AI After the Bubble
Concede the financial critique. Then build for what comes next.
Picture a shop owner who spent eighteen thousand dollars on AI subscriptions last year. Three tools, one internal champion, no defined win. Then one spring morning, the vendor behind the tool she leans on hardest triples her rate on ninety days’ notice. Her choice comes down to paying whatever they ask or watching a workflow she has come to depend on stop working on a Tuesday.
If that scene sounds invented, ask around. Some version of it is playing out in businesses across the country right now, and the mechanics behind it guarantee it will get more common, not less.
You have probably noticed that the conversation around AI’s economics has changed. The industry’s financial critics spent two years being dismissed as cranks. They are not being dismissed anymore. Their argument, that the companies selling AI access are losing staggering amounts of money on every customer and have no credible path to stop, has migrated from contrarian corners of the internet to the mainstream financial press. The numbers stopped being disputable. The largest AI lab’s audited 2025 financials showed roughly thirteen billion dollars in revenue against thirty-four billion in costs, with net losses growing nearly eightfold in a single year. The Bank for International Settlements, an institution not known for drama, warned this June that data center spending is outrunning the cash flows meant to justify it. One infrastructure giant has taken on more than a hundred billion dollars in debt to build capacity for demand that exists mostly in projections, with its founder pledging personal shares as collateral.
I run a regulated SBA lender that was early to serious AI adoption, and I founded Main & Machine, an AI implementation firm, which means I get asked two questions almost weekly now. If the critics are right, is AI finished? And if I invest in this now, am I building on quicksand?
Here’s the truth: they are largely right about the money, and the technology is going to be fine. Both of those things are true at once, and the space between them is where every owner reading this should be doing their thinking.
So concede the critique. Assume the reset comes. What then?
The Bubble Is a Business Model, Not the Technology
Let me name what is actually happening in plain language, because the coverage keeps blurring two different things.
The thing inflating past reason is a specific business model: charging by the query for access to enormous general-purpose systems that cost more to run than customers will ever pay. The frontier labs priced their products below cost as a bet. Either running the models would get cheap fast enough to close the gap, or one winner would take the whole market and set prices at will. Neither outcome has arrived. Both look less likely with each earnings report.
Beneath the labs sits a second layer of exposure. The companies building the data centers have borrowed against the assumption that the demand curve holds forever. If the curve breaks, the debt breaks with it, and debt problems at that scale have a way of becoming everyone’s problem.
And then there is the third layer, the one that matters most to you: the customer. Owners have spent three years being trained to treat AI as a metered utility, priced per query, billed monthly, exposed to the balance sheet of a company they do not control. Every workflow built on that arrangement inherits its fragility. The shop owner from the opening scene learned this the expensive way. She thought she had bought a capability. She had actually rented one, at an introductory price, from a landlord who was losing money on the lease.
Now, the concession that keeps this piece honest. AI is real. The capability is real. I am confident enough in it that we are building our own lending operation around it right now, under the eyes of regulators and a board, and I will walk you through that decision shortly. The productivity potential inside well-designed workflows is as concrete as anything I have seen in twenty-five years of operating businesses.
What is fake is the pricing. AI is real. The bubble is a business model, not the technology. Hold those two sentences together and most of the panic, and most of the hype, resolves into something you can actually plan around.
What the Dot-Com Collapse Actually Taught Us
We have run this experiment before, and recently enough that most Ampersand readers lived through it.
Between 2000 and 2002, the dot-com crash wiped out roughly five trillion dollars in market value. Pets.com died. Webvan died. Hundreds of companies whose entire thesis was “the internet, plus our logo” died. Serious people wrote serious columns declaring the internet a fad, a mania, a tulip craze with modems.
The internet did not die. What died was a set of business models built on the assumption that traffic itself was worth money, that growth could substitute for revenue indefinitely, and that being early was the same as being right. The technology kept improving straight through the wreckage. The crash killed the pricing and the ownership structure and left the capability standing.
Two things happened next, and both matter for the decade in front of us.
First, the companies that treated the internet as infrastructure rather than as a business model inherited everything. Amazon existed before the crash and survived it because it was a retailer first and a website second. It had inventory, margins, logistics, and customers who paid actual money for actual things. The internet made its operations better; the internet was never the product. When the froth burned off, the businesses built that way kept compounding while their flashier competitors liquidated office chairs.
