All-In PodcastHow hyperscalers bet $725B on a grid that can't keep up
The $725B hyperscaler CapEx wave is chasing electricity, not model demand. No GPUs sit dark; OpenAI's user miss is a power problem, not a product one.
CHAPTERS
Cold open: Miss Thing Podcast clip and besties banter
The episode kicks off with a comedic detour as Jason shares a viral “Gay Name, Straight Name” bit, prompting playful roasting of the hosts. The group uses the clip to warm up the room and settle into the episode’s tone before moving into AI and business news.
- •Jason introduces a favorite podcast segment and plays a clip
- •The hosts react and joke about how accurately they’re characterized
- •Light banter sets the tone before the main topics begin
- •Transition into the show’s formal welcome and topic lineup
OpenAI misses growth targets—why it matters (users, revenue, compute commitments)
The besties react to reporting that OpenAI missed internal targets for weekly active users and revenue. They frame the concern around OpenAI’s massive compute/data-center spending commitments and the tension between IPO timing and public-company readiness.
- •WSJ reports OpenAI missed 2025 targets for WAUs and ChatGPT revenue
- •OpenAI’s large compute commitments are highlighted as a core risk
- •Discussion of CFO Sarah Friar’s concerns vs. Sam Altman’s IPO push
- •What a delayed IPO could signal to markets (and Polymarket odds)
Codex vs Claude: product momentum shifts and compute gating
Sacks argues OpenAI’s “bad press week” contrasts with a strong product stretch, especially for coding. The group discusses how Anthropic’s latest release is perceived as weaker and how compute constraints can force ‘rationing’ that changes user experience.
- •Positive developer feedback on ChatGPT 5.5 and Codex
- •Anthropic Opus 4.7 criticized; some users roll back to 4.6
- •Compute/token constraints lead to throttling (“compute gating”)
- •Coding emerges as the key battleground vs. general consumer chat
The real choke point: power and grid infrastructure (not just GPUs)
Chamath reframes the shortfall story as supply-side: power availability limits token generation, not demand. The conversation expands to grid bottlenecks, transformer shortages, and the way hyperscalers may extract economics/control in exchange for capacity.
- •Power is presented as the limiting resource behind compute
- •Delays in grid buildout: turbines, transformers, permitting/red tape
- •Hyperscalers gain leverage over frontier model companies
- •Equity/control trade-offs to secure capacity; opportunity for alternative providers
Market structure and efficiency: the ‘rule of three’ and model pruning
Friedberg lays out a market-structure view (consumer vs enterprise) and predicts a small number of dominant players. He then highlights technical pathways to efficiency—like pruning and dynamic model selection—that could reduce inference costs dramatically.
- •BCG ‘rule of three’ used to reason about eventual market shares
- •Consumer AI: ChatGPT vs Google; Anthropic and others fighting for position
- •Enterprise AI adoption through platforms like Google Vertex discussed
- •MIT-style pruning techniques could cut model size ~90% and lower inference costs ~10x
AI cybersecurity is about to explode: Mythos, GPT 5.5 Cyber, offense vs defense
The hosts discuss AI-driven cyber tools moving from research to commercialization, with new models demonstrating multi-step attack simulation capabilities. They emphasize that these systems reveal existing vulnerabilities and can be used defensively to harden infrastructure before adversaries scale up.
- •Anthropic’s cyber advances (e.g., Mythos) vs OpenAI’s GPT 5.5 Cyber
- •AI Security Institute testing and ‘end-to-end’ cyber simulation performance
- •AI doesn’t create vulnerabilities; it discovers bugs already present
- •Prediction of a major ‘one-time’ hardening cycle before a new equilibrium
- •Warning that frontier-level cyber capability will spread globally (including China)
Musk vs Altman lawsuit: nonprofit-to-for-profit fight and the ‘diary’ discovery
The episode pivots to the Elon Musk vs OpenAI trial, focusing on claims of charitable trust breach and the controversy of converting to a for-profit structure. The group zeroes in on the unusual discovery of Greg Brockman’s diary entries and what they imply about intent and governance.
