Coinbase gets 88% of its revenue from something other than Bitcoin trading, Brian Armstrong said, citing the only figures he said he was allowed to quote — the company's most recent earnings deck.
The share price has not followed. Armstrong said the stock still moves closely with the Bitcoin price, while the business underneath it has become an exchange for stocks, commodities, derivatives, prediction markets and stablecoin payments.
"We don't want the AIs to be unbanked. We want to bank the AIs."
Armstrong co-founded Coinbase and took it public in 2021. He also co-founded New Limit, a longevity company that he said has a lab of 50 to 60 people in South San Francisco and a first human trial due next year.
I listened to the full interview so you can skip it. 45 minutes of audio, 23 minutes of reading.
Here are the 11 takeaways that matter.
👤 Guest: Brian Armstrong, Co-founder and Chief Executive Officer of Coinbase, and a co-founder and board member of the longevity company New Limit
🎙️ Host: Elad Gil, a startup investor who co-founded Color Health and Mixer Labs and is the author of the High Growth Handbook
📰 Published: 10 September 2026 on the No Priors feed
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 45 min | ✅ Time saved: 22 min
Key Takeaways
88% of Coinbase's revenue already comes from something other than Bitcoin trading, while the stock still tracks the Bitcoin price
Armstrong said those are the only numbers he is permitted to quote, from the company's most recent earnings deck
Card fees make agent-to-agent payments uneconomic, which is the whole commercial case for crypto rails
He put 76% of the agentic commerce Coinbase sees at under 30 cents a transaction, against a flat card fee of about 30 cents
Coinbase is issuing financial accounts to software agents, not only to the people who run them
Two versions: a segregated agentic account tied to a human identity, and a self-custodial wallet any agent can open with one prompt
Coinbase keeps a written "brain" for each team and repository, and forces every human correction back into it
The measurable result he cites is a rising accept rate for code changes an agent gets right the first time
He rejects the idea that AI shrinks headcount
"People are not being eliminated. I think tasks are being eliminated"
Tokenized stocks are the next thing after stablecoins, aimed at the roughly 4 billion people with no brokerage account
Coinbase's product is live only outside the US, and he said it is a real security held one-for-one in custody rather than a derivative
Prediction markets hit a $100M revenue run rate inside months of launching in the Coinbase app
Armstrong thinks the bigger use is policy and unresolvable questions, not sports
New Limit's first clinical trial targets alcoholic liver disease, a market he put at about $20B on its own
The actual goal is rejuvenating healthy tissue, and he called skin alone a trillion-dollar market
He wants US federal land turned into deregulated build zones for nuclear power, data centers and drone delivery
His diagnosis is that entrepreneurs have no sandbox to iterate in, which he called "incredibly stifling"
1. The Everything Exchange
Gil opened on Coinbase's strategy, and Armstrong laid out three businesses rather than one.
The first is the exchange itself, and the pitch is that asset classes are converging onto one venue. "It's now possible to trade everything in one place with good liquidity and good cross margin," Armstrong said, listing stocks, commodities, crypto, perpetual futures and prediction markets. He credited emerging regulatory clarity with letting Coinbase offer that in more countries
The second is stablecoin payments, which he said grow regardless of the crypto price cycle. Even when Bitcoin prices fall, he said, stablecoin payment volume keeps rising
On cost and speed he said: "It's kind of like under 1 second, under one cent US to move money anywhere in the world."
"You can send money at the speed of information" was his summary of it
The third is what Coinbase calls Agentic Finance, or AIFI. Armstrong said the first piece is a consumer-facing AI adviser, already launched in the Coinbase product, doing the work a wealthy client gets from a human investment adviser — portfolio construction, dollar-cost averaging, tax-loss harvesting, staged buying into a 5%, 10% or 15% dip, fee reduction and yield on idle cash
He framed the adviser as an access argument rather than a technology one. Investment advice has been available to wealthy people; his claim is that everybody should now have it, and that the product is "getting better and better"
2. Banking The AI Agents
The second piece of Agentic Finance is giving the software itself an account.
