76% of Agentic Commerce Is Under 30 Cents — So Coinbase Is Banking the AIs

The minimum fee on a typical debit or credit-card charge is about 30 cents flat, plus a percentage. That number quietly kills a whole class of business: anything under a dollar barely works, and it does not make sense to pay 30 cents to send a one-cent payment. Brian Armstrong, co-founder and CEO of Coinbase, tells Elad Gil on No Priors that the last stats they saw on agentic commerce put about 76% of those transactions under 30 cents.

The clip that opens the episode is the thesis in miniature. Coinbase does not want AIs to be unbanked; they deserve financial services too. Agents will need their own accounts, and Coinbase has already shipped tools where a single pasted prompt gives an agent a financial account. Armstrong's forecast is blunt: there will be more agents than humans in the not-too-distant future, and it stands to reason that the agentic economy will be bigger than the human economy as well.

Gil frames the hour as two acts at once. Coinbase has been aggressive about adopting AI and is trying to become a new global payments layer. Armstrong also co-founded NewLimit, an anti-aging company using high-throughput screening and AI on biology. The conversation runs from crypto rails to agentic finance to a second company a public-company CEO decided the world still needed.

Armstrong splits Coinbase into three curves. The first is the everything exchange: every asset class coming on-chain so stocks, commodities, crypto, derivatives, perpetual futures, and prediction markets trade in one place, with liquidity and cross-margin, in more countries as regulatory clarity emerges. That remains the core business. The second is stablecoin payments, which keep growing even when bitcoin prices are down. Money, he says, starts to move at the speed of information: under one second, under one cent, anywhere in the world.

A number from the recent earnings deck shows how far the mix has already moved: 88% of revenue is from non-bitcoin trading. The stock still trades correlated with bitcoin, and Armstrong still loves bitcoin as digital gold and sound money for everybody in the world. What crypto is actually doing, in his telling, is updating every aspect of the financial system. Tokenization is the frame. Stablecoins were the first major case, a token one-to-one with a dollar in a bank account or a U.S. treasury. Stocks are next.

Coinbase just announced a tokenized-stocks product, available only outside the United States for now while they work with the SEC on a U.S. path. Armstrong stresses it is the first truly tokenized stock that is not a synthetic, a derivative, or a debt instrument: a security, one-to-one represented by the actual security held in custody. About four billion people have no brokerage or U.S. investment account, so they cannot buy Coinbase, NVIDIA, or Apple. Tether worked ex-U.S. because people wanted dollar accounts; the same billions want high-quality investments. After stocks come private credit, treasuries, bank deposits. Gifting a share to a nephew becomes sending it to a wallet, not an antiquated transfer process.

Gil mentions investing seven or eight years ago in Harbour, a real-estate tokenization company he now thinks was too far ahead. Armstrong: early, but right; that will happen too. Prediction markets inside the Coinbase app are the surprise velocity story. Within months of launch they hit about a $100 million revenue run rate, growing over 100% quarter over quarter, numbers he would want to recheck exactly. Sports is part of it, and he is happy to celebrate that as entertainment. The societal use he wants people to notice is policy: if we implement this, what is unemployment in one, two, three years? Further out are opinion markets on the big life questions — is there a God, evolution or intelligent design — with no resolution date, like a stock that never expires, where a paper or an attack can move the implied probability from 60% to 80%.

Agentic finance, as Armstrong names it, has two layers. The first is for people who already have accounts: an AI advisor of the kind wealthy clients get from a human, now for everybody — portfolio construction, dollar-cost averaging, tax-loss harvesting, buying the dip in 5%, 10%, or 15% legs, earning yield on cash, talked to like a human. Coinbase has launched that advisor in the product. The second layer is the agents themselves. "We don't want the AIs to be unbanked. We want to bank the AIs. They deserve financial services as well." A pasted prompt can give an agent a self-custodial wallet on crypto rails. There is no KYC, because agents do not have government IDs and cannot walk into a branch, at least not yet, without a humanoid-robot companion.

They are building it both ways. One path is a segregated agentic account on Coinbase, linked to a human identity, with funds the agent can trade and pay from. The other is self-custodial wallets for any agent, which could even raise money or issue a token. Today, agents get stuck at paywalls, AWS spin-ups, and hotel checkout pages that demand a credit card. Tomorrow they need their own spend accounts, the way a personal assistant or employee does, plus probably a stablecoin-backed card for backward compatibility. Much of the new flow is already on protocols such as X402, incubated at Coinbase and placed in the Linux Foundation, with Google, Cloudflare, and AWS working on it. A site called agentic.market shows payment stats and service providers. Coinbase Business lets companies accept stablecoins from humans and agents out of the box. The question Armstrong wants merchants to hear is whether the store is ready to accept agents as customers.

