Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters
Summary
- Demis Hassabis’s FINRA-style self-regulatory organization emerged as the panel’s least-bad answer to AI oversight. The proposed industry-funded, federally overseen body would receive frontier models 30 days before release, refresh risk tests quarterly and initially operate voluntarily. Sacks supported it only with five guardrails—including startup and open-source representation, catastrophic-risk scope and no additional agency—because the alternative could become a “DMV for AI.”
- The real regulatory contest is whether an SRO prevents capture or merely becomes Anthropic’s opening bid for tighter controls. Sacks accused Anthropic of encouraging progressively stricter state rules and argued concessions invite government to “come back for more and more and more”; Chamath similarly warned that well-capitalized incumbents could “pull the ladder up.” Sacks said the approach would need to substitute for new regulation and, if treated as the industry’s line, secure preemption.
- A reported $53 billion Stripe-led offer for PayPal could assemble a genuine challenger to Visa and Mastercard. The discussed structure combined Stripe, Advent and potentially Block at roughly $60 per PayPal share, with Chamath expecting another 10-15% before the clearing price. The strategic prize is not PayPal’s aging interface but its roughly 439 million consumer accounts joined to Stripe’s merchants, Braintree, Venmo, Cash App and stablecoin infrastructure: “a shot across the bow for Visa and MasterCard.”
- PayPal may inaugurate a broader acquisition cycle in which AI-native operators revive neglected internet franchises. Friedberg linked the bid with Ryan Cohen’s eBay approach and Bending Spoons’ roll-up of Evernote, Vimeo and other Web 2.0 assets: mature, non-founder-led businesses can be cut, automated and rebuilt by specialized operators. His deliberately crude formulation was a wave of “flaccid digital businesses” revived by “the blue chew of capital.”
- Apple’s trade-secret suit and xAI’s Grok Build data leak exposed two separate liabilities around AI talent and proprietary data. Apple alleged that former employees brought “actual parts” and accessed internal storage while OpenAI recruited more than 400 Apple employees; Chamath and Sacks withheld judgment but offered a categorical rule: take only “what’s in your head.” Meanwhile, Grok Build reportedly uploaded whole repositories despite contrary assurances, reinforcing Chamath’s warning that zero-data-retention promises cannot eliminate “trapdoors everywhere.”
- Token economics are becoming an earnings issue, not merely an engineering choice. Chamath cited prices ranging from $56 per million tokens for Lovable to $0.50 for Chinese models, while Ramp said customer token spending rose 21 times in a year. Sacks warned that if engineers default to frontier models for work that cheaper models can handle, a public-company CFO could eventually miss a quarter because AI OpEx became a “money-burning furnace.”
- Electricity—not model demand—was framed as the binding constraint on US AI growth. Chamath projected a 2050 deficit equal to 2.5 Californias’ current consumption, cited a PGM auction seeking 7-8 GW but attracting roughly 156 MW, and said about 40% of planned projects are being stopped or mothballed. New York’s hyperscale-data-center moratorium therefore makes presently energizable sites more valuable while pushing the rule of deployment toward “GPUs chasing energy.”
- The episode closed with a concrete AI-enabled biotechnology result rather than a speculative catastrophe. Friedberg described an AlphaFold-assisted enzyme engineered through five directed-evolution cycles that removed 52-97% of the aging-associated molecule CML in tested proteins and 55% from donated skin of people over 70, purportedly reducing its measured skin age to 31. Delivery remains unresolved, but Chamath immediately saw the first market: “On your face as a cream. Game over.”
Deep dive
1. A FINRA for frontier AI wins support as the least-bad option
Jason laid out Demis Hassabis’s proposed US-led SRO: industry funded, staffed by independent technical experts and subject to federal oversight. Frontier labs would submit models 30 days before release, benchmarks would change quarterly, participation would begin voluntarily and could later become mandatory, and the body could coordinate a slowdown if severe risks emerged.
Friedberg thought the analogy to FINRA and the National Futures Association fit because market participants need protection from shared risks without freezing technology in statute. California’s attempted rules, he argued, already failed to map onto the technology a year later; an SRO can instead change tests, evaluators and expertise as cyber, biological, weapons and manipulation risks evolve.
The distinction Friedberg emphasized was “federal government oversight, but not control.” Congressional committees ultimately oversee existing financial SROs, but practitioners write and administer the rules—an arrangement he considered faster and more technically credible than creating a new agency.
Industry support, as summarized by Jason, extended across Elon Musk, Sam Altman, Jack Clark, Sundar Pichai, Satya Nadella, Jack Dorsey and the Collison brothers. Friedberg read that coalition as recognition that checkpoints are coming and that self-regulation offers a workable way to supply them.
