Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters

2026-07-18 - 89 min - source - Read full transcript

Jason Calacanis (host)Chamath Palihapitiya (host)David Sacks (host)David Friedberg (host)

Key insights

Demis Hassabis proposed a FINRA-style self-regulatory organization (SRO) for frontier AI, and it drew rare cross-industry endorsement.

The proposal would have frontier labs submit models 30 days before release to an industry-funded, federally-overseen body assessing cyber, national security, and biological risk, voluntary at first and mandatory later. Elon Musk, Sam Altman, Jack Clark, Sundar Pichai, Satya Nadella, and Jack Dorsey were all reported as supportive.

ai-self-regulation

David Sacks said he could back the SRO only under five conditions: broad representation including startups and open source, review limited to true frontier models, scope limited strictly to catastrophic risk (not speech or disinformation), a voluntary-first rollout, and it must substitute for new regulation rather than add to it.

He warned that without these guardrails the SRO would become an opening bid for further regulation rather than a stopping point, and contrasted it with the far heavier 'FAA for AI' model some in the industry (including Anthropic's Dario Amodei) have floated, which he said would impose 3-9 year certification timelines per model and hand China the AI race.

ai-self-regulation

Sacks and Chamath argue Anthropic is running a regulatory-capture strategy, citing a Politico investigation showing it is pushing states to adopt progressively stricter AI rules rather than converge on one national framework.

Sacks said this repeats a claim he made in October of the prior year that was dismissed as attacking a startup; he argues Anthropic, now valued near a trillion dollars, is using compliance costs and fear of catastrophic risk to entrench itself against smaller and open-source competitors, and that once a company starts trading concessions for goodwill, government pressure for more only escalates.

ai-self-regulation

Stripe, private equity firm Advent, and reportedly Block jointly offered about $53 billion (roughly $60/share) to acquire PayPal, aiming to combine Stripe's merchant-API network, Block's point-of-sale infrastructure, and PayPal's 430 million consumer accounts (plus its existing Braintree and Venmo assets) into a full-stack payments competitor to Visa and Mastercard.

Chamath argued the deal's only real logic is vertical integration: pairing Stripe's merchant relationships with PayPal's consumer base and Block's rails would let the combined entity route transactions on its own low-cost infrastructure and pass merchants a 2-5% discount, undercutting card-network fees entirely.

fintech-consolidation

The panel frames the PayPal bid as part of a broader pattern of AI-native operators acquiring mature, founder-less internet businesses (also citing Ryan Cohen's eBay bid and Bending Spoons's roll-up of AOL, Vimeo, WeTransfer, Eventbrite, and Evernote) to modernize them with automation and AI-driven efficiency.

Friedberg argued these targets are underutilizing their networks, overspending on operations, and haven't yet applied AI to their business - creating an obvious opportunity for capital providers willing to partner with proven, non-consultant operators to revive them.

fintech-consolidation

The panel says the AI industry has a severe inference-cost disparity - roughly $56 per million input tokens for top closed models versus about $0.50 for comparable Chinese models - a 50-100x gap that is starting to hit corporate budgets.

Ramp CEO Eric Glenn reported customer AI token spend has grown 21x over the past year and launched a token-spend-management product because CFOs can no longer see or control the bill; the panel warned that public companies will eventually miss earnings because of unmanaged token spend, since engineers have no incentive to economize on model choice.

ai-inference-economics

Cheap open-weight models and fine-tuning platforms are undercutting the case for paying premium prices to closed frontier labs for most tasks.

Sacks estimated 95%+ of everyday prompts don't need frontier-level intelligence and could run on far cheaper open models; the panel noted Anthropic has a strong incentive to see open-weight competitors curtailed, since they are what's driving the price collapse.

ai-inference-economics

New York Governor Kathy Hochul signed the first statewide moratorium on hyperscale data centers, and Sacks rebutted each of her stated justifications point by point.

Sacks argued utility costs only rise if data centers draw from the shared grid without adding power (solved by behind-the-meter generation); that New York State is roughly 70-80% undeveloped land outside NYC; that modern closed-loop water systems use about as much water as two-and-a-half In-N-Out burgers; and that natural gas, the typical data center power source, is relatively clean-burning.

data-center-energy-policy

Chamath said the U.S. faces a projected energy deficit equal to 2.5 Californias by 2050, evidenced by a PJM capacity auction seeking 7-8 gigawatts of forward supply that drew only about 150 megawatts of bids.

This shortage is pushing data center developers toward 'behind the meter' power generation on their own property, which runs into clean-air permitting rules; Chamath described Elon Musk's approach at the Memphis Colossus cluster of using mobile turbine generators classified as personal-use equipment to qualify for a permitting exemption.

data-center-energy-policy

The panel connected rising anti-data-center sentiment to a suspected foreign influence campaign, drawing a direct historical parallel to Russia Today's role in seeding U.S. anti-GMO sentiment starting around 2010.

