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

This All-In Podcast episode covers six stories: frontier AI regulation, a reported PayPal bid, Apple's lawsuit against OpenAI, enterprise AI privacy and token costs, New York datacenter policy, and an enzyme designed to remove age-related protein damage.

Demis Hassabis proposes a US frontier AI standards body modelled on FINRA. Government would supervise it. Industry would fund much of it. Independent technical experts and open-source representatives would help govern it. Frontier labs would initially submit models voluntarily for review up to 30 days before release. Passing could later become a condition for deployment in the US.

The hosts like the prospect of technical staff who can update tests faster than Congress or a conventional agency. They contrast it with aircraft certification, which can take years. FAA timelines do verify the aviation example, but they do not predict how long an AI review would take.

Their support is conditional. The body would need broad representation, a narrow frontier threshold, exemptions for smaller developers and open-source work, and federal pre-emption of overlapping state laws. Several hosts worry that large labs could write rules that burden smaller competitors. They point to Anthropic's support for state AI laws as evidence of regulatory capture. Anthropic supports laws including California SB 53. Anthropic says it prefers a federal standard and that those state laws target the largest developers. Regulatory capture is the hosts' interpretation.

FINRA also remains a government-backed system. The SEC supervises FINRA, while the CFTC oversees NFA. Congress sets and oversees the wider framework rather than running either body's daily work.

Foreign cooperation is unresolved. China and 28 other countries established the World Artificial Intelligence Cooperation Organization in July 2026. The UN runs an AI scientific panel and governance dialogue. A US-led body could not assume authority over foreign labs or a coordinated slowdown.

A reported offer valued PayPal at about $53 billion. The hosts describe a consortium involving Stripe, Advent and Block. The latest reporting said Stripe and Advent made the $60.50-per-share offer, backed by about $50 billion of bank financing and $17 billion of equity. Block had exited after an April approach.

The deal was not signed. PayPal had not formally accepted the proposal, and its board reportedly viewed the price as inadequate.

PayPal reported 439 million active accounts and $1.79 trillion in 2025 payment volume. Those accounts include consumers, merchants and platforms. Stripe reported $1.9 trillion in volume and more than five million businesses. PayPal brings Venmo, Braintree and PYUSD. Stripe owns the stablecoin infrastructure company Bridge.

They imagine Stripe operating the combined company, moving more payments onto cheaper internal rails and building a stronger alternative to Visa and Mastercard. They also speculate about a higher offer, another bidder and large AI-driven cost cuts. None of those points was part of a disclosed agreement.

Antitrust treatment would depend on the market regulators examine. Competition with card networks may support the deal. Combining Stripe with Braintree, consumer payment products and a multi-sided platform may create narrower concerns. The hosts argue the first point and acknowledge the second without resolving it.

Apple filed a 41-page federal complaint on July 10, 2026 against former employees, OpenAI entities and io Products. It alleges that confidential Apple hardware information was taken and used in OpenAI's device programme.

The complaint says former Apple engineering executive Tang Tan directed job candidates to bring physical parts to interviews. It says former engineer Chang Liu retained an Apple laptop and used an authentication flaw to reach shared folders after joining OpenAI. Apple also says more than 400 former Apple employees work at OpenAI. That number does not establish that all were recruited improperly or involved in the alleged conduct.

OpenAI denied interest in other companies' trade secrets. No court has ruled on Apple's allegations. The hosts say this several times, although some still infer that Apple would not sue without strong evidence.

California generally rejects employment noncompetes, so employees can change jobs. That does not permit taking or using protected information. The hosts' rule—leave with nothing except what is in your head—is sensible operational advice, but federal law also protects qualifying intangible information. One defendant admitted theft and received a conviction in the Waymo-Uber litigation; that result does not predict how a court would decide Apple's claims.

The companies did have a public relationship. ChatGPT is available through Siri, Writing Tools and visual intelligence, but users choose whether to use it. It is not the default for every iPhone interaction.

A researcher testing Grok Build `0.2.93` found that the coding tool uploaded a full Git bundle, including files it was told not to read and repository history. Turning off model improvement did not stop that storage path. The upload stopped on July 13 after a server flag changed.

The hosts use the incident to argue that enterprise AI privacy is brittle. A vendor can promise zero data retention while an auxiliary client or storage path behaves differently. SpaceXAI's documentation says ZDR is enforced for enabled enterprise teams. The Grok incident proves a specific implementation failure, not that every vendor or ZDR service leaks.

They want enterprises to control private evaluations, learning loops, orchestration, data and model outputs. They also expect price competition from open models. SpaceXAI lists Grok 4.5 at $2 per million input tokens and $6 per million output tokens. Anthropic uses different input, output, cache, batch, context and fast-mode rates. The podcast's quoted provider prices cannot be compared cleanly without those dimensions.

Ramp's AI Token Spend Management consolidates provider costs and lets teams set alerts and limits. Ramp's customer data shows volatile usage, premium-model migration and large differences in cache adoption. Those vendor figures do not establish the hosts' prediction that token bills will cause public companies to miss earnings.

Thinking Machines released Inkling as an open-weights multimodal model intended for fine-tuning and agentic applications. It does not prove that Inkling matches frontier closed models or costs less for every workload.

The conversation then moves to local and distributed inference. The hosts say future Apple systems with high-memory configurations will run most enterprise AI workloads on the device. No Apple announcement reviewed for this section confirms an M7 Ultra system with 1.5 TB of memory. They are forecasting unannounced hardware.

They finish with electricity constraints. A PJM planning deck lists about 156.4 GW of operational capacity and 7.8 GW of summer load management. It does not describe an auction that needed seven or eight gigawatts but received only 156 MW. The podcast appears to combine separate figures.