His starting position is that agents are currently locked out of the payments system. "We don't want the AIs to be unbanked. We want to bank the AIs. They deserve financial services as well," Armstrong said
The obstacle is identity verification, not willingness. He said there is no know-your-customer process an agent can pass: they have no government ID and cannot walk into a bank branch, at least not yet, "without their humanoid robots companion"
Coinbase is building it two ways at once. One is a separate agentic account on Coinbase with segregated funds, tied to a human's identity, which an owner can hand to an agent for trading and payments. The other is a self-custodial wallet available to any agent that wants to hold a stablecoin balance — and, he said, an agent could raise money or issue a token
The problem he says this solves is agents stalling mid-task. Armstrong described the familiar failure: the agent hits a paywall, or needs to spin up cloud resources, or reaches a hotel booking page, and comes back asking its user for a credit card
His model for the fix is an employee's spend account rather than a shared card
He expects a stablecoin-backed credit card to be needed for backward compatibility with merchants who cannot take anything else
Payment volume is already moving over a protocol Coinbase built and then gave away. Armstrong said x402 was incubated at Coinbase and contributed to the Linux Foundation, and that Google, Cloudflare and AWS are among the parties now working on it with Coinbase
Coinbase publishes the resulting activity on a site called agentic.market, which Armstrong said shows how many payments are flowing and which services are selling into the agent economy. He also named Coinbase Business, the product that lets a merchant accept stablecoin payments from humans and agents alike
3. Why The Payments Are Tiny
Gil asked how much of future volume runs on crypto rails rather than conventional ones. Armstrong's answer was an argument about transaction size.
The economics of card payments exclude the transactions agents actually make. Armstrong put the minimum card fee at about 30 cents flat plus a percentage on top, and said of that floor: "So arguably it doesn't really work well for anything under a dollar."
His own data puts most agent payments below that floor. "I believe the stats we last saw on this are that about like 76% of the Agentic Commerce transactions we're seeing are under 30 cents"
Asked what a sub-30-cent transaction buys, he said information. His examples were venture firms compiling research from paywalled data, recruiters scraping LinkedIn, and — most often — agents paying other specialized agents for source data, which he compared to a tool call
He expects specialist agents to become sellable products. Armstrong said frontier models are trained as strong generalists, but that smaller open-weight models can be fine-tuned or trained with reinforcement learning to beat them on a narrow task
His in-house example: train a small open-weight model on 100,000 Coinbase compliance cases and it can outperform a frontier model on the compliance review queue
His market examples were an agent that is simply good at ordering pizza, and a designer agent trained on proprietary Figma data
He treats the scale of that economy as close to arithmetic. "There'll be more agents than humans in the not too distant future," Armstrong said, and on that basis he expects the agentic economy to be larger than the human economy
4. Will There Be Money?
Gil raised the AI-safety worry that agents accumulate resources and use them to influence society. Armstrong said he does not hold that concern as strongly, and turned instead to whether money survives at all.
He named Elon Musk as the source of the argument he was answering, saying Musk has suggested there will be no money in the future because energy and mass become the scarce inputs
He conceded the limit case and rejected the timeline. Armstrong said that in the limit Musk has a point, but that scarcity persists for a long time — specifically land, and usable land in particular, plus energy and chips
His illustration was that "Dyson spheres are going to be expensive I think pretty far into the future"
He allowed both directions on prices at once: robots and AI could make commodities, food and housing very cheap, while government money printing could create inflationary pressure
The conclusion is that a medium of exchange is still needed, and that crypto rails are what make the agent economy work: crypto "was really good for humans and it's going to be essential for AI"
Gil's own observation was about people, not technology. He said the early AI community overlapped heavily with the crypto community — the same computer science graduates went into crypto or AI depending on the year they left school, splitting careers by graduation date rather than by interest
Armstrong's reply was that the choice is false. "You can do both. They're going to intersect here."
5. Coinbase's Company Brain
Asked what Coinbase is doing with AI internally, Armstrong went past coding tools to a system he says compounds.
