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

Episode Summary

Executive Summary: The episode centered on AI governance, arguing for a self-regulatory organization (SRO) over a slow government “FAA for AI,” while warning about regulatory capture, state-by-state AI rules, and foreign influence. The hosts also debated a Stripe/Block-led bid for PayPal, Apple’s lawsuit against OpenAI, a Grok code-leak incident, data-center moratoriums, and a breakthrough AI-designed enzyme that reverses skin-aging markers.

Main Topics: AI regulation: SRO vs. government agency (Priority: 5/5): The hosts endorsed Demis Hassabis’s proposal for a federally overseen but industry-run AI standards body, arguing it is preferable to a new government agency that would create delays, capture risk, and kill innovation. Regulatory capture and Anthropic’s role (Priority: 5/5): A long segment accused Anthropic of promoting tougher state-level AI rules and fear-based regulation, with the hosts arguing this strategy could entrench incumbents and constrain open source and startups. Stripe/Block/Advent bid for PayPal (Priority: 4/5): The panel discussed a reported acquisition bid for PayPal, framing it as part of a broader wave of mega-deals to revive legacy digital businesses and potentially build a payments network that competes with Visa and Mastercard. OpenAI legal and talent disputes (Priority: 4/5): The hosts covered Apple’s lawsuit alleging OpenAI misused trade secrets related to consumer hardware and discussed the broader pattern of lawsuits and employee poaching around OpenAI. AI privacy, leaks, and enterprise control (Priority: 5/5): A Grok Build incident where code was allegedly uploaded despite privacy assurances led to a broader discussion of brittle AI privacy, enterprise trust boundaries, and the need for third-party orchestration layers. Data centers, energy, and anti-AI politics (Priority: 5/5): The episode criticized New York’s moratorium on hyperscale data centers and argued that energy constraints, permitting, and anti-data-center activism are major bottlenecks to AI progress and economic growth. Science Corner: AI-designed enzyme for age reversal (Priority: 4/5): Friedberg explained a Calico/Revel Pharma study using AlphaFold and directed evolution to create an enzyme that degrades CML glycation products, reversing aging markers in human skin samples.

Key Arguments: An SRO is preferable to a government AI agency because the government lacks the expertise to evaluate rapidly changing models. The SRO should only cover frontier models, catastrophic risks (cyber and CBRN), be voluntary at first, and substitute for rather than add to new regulation. Broad industry representation, including startups and open source, is necessary to avoid regulatory capture. The hosts argued Anthropic is driving a patchwork of stricter state AI laws, which could ultimately lock in incumbents and harm open innovation. Lower-cost open models and AI cost controls will pressure expensive frontier offerings and reshape enterprise buying behavior. PayPal’s underperformance and huge account base make it a target for operational revival and possible payments-network consolidation. The reported Stripe/Block/Advent deal could be pro-competitive if it creates stronger alternatives to Visa and Mastercard through merchant-consumer integration. Apple’s lawsuit suggests OpenAI may have crossed a line in talent and IP acquisition, even though the facts remain alleged. AI privacy is fragile; zero-data-retention promises may not fully protect users because hidden upload pathways and trapdoors can exist. Data centers are economically valuable and generally cleaner and more efficient than critics claim; moratoria risk slowing AI, jobs, tax revenue, and infrastructure buildout. The Science Corner emphasized that AI is already enabling real biomedical advances, not just speculative risks. The enzyme study shows a plausible path from AI-generated protein design to cosmetic and therapeutic anti-aging products.

Data Points: Frontier model review delay: 30 days before release - Demis Hassabis proposal would require frontier labs to submit models before launch. Initial AI standards approach: Voluntary first, then mandatory later - SRO would begin as optional before possible legal enshrinement. AI risk categories: Cybersecurity, national security, biological threats, CBRN - The proposed body would benchmark models against catastrophic domains. Possible FAA certification timelines: 5-9 years for new aircraft; 3-5 years for major amendments - Used as a cautionary analogy for how slow AI regulation could become. PayPal consumer accounts: 439 million - Used to illustrate the company’s scale and why it remains strategically valuable. PayPal stock peak value: About $322 billion - Referenced as the company’s prior valuation high before its decline. Reported Stripe/Advent bid: About $60 per share - Initial reported offer for PayPal. Potential higher clearing price: $70 per share or 10-15% higher - Hosts speculated the final price could rise above the reported bid. Stripe annual transaction volume: About $2 trillion - Compared with PayPal to show Stripe’s scale in merchant payments. PayPal annual transaction volume: About $1.7 trillion - Used to compare PayPal’s scale to Stripe’s. OpenAI/Apple employee poaching: Over 400 Apple employees - Mentioned as part of the dispute over talent and IP migration. Token price on Fable: About $56 per 1 million input tokens - Used to illustrate high-cost frontier model usage. Token price on Seoul/Quad 4.8: About $26 per 1 million input tokens - Compared against frontier models to highlight cost differences. Token price on Grok: About $1.50 per 1 million input tokens - Example of dramatically cheaper model usage. Token price on another Grok reference: About $1 per 1 million input tokens - Used to show how inexpensive some models have become. Token price on Chinese models: About $0.50 per 1 million input tokens - Used to argue U.S. firms must lower prices or lose market share. Token spend growth at Ramp customers: 21x - Eric Glyman said AI spend is surging, creating CFO control needs. Data-center energy deficit by 2050: 2.5 California equivalents - Friedberg argued the U.S. will be massively short on electricity. PGM forward auction need vs. supply: 7-8 GW needed vs. 156 MW supplied - Illustrated the severe shortfall in available power capacity. NYSE moratorium: Statewide moratorium on hyperscale data centers - Kathy Hochul announced a pause on new large-scale data centers. Human skin sample result: 55% CML eliminated - Calico/Revel enzyme reduced glycation markers in elderly donated skin. Age-reversal comparison: Skin age reduced to roughly 31-year-old levels - Based on the treated skin’s biochemical marker profile. Model frontier concern: 10^25 FLOPs threshold discussed - Referenced as a prior policy benchmark that was already surpassed by later models. Predicted job-loss claim: 50% of entry-level knowledge-worker jobs in 1-5 years - Attributed to Dario Amodei as an example of alarmist AI rhetoric.

Pivotal Quotes: "The industry should have an industry certification like they do for countless other things." — Sacks: Arguing for self-regulation instead of a government AI agency. "This is not a little startup. They already have a trillion-dollar market cap valuation." — Sacks: Describing Anthropic as a major force shaping regulation and policy. "The whole thing is a money-burning furnace." — Sacks: Warning that uncontrolled AI token spend could hurt enterprise earnings.

Implications: The episode argues AI’s future will hinge on regulation design, compute/energy access, and cost control. Listeners should expect more battles over open source, state rules, data-center buildout, and enterprise AI economics, alongside rapid scientific spin-offs.