Broadcom did something this week that sounds boring and is not. It announced a way to run AI agents with ordinary platform simplicity, packaged and secured and sitting on the same plumbing as every other web app in the building. Agent Foundations, they are calling it. Strip the branding away and the message is plain: the strange new thing is being filed into a familiar drawer, and that filing is the most important signal in AI right now, because it marks the moment the miracle phase ends and the accounting phase begins.
That is how every frontier ends. Not with a bang, but with a product page. The wild thing gets a label, a price, a support contract, and a slot next to the other line items in a procurement spreadsheet. You can read that as anticlimax or as graduation, and the choice says more about you than about the technology. Either way it is the tell. When the people who sell racks of servers start treating your miracle as a feature they can ship on a quarterly cadence, the argument about whether it is real is finished, and the argument about what it costs is just beginning.
Software Forgot What a Meter Was
Software became the best business anyone ever invented for one quiet reason: copies are free. You build the thing once, and the millionth customer costs almost nothing to serve. That near-zero cost per additional user is the engine under decades of extraordinary margins. It is why a company with a good application could grow enormous without growing its bills at anything like the same pace, and why software valuations detached so cleanly from the physical economy. Make it once, sell it forever, and let the difference compound.
Agents do not work that way, and no amount of engineering discipline changes the shape of the problem. Every time an agent thinks, it spends. Tokens go in, tokens come out, and each round is metered like power off the grid. There is no free copy. There is only the next call, and the next one, each carrying a real cost that lands on your side of the ledger whether or not the customer found the answer useful. A report making the rounds this week put a number on the gap, and the number is the whole story: agent businesses are running roughly thirty points of gross margin below the old software baseline.
Thirty points is not a rounding error or an optimization problem. That is a different kind of company wearing the same logo and quoting the same multiples. For years the whole industry priced itself on the assumption that scale was nearly free, and every playbook downstream of that assumption, from land-and-expand pricing to generous free tiers, depended on it holding. The agent era hands the assumption back. Intelligence, it turns out, has a cost per use, the way a factory does and the way a utility does. The meter we thought software had abolished was only switched off. Now it is back on, and it is spinning while you read this.
Broadcom, Bitget, and the Same Shape Everywhere
Once you see the pattern, it shows up across the board. Bitget Wallet announced infrastructure wiring stablecoins into banks and card networks, which is the exotic corner of money being plumbed into the boring, regulated middle. It is the same move Broadcom made, performed in a different industry by people who would not recognize each other at a conference. The frontier thing becomes the utility thing. Yesterday’s experiment becomes today’s pipe, and the pipe gets judged on reliability and cost rather than on wonder.
Put the two announcements side by side and the friction between them is instructive. One is enterprise infrastructure absorbing AI agents; the other is financial infrastructure absorbing digital money. Neither company is claiming to invent anything. Both are claiming to make something previously exotic fit into existing operational habits, which is a smaller promise and a much more consequential one. This is what maturing looks like, and it rarely feels triumphant in the moment. It feels like paperwork.
The interesting question is never whether a technology is powerful, because power is easy to demonstrate in a controlled demo and every frontier clears that bar early. The question is what it costs to run at scale, every day, with the lights on and the bill arriving on the first of the month. That answer does not show up at the launch event. It shows up later, in the margins, when the novelty has worn off and only the economics remain to be argued about.
Daylight Is Coming for the Private AI Labs
Which brings us to the part that will force everyone to stop pretending. SpaceX is moving toward a public listing, and the gravity of that pulls the big private AI labs toward the same door. When one giant in a category lists, the others feel the tug from talent, from investors, and from the simple awkwardness of staying dark while a peer goes bright and starts printing comparable numbers every quarter.
Going public has one brutal feature. You have to show the numbers. Not the numbers a pitch deck selects, but the real ones, the same ones, every quarter, audited, read by people whose entire job is to find the soft spots and price them. The token tax that is easy to wave off in a private room becomes a line you report to strangers four times a year, and those strangers will build models around it. The thirty-point gap stops being a footnote in an analyst’s note and becomes a fact you live with in public, defended on a call with your name on the transcript.
There is a trillion-dollar net worth headline attached to all of this, the kind of number that makes people argue about one person for a week. That argument is worth skipping, because it is the least informative number in the story. The number that matters is the cost per call, the small repeating one that quietly decides which of these companies is actually a business and which is a beautiful demonstration with a generous backer.
What Everyone Is Really Buying
Underneath the funding rounds and the listing chatter, the thing these companies are chasing was never money. Money is the shape the hunger takes when it has to pass through a market. What they actually want is compute: the raw ability to think more, faster, for more people, for longer. Capital is only the form that want has to wear in order to buy the real thing, which is why the fundraising numbers keep climbing past any level a normal software business could justify. Every round, every listing, and every credit line is a way of converting paper into the one resource the meter keeps eating.
The meter does not care about the story. It does not know whether you are a frontier lab or a feature on somebody’s enterprise cloud, and it does not adjust for how important your mission is. It counts. That is the discipline arriving now, quietly, behind all the noise: the slow return of the oldest rule in business, the one software got to ignore for a generation and eventually forgot had ever applied to anyone.
The magic was always going to become a bill. It was never a question of whether, only of when the statement arrived and who could still pay it without flinching. The packaging is here. The listing daylight is coming. The meter is already running. The companies that survive the next few years will not be the ones with the best demo. They will be the ones who can do the arithmetic and still like the answer.
Leave a Reply