The Stablecoin Future, Milei's Memecoin, DOGE for the DoD, Grok 3, Why Stripe Stays Private
Summary
Stripe grew out of the Collisons’ teenage-era startup Auctomatic into infrastructure processing more than $1 trillion annually, roughly 1% of global GDP by the Collisons’ qualified comparison. Its scope now spans payments, lending, card issuance, treasury, billing, cross-border movement, and stablecoins because “every kind of money movement is going from being manually orchestrated to being orchestrated by software.” Yet Stripe remains profitable and private: the Collisons optimize for decade-long compounding and customer time, not an IPO as a status symbol.
Stablecoins’ first decisive market is cross-border dollar access, not replacing Visa and Mastercard at the American checkout. Patrick Collison contrasted the old eurodollar system’s roughly $1 million entry point with an Ecuadorian consumer now holding a $1 U.S. balance, while citing Nigeria’s currency devaluation by a factor of three or four over the last couple of years. Stripe bought Bridge late last year after Bitcoin proved too slow, costly, and uncertain in dollar terms for payments; modern stablecoins are “finally happening” on rails such as Ethereum L2s and Solana.
The expensive problem in payments is increasingly fraud and back-office complexity rather than settlement alone. Jason argued that businesses can lose 1%-3% of revenue through accounts-payable and accounts-receivable processes, while Stripe Billing has passed $500 million in ARR; Stripe also sees a prior transaction history for 93% of cards presented to its merchants. The Collisons’ larger thesis is that a trusted network of known counterparties can lower fraud—which they said is worsening industrywide even as Stripe’s rate fell 80%—and eventually reduce fees.
The remote-work argument resolved into a segmented labor thesis: experienced outliers can thrive remotely, but early-career employees measurably struggle. John Collison warned against designing policy around “the bottom 5% of the company,” because Stripe has extraordinarily productive remote employees, yet its pre-COVID data showed remote work was bad for young workers professionally and personally—“they go mad” in what the hosts called solitary confinement. Patrick rejected universal rules, pointing to successful but divergent models at Nvidia, Coinbase, Shopify, and Jane Street.
Corporate efficiency depends as much on organizational and software architecture as on headcount cuts. Jamie Dimon argued that any operation with 100 people could run better with 90 and ridiculed a wealth-management decision requiring 14 committees; Chamath countered that off-the-shelf enterprise software creates rigid job boundaries and bureaucracy rather than curing them. Friedberg saw the broader shift as leaders again saying, “My job is not to coddle my employees,” while Shopify’s deletion of recurring meetings supplied the episode’s cleanest software-level intervention.
An 8% annual Pentagon cut over five years—roughly $300 billion cumulatively in the show’s estimate—only makes strategic sense if procurement follows technology. Chamath contrasted a flagged $1.2 trillion Navy frigate plan and multibillion-dollar, decade-long ships with autonomous systems from newer defense companies; Friedberg noted that a $10,000 drone can destroy $10 million of equipment. Their conditional case combined a more multipolar foreign policy with replacing legacy “big iron,” while Patrick cautioned that budget cuts alone do not repair a procurement system criticized for decades.
Milei’s promotion of $LIBRA converted a celebrated reform narrative into an avoidable governance crisis. The coin reached a reported $4 billion market cap, collapsed 95%, and left 74,000 traders nearly $300 million poorer, including 24 wallets losing more than $1 million each. The panel likened memecoins to gambling but emphasized their more pernicious “pump and rug” mechanics; Jason’s verdict was that Milei “rug pulled the people who put him in office” and then compounded the failure by taunting his followers instead of owning the mistake.
The episode’s AI calls were unusually concrete: biological models may open a new research stack, while Grok 3 revived Chamath’s belief in brute-force pre-training and Nvidia demand. Arc’s open-source Evo 2 was trained unsupervised on nine trillion base-pair tokens yet saw only one human genome before achieving state-of-the-art predictions of harmful human mutations. Separately, xAI assembled 100,000 GPUs in Memphis—going to 200,000 in 122 days—by treating time as the hard constraint, prompting Chamath to reverse his view that base models were asymptoting: “I was completely wrong.”
