The Biggest Bet Ever Made
The frontier labs are betting on curing us and replacing us, and the bills will not wait for the cures.
Could we see ten percent real GDP growth and twenty percent unemployment in the same year?
It sounds like a typo. Growth needs hands. When the economy runs hot, people get hired. That is how it has almost always worked, for as long as anyone alive has been paying attention.
I think the pairing is possible now. I do not know when. Here is why.
It started with an Instagram reel about the AI buildout. The data centers, the chips, the power, the hundreds of billions. I build on these models every day, so I did what an operator does when he sees a number that big. I ran the back-of-the-napkin math on who pays it back.
OpenAI has told lenders it expects to spend roughly eight hundred fifty-six billion dollars on compute and infrastructure by the end of 2030. Anthropic has signed more than half a trillion in commitments over roughly a decade, most of which it cannot cancel. Its own draft filing says it in one line: "If our actual spend falls short, we must pay Google the difference."
Nobody commits that kind of money without a story about how it comes back.
So I started with the customers everyone pictures. You and me, paying twenty dollars a month. OpenAI is projecting three hundred fifty billion dollars in revenue for 2030. To get there on subscriptions alone, it would need about one and a half billion of us. Charge five hundred dollars a month, a price almost no household pays for software, and it still needs fifty-eight million.
Consumers will carry some of it. Maybe a lot. But I kept staring at the napkin, and something else came into focus.
Every company in America already has a budget for buying work.
For most of them, it is the biggest line item they have.
There are a lot of ways this money could come back. Subscriptions. Advertising. Tools that make workers faster. Industries nobody has invented. Two are worth pulling apart, because they run on different clocks.
The first is the dream. Machines that find drugs faster, catch disease earlier, and stretch healthspan, the years you live healthy and not just alive. Grow the pie so big that a trillion-dollar buildout looks cheap.
The second is the floor. Find a budget companies already spend every quarter without being asked twice. Offer to do the same work for less.
The dream pays in full someday. The floor pays now.
They are not Plan A and Plan B. They run at the same time. The floor keeps the lights on while the dream is still in trials.
Here is the floor on a napkin.
A company spends a billion dollars a year on its people, everything included. AI lets it do the same work with half of them. Payroll drops to five hundred million. It pays a frontier lab one hundred million a year for the models that absorb the work.
The company keeps four hundred million, before transition and oversight costs.
Now ask who says no.
Not the CFO. Not the board. Not the shareholders. Not the lab, which just booked a nine-figure contract. Every person at that table has a reason to make it work. If the work holds up, the financial case is hard to refuse.
Scale it up. Americans earned nearly sixteen trillion dollars in total compensation last year. Most of that work is not going anywhere soon. Nobody is sending a chatbot up a utility pole. But the math does not need most of it. Replace a tenth of that payroll, give the labs the napkin's cut, and you have a market, split across every lab, nearly the size of OpenAI's entire revenue target for 2030.
From one country. From one line item.
And the pressure climbs with the salary. A model reviewing a contract does not bill like the partner who used to review it. The bigger the salary, the bigger the gap. The bigger the gap, the harder the push.
The partner will tell you he is protected. Licenses. Liability. Malpractice. Those are real, and some of them are law. But look at what they rest on. Part of why licensing exists is that human judgment fails. Malpractice is priced on how often it fails. The day a machine fails less than the doctor, and keeps failing less, the argument behind each of those protections starts pointing the other way. It will take years. It will be fought.
But a wall built because humans are safer gets harder to defend the day they are not.
Some work holds. Reading a room. Moving a jury. Hearing what a person will not say. I grant all of it. It is a smaller share of the work than the people doing it want to believe.
There is one more party at the table, and nobody mentions it. The government.
Hire a person and you pay Social Security and Medicare on top of his wages. Buy the model and you pay for the model. Economists who have studied the tax code have argued for years that it leans toward machines over people. The napkin does not need that argument. It only needs the paycheck.
Then there is the cost no tax table captures. A worker is a commitment. Recruiting, training, benefits, severance, the lawsuit nobody saw coming. A usage-based model is a dial. When sales soften, the CFO turns the dial down next month. No paperwork. No exit interview.
Boards pay for dials.
Then competition shows up, and everything speeds up.
The company that cut its payroll in half does not sit on the savings. It drops prices, outspends its rivals, or buys them. The company that waited is now competing against a cost structure it cannot match. Its question is no longer whether. It is how late.
The labs get squeezed too. The hundred-million-dollar contract does not stay at a hundred million. Another lab does it for fifty. An open model on a cheap cloud does it for twenty.
Now the strongest case against me.
If prices fall that far, the labs may never collect what their investors were promised. The employer keeps almost all the savings. The model companies fight over scraps, and some of them lose the bet outright. The floor may never pay for the buildout at all.
I think that is a real possibility. For some of these companies it may be the likely one.
It does not change the employer's math.
His math worked at a hundred million. It works better at fifty. It works best at twenty. Cheaper machines do not slow the substitution. They speed it up.
The labs can lose the bet and the payroll still loses.
The honest answer to all of this is history.