New York Executive Order 62 pauses incomplete discretionary state permit applications for new or expanded datacenters capable of using at least 50 MW. The order does not cover smaller projects, completed applications, local permits, or facilities mainly used for manufacturing, research, education or medical care. The review is expected to take up to one year. The hosts' five-year forecast is not part of the order.

New York reported almost 12 GW of datacenter load requests in the NYISO queue as of May 2026. Queue requests are not completed projects.

The hosts argue that operators should bring generation and pay for upgrades instead of being blocked. New York's order already directs regulators to consider upfront contributions, dedicated clean power, storage and protection against stranded assets. Seven hyperscalers also signed a federal pledge to build, buy or bring generation and cover required grid upgrades. Those policies share the hosts' basic cost-allocation idea.

Lawrence Berkeley National Laboratory estimated direct datacenter water use of about 0.36 litres per kWh through 2023 and a much larger indirect footprint from electricity generation. Closed liquid loops can still use cooling towers that lose water through evaporation. Water use varies by design and location; it is not a hoax.

Natural gas emits less carbon dioxide and fewer air pollutants than coal or petroleum per unit of energy. It still emits carbon dioxide, and production and transport can leak methane. Behind-the-meter gas can bypass a grid bottleneck without removing its environmental cost.

The hosts treat the dispute as part of a race with China. The IEA estimated that the US used more datacenter electricity than China in 2024 and will add more absolute consumption through 2030, while China's percentage growth is faster. The US has not lost its infrastructure lead.

OpenAI identified a PRC-linked influence operation that produced content about US datacenter costs, but found no meaningful reach beyond the operation's own activity. Gallup found that 71% of US adults opposed a datacenter in their local area. Foreign influence does not explain broad domestic concern.

Science Corner mentions Yamanaka factors, then turns to a different mechanism. A *Nature Communications* paper engineers an enzyme called CMLase to remove Nε-carboxymethyl-lysine, an age-associated modification found on long-lived proteins.

The researchers screened protein structures, found a weak bacterial enzyme and improved it through five rounds of directed evolution covering more than 500 million variants. The final enzyme reduced CML signals by 52–97% across treated protein substrates. CML staining fell in donated human eye, artery and skin samples after treatment. Elderly skin sections showed a reduction of more than 55%.

These were in vitro and ex vivo experiments. No living animal or person received CMLase. The claim that old skin was returned to “age 31” refers to one staining comparison with younger tissue. The study did not measure wrinkles, elasticity, wound healing, pain, mobility or clinical appearance.

Delivery remains unsolved. The experiments placed purified enzyme directly onto proteins, homogenates or thin tissue sections. A cream would need to cross intact skin. An injection or RNA treatment would need to reach the extracellular matrix at an effective dose. Immune reactions to a bacterial enzyme, off-target activity and reaction byproducts also need testing.

Yamanaka-factor research is separate. Partial reprogramming has changed molecular-age measures or improved function in cultured human cells and mouse models. Other mouse work produced teratomas when reprogramming reached pluripotency. Those results do not establish safe human rejuvenation, and that tumour mechanism is not evidence that CMLase causes cancer.

Revel Pharmaceuticals filed patent applications for CMLase and lists skin aging, gout and eye disease as development targets. The podcast's predicted cream, first cosmetic market and $1–2 trillion opportunity remain speculation.

Subject: AI rules, PayPal, datacenters and CMLase

Preview text: Six All-In discussions, checked against the source record.

The All-In hosts open with Demis Hassabis's proposal for a federally supervised, industry-funded body to test frontier models before release. They like technical expertise and faster-changing tests, but worry that large labs could shape the rules, burden startups and create another layer above state laws. Anthropic's state advocacy is real; the claim that it amounts to regulatory capture is the hosts' interpretation. A US body would also need cooperation from existing UN and China-led processes.

They next discuss a reported $53 billion offer for PayPal. The latest bid came from Stripe and Advent; Block had left after an earlier approach. No deal was signed. The hosts see strategic value in combining Stripe with PayPal, Venmo, Braintree, PYUSD and Bridge. Their operating plan, higher-price prediction and consumer-savings case are speculation.

Apple's July 10 complaint alleges that former employees took confidential hardware information to OpenAI. OpenAI denies interest in other companies' secrets, and the claims remain unproved. California protects job mobility, not trade-secret use. Apple's ChatGPT integration is optional rather than the iPhone's universal default.

A verified Grok Build incident anchors the privacy discussion. During a researcher's test, Grok Build uploaded complete Git repositories and their commit histories until a server flag stopped the uploads. That is a concrete failure, not proof that every zero-retention service leaks. The hosts then move through token-price gaps, Ramp's spend controls, Inkling's open weights and unconfirmed predictions about local Apple inference. Their PJM electricity anecdote appears to combine a 156.4 GW capacity figure with 7.8 GW of load management.

New York's order pauses incomplete state permit applications for some facilities of at least 50 MW during a review expected to last up to a year. It is not a blanket ban. The hosts' proposal that operators bring power and pay for upgrades overlaps with provisions already under review. Water and grid impacts vary by project, but they are real. The US still uses more datacenter electricity than China, although China is growing faster proportionally. A PRC-linked influence attempt had little reach; domestic opposition is widespread.

Science Corner covers an enzyme that removed the CML aging marker from isolated proteins and donated human tissue. It was not Yamanaka-factor reprogramming. No living animal or person was treated. “Age 31” refers to one staining comparison, not restored skin function. Delivery, safety and clinical benefit remain untested. A cream and trillion-dollar market are forecasts.

Watch the original episode