Deep dive
1. Stripe became a trillion-dollar layer by following software into money
Chamath’s opening regret carried the venture lesson: he met the Collisons during their teenage-era startup Auctomatic, could theoretically have invested early, then watched Stripe compound for 17 years. Patrick recalled offering him water or milk from a crowded two-bedroom apartment—and John gingerly washing a glass for him.
Auctomatic eventually shut down, but the Collisons quickly spun up Stripe. Jason said he could have invested a single dollar and made $1 billion, then regretted not calling them at any point during the following 17 years.
The company now processes more than $1 trillion per year. Patrick qualified the comparison to roughly 1% of approximately $100 trillion in global GDP because Stripe transactions are not perfectly equivalent to final-goods output, but argued it remains reasonable given that most Stripe volume does finance final goods.
Payments remain the largest business, but the structural thesis widened: software is replacing manually orchestrated money movement. Customer pull took Stripe into lending, card issuance, treasury and money storage, billing, cross-border transfers, and stablecoins; its clientele likewise expanded from startups to enterprises including Hertz, Amazon, and Ford.
2. Stablecoins solve the dollar-access problem before the checkout problem
John’s crypto history was candid: Stripe tried to make Bitcoin work as a payment method, but it was slow, expensive, and uncertain in dollar terms. Bitcoin may function as “a store of value” or gold substitute; stablecoins on Ethereum L2s or Solana are now good enough for actual payments.
Stripe bought Bridge late last year, described as building “the Stripe of stablecoins.” Companies such as SpaceX use stablecoins for treasury management, while others use them to offer dollar services or move money internationally. The strongest present use cases involve a border: remittances, global contractor payments, multinational treasury, and dollar balances outside America.
Patrick’s sharpest example was currency preservation. Nigeria’s naira had devalued by a factor of three or four over the last couple of years, while stablecoins let someone in Ecuador hold a $1 U.S. balance—a consumer-scale version of the 1970s and 1980s eurodollar system, whose minimum transaction was roughly $1 million.
That adoption may deepen dollar reserve status, but it does not automatically dismantle Visa and Mastercard. Patrick stressed that most merchant interchange flows to issuing banks and then funds credit and card rewards; removing the networks therefore creates trade-offs involving rewards, consumer protection, and credit availability, not merely the elimination of rent extraction.
3. Fraud and reconciliation are the real costs hiding behind transaction fees
The All-In business could accept stablecoin payments through Stripe and can pay people in stablecoins. Yet domestic bank transfers are merely slow and annoying; Bridge’s “hair on fire” cases include companies such as Scale AI paying contractors in places like the Philippines, where conventional transfers become genuinely expensive and difficult.
Jason argued that companies often lose 1%, 2%, or 3% of revenue to accounts payable and receivable. Humans issue invoices, reconcile transfers, and match bank-statement entries through bureaucratic, inefficient processes; Stripe Billing, built to automate that stack, had passed $500 million in ARR.
As more counterparties sit inside Stripe, some transfers can become ledger entries, but the Collisons expect the largest savings from identity and risk rather than raw routing. A payroll company can lose millions in one fraudulent-company attack; a network of trusted nodes can do more than pass along an account number.
Stripe has previously seen the card used in 93% of purchase attempts across its merchants, turning the internet economy into a reputation network. An unfamiliar card paired with a new email or phone number is “ipso facto” suspicious; the Collisons said fraud metrics are rising across the industry, while Stripe’s were down 80%.
4. Stripe’s economic data is powerful but structurally biased
Asked whether Stripe could publish a faster economic-sentiment gauge than frequently revised payroll and GDP releases, John admitted some ruefulness: “On some level we should have done it.” The obstacle is interpretation, not data availability.
Stripe overweights online and innovative businesses, as COVID made obvious when online commerce looked healthy while the offline economy did not. Stripe itself is also growing and changing too quickly for a year-over-year increase in its volume to map cleanly onto the broader economy.