He named the basics first and called them table stakes. Coding tools hooked into the company's data repositories, and a long-standing use of machine learning for risk, fraud prevention and customer support inquiries
The ambition he stated is explicit. "We basically want to get to as close as we can and actually achieve recursive self-improvement," Armstrong said
The mechanism is a written body of context he calls a brain, kept at several levels. There is a brain for the company, and others for each team, each service repository, and each individual — and he said an individual's brain should leave with them when they leave Coinbase, which Coinbase is writing policies to allow
What a brain contains is operational history, not documentation. "It's basically it's a history of all the incidents that have ever happened with that service," Armstrong said, plus the financial controls that must be enforced, every A/B test ever run on it, and the full GitHub history of accepted and rejected pull requests
He was blunt about the implementation: "Today it's just mark it's like markdown files"
The step he says matters is what happens when a human corrects the agent. When a reviewer finds something the agent missed, the instinct is to patch it and ship. Armstrong said the rule Coinbase is enforcing is that the edit and the reasoning behind it go back into the brain, so the same miss is fixed in every future case
The metric he cites is the accept rate for pull requests an agent gets right first time, which he said "just starts to tick up over time"
Coinbase's internal agent harness is called Toshi, and Armstrong said it can now call external vendors and pay them over the x402 protocol
He described using it himself that morning. Armstrong said he asked an expensive high-end model to plan a complex feature in three phases of ten pieces each, then told it to execute phase one by spinning up ten agents in parallel and choosing cheaper models where they would do — a mix of open-source models and Grok
Two minutes before the recording, he said, it reported back: "By the way, all 10 of those things from phase one are now done and ready for you to review."
He described the effect on his own behavior as addictive, because instead of messaging the team with every idea or bug he now sends them a finished pull request to review
6. Tasks Die, Not Headcount
Gil put the common claim to him that AI makes the average company smaller. Armstrong did not accept the framing.
He separated productivity from headcount. "Productivity is definitely increasing," he said — but that is not the same claim as average company size falling
His distinction is the load-bearing one. "People are not being eliminated. I think tasks are being eliminated"
For an existing company he expects acceleration, not shrinkage. Armstrong said he does not think Coinbase or companies like it cut to ten people; the people already there get much more done and everything moves faster
He agreed the very small company becomes more capable. Two- or five-person companies will do more than was historically thought possible, in his account, but that coexists with large organizations rather than replacing them
The financial consequence he draws is margin. If revenue grows faster than headcount, he said, margins should rise over time — while allowing that some companies will keep growing both at the same pace
7. Tokenized Stocks Arrive
Gil asked what share of Coinbase's volume or revenue now sits outside crypto. That produced the interview's headline number and then a product argument.
The revenue mix has already moved. "Actually 88% of our revenue is from non Bitcoin trading at this point," Armstrong said, noting that this came from the recent earnings deck and that those are the only numbers he is allowed to discuss
The share price has not caught up, on his own reading. "If you look at our stock price it does trade like fairly correlated with the price of Bitcoin," he said — while adding that he loves Bitcoin, expects it to last, and called it "digital gold"
Coinbase's tokenized stock product is the part he says is genuinely new. It is available only outside the US for now, and he said Coinbase is working with the SEC and others on a US path
His distinction from earlier attempts is legal, not technical: "It's not a synthetic or derivative or some debt instrument. It's actually a security and it's one to one represented by the actual security held in custody"
The market he is aiming at is people with no brokerage at all. Armstrong said about 4 billion people in the world have no access to any brokerage or US investment account, and therefore no way to own Coinbase, Nvidia or Apple
His precedent is Tether's success outside the US among people who wanted dollar-denominated accounts
He expects tokenization to run down the asset list. Stablecoins were the first case — a token representing one dollar in a bank account or a Treasury bill — then stocks, then private credit, Treasuries and bank deposits
His consumer example is transfer rather than trading: gifting a share of stock to a nephew by sending it to their wallet instead of going through a stock transfer process he called "antiquated"
Gil's contribution was the historical parallel and a failed bet of his own. He noted the funds set up so foreign investors could get exposure to Hong Kong-listed stocks, and said he invested seven or eight years ago in Harbor, a company trying to tokenize real estate. Armstrong's response was that it will happen, and that Gil was "early, but too early"
8. Markets For Big Questions
Asked where prediction markets end up, Armstrong gave a growth figure and then a much broader claim.
