Jason Gardner was sitting in a sushi restaurant in San Francisco in 2010 when he pulled out his wallet and found a stack of coupons, gift cards, and Groupons. The friction didn’t come from not having money. It came from having too many ways to move it, each one a separate little protocol with its own rules about where it worked and when it expired. That moment became Marqeta. Fifteen years later the payment layer is being rebuilt again, from a different direction and by different hands, and the friction problem is the same one wearing a new outfit. The question that matters isn’t whether the rebuild works. It’s what the rebuild costs, and specifically which part of the stack gets cheap while the part sitting on top of it gets capitalized at a trillion dollars.
Look at what landed in a single week and the direction stops being ambiguous. OpenAI connected ChatGPT to bank accounts through Plaid, which means the assistant you already talk to about everything else can now see where your money goes. Bitget Wallet announced an onchain payments layer built to connect banks, card networks, and blockchains in one piece of infrastructure, which is a sentence that would have read as science fiction when Gardner was staring at his gift cards. And Crypto.com’s onchain wallet sits in the Google Play store as an ordinary app among your other apps, no mystique required, no hardware key, no invitation. AI is moving toward money and money is moving toward chain. These are not separate bets. They are the same architectural instinct placed from three different starting positions, and the instinct is that whoever sits closest to the moment of payment gets to define what happens next.
The Hidden Variable in the System
There’s a passage in Tools of Systems Thinkers by Albert Rutherford about a self-published author who kept his publishing pace, hit his launch targets, and watched his sales hold steady, and who still saw profits drop for the first time. Nothing on the dashboard he was watching had moved. Sales had been a perfectly good proxy for the health of the business right up until the moment it wasn’t, because the underlying economics had shifted underneath a surface metric that hadn’t caught up yet. The lag between those two numbers is where the whole story lives, and by the time the surface metric finally moves, the decisions that caused it are a year old.
That lag is the part of the payments infrastructure story that isn’t getting said clearly enough. The cost of moving money is genuinely collapsing, and not in the vague way that technology press uses the word. Stablecoin rails don’t charge 2.9 percent plus thirty cents. An AI budget assistant doesn’t bill by the hour or require a relationship manager. Open protocols don’t have quarterly earnings calls, which means they don’t have the institutional pressure to hold a price. Every one of those is a real, measurable reduction in the toll charged for the act of moving value from one place to another. But the companies building the intelligence layer on top of that collapsing cost base, the ones connecting the assistant to the bank account and training the models that learn your spending behavior, are preparing to go public at valuations near or above a trillion dollars.
SpaceX. OpenAI. Possibly Anthropic. Market analysts are already reading a flurry of record-size listings as a signal that the market itself may be near a top, and their reasoning has less to do with the quality of the businesses than with the timing. CNBC gathered the case that mega-IPOs tend to cluster at market peaks, not because the companies going public are bad, but because record offerings at record prices have historically arrived at exactly the moment when the people who had been waiting the longest decide they can’t afford to wait any longer. That’s a statement about patience running out, not about fundamentals. It’s the sales line looking fine while the profit line has already turned.
Which is the same thing Gardner noticed in his wallet, scaled up by fifteen years and several orders of magnitude. More mechanisms did not mean more value. More mechanisms meant more complexity, more points of failure, and more friction dressed up as convenience. A wallet with seven ways to pay is not seven times as useful as a wallet with one.
Access Without Ownership
Meanwhile, the cost of building with these tools keeps falling, and it’s falling faster than the narrative about building them has adjusted. A developer today can ship an AI product without a team of machine learning engineers, without a large cloud bill, and without a funding round. There’s a running argument on the timeline about what building an AI product still seems to require versus what it actually requires now, and the gap between those two beliefs is where a lot of hesitation lives. People are still pricing in a barrier that quietly came down. The infrastructure that once demanded institutional resources is being commoditized underneath the institutions that built it, and those institutions are not in a hurry to announce it.
That decoupling is the actual story: access widely distributed, ownership highly concentrated. OpenAI will go public at a price that assumes it owns the intelligence layer in the way a railroad owns track. But the intelligence layer doesn’t hold the way a railroad or a phone network holds. Track is expensive to lay twice. Weights are not. The open-weight models get better every quarter, local inference gets faster and cheaper, and the stablecoin rails that don’t need a bank charter keep expanding into places where charters were the whole moat. A trillion-dollar listing is a bet that centralization wins and keeps winning. The entire rest of the architecture being built this same week is a bet that it doesn’t.
Both bets can be right on different clocks, and that’s the part that makes this hard to trade and harder to reason about. Centralization usually wins the first five years, because distribution and capital compound faster than open alternatives mature. Decentralization tends to win the tail, because the cost curve only moves one direction. An investor buying the IPO and a developer building on open weights are not disagreeing about the facts. They’re disagreeing about the horizon.
What the Wallet Actually Knows
The ChatGPT and Plaid integration is being read as a product feature, and it’s better read as a land claim. If your assistant understands your spending, it doesn’t need to be your bank. It just needs to be the thing you consult before you move money, which is a different and quieter kind of leverage. Not the infrastructure leverage of owning the rails, where you take a cut of every transaction that crosses them, but the attention leverage of being the first thing asked. The rails can commoditize completely and the position at the front of the decision still has value, because the decision is where the margin was hiding all along.
Gardner’s insight in 2010 was that friction in payments is rarely visible from the outside. It shows up as hesitation. It shows up as the abandoned cart, the unredeemed card, the moment you look at your wallet and decide not to bother. The next version of that friction won’t be too many gift cards. It’ll be too many assistants, too many onchain wallets, and too many apps each claiming to be the last hop between you and your money. Every one of them will be cheaper than the card networks were. Collectively they will reproduce the exact problem that made the card networks feel like a relief.
Why the Surface Metric Holds Longest
The companies going public at record prices are betting they will be the one you keep when the consolidation comes, and that bet is not obviously wrong. Somebody does get kept. But the systems thinker’s habit is to watch the profit line rather than the sales line, and to treat a steady surface as evidence about measurement rather than evidence about health. Right now the surface metric is holding beautifully. Valuations are up, listings are queued, and the infrastructure announcements keep arriving. The underlying economics have already started moving in a different direction, and nothing in the current numbers is obligated to show it yet.
What goes into a prospectus is confidence, priced and formatted and audited. What comes out of it a year later is either vindication or a very expensive lesson about mistaking the surface for the structure. The difference between those two outcomes usually isn’t visible on the day of the listing, which is precisely why the listing is when everyone is most certain.
Leave a Reply