Even so, its team constructed what John described as a fairly reliable leading indicator for inflation. They would like to publish such analysis because better and more timely economic data is a public good, provided users understand the sample and growth biases.
5. Remote work is an experience-level trade, not a moral category
Jamie Dimon’s objection was about attention and apprenticeship: employees multitask on Zoom, text one another, and fail to read material, slowing efficiency and creativity. His strongest warning was that younger workers are “being left behind socially”—missing ideas, relationships, and exposure to colleagues unlike those in their home communities.
Stripe largely returned to its pre-pandemic pattern: most employees use offices in cities such as San Francisco, New York, Dublin, and Singapore, while a meaningful cohort remains remote. John credited remote work with expanding the talent pool and solving the “two-body problem” when one partner’s career fixes a couple’s location.
John’s empirical split predated COVID. Stripe found that remote work was poor for early-career employees both professionally and personally; at 23, isolated workers can “go mad.” Yet he warned executives not to build policies around quiet quitters or “the bottom 5%,” because some top remote performers are extraordinarily productive.
Chamath framed most new workers as initially residing below zero on a J-curve, requiring in-person mentoring before contributing; engineering may be an exception. Patrick preferred “many paths to heaven”: Nvidia tolerates location flexibility, Coinbase and Shopify are remote-first, while Jane Street prizes a shared trading floor. Jason added that U.S. labor productivity rose roughly 20% over ten years.
6. Bureaucracy is encoded in workflows as much as org charts
Dimon’s zero-based challenge was blunt: if an area employs 100 people, he could run it with 90 “in my sleep.” Learning that one wealth-management approval touched 14 committees made him want the names of all 14 chairmen—not another justification for staffing.
Friedberg interpreted the rant as part of a leadership change visible in Zuckerberg’s buyout offer, Elon Musk’s Twitter restructuring, and Brian Armstrong’s stance at Coinbase. Leaders are again stating the mission directly: “My job is not to coddle my employees”; it is to organize a team that wins.
Chamath’s mechanism was more structural. Enterprise software promises efficiency but rigidly defines where marketing, sales, and other jobs begin and end, thereby manufacturing bureaucracy. JPMorgan spends roughly $6 billion annually on IT, he said; internally built systems at Facebook, Google, Tesla, SpaceX, and likely Stripe can instead show up as superior revenue per employee.
Shopify offered the cleaner operational experiment: Toby deleted recurring meetings across the company through software. John noted that one might expect the meetings to return, but Shopify measured the result and reportedly found that many did not—evidence that some organizational problems really can be attacked with a script.
7. Pentagon savings require a new doctrine, not smaller legacy invoices
The proposed target was an 8% Pentagon reduction in each of five years, compounding to nearly $300 billion in the show’s estimate. Patrick noted that the magnitude resembled the defense-budget decline from 2010 to the present, making it substantial but not historically unprecedented.
Chamath argued that military spending must sit “downstream from technology.” A CBO-flagged Navy frigate project was heading toward $1.2 trillion, with ships costing $3 billion-$4 billion and taking eight to ten years, while autonomy and AI are already producing alternative systems through companies including Saronic, Saildrone, and Anduril.
Friedberg’s conditional strategy began with a multipolar world in which America, China, and Russia recognize shared power rather than continually funding primacy. Technology reinforces the case: a $10,000 drone can destroy $10 million of equipment, and China now has drone factories that can output millions of drones each month—raising direct questions about aircraft carriers and tanks.
Patrick kept the procurement distinction intact: Washington has criticized defense acquisition across parties and across decades, so cutting budgets does not itself fix purchasing. His recommended historical lens was Robert Coram’s Boyd, about the reform struggle behind aircraft including the F-16, A-10, and F-15 against generals committed to inferior planned systems.
8. Milei turned a meme coin into a crisis of leadership
Milei promoted $LIBRA as a private project supporting Argentina’s economy, then deleted the post and said he had not known the details. Its market capitalization surged to a reported $4 billion before collapsing 95%; 74,000 traders lost almost $300 million, with 24 wallets down more than $1 million each.
