’s endorsed framing via Matthew Iglesias: “I want professionally managed software companies to use AI coding assistants to make more, better, cheaper software products that they sell to me for money.”
• Jason’s darker prediction: some public-company “goofball” will vibe-code into a disaster and “torch the enterprise value. It’s gonna be glorious to watch.”

9. Retatrutide: Lilly’s premium peptide and the upgrade-pricing playbook

• The mechanism, per Friedberg: unlike dual-agonist tirzepatide, Retatrutide adds a third receptor — glucagon — which raises cellular metabolism and favors fat burning over muscle breakdown, countering the starvation-mode muscle loss of pure appetite suppression. That’s why the fitness crowd wants short low-dose cycles to “cut weight and maintain muscle and get ripped.”
• Phase 3 data driving the craze: average participant went from 214 lbs to 37 lbs lost vs 6 on placebo in 40 weeks; non-HDL cholesterol -27%, triglycerides -41%, liver fat -80%, A1C from 7.9% to 6% (“literally a life-saving product” for diabetics); 20% of people felt more nauseous than those on placebo. Separate studies show reduced inflammatory signaling — feeding talk of Reta as “kind of a de-aging drug.” Trials ran at 12mg, but Friedberg suspects 2mg could retain much of the benefit.
• Timing and pricing: approval projected mid-2027, though “could happen sooner” — the data’s in. The portfolio logic: Lilly’s November 2025 deal with the Trump administration put tirzepatide around $50 on Medicare, making it the Honda while Reta becomes “the Mercedes… the Model S Plaid” — “if I’m Lilly looking at this data, I’m like, my God, people will pay for this.”
• Chamath is openly in: “I want it… my liver health is important to me, my cardiac health ‘cause I’m South Asian, and it just looks like a wonder drug” — citing X anecdotes of 65-year-olds going “from a dad bod to looking like an incredibly ripped athlete in weeks.”

10. Friedberg at the Supreme Court: Bayer’s $20B question became a coin flip

• The case: whether EPA federal preemption under FIFRA — the EPA writes the pesticide label and determined Roundup doesn’t cause cancer — overrides state failure-to-warn laws like California’s, under which plaintiffs’ lawyers have won repeatedly. The stakes: Bayer has paid out $10B, reserved another $10B, and faces 90,000 outstanding cases. The White House solicitor general asked the Court to take it.
• The strategic flip Friedberg witnessed: going in, “we were like, oh, 6-3 Bayer is gonna win.” Then opposing counsel — “literally like watching LeBron James” — invoked the Chevron doctrine overturn: if federal agencies no longer get deference, why must states defer to the EPA’s label? Exit read: “a 50/50 coin flip, 5-4 either way.” Ketanji Brown Jackson pressed the other direction: if the EPA learns of a cancer risk after issuing the label and doesn’t act, “shouldn’t the states have a right to protect their people?”
• The wider implication he flags: if states can reinterpret federal law past federal regulators, “it opens up a whole new can of worms” — the same logic reaches the FDA, the USDA, “and on and on.”
• The institutional coda: Friedberg found the Court “one of the most amazing experiences I’ve ever had” — two one-hour cases a day, 30 minutes a side, arguing only questions of law. Sacks’s warning: “one of the last highly functional institutions in the United States… enjoy it while it lasts” — citing James Carville’s “when we get power, we’re packing the court” to 13 justices. Favorability: a 35% low in 2024 Gallup, back to 44-50% now.