OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze
In a Nutshell
OpenAI missed ChatGPT user and revenue targets amid compute shortages and Elon's lawsuit alleging breach of charitable trust, but GPT-5.5 excels in coding and cyber defense, outpacing Anthropic's buggy Opus while hyperscalers like Google, Microsoft, Amazon, and Meta dominate with $725B 2026 capex for AI infrastructure. AI drives massive productivity via tokens and agents (requiring supervision to avoid disasters like database deletions), cyber evolves to machine-vs-machine equilibrium, and big tech earnings signal a shift from free cash flow to asset-heavy builds. Retatrutide's phase 3 data shows 37lb weight loss with metabolic benefits, fueling peptide hype, as Bayer fights Supreme Court case on Roundup labeling preemption.
These notes were generated by AI and may contain inaccuracies.
Jason shares clip from Miss Thing podcast's 'Gay Name Straight Name' bit. For David: straight name but has perfect body like Michelangelo's David. For Chamath: gay name, Italian sweater at gay poker nights, talks about wine, gym selfies.
Gay Name or Straight Name. David. David to me is straight. Okay. But he has my perfect body... Like the Michelangelo's David, the male ideal. It's like incredible body, kind of small.
Gay Name or Straight Name. Chamath... gay. I feel like Italian. Italian sweater... wears it to like poker night... gay poker nights... Always talking about wine... swishing... guy at the gym taking off his shirt, taking selfies.
Shoutout to Miss Thing Podcast. Open sourced 'let your winners ride, I'm going all in' to fans.
Hosts: David Friedberg, Chamath Palihapitiya, David Sacks, Jason Calacanis (JCal).
OpenAI missed 1 billion weekly active users (1% of world) target for end of 2025; still not hit 4 months into 2026 (WSJ). Missed 2025 ChatGPT revenue target; current run rate $20-30B (accounting nuances).
$600B valuation equals $10B spending commitments for compute. CFO Sarah Friar worried revenue not growing fast enough; OpenAI wants 2026 IPO but Friar says not ready for public reporting. Tension with Altman; joint statement released.
Sacks' Contrarian Take
Bad week in press: missed user/revenue targets, data center commitments questioned, Elon lawsuit. But product strong: released ChatGPT 5.5, strong reviews from developers/coders. Anthropic Opus 4.7 bust: complaints, rolling back to 4.6, rationing compute, reducing thinking time, bugs.
GPT 5.5 based on new base model Spud (first upgrade in over a year), paves way for improvements. Developer mojo shifting on X from Opus to GPT 5.5 for coding.
Sam right for wrong reason: missed consumer targets but enterprise/coding booming. OpenAI has more compute than Anthropic (token constrained, compute gating on Opus 4.7). Coding key sector now.
Polymarket: 32% chance OpenAI public by end 2026 (down from 60% Dec).
Chamath on Compute/Power Constraints
OpenAI/Anthropic multi-trillion dollar companies. Choke point: power for tokens. Misses due to insufficient compute capacity today, worsening.
Anthropic induced Amazon capacity (bypass Bedrock); differentiated deals with economic participation (e.g., Google).
New issue: backlogs in components (resips, nat gas turbines, transformers, grid infrastructure).
Gigawatts under construction mismatch: <50% actually built (red tape, supply chain delays), no credible on-ramp.
Hurts OpenAI/Anthropic most; benefits hyperscalers (Oracle, Amazon, Meta, Microsoft, Google). Negotiation: equity/control for compute vs. missing growth.
Opportunity for Grok/SpaceX (excess capacity); Cursor deal appetizer. Elon could deal with Anthropic (not OpenAI due to baggage/lawsuit).
Limiting resource: power → compute → tokens for developer/cowork/consumer/enterprise demand.
40% announced projects canceled (poor planning, bad feelings on data centers/AI/jobs; violence blaming them).
Friedberg on Market Dynamics
BCG rule of three: mature markets evolve to 4:2:1 market share ratio for 90% market (leader 4x #2, #2 2x #3).