Chamath could not reconcile the episode with Milei’s prior momentum. He called the subsequent distinction—Milei merely “shared” rather than endorsed—the kind of cover-up worse than the original act, and cited Hayden Davis’s Coffeezilla interview plus texts appearing to claim influence over Milei and implicate Milei’s sister, while acknowledging the full story remained unclear.
Friedberg compared memecoins to digital collectibles and gambling but argued that software removes physical friction and amplifies the social feedback loop “by like a thousand X.” Patrick noted that tickers, charts, exchange listings, and price forecasts encourage buyers to treat them more like financial assets than objects with intrinsic aesthetic value.
John focused on the pump-and-rug structure and compared it with the normalization of negative-expected-value state lotteries. Jason’s leadership standard was harsher: “The appearance of impropriety is impropriety.” Milei had “rug pulled the people who put him in office,” then taunted his followers instead of owning the error.
9. Fiscal discipline needs debate that survives partisan ownership
Jason’s West Wing takeaway was the missing “great debate”: test DOGE, abortion, states’ rights, and spending on their merits rather than attacking any proposal associated with the other party. Both sides can hold valid pieces of an answer, but election-cycle point scoring prevents resolution.
Patrick challenged the nostalgic picture, noting that The West Wing embodied a particular worldview and that the Clinton years were not necessarily an age of ideological ferment.
Friedberg highlighted Clinton’s early deficit-reduction act and estimated that federal spending fell by five percentage points of GDP during the presidency—a meaningful fiscal contraction, even with defense cuts. The group also noted technological tailwinds from the PC and internet booms.
Chamath contrasted earlier surplus windows with the roughly $16 trillion added to U.S. debt during the last two administrations despite a strong stock market: “What’s going to happen during a bad time?”
10. Arc Institute is a wager against consensus-driven basic research
Patrick Hsu described Arc as a Palo Alto nonprofit in partnership with Stanford with roughly 230 people, spending about $100 million annually. He and John are among its funders alongside other donors, with the goal of supporting basic biology through curiosity-driven research rather than narrowly prescribed grants.
Scientists can spend about 40% of their time on NIH grant overhead and related work, while consensus scoring penalizes work outside established fields. In Arc’s survey, 79% of leading scientists said they would substantially change their research agendas if free to allocate funding themselves—the analogy was a startup world with one government-run VC firm.
Arc organizes its disease thesis into infectious, monogenic, and complex conditions. Medicine can broadly generate cures or treatments for many infections and screen for some single-mutation diseases, but Patrick said humanity has never cured a complex disease—the category containing most cancers, autoimmune and neurodegenerative diseases, and Alzheimer’s.
The opportunity is a newly available “read, think, write loop”: single-cell DNA and RNA sequencing, CRISPR and functional-genomics perturbations, plus transformers and machine learning. Patrick repeatedly hedged the outcome—“we’ll see”—but framed the stack as a plausible way to illuminate previously intractable gene-environment diseases.
11. Evo 2 learned human mutation risk from the wider tree of life
Arc’s Evo 2 was presented as the largest biology machine-learning model yet and, Patrick believed, the largest fully open-source AI model: not just weights, but public training code. It was trained on nine trillion base-pair tokens, treating DNA as “the language of life.”
Only one human genome appeared in training, and that person did not carry the pathogenic mutations being tested. Nevertheless, Evo 2 achieved state-of-the-art prediction of harmful human mutations, including BRCA-associated breast-cancer variants—suggesting it learned transferable structure across species rather than memorizing human labels.
Friedberg asked where the phenotype labels entered; Patrick’s answer was that training was entirely unsupervised. The model sees genomes, learns latent structure, and scores how likely a sequence is relative to the genetic universe; small task-specific models can then use its upper-layer embeddings and learn quickly from a few examples.
