Featuring
No guests this week: the four besties moved from OpenAI's missed targets and Codex's rebound to the hyperscalers' $700 billion-plus infrastructure wager, then let Friedberg explain why retatrutide has the fitness crowd buzzing. The sharpest exchanges asked whether OpenAI has a demand problem or a power problem, and whether today's fully booked GPUs can still become tomorrow's stranded capital. Sacks had the strongest episode by separating current demand from long-term returns; Chamath supplied the useful warning label on both debates.
Spice rack
Did OpenAI miss its targets because demand softened or because power constrained supply?
Original point: The miss reveals insufficient compute capacity, not weak demand: power, grid equipment, and project delays are the choke points limiting the tokens OpenAI and Anthropic can sell.
What everyone argued
Chamath Palihapitiya
Chamath argues that demand is effectively unlimited and that power is the binding constraint behind OpenAI's miss. Grid delays and scarce capacity will force frontier labs to trade economics or control to hyperscalers, while infrastructure-rich rivals gain room to attack.
David Sacks
Sacks says OpenAI's consumer business missed because Google took meaningful share, while arguing that the picture improved in coding and enterprise. In his account, Sam Altman's compute bet may work for a different reason than planned: Codex growth can absorb capacity that the consumer forecast did not justify.
David Friedberg
Friedberg treats the market as an unsettled three-player structure and argues that efficiency can relax today's bottleneck. He points to model pruning and routing among smaller subnetworks as a path to far more inference from the same energy footprint.
Winner circle
Sacks wins by separating consumer competition from enterprise and coding demand. Later Codex growth supports his claim that OpenAI's capacity bet could find a second justification, while continued industry-wide capacity constraints preserve an important part of Chamath's case. Chamath loses on burden of proof because 'entirely 100%' requires evidence that rules out competition, and the public record does not do that.
Commentary
Chamath Palihapitiya
Assumptions and fact checks
OpenAI's missed targets had nothing to do with demand and were entirely caused by insufficient power and compute.
Why it mattersCapacity constraints are well documented, but the claim is too absolute. Missing a weekly-user target is not mechanically explained by an inability to serve tokens, and reported gains by Gemini provide a plausible competing demand explanation.
Scarce power and compute will increase hyperscaler bargaining power over frontier labs.
Why it mattersWhen capacity is scarce and concentrated, suppliers can demand longer commitments and better economics. OpenAI's and Anthropic's large AWS capacity commitments fit that mechanism, even if the eventual division of value remains unsettled.
David Sacks
Sacks wins by allowing two things to be true: OpenAI can lose consumer momentum and still improve its product position elsewhere. The analysis would be tighter if he treated Google's causal share as a hypothesis instead of a settled attribution.
Assumptions and fact checks
Google taking consumer share was the single main reason OpenAI missed its user and revenue targets.
Why it mattersGoogle's consumer and cloud AI products grew rapidly, so competition is a credible contributor. Public disclosures do not isolate the counterfactual or show how much of OpenAI's miss came from Google rather than pricing, conversion, capacity, or an aggressive internal forecast.
Coding and enterprise demand can vindicate OpenAI's large compute commitments even though the original consumer forecast missed.
Why it mattersCodex's subsequent weekly-user growth supports the substitution mechanism. Whether that adoption produces returns commensurate with the full infrastructure commitment remains unresolved.
OpenAI missed an internal target of one billion weekly ChatGPT users and was at roughly 900 million weekly users.
CheckThe internal target miss was reported by The Wall Street Journal, and OpenAI publicly stated that ChatGPT had 900 million weekly users in March 2026.
David Friedberg
Friedberg correctly adds algorithmic efficiency to a debate dominated by construction forecasts. His specific research claim conflates an influential sparsity result with a mature inference-routing system, so the directional insight survives but the numerical promise does not.
Assumptions and fact checks
Model pruning and routing will materially soften power constraints before demand outruns the efficiency gains.
Why it mattersPerformance-per-dollar improvements are likely, but efficiency often lowers prices and induces more usage. The net effect on total power demand depends on deployment speed and demand elasticity.
The cited MIT pruning paper showed that deployed large models can be reduced by 90% and dynamically routed to deliver ten times the inference with no loss of accuracy.
CheckThe lottery-ticket paper found that dense networks can contain sparse subnetworks which, when reset and retrained, can match the original network's test accuracy. It did not demonstrate dynamic routing among pruned large-language models or a general production guarantee of ten times the inference per unit of energy.