2. Sacks attaches five conditions before accepting the lesser evil
Sacks said the SRO must represent startups and open source, not merely the three largest labs. He suggested a coalition broad enough to include Jensen Huang, Elon Musk, Mark Zuckerberg and Mira Murati, whose Thinking Machines platform was described as an open-weight model platform based on open source: diversity of interests is the defense against regulatory capture.
Review should apply only to models that produce a genuine step-change beyond the existing frontier. Holding back inferior or incremental models creates no corresponding reduction in catastrophic risk and gives incumbents an easy mechanism for tying up smaller competitors.
Its remit should stop at catastrophic cyber and CBRN—chemical, biological, radiological and nuclear—risks. Sacks explicitly excluded disinformation, microaggressions and other speech questions: “This should not become a speech regulator.” It should also prove itself voluntarily before receiving legal force.
His fifth and decisive condition was substitution: no new SRO layered atop a separate federal regulator. An “FAA for AI” could turn model releases from months into years; Sacks noted that new aircraft type certification takes five to nine years and amendments three to five. Between an FAA or “DMV for AI” and a tightly bounded SRO, he chose the SRO.
3. Regulatory capture remains the unresolved threat
Chamath expects “a torrent of money” to shape regulation for incumbents on both sides of the aisle. Moving quickly on credible industry rules could block the off-ramp toward a duopoly, while the Justice Department, Commerce Department and federal oversight would still prevent a genuine Wild West.
Sacks argued that even FINRA ultimately reports to the SEC, meaning an AI SRO’s governmental home would provoke a political fight. Software has never had a dedicated regulator; importing that structure could therefore create pressure and capture even if its initial mandate looks narrow.
His larger accusation was that Anthropic is executing a fear-driven capture strategy. Sacks characterized it as already carrying a $1 trillion valuation, cited Gavin Baker’s post-IPO estimate of $3 trillion, and pointed to a Politico account of “one-upmanship” in which rules such as California’s SB 53 become models for progressively tougher state legislation.
The government-ratchet problem was his core warning: when a company asks officials for more authority, very few respond, “We’re not qualified.” Sacks said an SRO could work only if the industry treated it as the line, secured preemption and fought additions; offered “for free,” it becomes merely the opening bid.
4. Stripe’s PayPal bid is really an attack on card-network economics
Reporting was still inconsistent on air, but Jason described Stripe and Advent offering approximately $53 billion, or $60 a share, for PayPal, with Block participating through roughly $17 billion of equity. Market expectations centered nearer $70, while Chamath anticipated a clearing price another 10-15% above the initial bid.
The panel’s strategic thesis was that Stripe’s merchant relationships could be combined with PayPal’s consumer base to create more end-to-end payment rails and challenge Visa and Mastercard. Block would add point-of-sale infrastructure and Cash App.
The asset map supports that thesis: PayPal brings 439 million consumer accounts, Venmo, Braintree and PYUSD; Stripe brings merchant APIs, Bridge—acquired for $1 billion in 2025—and approximately $2 trillion of annual transaction volume; Block adds point-of-sale infrastructure and Cash App. PayPal itself was put at roughly $1.7 trillion of annual volume.
Sacks worried that combining merchant and consumer relationships does not automatically make consumers choose a new payment method. Chamath’s rebuttal moved the decision to merchants: offer Stripe or other participating sellers a 3-5% discount through cheaper internal rails, and they will favor the option that puts another 2-4% in their pockets.
5. Market definition determines whether the combination is a monopoly or a challenger
The antitrust argument turns entirely on the market selected. Define it as merchant-payment APIs and Stripe’s combination with PayPal-owned Braintree looks consolidating; define it as payment networks and the transaction becomes a new competitor to the Visa-Mastercard duopoly. “It’s the game,” Chamath said of that definition.
PayPal’s attraction is partly distress. Sacks called the 25-year-old interaction model “legacy,” noted growth of only about 7%, and said its valuation had fallen from a cited $322 billion peak to roughly $30-40 billion before the offer. Efficiency gains alone would not solve the need for a modern product vision.
Chamath’s simpler answer was, “I think they’re buying the accounts.” Stripe already owns the merchant relationship PayPal lacks, while PayPal owns the consumer relationship Stripe lacks; more “on-us” transactions could bypass card-network fees even if the original PayPal interface is gradually retired.
Sacks traced the stagnation to eBay eliminating PayPal’s founding DNA after the 2002 acquisition and replacing it with a corporate, consulting-led mindset. He prefers “PayPal diaspora” to PayPal Mafia: “Our homeland was taken over and they burned our temple and then kicked everybody out.”