Growth inside the Coinbase app has been fast enough to disclose. Armstrong said that within months of launching, prediction markets "had hit like $100 million revenue run rate," and that growth is "over 100% quarter over quarter, something like that," with the caveat that he would have to check the exact number
He accepts that sports is the early use and says the category is much bigger. Sports betting is entertainment and worth celebrating, in his framing, but not the reason he is interested
His policy case is that a market can price a counterfactual. Armstrong's example: ask what the unemployment rate would be one, two and three years after a given policy is implemented, let a market form on that, and use it to inform which policy gets chosen
He went further, to questions that never resolve. Armstrong raised betting on whether there is a god, or on evolution against intelligent design — what he said some people would call opinion markets, because there is no resolution date
His mechanism is repricing rather than settlement: if 60% of people believe something and a new scientific paper or a new terrorist attack shifts that to 80%, there is a tradable move
His analogy is equities: "It's kind of like when you invest in the stocks, there's not some resolution date of the stock where it's done and you get your payout from Nvidia. You're just betting on is it going to do better or worse in the future."
He put the whole category at an early stage, saying of prediction markets for life's big questions that "we're just scratching the surface"
9. Why He Started New Limit
Gil asked how Armstrong started a company while running a public one. The answer was a long account of how he chose the field.
He ruled out leaving Coinbase first. After the 2021 listing Armstrong said he asked whether he wanted to do something different, and concluded he was at the beginning rather than the end: "About half of 1% of global GDP is running on crypto"
He is candid about which role suits him. "I'm probably better as an entrepreneur than I am as an investor," Armstrong said, describing some of his own investments as things he wanted to exist rather than things that would make money — and being reminded by others that making money is the point of investing
His stated obligation is to move software money into hard technology. Armstrong said someone who makes money in software has an obligation to put it into capital-intensive hard tech that advances society, and cited Musk's use of the PayPal proceeds as the model
He eliminated the obvious fields because they were already covered. Space, AI, fusion energy and brain-machine interfaces all had good teams, he said, and he had no differentiated thesis about any of them
Longevity was the exception because he thought the field was weak. Armstrong said he did not see many great teams working on it and thought much of it was "kind of snake oil," and that Gil — who came from a biology background — had told him the science was largely bogus, pointing to decades of interventions like caloric restriction that extended life in mice and failed in humans
The dinners and the rabbit hole
His method was to convene people and ask one question. Armstrong hosted dinners for chief executives and scientists, Gil among the guests, asking each what was most exciting, high-potential and underfunded in biology
The answer that stuck was epigenetic reprogramming, which Armstrong said some labs had already shown early results with, naming Shinya Yamanaka's lab. Those experiments changed a cell's type; Armstrong said New Limit does not want to do that
He compared the obsession directly to his first one. Armstrong said he could not stop thinking about epigenetic reprogramming for about three months, the same way he could not stop thinking about Bitcoin in 2010 or 2011, and said a moment of motivation like that is perishable and has to be acted on
The team came out of those dinners. Armstrong named Jacob Kimmel, now New Limit's chief executive, and co-founders Blake Byers and Greg Johnson. He put in initial money, helped with company building, and described himself as a sponge on the science
His current role is deliberately limited. Coinbase is his full-time job, he said, and he is "basically just an investor/board member at New Limit"
He named the risk in doing this too early. Armstrong said he tried about ten ideas before Coinbase worked, that distraction can be dangerous, and that starting a second thing too early is "arguably like irresponsible"
Livers first, then skin
The lab is a screening operation, not a hypothesis shop. Armstrong said New Limit has 50 to 60 people in South San Francisco, and that the search space of transcription factor combinations is large enough to need AI to test millions of hypotheses
He says New Limit built its own model for this. Armstrong described it as the leading frontier model for epigenetic reprogramming, whose job is to recommend the next set of lab experiments — large pooled screens, run at the biggest scale the lab can manage