Those sub-30-cent payments are mostly information: venture firms compiling paywalled data, recruiters scraping LinkedIn, agents paying other specialized agents almost like a tool call. Frontier models are generalists. Smaller open-weight models, fine-tuned or reinforced on proprietary data, can beat them on a narrow task — Coinbase's own example is a compliance-review queue trained on 100,000 internal cases. The marketplace, he thinks, fills with specialists: an agent that only orders pizza, a designer trained on tons of proprietary Figma data. Gil floats the safety concern that AIs will aggregate resources and manipulate society. Armstrong does not share it strongly. On Elon Musk's claim that there will be no money, he thinks he knows the limit-case point — energy and mass as the scarce units — but scarcity of land, energy, and chips lasts a long time, and Dyson spheres will be expensive far into the future. A medium of exchange remains. "Crypto was really, really good for humans. And it's going to be essential for AI."

Inside Coinbase the table stakes were coding tools, data warehouses, ML for fraud and risk, AI in support. The boundary they are pushing is recursive self-improvement: a brain for the company, and for each team, each service, and each person. Individuals, he believes, should take their personal brain with them when they leave; policy is being written for that. A service brain is markdown: every incident, every financial control, every A/B test, the full history of accepted and rejected PRs. An agent ingests it, opens a PR, a human catches a miss. The naive move is to patch and ship. The rule they are enforcing is that the correction goes back into the brain so it applies to all future cases. One-shot PR accept rates start to tick up. Their internal harness, Toshi, can already call external vendors and pay over X402. Two minutes before the taping, a high-end model Armstrong had asked to plan a complex feature — three phases, ten pieces each, ten agents in parallel, cheaper open-source models and Grok recommended to save money — pinged him that phase one was done and ready for review. He used to ping the team in Slack all day. Now he tags an agent or sends a finished PR. He calls it addictive. Gil asks whether every company shrinks to ten people. Armstrong is careful: tasks are being eliminated, people are not. Existing headcount will not collapse to ten; the same people will get more done, faster. Some two- or five-person companies will do what used to take dozens. If revenue outruns headcount, margins rise.

After Coinbase went public in 2021, Armstrong asked whether he was done. He decided he was at the beginning: the mission of economic freedom, with about half of 1% of global GDP on crypto rails, was unfinished. He still wanted the ability to incubate other companies. A principle he had written down is that people who make money in software have an obligation to put it into hard tech, the more capital-intensive work that advances society. Elon Musk's path out of PayPal was part of the inspiration. With friends he listed space, AI, fusion, brain-machine interfaces — good teams already on those, no unique thesis from him. Longevity looked different: not many great teams, a lot of it snake oil. He cornered Gil, who has a bio background, and heard that much of the science was bogus, caloric restriction included: mice live longer, humans do not.

A series of dinners changed the map. He invited the smartest CEOs and scientists he could get, Gil among them, and asked what on the biology horizon was high potential and underfunded. Epigenetic reprogramming kept coming back. Shinya Yamanaka's lab had shown you could restore younger function in cells, while also changing cell type. NewLimit wanted only the age axis, identity intact. Armstrong describes the feeling as the same as Bitcoin in 2010 or 2011: he could not stop thinking about it for three months. Inspiration, he says, is perishable; act before it fades. The group that formed was Jacob Kimmel, now CEO of NewLimit, plus Blake Byers and Greg Johnson. Armstrong put in initial capital and helped with company-building. Coinbase is the full-time job; at NewLimit he is an investor and board member.

The science moved faster than he expected. A South San Francisco lab of about 50 or 60 people starts with AI: transcription-factor combinations are a massive search space, and they have built what he calls the leading frontier model for epigenetic reprogramming to recommend the next experiments. Then come large pooled screens for phenotypic hits, then slower, costlier functional assays in animal models, then human trials. They have demonstrated successful reprogramming of at least one human cell type in humanized mouse models, are testing in non-human primates, and plan to launch a phase-one trial next year, with three to ten other candidates coming off the platform in the coming years.