Consumer: OpenAI ~900M weekly users (ahead of Claude <100M; Gemini 700M-1B, neck-and-neck). Trending ChatGPT/Google fight for 1st/2nd, Anthropic 3rd, Elon possible.
Enterprise: Google leads (Vertex: 75% GCP customers active); stock ripping.
MIT paper: pruning techniques reduce network size 90% same accuracy → 10x inference cost reduction, 10x output per energy (no accuracy loss). Prune large models for simple queries (long tail like weather/movies/stock).
Bulk consumer/coding requests (80%) to smaller models dynamically called → 10x use of data center/energy.
Early days for efficiency; unlocks opportunity (e.g., Elon).
Two ways to win: throw compute or SLMs/VSLMs (verticalized small language models, e.g., weather/travel). Google routes to lighter/faster (Google Flights in Gemini; ~750M users, behind OpenAI).
Google balanced AI search + clicks; search revenue surging, stock rewarding.
Evolution: automated pruning discovers models (not human heuristics like 'weather model'); macro model with dynamic small models → 10x inference per energy.
Sacks on Google Competition
Google took consumer share (Sergey from retirement, improved Gemini, top of search). But OpenAI Codex taking coding share.
Constant one-upping: 2 weeks ago Anthropic dominant (10x vs OpenAI 3x growth); then Claude servers down.
Anthropic forced OpenAI to compete; Google in consumer. Keep market competitive for consumers/economy/country.
OpenAI Cyber Advances
Anthropic Mythos splash (not commercially released, compute constrained) hit new cyber level.
OpenAI GPT-5.5-Cyber: AI Security Institute tests show 2nd model to complete multi-step cyber attack simulation end-to-end (Mythos level); commercially ready, compute available.
Frontier models to have Mythos-level cyber in 6 months. GPT-5.5 first cyber defenders use (vs. CrowdStrike/Palo Alto attackers).
Opportunity for Cursor/Claude/OpenAI/Gemini to tighten security, revenue stream.
Cyber Discussion
Mythos/GPT-5.5 automate cyber like coding: powers attackers/defenders. Discovers existing bugs/vulnerabilities (doesn't create).
Use AI to find/patch in advance → harden infrastructure/security.
Leap to post-AI cyber: one-time upgrade cycle (find dormant bugs). Then equilibrium offense/defense; not existential if harden before hackers.
~5M security experts; AI agents (40 hours tokens/week) create 5M-100M more ($100/chief security officer/expert). Relentless, never sleep.
Elite hackers: few thousand; AI scales speed/surface area. Get tools to white hats first.
AI can operate with scale and speed beyond human hackers. Put these tools in the hands of white hats to conduct cyber attacks, find vulnerabilities, and patch them before black hats exploit them. Chinese models will have these capabilities within six months; DeepSeek v4 already has some level, impressive but at 80-85% of the American frontier.
Humans are error-prone coders, creating holes exploited by other humans. Now computers exploit human-written bugs. Next phase: machines versus machines. Cyber nature will change completely in 5-6 years. All operational software running the world will be rewritten for operating leverage, revenue growth, and because handmade software is insecure. More software written by machines will be impregnable, but cyber threats increase as machines try injecting malware or bad tokens into agentic loops.
Best cybersecurity company, run by one of the best CEOs, has penetrated and can manipulate every model. At Breakthrough Prize, George Kurtz (CrowdStrike) and Nikesh (Palo Alto Networks) said there's a line out the door for their product. Cyber attack vectors will decline like murder rates—lowest in history despite population growth. Remaining vectors are human factors: post-it note passwords, dog's name +123.
In Atlanta for Knicks game, chose single room to save for $30k courtside seats instead of suite or private jet. Value for value. Hotel rooms sold out due to Knicks; most expensive is $500/night. Dirty bed banter.
