Coinbase’s Everything Exchange: Agentic Finance, Stablecoins & Tokenization with CEO Brian Armstrong

Summary

• Armstrong’s headline claim is that agentic commerce is already breaking card economics: roughly 76% of the agentic transactions Coinbase is seeing are under 30 cents, below the ~30-cent flat minimum fee on debit/credit cards, making conventional card rails uneconomic for many machine payments. His framing — “we don’t want the AIs to be unbanked… they deserve financial services as well” — is backed by product: single-prompt self-custodial wallets for agents (no KYC, no government ID) and the X42 protocol Coinbase incubated and put into the Linux Foundation with Google, Cloudflare, and AWS. “There’ll be more agents than humans in the not too distant future… the agentic economy will be bigger than the human economy at some point.”
• Coinbase disclosed that 88% of revenue is now non-Bitcoin trading, even as the stock “trades fairly correlated with the price of Bitcoin.” The broader thesis is an everything exchange: tokenized stocks (a true one-to-one custodied security, not a synthetic — available outside the U.S. for now, with an SEC path in progress), then private credit, treasuries, and bank deposits, targeting ~4 billion people with no brokerage or U.S. investment-account access, “much like Tether was very successful ex-US.”
• Prediction markets hit a ~$100M revenue run rate within months of launch inside the Coinbase app, growing over 100% quarter-over-quarter, and Armstrong thinks the category extends past sports into policy and potentially open-ended “opinion markets” on questions like “is there a god” — bets without a clear resolution date, somewhat like holding a stock.
• Internally, Coinbase is building toward “recursive self-improvement”: every service and team has a “brain” (incident history, financial controls, A/B tests, full PR history — today just markdown files) that agents ingest before making changes, with human corrections fed back so “the accept rate for one-shot PRs just starts to tick up over time.” Armstrong himself now spins up 10 parallel agents for a feature phase via the internal Toshi harness and sends the resulting PRs for review instead of pinging teams in Slack — “that’s a magical moment… it’s very addictive.”
• On the view that money may disappear, Armstrong says scarcity persists far into the limit — land, energy, chips (“Dyson spheres are going to be expensive pretty far into the future”) — so a medium of exchange survives, and “crypto was really good for humans and it’s going to be essential for AI.” On headcount, he rejects the shrink-to-10-people thesis: “tasks are being eliminated, people are not being eliminated” — same teams, more velocity, and potentially rising margins as revenue outgrows the team.
• New Limit, his epigenetic-reprogramming company, has demonstrated reprogramming of at least one human cell type in humanized mouse models, is now in non-human primates, and plans its first phase-1 trial next year in alcoholic liver disease — a market he estimates might be ~$20B for that drug alone, with skin rejuvenation “probably a trillion dollar market.” The platform (a frontier AI model recommending pooled screens across transcription-factor space) targets liver, vascular, and immune cells first, attacking “the meta problem” of restoring youthful cell function rather than individual diseases.
• His next frontier obsessions: cognitive enhancement (“I haven’t found a great thesis… if somebody has one, let me know”), and U.S. “freedom cities” — federal-land special economic zones with reduced EPA/FAA red tape for nuclear, drones, and data centers. Elad, not Armstrong, discussed his investment in Preventive, cited ~80% American support for embryo editing for disease prevention, and raised the concern that standardized “intelligence packages” could diminish neurodiversity; Armstrong argued diversity could instead increase, potentially producing children with “gills” or wings.

Deep dive

1. Bank the AIs: sub-30-cent commerce breaks card rails

• Armstrong’s core stat: card transactions carry a ~30-cent flat minimum fee plus percentage, so “it doesn’t make sense to pay 30 cents to send a 1 cent payment” — and roughly 76% of agentic commerce transactions Coinbase observes are under 30 cents. Stablecoin payments move money “at the speed of information” — under 1 second, under one cent, anywhere.
• The mission statement, verbatim: “We don’t want the AIs to be unbanked. We want to bank the AIs. They deserve financial services as well.” A single pasted prompt gives any agent a self-custodial wallet — no KYC, since “our AI agents don’t have a government ID… they can’t walk into a bank branch, at least not yet, without their humanoid robot companion.” Coinbase will offer both segregated agentic sub-accounts tied to a human’s identity and fully independent wallets that could hold stablecoin, raise money, or issue a token.
• Infrastructure: the X42 protocol, incubated at Coinbase and put into the Linux Foundation, now has Google, Cloudflare, and AWS working on it; stats live at agentic.market. Answering Elad’s question on what trades at a few cents: paywalled research pulls, LinkedIn scrapes, and mostly “agents talking to other agents… almost like a tool call.”
• Where it goes: Armstrong says smaller open-weight models can, in some instances, outperform frontier models on specialized tasks — for example, a hypothetical compliance-review model trained on 100,000 Coinbase cases. He expects a marketplace of specialist agents, from “really good at ordering you pizza” to a designer trained on proprietary Figma data. “It kind of stands to reason that the agentic economy will be bigger than the human economy at some point in the not too distant future.”