They cited an OpenAI blog post on PRC-linked influence operations targeting U.S. AI debates and a recent poll showing over half of Americans believe data centers raise water and electricity costs even when a facility fully recirculates its water and generates its own power, arguing China has a direct incentive to slow U.S. data center buildout to win the AI race.

data-center-energy-policy

A new paper from Calico (Google's aging-research arm) and a partner biotech engineered a novel enzyme, using AlphaFold and directed evolution, that degrades CML, the sugar-and-fat buildup (glycation) responsible for stiffening the extracellular matrix and driving wrinkles, joint immobility, and inflammation with age.

Researchers used AlphaFold to identify a candidate binding protein, then ran five rounds of directed evolution, producing thousands of variants to optimize its ability to break down CML, first in a test tube against proteins like collagen and hemoglobin, then on donated human skin.

longevity-science

Applied to skin from patients over 70, the engineered enzyme eliminated 55% of accumulated CML and reversed the skin's biological age to roughly that of a 31-year-old.

Friedberg predicted the first commercial application will be a cosmetic cream (a market he estimated at over a trillion dollars) well before any injectable or systemic therapeutic use, with open questions remaining about delivery mechanism, including a possible future RNA-based approach that produces the enzyme inside the body.

longevity-science

Media referenced

Demis Hassabis's X post proposing a FINRA-style AI self-regulatory body - other - The proposal that opens the episode; panel debates its merits and conditions for supporting it.
Politico: "Inside Anthropic's State-by-State Plan to Ratchet Up AI Rules" - article - Cited by Sacks as evidence Anthropic is pushing incrementally stricter state AI regulation rather than a single national framework.
Satya Nadella blog post on the "reverse information paradox" - article - Cited as building on Alex Karp's argument that enterprises need to retain control over their compute, models, weights, and data rather than cede it to closed model vendors.
OpenAI blog post: "PRC-linked influence operations are targeting AI debates in the U.S." - article - Cited as evidence that Chinese-linked influence campaigns are shaping U.S. public opinion against AI data center construction.
Calico / Retro Biosciences extracellular-matrix aging paper (Science Corner) - paper - Friedberg covers a new paper using AlphaFold and directed evolution to engineer an enzyme that degrades glycation end-products (CML) and reversed donated elderly skin's biological age to roughly 31 years old.

Companies

Google DeepMind - Demis Hassabis's proposed SRO for frontier AI models opens the episode and anchors the regulation debate.
Anthropic - Accused by Sacks and Chamath of running a regulatory-capture strategy via state-by-state AI rules and of funding anti-data-center advocacy groups like Public First.
OpenAI - Signed onto the SRO proposal; separately sued by Apple over alleged trade-secret theft tied to OpenAI's hardware device and hiring of former Apple staff.
xAI - Its Grok Build coding tool was found to be uploading users' full codebases to xAI's servers despite a privacy setting meant to prevent it; xAI disabled the upload and open-sourced its harness in response.
Apple - Filed a 41-page trade-secret lawsuit against OpenAI on July 10th; separately, a rumored M7 Ultra chip with up to 1.5TB of RAM is framed as a threat to cloud AI providers by enabling local frontier-class inference.
Stripe - Jointly bid roughly $53B alongside Advent and Block to acquire PayPal, aiming to combine Stripe's merchant network with PayPal's consumer base to bypass Visa/Mastercard.
PayPal - Target of the Stripe/Advent/Block acquisition offer at about $60/share; panel frames its 430M consumer accounts and Braintree/Venmo assets as the strategic prize.
Block (Square) - Reported to be contributing $17B in equity to the PayPal acquisition offer alongside Stripe and Advent.
Ramp - CEO Eric Glenn announced a new token-spend management feature after customer token spend grew 21x, used as evidence AI inference costs are becoming a real corporate cost-control problem.
PJM Interconnection - Its capacity auction sought 7-8 gigawatts of forward power supply and received bids for only about 150 megawatts, cited as proof of a severe U.S. electricity shortfall.
Bloom Energy - Cited as a provider of behind-the-meter power installations that can qualify for personal-use clean-air permitting exemptions needed to power data centers.

Techniques and frameworks

Self-Regulatory Organization (SRO) model - Hassabis's proposal to regulate frontier AI the way FINRA regulates finance: industry-funded, expert-run, federally overseen, reviewing only true frontier models for catastrophic risk.
Behind-the-meter power generation - Building power generation directly on a data center's own property to bypass grid interconnection backlogs; complicated by clean-air permitting rules, as in Musk's mobile-turbine approach for the Memphis Colossus cluster.
Directed evolution + AlphaFold protein design - Used in the Science Corner paper to iteratively engineer a novel enzyme that degrades glycation end-products (CML) in the skin's extracellular matrix.
Zero Data Retention (ZDR) - Industry term for a vendor promise not to retain customer data; Chamath argues it is functionally unverifiable given the xAI Grok Build leak.

