Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters
In a Nutshell
Demis Hassabis proposes a FINRA-style AI self-regulatory organization that frontier labs would join voluntarily before mandatory enforcement, with industry experts reviewing only catastrophic risks like cyber and biological threats. Stripe's $53B PayPal acquisition bid signals a wave of AI-native operators buying stale digital businesses to revive them through AI implementation, while New York's data center moratorium and similar state-level restrictions threaten to cede AI infrastructure advantages to China. The core tension is that companies like Anthropic are strategically pushing for heavy regulation that could create 50-100x cost disadvantages for American firms versus open-source Chinese models, potentially handing China the AI race despite US technological lead.
These notes were generated by AI and may contain inaccuracies.
DeepMind CEO Demis Hassabis published an article proposing a US-led international AI standards body modeled after FINRA, the Financial Industry Regulatory Authority. The proposed self-regulatory body would be federally overseen but industry-funded and run by independent technological experts. Frontier Labs would submit their models 30 days before release, initially on a voluntary basis with mandatory requirements to follow. Models would be assessed on risks to cybersecurity, national security, biological threats, and other high-risk domains, with benchmarks updated quarterly. The body would have authority to coordinate a slowdown in development if the situation demands it.
The proposal has received support from Elon Musk, Sam Altman at OpenAI, Jack Clark at Anthropic, Sundar Pichai, Satya Nadella, Jack Dorsey from Block, and the Carlson brothers.
An SRO is a self-regulatory organization like FINRA and the National Futures Association that exists in financial markets. These organizations allow financial institutions to set their regulatory rules, check each other, and ensure safety because they trade risk with one another. The industry wants to avoid exposure and prevent things from getting slowed down, which would make markets inefficient.
The AI analogy is appropriate because multiple players are trying to progress AI technology without a single regulatory body from government or outside dictating tests. California's attempt to pass AI legislation a year and a half ago showed that none of the rules and requirements made sense or mapped to the technology of the day. An SRO can adjust how tests are run and ensure the right experts are involved, with federal government oversight but not control.
FINRA and NFA report to Senate and House committees that provide oversight. The SRO concept would bring experts from industry to assess models for cyber risk, bio risk, weapons risk, and social manipulation. This independent body could be voted on and changed over time, operating faster than a new government agency because of expertise in running software tests.
The industry recognizes the need for oversight and checkpoints. The SRO approach solves the problem by getting the right people to review models while maintaining government oversight, rather than handing control to government. Anthropic and OpenAI have both signed on to the proposal, with Dario Amodei's president giving thumbs up and Sam Altman also supporting it.
David Sacks spoke with Demis Hassabis and outlined five conditions he believes are essential for the SRO to work effectively:
-
The SRO must have broad representation from within the AI industry, including startups and open source, not just the three biggest labs, to avoid regulatory capture. This would include Jensen Huang, Elon Musk, Mark Zuckerberg, and potentially Mira Murati because of her open platform launch.
-
The body should only review frontier models representing true advances in the state-of-the-art. Models below this level should not be held from market, as regulations could be used by market leaders to tie up lesser models.
-
The body should deal with catastrophic risk only, such as cyber and CBRN risks (chemical, biological, radiological, nuclear). It should not address disinformation, microaggressions, or become a speech regulator.
-
The organization should be voluntary first, proving it works before becoming legally mandatory.
-
The SRO should be a substitute for a new regulatory agency, not additive to new regulatory structures.
Even if the SRO is established, concerns remain that it could be an opening bid for more regulation. Anthropic has expressed support for this approach, but it may be a stepping stone to what Dario has called for: an FAA for AI. The SRO must be kept honest and pure, otherwise it becomes the opening bid in a wave of regulation and a vehicle for massive regulatory capture.
Anthropic is pursuing a strategy of one-upmanship that encourages states to impose increasingly tougher AI guardrails rather than aligning around a single set of regulations. The company gets regulations passed in one state like California's SB53, then makes regulations more strict and all-encompassing with each new state. This creates a patchwork rather than a single national framework, using state-level pressure to impose more regulations.
Anthropic has won in California, Illinois, New York, and other blue states, and possibly some red states. States have sovereignty rights and will decide on regulations, as seen with self-driving cars.
