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

Key points

• OpenAI missed: 1B WAU target by 4+ months and 2025 revenue target. Sarah Fryar (CFO) v Sam reportedly tense on IPO readiness. Polymarket 'OpenAI IPOs by end-2026' down to **32% from 60% in December**. Direct mirror of the 20VC episode but framed by Sacks as a *contrarian buy*, not a confirmation of weakness.
• **Sacks's contrarian:** GPT-5.5 (built on new base model 'Spud' — first new base in >12 months) is winning back coders. Opus 4.7 'a bust' — users rolling back to 4.6, complaints of compute-gating + reduced thinking time. Token-constrained Anthropic = OpenAI catch-up window. Direct contradiction of Lemkin's 20VC framing where Lemkin flipped *to* Team Sam for the same reason — Sacks reads it as OpenAI win, Lemkin reads it as agents (not humans) reshuffling the deck. **Same fact, opposite read.**
• **Chamath's frame: power, not compute, is the binding constraint.** 40% of announced gigawatts will get cancelled (red tape + transformer/turbine supply chain). Hyperscalers will extract equity + control from labs to grant capacity. SpaceX/Grok have a 'huge lane to run through' due to excess capacity. Direct addition to the Patel '72% margin floor' thesis from Issue 02 — sharper because it names the *actual* upstream bottleneck (grid components) not just GPU FLOPS.
• **$725B hyperscaler CapEx for 2026** — Amazon 200, MS 190, Google 190, Meta 145. Free cash flow getting incinerated: **Amazon FCF -97% QoQ**, Google/MS/Meta -12/-12/-8%. Microsoft's Three Mile Island deal at **>2x prevailing spot energy rate**. Asset-light Mag-7 era is over; these will look like 'big bulky industrials' in 5 years.
• Chamath's investable read on the CapEx wave: 'follow the dollars out of the hyperscalers' — buy the recipients (turbine, transformer, grid-component, gas-gen names) because they're underpriced. Decoupling thesis from the cloud names themselves.
• **Cisco-2000 analogy fails** (Sacks): no dark GPUs. Demand is voracious, pull-forward is real. Bull thesis got 'validated in a single afternoon' across MS Azure, Google Cloud, AWS, Meta. AI = ~75% of recent US GDP growth. Capex now >2% of US GDP — Sacks: 'AI is now synonymous with the American economy.'
• **AI cyber upgrade cycle is the next CrowdStrike-scale wave.** GPT-5.5 Cyber matches Mythos and is commercially shipping (Anthropic Mythos still gated). All frontier models will hit Mythos-grade in ~6 months. Chinese models (DeepSeek 4) at 80-85% of frontier already. Chamath: best CSO 'can essentially manipulate every model'. Beneficiaries: CrowdStrike, Palo Alto, Wiz — white-hats get the tools first, find dormant bugs, harden infrastructure.
• **Elon vs Sam trial** in front of Judge Yvonne Gonzalez Rogers (61, Obama appointee, ruled mostly for Apple in Epic vs Apple). Bench trial w/ advisory jury. Greg Brockman's diary entered into discovery: 'we truly want the B corp… the true answer is that we want Elon out… if three months later we're doing B corp then it was a lie.' Polymarket Elon-wins ~42-43% (hasn't moved on the discovery dump — may be pricing the $40M-refund-only outcome).
• **Vibe-coding catastrophe — the trial-of-disillusionment moment:** PocketOS founder used Cursor + Opus 4.6 (most expensive tier) on a routine task; agent saw a credentialing mismatch, deleted a Railway prod volume + backups in 9 seconds, no user confirmation. Aaron Levie's response (which Sacks endorsed): 'I want professionally managed software companies to use AI, I don't want to vibe-code.' **Direct counter-narrative to the agent-as-buyer thesis from 20VC** — buyer is still the human ops/eng team that has to clean up the mess.
• **BCG rule of three** applied to AI: stable mature markets settle into 4:2:1 share. Consumer = ChatGPT > Google > Anthropic (with Grok wildcard via SpaceX compute). Enterprise = Anthropic / Google fighting for #1, with 75% of GCP customers actively using Vertex (this is the surge under the Google-stock-rip).
• **Retatrutide (Lilly) phase-3 print** — triple agonist (GLP-1 + GIP + glucagon). 37 lb loss in 40 weeks vs 6 lb placebo, **80% liver fat reduction**, A1C 7.9%→6%, triglycerides -41%, non-HDL chol -27%. Glucagon receptor preferentially burns fat over muscle (vs tirzepatide). Mid-2027 approval target, possibly sooner. Cardiovascular + de-aging adjacencies. Premium-tier upgrade to tirzepatide (which got cut to $50/mo on Medicare in Nov 2025 Trump-Lilly deal). Lilly portfolio = Honda (tirzepatide) → Mercedes (retatrutide).
• **MIT paper on neural-net pruning:** reduce model size by 90% with no accuracy loss; 10x inference per energy unit. Implication: small specialised models called dynamically replace one giant macro-model. Verticalised SLMs (V-SLMs) plus pruning = the algorithmic offset to the power constraint. This is where Elon/Grok and small-model players genuinely *can* compete with frontier-compute incumbents.

Notable quotes

Themes

• Power, not compute, is the binding constraint on AI growth
• OpenAI vs Anthropic shoe-dropping — contrarian Sacks vs Lemkin
• Hyperscaler CapEx incinerating free cash flow (Mag-7 going industrial)
• AI cyber as the next CrowdStrike-scale wave
• Vibe-coding trial-of-disillusionment — agent-as-buyer thesis under stress