Second, the overbuild itself became the foundation. The telecom companies of the late nineties laid staggering amounts of fiber optic cable at ruinous cost to the people who financed it. Most of those companies went bankrupt. The fiber stayed in the ground. That dark fiber, bought out of bankruptcy for pennies, became the physical substrate for everything the next two decades delivered: streaming, cloud computing, the modern web. The people who paid for the infrastructure and the people who profited from it were two different groups.
Think about that for a moment. The bubble’s investors funded the future and then handed it, at a discount, to whoever was still standing.
That is the pattern, and it is remarkably consistent across technology bubbles going back to the railways: the bubble destroys the pricing and the ownership, never the technology. The next generation of winners buys the wreckage cheap and builds on it. The question for an owner is simply which side of that handoff you want to be on.
The Six Things That Change When the Bubble Pops
If the pattern holds, and I believe it will, here is what the morning after looks like for business AI. Six predictions, each one already visible at the edges if you know where to look.
1. The token meter dies. Metered, per-query pricing exists because it lets vendors pass their unpredictable costs on to you. It survives only as long as customers tolerate it. After the reset, owners will pay for a workflow that works, priced as a fixed cost against a defined outcome, the way they pay for every other piece of operational infrastructure. Vendors who cannot commit to a fixed outcome will lose the customer segment that pays its bills on time, which is the only segment that matters in a downturn.
2. Small models eat most of the work. The industry spent three years chasing frontier scale because scale was the story investors funded. But sit inside an actual business and count the tasks. Ninety percent of the daily work in a lending office, a dental practice, or a fabrication shop involves reading documents, summarizing them, checking them for completeness, classifying them, and drafting routine language. None of that requires the largest system ever built. It requires a competent one that runs cheaply, predictably, on hardware someone you trust controls. The economics of the reset will force the industry toward what the work actually needs.
3. The workload moves in-house. Regulated industries figured this out first because they had to. When the system doing the work runs on a machine in your own building, three hard conversations get simple at once: the compliance conversation, the vendor conversation, and the pricing conversation. The direction of travel is toward local capability handling the bulk of the work, with selective use of larger outside systems for the small fraction of tasks that genuinely demand them. The frontier becomes a specialist you consult, and stops being a landlord you pay rent to.
4. Judgment stays human, on purpose. The bubble narrative sold “AI does the work.” The durable version draws a harder and more useful line. The machine drafts, summarizes, retrieves, and classifies. A person decides. Every business built around that boundary keeps running after the reset, because the reset touches the price of drafts and leaves the value of judgment exactly where it was. Businesses that handed decisions to the machine will spend the next few years discovering what that boundary was for.
5. Implementation beats subscription. The subscription era rewarded distribution: whoever could sign up the most users fastest won the funding round. The implementation era rewards delivery. The winning firms on the other side will be the ones who come inside your operation, learn your workflows, do the unglamorous integration work, and hand you a system that keeps working when the market convulses. Selling access was the old business. Building capability is the next one.
6. The most interesting companies after the reset will be mid-size operators, in every industry, who built AI into their operations quietly and now hold a durable cost advantage their larger competitors cannot match without ripping out the plumbing. The frontier labs will dominate the headlines through the correction, the way the telecom giants dominated the headlines through 2002. The compounding will happen somewhere much less glamorous: in the fabrication shop and the regional lender and the forty-person logistics firm that spent the bubble years building instead of subscribing.
Run all six predictions through a single filter and they say one thing. The bubble rewarded scale. The reset will reward control.
How We Are Building for the Morning After at B:Side
I want to show you what control looks like in practice, and I can do it from the inside, because we are in the middle of building exactly this kind of system right now.
B:Side Capital, the company I lead, is a nonprofit SBA lender operating across Colorado, Arizona, New Mexico, and Utah. We are regulated, examiner-audited, and board-governed. Every borrower file we touch contains sensitive financial data, and every credit decision we make is a decision a specific person has to own, in writing, in front of a regulator if it comes to that. When AI got hot in 2024, the temptation was the same one every executive team felt: buy the flashy vendor tools everyone else was buying and announce an AI strategy.