The FAA for AI is an extreme proposal. The FAA requires type certification for new aircraft designs or major changes. For entirely new aircraft designs, certification takes 5 to 9 years. For amendments to certificates for major changes, it takes 3 to 5 years. The Boeing 737 Max took about 5 years. This permission-based regulation would replace a system releasing new AI versions every couple of months with one fully under government control, where everything must be certified. Timelines would extend from months to years, causing the US to lose the AI race since China would not abide by the same rules.
When companies go to government asking to be regulated, few people in government will say they are not qualified or do not want power. Most will accept the offer and come back for more. The mistake tech companies make is thinking they can buy off politicians through concessions, which will lead to ratcheting up of pressure. Government will take what is offered and return for more until achieving full control. Companies need to grow a spine, fight, and decide where to draw lines. If the SRO is positioned as the line with preemption and other protections written into law, it could work. Offering it up freely will not be the end, just the opening bid.
Stripe, still a private company, is bidding $53 billion for PayPal along with private equity firm Advent. The offer is approximately $60 per share, a small premium, though most expect the final price to be closer to $70 per share. PayPal stock jumped on the news. Block, formerly Square and Jack Dorsey's payment company, is contributing $17 billion in equity to the combined offer.
PayPal has 439 million consumer accounts and has maintained its position 25 years after founding. However, the product is growing only 7% per year on a large base, and the product has become somewhat obsolete as a legacy product. The question is how Stripe would fix this, or whether they would run it more efficiently as a private equity play.
The combination of Stripe, Advent, and Block could create a competitor to Visa and Mastercard. The combined entity would have 600-700 million accounts, massive stablecoin infrastructure, and all of Stripe's risk management infrastructure. Stripe owns Bridge, its stablecoin infrastructure company acquired for $1 billion in 2025. PayPal has PYUSD, its stablecoin already in circulation. Stablecoins are part of this deal.
The deal would allow vertical integration and access to ultra-low-cost payment rails, bringing Stripe's infrastructure everywhere in the world. Advent can pour money into the combined entity. This addresses the early critique of Stripe's business model, which required building many value-added services because the take as a middleman on traditional rails could get competed away.
A new wave of AI-native operators is targeting mature first-generation digital native businesses that have become stale. These companies are no longer run by founders, have not realized their AI potential, and are overspending. When viewed through the lens of a modern AI operator, these businesses show clear underutilization of networks, poor operations, overspending, and inadequate AI implementation.
Capital markets have already seen examples like Josh Kushner's rollup of accounting firms and General Catalyst's project of using capital to buy traditional services businesses and "AIify" them. The same opportunity exists for traditional digital businesses including software companies and network businesses from the early internet era that have stalled out in public markets.
The key challenge for capital providers is finding the right operators to execute operational revival, as traditional McKinsey consultants are not suitable. Ryan Cohen demonstrated this capability with Chewy and GameStop. For payments businesses specifically, Stripe and potentially Jack Dorsey are considered ideal partners.
A $17 billion equity contribution is proposed where Stripe and Block sell $17 billion of equity to cash investors. The cash then buys PayPal, resulting in Stripe and Block owning pieces of PayPal while private equity investors own pieces of Block and Stripe. The deal documents do not specify who will operate PayPal post-close, but the prediction is that Stripe operators will take control given their qualifications and stake.
This transaction represents the beginning of a wave of mega deals involving flaccid digital businesses that can be revived with capital and the right operator. The current price is not considered final, with expectations of another 10-15% increase.
Bending Spoons exemplifies the operational approach, having gone public after acquiring non-founder-led assets including AOL for $1.4 billion, Vimeo for $1.4 billion, WeTransfer, Eventbrite, and Bright Code. The company also acquired Evernote and operates from Milan. Their methodology involves diagnosing overspending and under-spending, fixing product and marketing issues, and revitalizing Web 2.0 businesses through AI-first executives.
Venture capital faced challenges under Lina Khan's regulatory approach, but following the Trump election, corporate development executives have pursued M&A deals aggressively. Recent examples include Uber's acquisition of Delivery Hero, which will increase revenue by approximately 20% despite 10% dilution. Liquidity providers including LPs and family offices are inquiring about new funds and deals, driven by SpaceX distributions and increased M&A activity.