6. AI-native operators may turn stale internet companies into a new buyout category
Friedberg connected PayPal with Ryan Cohen’s eBay bid: modern operators can inspect mature, non-founder-led digital businesses and immediately see unused networks, bloated costs and unimplemented AI. The capital provider’s challenge is finding an operator capable of rebuilding the product, not hiring “some McKinsey consultant.” Friedberg said Cohen had proved his mettle with Chewy and GameStop, while acknowledging that the latter assessment is debatable.
He expects “a wave of mega deals” involving “flaccid digital businesses” revived with “the blue chew of capital and the right operator.” Existing examples included Josh Kushner’s accounting-firm roll-up and a similar General Catalyst effort, both using acquisition capital to AI-enable traditional services.
Bending Spoons was the operating specimen. Jason listed AOL and Vimeo at $1.4 billion each, plus WeTransfer, Eventbrite, Brightcove and Evernote; Friedberg described the Milan-based team diagnosing overspending, underspending, product and marketing, then using leverage to make old Web 2.0 properties produce cash. Jason said he was told the company uses young, AI-first executives.
Jason framed the macro backdrop as M&A returning after Lina Khan’s tenure and Trump’s election. He cited Uber’s $15 billion Delivery Hero purchase, saying it involved roughly 10% dilution for an estimated 24% revenue increase “or something crazy like that,” and said acquisitions plus SpaceX distributions were restoring LP and family-office appetite for venture exposure.
7. Apple’s suit turns talent mobility into a trade-secret boundary test
Apple’s 41-page complaint, filed July 10, alleged that OpenAI used stolen trade secrets in developing consumer hardware. Former iPhone-design vice president Tang Tan allegedly asked candidates to bring “actual parts” for interview “show and tell,” while former engineer Chang Liu reportedly texted, “LOL. I found out I can access the network storage. So funny.”
Jason highlighted the scale—more than 400 Apple employees recruited over a year or two—and the damaged relationship, given that ChatGPT had been expected to serve as the iPhone’s default AI. Chamath nevertheless refused to gossip before adjudication, noting only that Apple’s unusual willingness to litigate suggested it was deeply upset.
The panel’s rule survived all factual uncertainty: employees may change jobs in California and carry accumulated knowledge, but no physical parts, drives, documents or files. Sacks’s formulation was absolute: “The only thing you can bring to your new job is what’s in your head. That’s it.”
8. Grok Build shows why zero retention is not a trust boundary
xAI’s Grok Build, powered by Grok 4.5, reportedly sent entire developer repositories to SpaceX cloud servers despite assurances that no codebase data was transmitted. That potentially included passwords, API keys and change logs; the privacy setting failed, the upload was disabled server-side on July 13, and Elon Musk said previously uploaded data had been deleted.
Chamath’s conclusion was broader than one apparent implementation error: “Privacy in AI is very fragile, and it’s very brittle.” Even providers sincerely offering zero data retention cannot guarantee that unknown leak vectors do not exist; “It’s not gonna be okay” merely because a ZDR switch is enabled.
He advocated an independent layer between enterprises and models, while disclosing that this is part of what 8090 does for large enterprises through its software factory. The function is to control exposure across a “stratified ecosystem” instead of trusting every provider’s implementation.
Sacks cited a blog post by “Sacha” titled “The Reverse Information Paradox.” The post extended Alex Karp’s argument that enterprises want control over compute, models, weights, data and alpha, recommending private evaluations, proprietary learning loops inside the tenant, decoupled orchestration and an explicit right to fine-tune their own output.
9. Token spending is heading from engineering budgets into earnings calls
Chamath’s cited pricing table put one million tokens on Lovable at about $56, Bolt at about $26—the same figure he gave for Claude 4.8—Grok and Zuck’s model at roughly $1.50, Elon’s at about $1, and Chinese models at $0.50. His objection was the combination of premium pricing and the risk of surrendering proprietary knowledge.
Ramp CEO Eric Gleiman said token spending among customers had grown 21 times in one year and introduced Token Spend Management for both Ramp and non-Ramp customers. AI providers have effectively created an uncapped “tab,” he said, while CFOs struggle to see or control decentralized employee consumption.
Sacks connected that 21-times curve directly to earnings: a company could eventually miss a quarter because token OpEx ran ahead of controls. Engineers naturally select “the latest, greatest model” without owning ROI; Chamath said CFOs must decide whether 95% of work belongs one tier below the frontier at perhaps 1/100 of the cost.
Jason’s counter-position was that cheap local inference could make Apple “a screaming buy.” He cited an M7 Ultra configuration with as much as 1.5 TB of memory and imagined 90-99% of workloads running on a $4,000-$5,000 Mac Studio; Sacks and Chamath immediately rejected his claim that the original iPhone had been widely laughed at.