2. Scarcity — and money — survive the AI transition

• Elad raised the concern that AIs could aggregate resources and use them to manipulate society, while saying he does not share that concern strongly. Armstrong then discussed Elon Musk’s view that money may disappear: in the limit, measuring things in energy and mass gives Musk “a point,” but scarcity persists — usable land, energy, chips — and “Dyson spheres are going to be expensive pretty far into the future.”
• Robots and AI could drive price deflation in commodities, food, and housing, while government printing might add inflationary pressure, but a medium of exchange remains necessary — “crypto was really good for humans and it’s going to be essential for AI.” Elad’s observation on the talent overlap: the same MIT CS-type graduates went into crypto around 2020 and AI around 2022, and the two fields “just might” intersect now.

3. Recursive self-improvement as an operating system for Coinbase

• Past the table stakes (coding tools, ML for fraud and support), the big internal idea is a “brain” for every team, service, and individual — incident history, financial controls, every A/B test, the full record of accepted and rejected PRs, “today it’s just markdown files.” Agents ingest the relevant brain before touching a repository; notably, Armstrong believes individuals “should be able to take [their brain] with them when they leave Coinbase.”
• The load-bearing discipline: when a human reviewer catches an agent’s miss, the fix must flow back into the brain — not just a manual edit — “so that you not only fix it in this case but in all future cases going forward… the accept rate for one-shot PRs just starts to tick up over time and you’ve now got a recursive self-improvement system.”
• The lived demo: earlier that day, Armstrong had a high-end model plan a complex feature in three phases of ten pieces, then spin up 10 parallel agents for phase one, using open-source models and Grok among the selected agents; “2 minutes before I hopped on this podcast, it pinged me… all 10 of those things from phase one are now done.” The internal harness, Toshi, can also pay external vendors over X42 rails. Instead of pinging teams in Slack, “I just send them a PR for review and it’s already done… it’s very addictive.”
• On Elad’s question whether companies shrink to 10 people: “tasks are being eliminated, people are not being eliminated” — existing headcount gets more done and the pace accelerates. Armstrong added that margins should rise if revenue grows faster than the team.

4. The everything exchange: 88% non-Bitcoin, and tokenize everything

• The disclosed number: 88% of Coinbase revenue is now non-Bitcoin trading — “funny because if you look at our stock price it does trade fairly correlated with the price of Bitcoin.” He still calls Bitcoin “digital gold… sound money for everybody in the world,” but Coinbase’s broader exchange brings stocks, commodities, derivatives, perpetual futures, and prediction markets together with cross-margin, enabled by emerging regulatory clarity.
• Coinbase has announced a tokenized-stocks product available outside the U.S. for now — “the first truly tokenized stock product,” an actual security held one-to-one in custody, not a synthetic, derivative, or debt instrument — while working with the SEC and others on a U.S. path. The demand case: ~4 billion people lack access to a brokerage or U.S. investment account and therefore cannot invest in Coinbase, NVIDIA, Apple, or similar assets, “much like Tether was very successful ex-US because a lot of people wanted dollar-denominated accounts.” Next down the list: private credit, treasuries, and bank deposits; gifting a share could become as simple as sending it to a wallet.
• Elad’s caveat from experience: he backed Harbor’s real-estate tokenization seven or eight years ago — “early but too early but right.” Armstrong replied that he thought it was “a little bit too far ahead.”