Private jet story: Flew team to Atlanta event (4-hour flight from Bay Area) with vintage Pappy Van Winkle bottles ($4k each, antique). Fell asleep after drinks; woke to all three bottles cashed. Charged team $8k jet fuel + $12k Pappy. "Fuel the vibes as well as the plane." Team still works there. Atlanta has rebound opportunity like Detroit; went to Waffle House at midnight, no shootings. Loyalty points at Best Western via Visa card.
Elon accuses OpenAI of breach of charitable trust, unjust enrichment, flipping non-profit to for-profit. Seeks $150B damages, revert to non-profit, remove Altman and Brockman. Fireworks with lawyers. Elon: "If we make it okay to loot a charity, the entire foundation of charitable giving in America will be destroyed."
Greg Brockman kept diary journal-maxing plans like a Bond villain:
"Conclusion: we truly want the beacons we truly want the beacons... The true answer is that we want Elon out. If three months later we're doing B Corp, then it was a lie. Can't see us turning this into a for-profit without a nasty fight... In the end it's still about wanting a for-profit just without him."
Brockman's diary is discovery-maxing, smoking gun-maxing. Reference to The Wire scene: "Is you taking notes on a criminal conspiracy?"
Don't keep diaries—rumination leads to unhappiness. Advice: Go to dinner, courtside seats, work on new projects, retard max (9/10 bets lose, one wins). Don't write anything down.
Polymarket at 42-43% Elon win probability unchanged despite discovery. Possible: Elon technically wins, credited $40M. Life advice: Keep moving forward, don't ruminate—gateway to therapy/medication industrial complex fraud ($1200/hour, no incentive to stop).
Value in unpacking acute trauma, not endless self-improvement rumination.
Bench trial in Oakland; jury advisory, Judge Yvonne Rogers (61, Obama appointee) decides damages. Oversaw Epic vs. Apple (ruled for Apple, no monopoly). Hard to handicap: Bay Area judge, Elon's politics, Altman screwing people.
Not about politics; protecting charity left-coded. Elon reportedly turned down shares to keep charitable entity, principled stance against Microsoft keys to kingdom. Likely settlement; Elon put in first $40-50M, due 10-30% post-dilution. Worst for OpenAI: Unravel structure, delay IPO, chaos.
5th edition, LA, Sept 13-15. Top-tier speakers. Possible Vegas move next year. Live events: Liquidity, All-In Summit.
Google, Microsoft, Amazon, Meta smashed earnings Thursday. Story: $725B capex guidance 2026 (Amazon $200B, MSFT/Google $190B each, Meta $145B). Add Grok, OpenAI, others—nearing $1T buildout, driven by AI/cloud.
- Google Cloud (incl. Suite): +63% YoY to $20B/Q.
- Microsoft Cloud (Azure, Windows/SQL Server): +30% to $34.7B.
- Amazon AWS: +28%? to $37.6B (pure cloud).
All neo-clouds serving AI jobs/tokens. Demand from startups to frontier models. Shrinking free cash flow:
- Amazon: -97%
- Google: -12%
- Microsoft: -12%
- Meta: -8%
End of free cash flow deluge, buybacks, dividends—focus on infrastructure.
Structural shift: Mag 7 asset-light models absorbed investments. Now asset-heavy. Hyperscalers signing big checks, possibly >2x spot energy rates (e.g., Microsoft Three Mile Island PPA >2x spot, small % of needs). Will lever up with debt, look like bulky industrials. Valuation case weakens. Follow the $1T/year dollars to underpriced suppliers.
Reminiscent of 2000 Cisco peak (25 years to recover); massive fiber buildout, dark fiber unused. Today: No dark GPUs, demand pulls CapEx. Bull thesis validated—clouds exceeding guidance. 2024 hyperscaler CapEx: $700B+ (up from $350B 2023), ~2% GDP. AI 75% of last quarter GDP growth (CapEx only, not token economic impact). Tokens enable code explosion, bespoke software, productivity boost—businesses generate code without engineers.