PayPal's stagnation traces back to its 2002 acquisition by eBay under Meg Whitman, who had a corporate background from Procter & Gamble, Disney, and Bain. The acquisition led to the removal of founding talent, creating what is termed the "PayPal diaspora" rather than mafia. Unlike typical acquisitions that lock up talent, PayPal's founders were effectively locked out. Current CEO Enrique is described as performing well in tightening operations.
The primary synergy in contemporary acquisitions is AI implementation, enabling automation, product development, efficiency improvements, and user experience enhancements. However, PayPal faces an existential challenge with its 25-year-old product interaction model that requires fundamental resuscitation beyond mere efficiency gains.
The proposed combination would merge Stripe's merchant relationships (2 trillion annual transaction volume) with PayPal's 400 million consumer accounts and BrainTree ownership. Block's point-of-sale infrastructure and Cash App would complete the ecosystem, potentially enabling bypass of credit card networks through on-us transactions.
The deal would lower prices for consumers through competitive pressure on Visa and Mastercard. If defined narrowly as APIs for merchants, the consolidation of Stripe and BrainTree would appear anti-competitive. However, when viewed against the Visa-Mastercard duopoly, PayPal's addition creates pro-competitive outcomes in a vastly larger market.
On July 10th, Apple filed a 41-page lawsuit against OpenAI alleging stolen trade secrets for developing consumer hardware devices. Key allegations include:
- Tang Tan, Apple's former VP of iPhone design, now serves as OpenAI's chief hardware officer
- Alleged direction to Apple job candidates to bring "actual parts" to interviews for show-and-tell
- Changu, former Apple senior technical engineer, sent text messages about accessing network storage
- OpenAI has poached over 400 Apple employees
The lawsuit marks Apple's first major litigation in 25 years. The relationship breakdown is notable given ChatGPT's previous status as default iPhone AI.
California's employment-at-will system prohibits non-competes, but fundamental rules apply: employees may only bring their memories and knowledge, not physical documents, thumb drives, CDs, or any other materials. The only acceptable transfer is "what's in your head."
SpaceX's Grok Build, powered by Grok 4.5 and launched in public beta at end of May, experienced a data leak. Despite promises that codebase information would not transmit to xAI servers, the tool was sending entire codebases including passwords, API keys, and change logs. The privacy setting failed to prevent this transmission.
SpaceX disabled uploads on July 13th and Elon Musk promised deletion of previously uploaded data. The company subsequently open-sourced Grok Build's harness.
AI privacy remains fragile despite best efforts by trustworthy companies. Zero Data Retention (ZDR) policies cannot guarantee complete protection due to non-obvious data leak vectors and trap doors that may not be discovered until identified by external parties.
This necessitates a stratified ecosystem with third-party layers to interface with models and manage exposure. Enterprise recommendations include establishing trust boundaries with private eval, proprietary learning loops inside tenants, decoupled orchestration, and explicit rights to fine-tune outputs.
Token pricing per million input tokens varies dramatically across providers:
- Fable: $56
- Soul: $26
- Quad 48: $26
- Grok: $1.50
- Z: $150
- Elon/Grok: $1
- Chinese models: $0.50
Perplexity has begun supporting both Grok and GLM52, allowing users to compare model performance on identical tasks. A podcast player deep-linking feature was built for $11 using Grok versus significantly higher costs on other platforms.
RAMP launched token spend management tools available to both RAMP and non-RAMP customers. Token spend among RAMP customers has grown 21 times over the past year. CFOs are often surprised by AI bills because the current model functions like an open tab with no visibility into proactive spending. New models frequently increase rates, creating misaligned incentives between AI companies and their customers.
Engineers running unguided systems can burn through millions of tokens at $56 per million tokens, creating downstream impacts on company earnings. Public market CFOs may miss earnings quarters if token costs continue growing at current rates, potentially impacting earnings by dollars rather than pennies.
The AI spending growth creates massive market opportunities. Companies like Sachs, Bit Tensor, Subnets, GLM52 hosting Grock 4.5, and Inkling are offering competitive pricing. People currently paying $26-56 per million tokens could be paying 50 cents instead.