- •Overview of allegations: charitable trust breach, unjust enrichment, governance remedies
- •Brockman diary excerpts become central ‘smoking gun’ discussion
- •Bench trial dynamics and the role of the judge vs advisory jury
- •Speculation about settlement scenarios and potential remedies
Therapy, rumination, and ‘keep moving forward’—a cultural sidebar
A side conversation spins out from the diary topic into broader commentary about rumination, self-improvement culture, and therapy incentives. While comedic and provocative, the segment returns to a shared idea: forward motion and specificity beat endless reflection.
- •Critique of excessive rumination as unproductive
- •Debate over therapy’s value: acute trauma vs endless self-analysis
- •Framing ‘moving forward’ as practical life advice
- •Hosts’ contrasting perspectives on mental health discourse
Hyperscalers smash earnings: the CapEx supercycle and ‘no dark GPUs’
The besties break down blockbuster results from Google, Microsoft, Amazon, and Meta, emphasizing that CapEx guidance is the real headline. They debate whether this resembles the dot-com infrastructure bubble, concluding demand for tokens/compute is real and immediate.
- •~$725B 2026 CapEx guidance across major hyperscalers discussed
- •Cloud growth numbers cited (Google Cloud, Azure, AWS)
- •Free cash flow compression as companies reinvest aggressively
- •Comparison to 2000-era buildouts (Cisco/dark fiber) and why this cycle differs
- •‘No dark GPUs’ argument: capacity is consumed as fast as it’s built
Vibecoding nightmare: an agent deletes a production system
A cautionary tale illustrates the risks of agentic coding when tools have direct permissions in production environments. The hosts argue this isn’t ‘AI scheming’ but classic software/process failure amplified by overconfident automation, reinforcing the need for supervision and accountability.
- •Story recap: AI deletes a Railway volume and wipes backups in production
- •Discussion of permissioning, credentials, and unsafe deployment pathways
- •Core limitation: models don’t reliably know when confidence should be low
- •Agentic coding as productivity tool for pros—not a substitute for software engineering discipline
- •Need for human supervision, validation, and clear accountability
Retatrutide and the peptide mainstream moment: weight loss, metabolic and liver benefits
Friedberg and Chamath dive into the ‘peptide craze,’ focusing on Lilly’s retatrutide trial results and why they’re generating hype beyond obesity and diabetes. They discuss multi-agonist mechanisms, cardiometabolic markers, liver fat reduction, muscle preservation, and timelines for approval.
- •Retatrutide described as a triple-agonist (GLP-1/GIP/glucagon)
- •Trial signals: major weight loss, improved A1C, triglycerides, non-HDL cholesterol
- •Liver fat reduction and potential anti-inflammatory/‘de-aging’ implications
- •Fitness community interest due to reduced muscle loss vs earlier GLP-1s
- •Pricing/positioning: tirzepatide as mass-market vs retatrutide as premium upgrade; approval timeline speculation
Friedberg goes to the Supreme Court: inside the Monsanto/Roundup preemption case
Friedberg recounts attending a Supreme Court oral argument, describing the courtroom’s rituals and the intensity of elite advocacy. He summarizes the legal core: whether EPA labeling under FIFRA preempts state ‘failure to warn’ claims, and how post-Chevron doctrine changes the analysis.
- •Atmosphere and procedure: security, silence, seating, strict decorum
- •Oral argument format: 30 minutes per side plus justice questioning
- •Case summary: EPA-approved label vs state failure-to-warn lawsuits; Bayer/Monsanto liabilities
- •Federal preemption under FIFRA and implications for regulatory authority
- •Chevron being overturned reshapes deference to agencies; case outcome feels closer than expected
Wrap-up: Supreme Court access, institutional trust, and closing jokes
The group closes with musings on court legitimacy, ‘court packing,’ and how public perception tracks recent decisions. Jason ends by joking about monetizing access and the show signs off with the usual besties banter.
- •Debate about institutional durability and popularity of the Supreme Court
- •Concerns about potential court expansion and politicization
- •Clarification that public access exists via lottery/tickets
- •Final comedic riffs and sign-off