So far, every wave of automation killed jobs and created more. The tractor emptied the farms and filled the factories. The spreadsheet shrank the bookkeeping office and built the financial analyst. I expect the same at first. I think the early years produce net new work. New companies. New roles. Jobs nobody has named yet.
Grant all of it.
Then look at what those new jobs are.
Every one of them is a new line on somebody's payroll. And every line faces the same napkin.
A new job is not exempt because it is new. The incentive waits. It waits for the role to mature, the work to standardize, the growth to level off. Then the CFO runs the same arithmetic on the new job that she ran on the old one.
That is what the napkin proves. No job escapes the calculation. Some will win it. What the napkin cannot prove is that the total falls. That part is my forecast, and I hold it because I watch these models get better every month and I cannot find the new work the next one will not learn.
Agency can give one person an edge. It cannot guarantee enough work for everyone.
The old waves had somewhere to send people. The machine could not follow them there.
This one is being built to follow.
Now give the dream everything it deserves.
The labs are pointing their best tools at medicine. Anthropic has built a workbench for drug discovery and started hunting treatments for rare diseases the drug companies rarely touch. OpenAI has signed with some of the biggest names in pharma. If it works, it belongs next to antibiotics and vaccines. Diseases we have carried for ten thousand years, gone. Years of healthy life added. A pie growing so fast nobody fights over slices.
I want that. I am an optimist about it.
The labs can charge for research tools today. The full medical payoff takes longer. A molecule a model finds this year still has years of trials ahead of it. And when a lab sells its tools to a drug company, the drug company usually keeps the patent. The labs' own drug programs could change that. They are early. Until then, the lab sold the shovel. Somebody else owns the mine.
The compute bill does not wait for the trial results. It comes due on schedule.
That is why the floor matters. The dream is still in the lab. The payroll is already being spent.
In 1850 a French economist named Frédéric Bastiat wrote that a bad economist sees only the visible effect of a decision. A good one follows the effects nobody sees.
Here is what will be seen. GDP climbing. Prices falling. Margins fat. A stock market that loves every earnings call.
Here is what will not be seen unless you go looking. Output, jobs, wages, and the share of income that goes to people who work are four different numbers. People talk about them as if they were one. Nothing forces them to move together.
That is how ten percent growth and twenty percent unemployment could fit in one sentence.
Follow the payroll one more step. Social Security and Medicare's hospital insurance fund run on payroll taxes. Shrink the payroll and you shrink the base they stand on.
Now put the two bets together. If the dream works, people live longer. Some of those years will be working years, and a healthy sixty-eight-year-old who keeps working keeps paying in. That helps, if there is work to keep doing. The rest are years drawing on programs funded by payroll.
If taxable payroll shrinks while retirements lengthen, those programs face pressure from both sides. From the same technology.
The free market has answers, and they are real. Cheaper goods stretch every paycheck that is left. New companies form. People who own the capital spend it. Any of that can carry demand. None of it is automatic, and none of it moves at the speed of a CFO.
The risk is not that production fails. Production is about to get easier than it has ever been. The risk is the gap between how fast the paychecks leave and how fast anything takes their place.
Go back to the napkin.
A billion-dollar payroll. Five hundred million gone. A hundred million to the lab. Four hundred million kept.
Every decision on that napkin was rational. Nobody lied. Nobody conspired. Nobody broke a rule. The competitor who hesitated got punished, exactly the way markets punish hesitation.
Thomas Sowell wrote that there are no solutions, only trade-offs. This is a trade-off the market will make without asking anyone. Prices will find the cheapest way to get the work done. They always have. If the cheapest way stops needing most of us, the price system does not promise anyone another way to earn.
So ask it again. Not as a forecast. As the question nobody on that napkin had a reason to ask.
Ten percent growth. Twenty percent unemployment.
The market is not broken.
That is the problem.
Sources
OpenAI projected compute spending and revenue through 2030, Reuters citing the Financial Times, Sept 18, 2026, https://gvwire.com/2026/09/19/openai-forecasts-cash-burn-near-280-billion-by-2030-ft-reports/
Anthropic's $518 billion infrastructure commitments, Reuters, Sept 29, 2026, https://www.reuters.com/business/anthropics-518-billion-ai-buildout-hinges-largely-deals-that-cannot-be-canceled-2026-09-29/
Compensation of employees, 2025, U.S. Bureau of Economic Analysis via FRED, 2026, https://fred.stlouisfed.org/series/A033RC1A027NBEA
"Does the US Tax Code Favor Automation?" by Acemoglu, Manera and Restrepo, Brookings, 2020, https://www.brookings.edu/articles/does-the-u-s-tax-code-favor-automation/
Anthropic launches Claude Science; OpenAI pharma partnerships, Pharmaceutical Technology, July 1, 2026, https://www.pharmaceutical-technology.com/news/anthropic-launches-claude-science-ai-tool-drug-discovery/
"That Which Is Seen, and That Which Is Not Seen" by Frédéric Bastiat, Library of Economics and Liberty, 1850, https://www.econlib.org/library/Bastiat/basEss1.html
A Conflict of Visions by Thomas Sowell, 1987, https://en.wikipedia.org/wiki/A_Conflict_of_Visions