Inkling's value proposition focuses on being a platform for fine-tuning open models that are significantly cheaper than frontier models while achieving desired results. However, Anthropic has incentive to stop open models because they sell most products at 50 cents per million tokens versus their own pricing at $56.
Despite open model competition, frontier AI companies are still the industry leaders in revenue growth. Eric Glenn would not have released RAMP's spend management product unless CFOs reported inability to control AI spending.
Engineers prioritize using the latest, most advanced models regardless of ROI, while CFOs are directly tied to financial outcomes. For 95% of tasks, users could operate at one level lower model costing 1/100th the price, but engineers never make this cost trade-off.
The situation resembles corporate travel where employees don't see prices and default to business class and premium hotels because the travel department handles costs.
Mark German predicts M7 Ultra chips will support up to 1.5 terabytes of RAM, doubling current M5 chip capabilities. This would enable running frontier-level models like Claude Opus locally on Mac Studios. Employers could run 90-99% of workloads on local Mac Studios rather than cloud services.
Apple's hardware monetization strategy through devices could put significant downward pressure on Claude and OpenAI by supporting local models with advanced memory architecture, creating unlimited local tokens on desktop computers.
Sunun announced distributed data center blocks for residential installation. SPAN partnered with Nvidia on similar distributed solutions. This represents fragmentation and distribution of edge compute as a major theme.
The US faces a massive energy deficit projected at 2.5 times California's consumption by 2050. PJM (Pennsylvania, New Jersey, Maryland utility serving 13 states) held an auction needing 7-8 gigawatts but received only 156 megawatts in responses.
Elon Musk purchased 18-wheeler engines to power data centers behind the meter, avoiding grid connection delays. This approach circumvents clean air permitting requirements when engines operate as individual units rather than blocks. Bloom Energy provides similar solutions allowing large installations under personal use clean air permits.
Governor Kathy Hochul signed the nation's first statewide moratorium on hyperscale data centers, citing power consumption, land use, agricultural displacement, utility bills, water supply, and noise pollution concerns.
Data centers are being scapegoated for broader AI concerns. The moratorium represents a clumsy attempt to slow innovation. New York State is 80% underdeveloped, with 70-80% of upstate land undeveloped, contradicting claims of insufficient space.
Data centers are models of land use efficiency with excellent economic ROI relative to land consumption. Modern data centers use closed-loop water recirculation systems, consuming equivalent water to 2.5 In-N-Out burgers. Natural gas powering most data centers is among the cleanest burning power sources available.
North Dakota data centers generated sufficient tax revenue to provide teachers $30,000-40,000 bonuses. Data centers create ongoing blue-collar jobs beyond initial construction phases.
Senator Dave McCormack's defense and innovation summit revealed common funding patterns between current data center protesters and previous fracking protesters. These appear to be professionally paid protesters using AI concerns as hobby horses for fundraising and employment.
Anthropic continues funding groups like Public First seeking to halt data center construction through new regulations. Dario Amodei made seven-figure contributions to such organizations, despite compute availability being the primary constraint on Anthropic's revenue growth.
Democratic strategy appears to involve pausing data centers until achieving sufficient control to dictate regulatory terms. A moratorium lasting until a future Democratic administration could result in 5+ years before new data centers become operational in affected states, accounting for project ramp-up time.
Data center assets with verifiable energizable power today command extreme premiums due to 40% of projects being mothballed. This creates massive deficits in available energy for AI applications including drug discovery, cancer diagnosis, healthcare, and legal services.
Opposition to data centers extends beyond the US to allied countries through export controls on chips. Countries with unlimited energy like Middle East nations represent attractive alternatives. UAE can now import best-in-class chips, potentially restarting Middle East data center development.
Australia saw unexpected data center growth despite being further left politically. Countries embrace data centers when debt-to-GDP ratios, productivity, or foreign direct investment require the economic benefits.
Larger Starlink enterprise versions can bundle to achieve 10-20 gigabit setups, enabling data center placement in previously inaccessible locations.
Russia Today launched in the US in 2010 and was expelled in 2022. Prior to Russia Today's existence, there was no anti-GMO sentiment despite GMOs being commercially available since 1996. Anti-GMO sentiment originated from Russian media influence campaigns and declined as Russia Today coverage was reduced across media outlets.