10. The AI buildout is colliding with an acute shortage of electrons
Chamath projected that by 2050 the US will lack energy equal to 2.5 times California’s current consumption. At a PGM auction spanning Pennsylvania, New Jersey, Maryland and other states, he said the system sought roughly 7-8 GW but received only about 156 MW.
In his data-center portfolio, sites with verifiable power available today command extreme front-end prices because roughly 40% of prospective projects are being mothballed or stopped. The constraint could prevent demand for drug discovery, cancer diagnosis, healthcare and legal services from being served even when models and customers exist.
Behind-the-meter generation avoids waiting for a utility interconnection by producing power on-site. Chamath described Elon Musk using mobile turbines for Colossus in Memphis, navigating clean-air permitting, and cited Bloom Energy as another route to large on-site installations; the regulatory complexity remains even when a facility brings its own power.
Sunrun and Span, the latter partnered with NVIDIA, were cited as building distributed residential data-center blocks. Combined with local models, solar and batteries, that suggests fragmentation toward edge compute—but Jason’s simpler near-term rule was “GPUs chasing energy.”
11. New York’s moratorium makes powered sites scarcer and moves investment elsewhere
Governor Kathy Hochul announced what she called the nation’s first statewide moratorium on hyperscale data centers, citing fossil-fuel emissions, land displacement, higher utility bills, depleted water and noise. Sacks disputed every premise and argued that behind-the-meter power can avoid competition with residential ratepayers.
His rebuttal cast data centers as unusually efficient land use with manageable noise, closed-loop water systems and substantial construction, operating and tax revenue. He cited a study equating typical water consumption to roughly 2.5 In-N-Out Burger locations, while Chamath recalled teachers receiving $30,000-$40,000 bonuses from local tax windfalls.
Sacks interpreted the moratorium as potentially temporary leverage: pause construction until a future administration can dictate a regulator, speech controls and other conditions. Even if lifted in 2.5-3 years, he estimated project restart times could leave New York without another operating facility for at least five years.
Export controls compound the domestic constraint by limiting chip deployment in allied countries. Chamath had expected more Middle Eastern construction but instead described rapid growth in Asia, specifically Australia; he noted that the Middle East could serve roughly four billion people in under 200 milliseconds, given its geography and energy availability.
12. The panel sees a moral panic arriving before measurable catastrophe
Rupert Darwall, in a clip played on the show, used anti-GMO sentiment as an analogy, noting its rise after Russia Today entered the US in 2010 and decline as that outlet lost distribution. He stopped short of proving causality but worried that NGO funding, media repetition and social amplification could similarly reflect foreign interest in anti-data-center activism.
Sacks went further, citing an OpenAI post titled “PRC-linked influence operations are targeting AI debates in the US.” Chamath argued that if China could encourage Anthropic to pull up the ladder, kill open source and leave US users paying 50-100 times more per token, a merely adequate foreign competitor could overcome a technical disadvantage. Sacks agreed that the incentive for foreign governments to influence US policy was obvious.
Sacks’s urgency came from Kimi K3, which he said was “very, very close to the frontier”: America may have “months on China, if that.” Meanwhile, the previously proposed 10^25-FLOPS danger threshold has been crossed by every AI model, according to Sacks, without the predicted catastrophe, and Dario Amodei’s warning that 50% of entry-level knowledge jobs could disappear within one to five years remains unproven.
Friedberg’s preferred posture was “monitoring,” not panic—watch self-driving employment effects, hacking capability and job losses as evidence develops. Sacks reduced Anthropic’s alleged commercial strategy to three steps: “Brand yourself as a safe AI company. Number two: ban unsafe AI. Three: profit.”
13. AlphaFold helps engineer an enzyme that clears an aging byproduct
Friedberg shifted from Yamanaka-factor cell rejuvenation to the extracellular matrix. With age, sugars and fats bind to proteins through glycation, changing collagen and other structures, limiting repair and provoking inflammation; CML is a predominant advanced glycation end product that, he said, nothing breaks down.
A Calico and Reval Pharma collaboration, as named on air, used AlphaFold to identify a bacterial protein able to bind CML and start degrading it. Researchers then altered its DNA, produced hundreds and thousands of variants, screened their activity and repeated this directed-evolution process through five cycles.
The resulting enzyme reportedly removed 52-97% of CML from casein, collagen, retinal proteins and hemoglobin, with several sites exceeding 90%. On donated skin from people older than 70, it eliminated 55%, which Friedberg described as reversing the tissue’s measured age to that of a 31-year-old.
The unresolved question is delivery: cream, injection, supplement or an RNA shot that produces the enzyme inside the body. Chamath predicted cosmetics would precede joint or systemic treatment and called a successful facial cream a potential $1 trillion—and moments later $2 trillion—market; Friedberg’s larger point was that this was “AlphaFold used to discover this thing and evolve it.”