AI will unlock huge productivity across the economy with new use cases like co-working and agents for workflow automations. It's still early. Agents won't replace humans. Recent case: agent deleted a production database in nine seconds due to a bug. Agents are valuable but must be supervised. Someone has to be accountable—not the CEO for thousands of agents. Need IT people savvy enough to supervise, ensure it works, and be accountable to the CEO. AI is not end-to-end; it's middle-to-middle: someone prompts, validates, supervises, and ensures accountability.
No huge job loss from this productivity boom. AI is becoming synonymous with the American economy: generates 75% of GDP, capex explosion, energy explosion, beginning of applications from new token factories. Positive development. Doomers throwing wet blankets don't want progress—stopping AI means stopping the American economy and economic growth. Without growth, no money for social programs, national debt, national defense. Strong economy is popular and now synonymous with AI.
President Trump set the table with good advisors. President wants America to win. Analogy: 100 years ago, people opposing highway system (federally funded, most expensive inflation-adjusted project in US history) or railroads—you can't stop these things. High demand for tokens/intelligence resource, unlike fiber where capacity built without applications.
Story: Founder of Pocket OS (software for rental car companies) vibe coding with Opus 4.6 via Cursor AI platform (most expensive tier), configured safety rules. Agent on routine task saw credentialing mismatch, fixed by deleting railway volume without confirmation, pushed code from repo to live app, deleted everything including backups.
Reference to Silicon Valley HBO clip: "You gave your AI permission to overwrite code in the internal file system? Were you gonna tell me about this? No, I thought that was company policy." Possible AI logic: most efficient way to eliminate bugs is to eliminate all software—technically correct, but black box neural nets.
Question on blame for blowing up stack (e.g., Ohalo): blame Dario (17th invite to podcast).
Not AI scheming (paperclip maximizer). Old-fashioned bugs at edge case: API not designed for permissioned usage, credential left lying around, perfect storm. Systemic issue: AI doesn't know what it doesn't know—overconfident, lacks low confidence pause like human before deleting production database. Longer time horizon tasks more likely to drift off rails. Must be supervised.
Peak of inflated expectations: eliminating software developers. Tweet by Aaron Levy retweeting Matthew Iglesias: "Five months in, I think I've decided I don't want to vibe code. I want professionally managed software companies to use AI coding assistance to make more, better, cheaper software products that they sell to me for money. Just lower your prices."
Aaron Levy: Agentic coding boon for developers wanting to get more done, great for learning coding. Less great for casually building complex software to maintain ongoing: upgrades, maintenance, security, bugs, cyber—huge risk that must be managed. People will get fired; public companies trying vibe coding will torch enterprise value.
Tech adoption chart: trigger → peak of inflated expectations → trough of disillusionment → slope of enlightenment → plateau of productivity.
Lilly phase 3 data: retatrutide (triple agonist: GLP-1, GIP, glucagon receptors). Glucagon increases metabolism, accelerates fat over muscle consumption, quick fat loss, reduces muscle loss.
Data: non-HDL cholesterol down 27%, triglycerides down 41%, liver fat down 80%, A1C from 7.9% to 6% in 40 weeks (life-saving for diabetics). Average: 214 lbs to 177 lbs (37 lbs loss vs. 6 lbs placebo) in 40 weeks. Modest side effects: 20% more nausea than placebo.
Reductions in inflammatory signaling (de-aging potential). High-dose 12mg studies, but 2mg could give anti-inflammatory, maintenance benefits.
Projected: mid-2027, possibly sooner. Swim (Reddit term): cycled retatrutide, spectacular muscle gain, no loss, fat lowering. X search "Retta": 65-year-olds from dad bod to ripped athletes in weeks. Liver/cardiac health benefits.
GLP-1 turns off appetite → starvation mode favors muscle breakdown (denser than fat). Retatrutide glucagon favors fat burning → short-term low-dose for cutting weight, maintaining muscle. Hype in fitness community.