Our board approved a different path. We are building an internal system, named MARCUS, with an executive owner and a set of constraints that came before a single line of anything else. No borrower data leaves the building. No machine ever approves or denies a loan. No vendor sits between us and our regulator. The system runs on hardware we own, inside our own walls, using small models that fit on that hardware, with a deterministic rules engine making every eligibility and pricing determination the same way every time. The models draft. A person decides.
Here is what the system is being built to do. It reads incoming borrower files. It summarizes them. It runs a completeness check against what the file should contain. It drafts the first pass of internal memos, and it traces every claim it makes back to a source document, so the human reviewing the draft can verify rather than trust.
And here is what humans kept, deliberately and permanently. Every credit decision. Every hardship conversation with a borrower going through the worst stretch of their business life. Every submission to a regulator. The operating rule underneath all of it is simple enough to fit on an index card: the machine can hold knowledge, and it cannot hold responsibility.
Notice what building this way does to the economics. The cost of the system is knowable before it runs: the hardware, the electricity, the people who maintain it. Every one of those numbers sits on our side of the ledger, which means the cost does not move when someone else’s board decides to reprice. That is the part that matters for this article. The shop owner from the opening absorbed a tripled rate on ninety days’ notice. Our exposure to that kind of Tuesday is the electric bill.
I will offer the honest coda, because a case without mistakes is a brochure, and we are far enough into the build to have already collected a few. When we announced the project to our staff, I led with what the technology could do. I should have led with what would not change. People hear “AI initiative” and privately calculate their own odds, and I let that anxiety run longer than better leadership would have. We also learned that the operating rhythm around a system like this, the weekly review of what it drafts and where it fails, belongs on day one rather than on the list of things to formalize later. Whatever you build, at whatever scale, start there.
Three Questions Every Owner Should Ask a Vendor This Quarter
Now, I know what some of you are thinking. You run a twelve-person company, you are never going to build an internal system with a name and an executive owner, and a nonprofit lender’s architecture is not your Monday morning problem. Fair enough. You do not need our system. The design principles underneath it are fully portable, and they compress into three questions you can ask any AI vendor this quarter. Three, and only three.
One. Where does the work run, and who controls that hardware? If the answer is a vendor’s cloud, priced per query, you are exposed to the reset and to every repricing decision made between now and then. If the answer is hardware you or your implementation partner controls, at a fixed and knowable cost, the reset becomes a headline you read about rather than a bill you receive.
Two. What does the machine decide, and what does a human decide? If the pitch involves the machine making a decision your business will be held responsible for, a credit call, a medical judgment, a safety determination, walk away. The machine drafts. A person decides. Every time, in every workflow, with no exceptions carved out for convenience.
Three. What happens to your workflows if this vendor triples the price, gets acquired, or shuts down? If the honest answer is “we start over,” then what you have bought is a dependency wearing a workflow’s clothes, and dependencies fail on someone else’s schedule, usually at the worst possible moment.
That is the whole test. A vendor who answers all three cleanly is selling you capability. A vendor who cannot is selling you subscription-era plumbing at implementation-era prices, and the clock on that arrangement is already running.
Build So the Next Morning Belongs to You
Come back to the shop owner one last time. Eighteen thousand dollars, three tools, one tripled rate, and a workflow held hostage on a Tuesday. What she needed was a workflow she owned, running on infrastructure she controlled, priced against a result she could measure. A bigger subscription solves none of that. A better-designed operation solves all of it, and that option is available to her right now, at her scale, for owners willing to treat AI as an operating decision rather than a shopping decision.
The correction, when it comes, will be loud. Write-downs, bankruptcies, columns declaring the whole thing a mania. Remember the fiber in the ground. The technology will keep working straight through the noise, and it will get cheaper on the other side, and the owners who spent this era building workflows they control will wake up with a cost structure their competitors cannot touch.
AI is real. The bubble is a business model, not the technology. The bubble rewarded scale, and the reset will reward control, and control is built the same way it has always been built in a small business: deliberately, unglamorously, one workflow at a time.
The bubble will do what bubbles do. The businesses built on top of it, thoughtfully and quietly, will keep running the next morning. Build so the next morning belongs to you.
To learn more about MARCUS or to see what human-centric AI adoption could look like in your business, check out Main & Machine.