The first three cell types on the published roadmap are hepatocytes, vasculature, and immune T cells. The playbook copies GLP-1s such as Ozempic: start with a high unmet-need indication. The liver program's phase one is alcoholic liver disease, where the alternative is a transplant most patients do not get in time and 12-month survival after diagnosis is very low. That indication alone, if the drug works, they estimated might be worth $20 billion. The real goal is a 40-year-old with a 20-year-old liver, immune system, and vasculature; skin rejuvenation by itself he calls a probable trillion-dollar market. Over time, a cocktail of therapies, different payloads, same theme: restore younger function. Most of the diseases that kill people are highly correlated with age. Instead of chasing each as a symptom, they are going for the meta-problem, the root cause. On running two companies, his warning is that distraction too early is dangerous, arguably irresponsible — he tried about ten ideas before Coinbase worked. At scale, doing only one thing forever also limits potential. He names Gil, Joe Lonsdale, Sam Altman, Palmer Luckey, and Musk as people who have walked that blurry line between CEO and capital allocator.

Asked what he is most excited about over five years, Armstrong starts with the job: agentic finance, stablecoin payments, the everything exchange. Crypto still powers about half of 1% of global GDP. Maybe 700 million people have used or hold crypto, with monthly users more like 50 to 100 million. The target is a billion or more, on two legs, humans and agents. He also wants more people to try being builders rather than critics. Then the frontier ideas. One is cognitive enhancement. Folate and certain nutrients in childhood development, in some countries, actually raise IQ. He may host a dinner on therapies in adults, children, or embryos to lift baseline cognition. He has not found a great thesis yet; if someone has one, he wants to hear it. AI is getting smarter. Humans, he thinks, need at some point to take control of their own evolution. Some will merge via brain-machine interfaces; some will not. Humans still have to advance.

He has already invested in Preventive, a company doing embryo editing. The topic is controversial, *Gattaca* and all. Pew research, as he cites it, shows about 80% of Americans support embryo editing for disease prevention. Nobody wants rare genetic disease for their children. He thinks it looks out-there today and will later look like driving without a seatbelt. Gil worries that intelligence and health "packages" could flatten neurodiversity, a Ford Model T of intelligence for everyone. Armstrong has heard that argument and thinks the opposite: unlimited diversity of parental preference — artist, general, engineer, politician — and much of it may only raise the floor, the child still who they were, minus schizophrenia or depression. In the limit he thinks diversity could get weirder, kids with gills or wings, if humans really take control of evolution.

The last meta-problem is what slows progress. Over-regulation is one answer; Europe is the ready example, the U.S. has its own version. Crypto and libertarian circles have long talked special economic zones. Shenzhen, Singapore, and Dubai are the success cases. Coinbase invested in Prospera in Honduras to learn; the local-government story has been mixed, closer to a resolution, and he thinks they will make it. Interest in exit options was higher before the current Trump administration's deregulatory turn. Campaign talk of "freedom cities" is the American version he likes: federal land, 10 or 100 square miles, where EPA and FAA red tape do not apply, so you can build data centers, nuclear plants, and unlimited drone delivery in a contained sandbox, then export what works. Entrepreneurs cannot iterate without a sandbox, and without iteration they cannot solve the problem. Federalism already makes the U.S. 50 competing zones. He thinks it can go further.

The hour is a tour of rails, cells, and regulation, but three moves sit close enough to copy. The first is operational. Give every core service a brain in markdown: incidents, financial controls, A/B tests, the PRs that were accepted and the ones that were rejected. The load-bearing rule is the closed loop. When a human corrects an agent, the reason goes back into the brain so the miss does not recur. One metric tells you if it is working: whether the accept rate on one-shot PRs is climbing.

The second is a product question. Can the service accept an AI agent as a customer? Most agentic-commerce orders Coinbase is seeing are under 30 cents, which card networks cannot clear, and the X402 plus stablecoin path is already being built with Google, Cloudflare, and AWS in the Linux Foundation. Coinbase Business is their own answer for merchants who want humans and agents on the same rails. The store that cannot take a few-cent tool-call payment is not in that economy.

The third is how he screens ideas. A notion has to be something he cannot stop thinking about for three months, and he has to bring an advantage nobody else will supply. Both conditions, then move, because the heat fades. That filter is why he skipped space and fusion, cornered Gil about snake-oil longevity, and ended up in epigenetic reprogramming. Crypto, he keeps saying, was already a good complement for humans. For agents that need to pay a few cents to one another, it is not a complement. It is the account.