5. Prediction markets past sports: policy signals and “opinion markets”

• The numbers: within months of launch, prediction markets in the Coinbase app hit roughly a $100M revenue run rate, growing “over 100% quarter over quarter, something like that — I’d have to go look at the exact number.”
• Sports is fine entertainment, but Armstrong’s bigger claim is societal: markets could ask, “if we were to implement this policy, what would the unemployment rate be within one, two, three years,” potentially informing government choices — and could support “opinion markets” on big questions (“is there a god… were we designed by evolution or intelligent design”) without a definite resolution date. New papers or events could shift the market, “kind of like when you invest in stocks… you’re just betting on is it going to do better or worse.” “We’re just scratching the surface.”

6. New Limit: from dinner-party skepticism to a phase-1 trial next year

• The origin as told: after Coinbase went public in 2021, Armstrong scanned big trends (space, AI, fusion, and brain-machine interfaces) but saw “not that many great teams working on longevity — I thought a lot of it was snake oil,” a view Elad had reinforced (“we found a bunch of ways to make mice live longer and it doesn’t work in humans”). Dinners with CEOs and scientists surfaced epigenetic reprogramming associated with Shinya Yamanaka’s lab, and it gripped him like Bitcoin in 2010-11: “I just couldn’t stop thinking about it for like 3 months… you kind of want to hold on to that moment of inspiration… because it’s perishable.” The group that came together included CEO Jacob Kimmel, Blake Byers, and Greg Johnson; Armstrong is “basically just an investor/board member” — Coinbase remains the full-time job.
• The platform: a 50-60-person South San Francisco lab running what Armstrong calls the leading frontier model for epigenetic reprogramming, recommending transcription-factor combinations across a massive search space, validated through large pooled screens, then animal models, then humans. Status: successful reprogramming of at least one human cell type in humanized mouse models, now in non-human primates, phase 1 launching next year, with “three to ten other drug candidates coming off this platform in the coming years.”
• The roadmap, published at NewLimit.com: hepatocytes, vascular cells, and immune T cells, following a pattern similar to GLP-1 development by starting with high unmet need. The initial trial targets alcoholic liver disease, where patients have few options beyond a liver transplant and 12-month survivability after diagnosis is very low; that indication alone “might be worth $20 billion.” The endgame is a cocktail restoring youthful function across cell types — skin alone is “probably a trillion dollar market” — attacking “the meta problem” since most diseases that take people out are correlated with age.
• Meta-lesson on serial founding: he tried ~10 ideas before Coinbase worked, calls premature multi-company distraction “arguably irresponsible,” but says that if someone has the potential to incubate more companies, “if you’re just doing one thing forever… you’re just limiting your potential” — citing Elad himself, Joe Lonsdale, Sam Altman, Palmer Luckey, and Elon as examples.

7. Frontier ideas: cognitive enhancement, embryo editing, freedom cities

• Crypto still powers only “about half of 1%” of global GDP — roughly 700 million people have used or hold crypto, while monthly users might be 50-100 million — and the goal is a billion-plus. Armstrong’s open problem is cognitive enhancement (folic acid and certain nutrients have, he says, raised IQ or improved development in some countries): “I haven’t found a great thesis about how to exactly tackle that, but if somebody has one, let me know.”
• Elad framed the broader stakes: “AI is getting smarter and smarter,” so humans may need to take control of their evolution; he also discussed his investment in Preventive, which is working on embryo editing. Elad cited ~80% American support for embryo editing for disease prevention despite the Gattaca framing, and suggested that eventually having children without this screening and editing could seem “like driving without a seat belt.”
• Elad’s concern was that standardized “intelligence packages” could diminish neurodiversity, creating “the Ford Model T of intelligence for everyone.” Armstrong’s rebuttal: unlimited diversity of parental preference plus floor-raising, such as removing schizophrenia or depression, could increase diversity; in the limit, “kids born with gills… or wings.”
• On special economic zones: Coinbase invested in Próspera in Honduras, where Armstrong said there had been a mix-up with the local government but that the parties were getting closer to a resolution and he thought Próspera would make it. He champions the “Freedom Cities” idea discussed during a campaign: 10 or 100 square miles of federal land where innovators would not face all the normal EPA, FAA, and local red tape for projects such as nuclear power, data centers, and drone delivery, because “you can’t really solve a problem as an entrepreneur unless you can iterate on it.” Federalism already makes the U.S. “50 special economic zones… but we could go even further.”