Whether DNA alone is sufficient remains open because proteins, RNA, cells, and phenotype sit downstream but contain useful additional information. Evo 1 had suggested protein-structure prediction could emerge from a DNA model; Patrick compared Evo 2 loosely to GPT-2 or GPT-3 and anticipated a similar “Cambrian explosion” of applications.
12. Programmable biology still runs into the phenotype bottleneck
Friedberg’s long-term aspiration is software that starts with a desired phenotype and resolves the genome needed to produce it—perhaps a plant adapted to Martian soil, a CO₂-heavy atmosphere at 1% of Earth’s pressure, unusual daylight, and high winds. Patrick agreed but said the most powerful models will need substantial environmental and phenotypic data.
Sequencing raced ahead because genomes are plentiful and cleanly digital; phenotype is harder even to define consistently. A model may predict whether a genome looks internally correct, but not yet fully connect that sequence to how an organism performs in a particular environment.
John made the problem personal through ash dieback in Ireland’s surviving ancient woodland, then cited California bark beetles, black pod fungus affecting cacao and coffee, and TR4 affecting bananas. Friedberg said a resistance trait may involve silencing a gene that suppresses immune function, but a genomic solution ultimately requires regenerating resistant plants and replanting the affected forest.
Agriculture supplies the economic specimen: commercial bananas descend from one Dwarf Cavendish clone, enabling TR4 fungus to specialize against a genetically uniform crop. Friedberg said roughly $0.60 of every banana dollar now goes to fungicide; Ohalo’s response is genetic diversity and targeted resistance, including a University of Florida project against a strawberry fungal pathogen.
13. Grok 3 made time—not capital or talent—the binding constraint
Chamath reversed an earlier conviction that base models were asymptoting and extra Nvidia capex might be unproductive. Colossus suggested that larger pre-training clusters still produce valuable gains, leaving him “a little bullish on Nvidia” and concluding: “I was completely wrong on a couple of my earlier thoughts.”
xAI found an old Electrolux factory in Memphis for 100,000 GPUs, with a stated path to 200,000 in 122 days. Starting with roughly 15 megawatts against a need near a quarter-gigawatt, the team bought generators, acquired about one-third of America’s portable liquid-cooling capacity, and rewrote Tesla Powerpack firmware to smooth power.
With capital and recruiting power treated as abundant, Elon Musk imposed time as the artificial constraint. Chamath’s general rule was that innovation needs one hard boundary—capital, talent, or time—and called the cross-company mobilization of Tesla engineers and hardware an American version of a keiretsu.
Friedberg compared Musk with industrialist Henry Kaiser, who challenged conventional bids, prioritized speed, built ships and the Hoover Dam, and expanded Richmond’s workforce from zero to 100,000 in a year. The panel treated Grok 3’s benchmark lead cautiously because leaderboards are disputed and leapfrogged, but found its quality-per-unit-of-time extraordinary.
14. Stripe will stay private while private ownership improves compounding
Patrick rejected IPO dogma in either direction. Public markets offer cheaper, deeper, more liquid capital; stable private markets now provide funding and shareholder liquidity too. Stripe therefore asks a pragmatic question—whether it is currently better as a private or public company—and has so far answered private.
The discipline argument drew his sharpest dismissal: “If you need a 25-year-old Fidelity analyst asking you to double-click on your capex…to run the company with discipline, something is horribly wrong.” Public status is neither spiritual nor moral, and weak internal management is not repaired by quarterly questioning.
Financial-services precedents include private Bloomberg, Fidelity, Vanguard, Jane Street, and Citadel, while Goldman Sachs, JPMorgan, and Visa reportedly waited 130, 70, and 50 years respectively. Patrick’s sector-specific concern is procyclicality: public financial companies must resist pressure to expand with exuberant markets and contract at the wrong moment.
Patrick said Stripe is profitable on a fully loaded GAAP net-income basis. He noted that private companies can provide yearly liquidity, then contrasted public-market outcomes: Square is 70% below its 2021 peak and PayPal 80% below. The governing test is to maximize ten-year compounding, serve customers, and spend the marginal hour with a customer. “This is our life’s work.”