ChatGPT had about 900 million weekly users at the time of the discussion.
CheckOpenAI stated in March 2026 that 900 million people used ChatGPT each week.
Is hyperscaler AI capex productive scarcity investment or an industrial-scale valuation trap?
Original point: The hyperscalers are becoming capital-heavy industrial businesses whose power contracts, debt, and financial engineering may consume the cash that once supported premium valuations.
What everyone argued
Chamath Palihapitiya
Chamath argues that the market is underpricing a business-model change. Hyperscalers are exchanging asset-light free cash flow for power contracts, equipment, debt, and long-lived infrastructure; even if they can fund it, suppliers receiving the spend may offer the cleaner investment than the buyers.
David Sacks
Sacks rejects the dot-com analogy because today's accelerators are being deployed into visible demand rather than sitting dark. Strong cloud growth, constrained capacity, and customer demand make the buildout productive; he broadens the upside case to software creation and economy-wide gains from the tokens those factories produce.
Winner circle
Sacks wins narrowly on the question the episode could actually answer in 2026: this was not a dark-capacity buildout without customers. Official results and commitments show real demand, constrained supply, and fast cloud growth. Chamath is right that utilization alone does not preserve an asset-light valuation, so the long-run investment ruling stays open; his extrapolations about power pricing and inevitable leverage do not yet clear the burden needed to overturn the current evidence.
Commentary
Chamath Palihapitiya
Chamath makes the strongest skeptical argument because he does not rely on empty 'bubble' rhetoric; he focuses on cash conversion and capital structure. The missing step is a return model showing that utilization and pricing cannot cover depreciation, energy, and financing costs.
Assumptions and fact checks
Hyperscalers will become highly levered industrial businesses within five years, eroding their premium valuation case.
Why it mattersCapital intensity and alternative financing are rising, but the companies still generate large operating cash flows and can pair infrastructure with high-margin software. The valuation effect depends on utilization, pricing, useful lives, and financing terms.
Buying infrastructure suppliers is a cleaner way to capture the AI buildout than owning hyperscalers funding it.
Why it mattersSuppliers receive the near-term dollars, but they also face cyclicality, customer concentration, and eventual normalization. Hyperscalers retain distribution and application-layer upside if the assets stay utilized.
Microsoft's Three Mile Island power purchase agreement was priced at more than twice the prevailing spot rate.
CheckConstellation disclosed a 20-year agreement for roughly 835 megawatts but did not disclose the contract price. Outside analyst estimates may imply a premium to some spot benchmarks, but the public primary record does not establish Chamath's precise 'more than 2x' claim.
David Sacks
Sacks wins the present-tense question: the assets are serving real demand, and customer commitments make this unlike speculative fiber laid without traffic. He overstates the macro evidence and does not fully answer Chamath's valuation question, which concerns returns after the capacity wave lands rather than utilization today.
Assumptions and fact checks
The absence of idle GPUs today means the AI buildout is fundamentally unlike the dark-fiber overbuild.
Why it mattersCurrent utilization and contracted demand are meaningful differences. They reduce near-term stranding risk, though they do not rule out future overcapacity after supply arrives or model efficiency improves.
AI infrastructure and token-driven productivity will generate returns large enough to justify current capex.
Why it mattersDemand and revenue growth support the thesis, but the burden is long term: realized pricing and utilization must cover rapid hardware depreciation, power, and financing across the cycle.
Google Cloud grew 63% to more than $20 billion, Microsoft's Intelligent Cloud segment grew 30% to $34.7 billion, and AWS grew 28% to $37.6 billion in the reported quarter.
CheckThose figures match the companies' Q1 or fiscal-Q3 2026 disclosures. Microsoft also said demand across workloads and regions continued to exceed available capacity.
AI accounted for 75% of U.S. GDP growth in the first quarter of 2026.
CheckThe 75% figure was an outside estimate, not a BEA statistic, and BEA does not publish an 'AI' GDP component. Official data show that information-processing equipment and software contributed materially, while later Federal Reserve analysis estimated a smaller contribution once software, hardware, data centers, and related inputs were separated.
Chamath contributes the debate's most useful mechanism—tokens ultimately need power—but converts a strong constraint argument into an unsupported single-cause diagnosis. He would have been stronger had he distinguished demand that could not be served from users who chose Gemini or Claude instead.