Summary

The episode opens on Demis Hassabis's proposal for a FINRA-style self-regulatory organization (SRO) to govern frontier AI, which drew rare cross-industry endorsement from Elon Musk, Sam Altman, Anthropic's leadership, Satya Nadella, and Jack Dorsey. Sacks says he could support it, but only under five conditions he laid out directly to Hassabis: broad representation including startups and open source, review limited to genuinely frontier models, scope restricted to catastrophic risks like cyber and CBRN (not speech or disinformation), a voluntary-first rollout, and it must substitute for new government regulation rather than add to it. He contrasts this with the much heavier "FAA for AI" model some, including Anthropic's Dario Amodei, have floated, warning that aircraft-style multi-year certification timelines would hand China the AI race. Sacks and Chamath spend much of this segment reiterating a charge Sacks first made a year earlier: that Anthropic is running a sophisticated regulatory-capture campaign, citing a Politico investigation showing the company is pushing states toward progressively stricter AI rules rather than a single stable national framework, using fear of catastrophic risk to entrench its position against smaller and open-source rivals.

The panel then covers Stripe, Advent, and reportedly Block's roughly $53 billion joint bid to acquire PayPal at about $60/share. Chamath frames the strategic logic as vertical integration: combining Stripe's merchant-API network and Block's point-of-sale infrastructure with PayPal's 430 million consumer accounts (plus its Braintree and Venmo assets) could let the combined company route transactions on its own low-cost rails and undercut Visa and Mastercard directly. Freiburg situates the deal within a broader pattern of AI-native operators acquiring mature, founder-less internet businesses to modernize them (also citing Ryan Cohen's eBay bid and Bending Spoons's roll-up of AOL, Vimeo, WeTransfer, and Evernote), while Sacks and Chamath note the deal likely wouldn't have cleared antitrust review two years ago, and that how regulators define the relevant market (payment APIs versus the card-network duopoly) determines whether it reads as anti- or pro-competitive.

A middle segment turns to AI inference economics and a fresh xAI privacy lapse: Grok Build was found uploading users' entire codebases to xAI's servers despite a privacy toggle meant to prevent it, which xAI disabled and then partly remediated by open-sourcing its harness. Chamath connects this to Satya Nadella's "reverse information paradox" framing (building on Alex Karp's argument that enterprises need to retain control of their own compute, models, and data) to argue that "zero data retention" promises from any vendor are functionally unverifiable. The panel then cites data showing frontier-model inference can cost roughly $56 per million input tokens versus about $0.50 for comparable Chinese models, a 50-100x gap, alongside Ramp's report that customer token spend has grown 21x over the past year, warning that unmanaged AI spend will eventually cause public companies to miss earnings.

The back half covers New York Governor Kathy Hochul's newly signed statewide moratorium on hyperscale data centers, which Sacks rebuts point by point (behind-the-meter power avoids raising grid costs, the state is mostly undeveloped land outside NYC, closed-loop cooling uses minimal water, natural gas is relatively clean). Chamath cites a PJM capacity auction that sought 7-8 gigawatts of forward power and drew only about 150 megawatts of bids as evidence of a looming U.S. energy deficit, projecting a shortfall equal to 2.5 Californias by 2050, and describes Musk's workaround at the Memphis Colossus cluster of using mobile turbine generators to qualify for personal-use clean-air permitting exemptions. The panel then ties rising anti-data-center sentiment to a suspected foreign influence campaign, drawing a direct parallel to Russia Today's historical role in seeding U.S. anti-GMO sentiment and citing an OpenAI blog post on PRC-linked influence operations targeting U.S. AI debates.

The episode closes with a Science Corner segment on a Calico-linked paper that used AlphaFold and directed evolution to engineer a novel enzyme degrading CML, the glycation buildup responsible for skin, joint, and inflammatory aging. Applied to donated skin from patients over 70, the enzyme eliminated 55% of accumulated CML and reversed the skin's biological age to roughly that of a 31-year-old, with the panel predicting a cosmetic cream will be the first, multi-trillion-dollar commercial application ahead of any therapeutic use.

Notable Quotes

"The whole industry is going to need to be regulated. And I think the industry needs to regulate themselves." - Jason Calacanis

"I thought that an SRO, again, a self-regulatory approach, should be infinitely better than creating a new government agency that I think would rapidly become a DMV for AI." - David Sacks

"Everything she's saying there is a false accusation on the data center." - David Sacks

"...which basically reversed the skin's age down to the age of a 31-year-old." - David Friedberg