Avalanche of peptides. Nov 2025: Lilly deal with Trump admin drops tirzepatide to $50 on Medicare. Portfolio: tirzepatide $50 (Honda), retatrutide premium (Mercedes/BMW/Model S). Tirzepatide like one-bedroom messy hotel; retatrutide two-bedroom suite.
Promo: roe.co/sparks (Roe Sparks pill). Joking spokesmen: Charles Barkley, Serena Williams. Story: tried it, out of control—requires planning, scheduling (10 PM or vacation 10AM-noon), no kids, well-rested, extended session, big commitment for extra 20% performance (double overtime). Gave pills to Stan Tang, dealt at poker.
Freedberg attended Supreme Court hearing on Monsanto/Bayer Roundup case—amazing experience. Massive protest, through marshals, sacred marble building, total quiet, no politics/freedom of speech—justices dictate behavior.
Courtroom: marble frieze above justices with historical figures (Moses etc.), nine justices mixed seating (Roberts center as chief). Justices jump in with questions; Roberts calls if quiet. Lawyers like LeBron James—mind-blowingly impressive, energy-sapping to watch.
Debate legal interpretation/constitutionality, not facts (determined lower court). Only questions of law. Full briefing/history. Hear 2 cases/day (1hr each), Mon-Wed last two weeks Oct-Apr, handful selected—shot clock, 30 min/side.
Case: EPA under FIFRA (Federal Insecticide, Fungicide, Rodenticide Act) sets pesticide labels (warnings, cancer etc.). EPA determined Roundup does not cause cancer. Must register, use exact EPA label.
All lost cases: state failure-to-warn (e.g., California). Bayer paid $10B, reserved $10B, 90,000 cases outstanding. White House Solicitor General urged Supreme Court to take for federal preemption (federal label over state laws).
California has a failure to warn law requiring manufacturers to warn consumers if a product carries a known risk. Lawyers argue that Monsanto or Bayer knew their product caused cancer but failed to warn consumers. They've won enough cases to create a multi-decabillion-dollar problem for the company.
Bayer's defense relies on EPA approval stating the product doesn't cause cancer and federal preemption, giving the EPA sole authority to determine safety.
The opposing attorney invoked the overturning of the Chevron doctrine last year, which ended deference to federal agencies. States should interpret the law directly without deferring to the EPA.
The case feels like a 50/50 coin flip, potentially 5-4 either way. Initially, expectations were 6-3 for Bayer.
Oral Arguments Process
Each side gets 30 minutes for arguments plus Q&A, totaling about an hour. Lawyers distill complex cases into briefing docs focused on the precise legal interpretation sought.
Ketanji Brown Jackson asked: what if evidence emerges post-EPA labeling that the product causes cancer? Shouldn't they update the label? The response: no, only EPA can issue labels; changing it without reporting to EPA is criminal. She countered: what if EPA doesn't act? States should protect their people.
Broader Implications
If states can interpret federal law and ignore agencies like EPA, FDA, or USDA, it opens a can of worms allowing states to bypass federal regulators.
Oral arguments focus on one specific legal question. After arguments, justices hold a private conference with a double-door airlock system—no one else enters. They discuss, determine the majority, and the Chief Justice assigns the opinion writer.
Ted Cruz, who clerked for William Rehnquist, shared insights on the Supreme Court.
Supreme Court Experience
The Court feels like England's institutions—highly coordinated security rotates every 20 minutes. About 150 people in attendance, all dressed respectfully. Access is through clerks or staff; there's an online ticketing lottery for four-packs. No public lineup visible; guests via Chief Justice pass.
The Supreme Court is one of the last highly functional U.S. institutions. Popularity tracks agreement with decisions—currently 44-50% favorability (Gallup: low of 35% in 2024), higher than Congress.
James Carville said Democrats will pack the court to 13 justices when in power, then create new states to add more.
Suggestions to monetize: auction courtside seats via Howard Lutnick.
Enjoy it while it still is in its current form.
The idiocracy version: justices in an arena with jerseys, sections, Cracker Jacks.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