The funnel narrows through three stages. Pooled screens look for phenotypic hits; the best of those go into functional assays, typically animal models, which he said are slower and more expensive; the best of those go to human trials
The results so far are in animals. Armstrong said New Limit has demonstrated successful reprogramming of at least one human cell type in humanized mouse models and is now testing in non-human primates, with a first phase one clinical trial launching next year and three to ten further drug candidates expected off the platform
The three cell types it has started with are published on its own site: liver cells, vascular cells, and immune cells, T-cells in particular
The first indication is chosen the way drugs usually are — highest unmet need, not biggest market. Armstrong compared the sequencing to GLP-1 drugs like Ozempic, and said the first phase one will be in alcoholic liver disease, where he said patients have no option other than a liver transplant most of them do not get in time, and survivability within 12 months of diagnosis is very low. On his own estimate that indication alone could be worth $20 billion
The real target is healthy tissue in healthy people. Armstrong's framing is a 40-year-old wanting the liver, immune system, vascular system, brain, muscles or skin of a 20-year-old, delivered as a cocktail of therapies with different payloads and one shared mechanism
"That's probably a trillion dollar market right there if you can get your skin rejuvenated"
The strategic argument is to attack age instead of diseases one at a time. Armstrong said the diseases that kill people "are all correlated with age," because a young body recovers from an errant cancerous cell or liver damage and an old one has lost that function
Gil restated the science back to him to check it. He described it as isolating the age component of the Yamanaka factors from the cell-type component, and reaching for the everyday evidence: young tissue heals wounds and injuries faster than old tissue does
10. Raising The IQ Floor
Asked what he is most excited about over the next five years, Armstrong gave the company answer briefly and then spent most of it on what he called meta-problems.
His main focus is the Coinbase roadmap already described — Agentic Finance, stablecoin payments and the everything exchange
He sized the runway in users, not price. "There's maybe about 700 million people in the world who have used crypto or hold crypto," he said, of whom monthly users might be 50 to 100 million; his target is a billion or more
He began with an unprompted argument for building over criticizing. Armstrong said everybody should try to contribute to civilizational progress, that it is easy to be a critic or an activist, and that "I think it's better to try to be a builder and you're going to fall flat on your face" — which he said produces both empathy and a more fulfilling result
The frontier idea he is contemplating is cognitive enhancement. Armstrong said adequate folic acid and certain nutrients in child development have been found to raise IQ in some countries, and that he might host a dinner on how therapies could target adults, children or embryos to lift baseline cognition
He was explicit that he has no thesis yet and asked anyone with one to contact him
He framed it as a race: AI is getting smarter, and he thinks humans need to take control of their own evolution and do the same, with brain-machine interfaces as a third, hybrid path some people will choose
He is an investor in embryo editing and expects the politics to flip. Armstrong named Preventive, a company doing embryo editing, called the topic controversial today, and said research shows 80% of Americans support embryo editing for disease prevention
His prediction is that not screening will eventually look reckless, like "driving without a seat belt"
Gil pushed back on genetic enhancement narrowing humanity. He raised the prospect of purchasable packages — an intelligence package, a health package — and asked whether defaulting parents into the same one would reduce the neurological diversity of the population, comparing it to buying the Model T of intelligence
Armstrong said he thinks the opposite happens. Human preferences are unlimited, in his account, so parents would choose differently — a great artist, a general, an engineer, a politician — and much of the editing may just raise the floor by removing conditions like schizophrenia or depression while leaving the child who they were going to be
He took it to the extreme case: children born with gills to breathe underwater, or with wings
11. Freedom Cities
Gil closed on special economic zones, noting Shenzhen's role as a manufacturing hub in China and proposals involving the Philippines and US interests.