KGB-designed "directed measures" during the Cold War created influence campaigns through foreign media targeting Western democracies. Similar patterns may apply to current nuclear energy opposition in Germany.
A series of recent pushes against data centers appear nonsensical when examined through rational, objective analysis of scientific and economic merits, yet activist communities have become heightened in opposition to them. Americans are being handicapped by unfounded, scientifically unsupported reasons for wanting to eliminate these technologies. A poll released the day of recording showed over 50% of Americans believe data centers increase the cost of water and electricity, even when data centers fully recycle water and produce their own electricity. This reflects a deeply sewn psychological shift in the United States characterized by anti-rich, anti-tech, and anti-AI sentiment.
There appears to be foreign interest driving these influence campaigns. One month prior to recording, OpenAI published a blog post titled "PRC linked influence operations are targeting AI debates in the US," which was covered by Politico and other outlets. The post and subsequent reporting indicate China is behind influence campaigns to shape US attitudes on AI and data centers. This strategy makes strategic sense because stopping US infrastructure development represents a path for China to win the AI race.
If China can stop necessary infrastructure development, incentivize companies like Anthropic to restrict access, kill open source in the United States, and constrain demand for cheaper models, the economic consequences are severe. At $56 per million input tokens versus 50 cents for the rest of the world, American companies face 50-100x cost disadvantages. The math becomes impossible to overcome when costs reach these multiples. Enterprises are not just paying for AI with money but also feeding frontier models their proprietary knowledge and competitive advantages, creating a double disadvantage of higher costs and potential future mortgaging of competitive position.
A significant moral panic exists regarding AI, focused on potential future catastrophes that have not yet manifested. Claims about cyber risks and job losses have not materialized despite widespread discussion. The concern is that America may destroy its system of free market innovation—the culture of rapid iteration where anyone with a good idea can raise risk capital and start a company. The Overton window has shifted so dramatically that creating a FINRA equivalent for AI might be considered the least bad alternative, despite FINRA representing stock brokers writing rules with minimal sector innovation beyond Robinhood's free trading model.
Kimmy K3 was recently released and is now considered very close to the frontier, potentially leaving the US with only months of lead time over China. The discussion questions whether vast regulatory apparatus is needed when only a handful of companies operate at the frontier, all conducting safety testing and red teaming. The regulatory push appears premature and driven by science fiction scenarios rather than demonstrated harm. Dario Amodei previously claimed 50% of entry-level knowledge worker jobs would disappear within 1-5 years, a prediction made one year prior that has not materialized.
One interpretation suggests Anthropic's strategy involves branding as the safe AI company, advocating to ban unsafe AI, and profiting from this positioning. This represents a coherent business strategy that may explain regulatory advocacy beyond genuine safety concerns.
Recent research from Calico (Google's age reversal startup) in partnership with Revel Pharma focuses on the extracellular matrix—the areas outside cells that age. Over time, sugars and fats bind to proteins between cells through a process called glycation, creating advanced glycation end products (AGEs). The predominant molecule CML accumulates and cannot be broken down naturally, causing stickiness, reduced mobility, wrinkles, joint immobility, and inflammation.
Researchers used AlphaFold to identify proteins that could bind to CML and activate enzymatic processes to break it down. Through directed evolution, they created hundreds and thousands of protein variants, testing them recursively across five cycles. The optimized enzyme successfully degraded 52-97% of CML from proteins including casein, collagen, retinal proteins, and hemoglobin, with several sites achieving over 90% degradation.
When tested on actual human skin donated by elderly patients over 70 years old, the enzyme eliminated 55% of CML, effectively reversing skin age to that of a 31-year-old. This represents a groundbreaking demonstration combining AlphaFold with directed evolution and high-throughput screening to create novel proteins not found in nature. Future delivery methods being considered include creams, shots, supplements, or RNA shots that would produce the protein internally.
The first market application is expected to be cosmetic skin treatments, potentially representing a $2 trillion market opportunity if a cream-based delivery system proves effective. This research demonstrates concrete AI applications that benefit everyone, using AlphaFold to discover and evolve proteins driving profound human health outcomes.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