Armstrong's interest is in the category of problem, not the specific place. He said he is drawn to meta-problems that unlock progress across many things at once
His diagnosis is regulatory. "I think if you look at what slows down progress in many parts of the world it is overregulation," Armstrong said, naming Europe as the clearest example and saying the US has its own imperfect version
Coinbase has money in one of these projects. Armstrong said Coinbase invested in Próspera in Honduras in order to learn what was happening there, described a mixed experience with the local government, and said his understanding is that a resolution is close and the project will make it
Interest in exit options fell as US deregulation picked up. He said there was more appetite for this before the most recent Trump administration, because the deregulatory environment has created opportunities domestically and made the US a good substrate to build on
The specific proposal he pointed to is "freedom cities," an idea he said was discussed during the campaign and can be found by searching for it: take federal land and designate it for a strategic capability
His examples of what such a zone would be for were nuclear power, drug discovery, crypto and drone delivery
The mechanism is exemption. On a 10- or 100-square-mile parcel, he said, a builder would not have to go through "all the normal red tape on EPA" or clear it with the local community — build data centers, build nuclear plants, fly drones without FAA approval
The purpose he gives is experimentation that stays contained, with anything that works then going through the ordinary process to be rolled out nationally
The absence of a sandbox is the part he calls damaging. Armstrong said an entrepreneur cannot solve a problem without iterating on it, and that having nowhere to try is "incredibly stifling"
He ended on the argument that the US already runs the experiment. Armstrong named Shenzhen, Singapore and Dubai as successes elsewhere, and said "the US is 50 special economic zones" through federalism and the state system — competition he credits with US leadership, and which he thinks could go further
Bonus Insights
Armstrong put a rough shape on how much of the financial system crypto has actually reached, returning twice to the same figure: about half of 1% of global GDP runs on crypto today
Gil's account of serial company-building was a list of the people doing it — Joe Lonsdale, Sam Altman, Palmer Luckey and his bank Erebor, with Musk as the most famous case — and his own view that the line between chief executive and investor blurs at scale, because both jobs become capital allocation
His warning was about concurrency rather than ambition: two things going wrong at once costs you sleep and your family
Armstrong added Daniel Ek to that list, describing him as having made the transition to incubating more companies in Europe, and said a founder's time is usually more valuable to a company than the capital
His own AI adoption argument was about non-engineers. Gil said the people who do best with these tools without being day-to-day engineers treat it like a video game with levels, and noted Armstrong's own engineering background at Airbnb before Coinbase
He does not think an agent needs a bank in the ordinary sense. The agent accounts Coinbase is issuing use self-custodial wallets on crypto rails precisely because there is no identity check to pass — which is the same property that makes them impossible to offer through a conventional bank
Armstrong's bottom line is that the same crypto rails he spent a decade building for people are what the agent economy will run on, because the payments it generates are too small for a card network to carry — and that the businesses he is most interested in now are the ones that remove a constraint on everything downstream of them, whether that is a payment floor, a regulator's approval, or the age of a cell.
Products, Companies & Tools Mentioned
Coinbase (Armstrong's company — the everything exchange, stablecoin payments and Agentic Finance, with 88% of revenue from non-Bitcoin trading)
New Limit (His longevity company, using epigenetic reprogramming to restore younger cell function; first phase one trial due next year)
x402 (The agent payment protocol Coinbase incubated and contributed to the Linux Foundation; Google, Cloudflare and AWS now work on it with Coinbase)
agentic.market (Coinbase's public dashboard for agent payment volume and the services selling into it)
Tether (His precedent for demand outside the US — people who wanted dollar-denominated accounts and had no bank offering them)
Preventive (The embryo-editing company he has invested in, and the basis for his claim that 80% of Americans support editing for disease prevention)
Próspera (The Honduran special economic zone Coinbase invested in to learn how such projects work; he described friction with the local government)
Figma and LinkedIn (His examples of proprietary data behind a specialist agent, and of data an agent pays to scrape)
Nvidia and Apple (Named as the kind of stock the 4 billion people without a brokerage account cannot buy today)
Grok (One of the model families his ten parallel agents ran on that morning, alongside open-source models)
Toshi (Coinbase's internal agent harness, now able to call and pay external vendors over x402)
Harbor (Gil's investment seven or eight years ago in tokenizing real estate — Armstrong's verdict was "early, but too early")
Books & Resources Mentioned
New Limit's therapeutics roadmap (Where the three starting cell types — liver, vascular and immune — are published, as Armstrong pointed listeners to)
Shinya Yamanaka's lab work on reprogramming cells (The early results that made epigenetic reprogramming look investable to him; the lab changed cell type, New Limit is after cell age)
Coinbase's most recent earnings deck and earnings call (The only source he said he was allowed to quote numbers from — the 88% revenue figure and the prediction-market run rate)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: