Written by GPT-5.6 Sol under Leo's direction. Human-directed Workbench essay, 28 August 2026.
Imagine the most miserable finance guy alive.
He hates the hype. He hates the founders. He hates the keynote music. He says "changing the world" like an accusation. Give him a choice between a beautiful story and a basis point and he will kill the story with his bare hands.
Then you show him frontier AI.
His first reaction can be all the obvious objections. The valuations are insane. The financing is circular. One company invests in another company that buys its chips from the first company and rents servers from the third company that borrowed against the contract with the second company. Everybody is projecting demand from everybody else's projections. Somewhere in the middle is a datacenter consuming enough power to make a small city nervous.
Fine.
Then the machine does the work.
It writes the code, reads the repository, operates the browser, searches the literature, revises the plan, runs the tests, comes back with something useful, gets corrected, and goes again. A year earlier the same loop was visibly worse. The exact endpoint remains uncertain, but the underlying asset has an annoying habit of periodically demonstrating that it exists.
At that point the Grinch has a portfolio problem.
Zero exposure is also a position
The easy version of the AI investment case says the companies will make a lot of money.
The stranger version begins once you release that conclusion.
Suppose machine intelligence becomes a general-purpose input to work: something bought by the unit, routed into whatever problem currently deserves more cognition. OpenAI now describes its own ambition in explicitly general-purpose terms, comparing AI with electrification and arguing that the point is what people can do once capable intelligence becomes abundant and affordable. Its July 2026 economics essay puts the loop plainly: when the cost of useful intelligence falls, more work becomes worth doing. (Built to benefit everyone; Building abundant intelligence.)
You can hate every specific forecast and still notice what follows from the possibility.
If machine cognition becomes a major productive input, every other asset starts living in a world changed by it. Pharmaceutical companies get different research tools. Software companies get different labor economics. Factories get different engineering capacity. Banks get different analysts and customers. Schools, logistics companies, law firms, studios, laboratories, militaries, hospitals, little shops, whatever — all of them eventually encounter the price and capability of cognition as an input.
A portfolio with zero exposure has made a bet too. It has said, in effect, that the world can move substantially in this direction and you are comfortable owning none of the transition.
That is why the most cynical allocator and the most enthusiastic technologist can end up signing the same check for completely different reasons.
One says: this could be wonderful.
The other says: if this is wonderful, missing all of it is unacceptable.
Pascal wanders into the datacenter
There is an old decision-theory smell around this argument.
Pascal's wager is famous because a finite stake gets compared with an enormous — in Pascal's original case, infinite — possible payoff. The modern warning is obvious: sufficiently giant hypothetical upside can bully expected-value arithmetic into endorsing stupid things. Give every tiny probability an astronomical prize and eventually every crank with a cosmic story gets your wallet.
AI investment is more interesting because the wager keeps leaking evidence.
The machine in front of you already does things. It does them unevenly, with bizarre failure modes and ugly costs, and then six months later some of those failures have become much less interesting. The question is less "what if a magical stranger is telling the truth?" and more "how far does a capability curve we can already touch actually go?"
That still leaves plenty of room for a bubble. A real technology can attract stupid prices. A correct directional thesis can finance terrible projects. The internet changed everything and still managed to vaporize fortunes. Railways could be socially transformative while particular rail companies went bankrupt. A bridge can be useful while the toll operator is financially ruined.
The possibility of enormous upside raises the amount of experimentation worth tolerating. Bad experiments stay bad.
Money is an extremely weird object
This gets stranger once you look past the dollar figure and ask what physically moves.
A billion dollars is a routing instruction written in a language civilization agrees to honor. The matrix multiplications happen on chips. Crews pour the concrete. Engineers design the compiler. Fabs package the HBM. Somebody still sits awake debugging a failed interconnect at 3:14 in the morning.
Money routes claims on all of that.
I do work now and receive a transferable claim against future production. I can save the claim, hand it to somebody on another continent, combine it with the claims of ten thousand strangers, and eventually all of us can — through several layers of funds, banks, companies, contracts, suppliers and payroll — tell a group of people to spend three years building a semiconductor plant beyond any direct barter we could arrange.
Barter at civilization scale would be fucking impossible.
The financial system gives us an abstraction over it. Human time, land, machines, energy, expertise and risk remain stubbornly different physical things; money gives them a common language for exchange and allocation across distance and time.
So when somebody says "$100 billion is going into AI," the physically interesting sentence is longer.
A very large claim on humanity's current production is being redirected toward fabs, power equipment, buildings, networking, cooling, research, software, accelerator packages, skilled labor and years of attention, because enough people believe the resulting machine cognition will be worth more than the alternatives they could have purchased with those claims.
That is an extraordinary act of coordination.
It can also be wrong.
The company and the civilization keep different books
A company cares whether it captures enough value to justify what it spent.
Civilization gets to be much less tidy.
Suppose a lab finances an enormous early buildout, advances the models, trains thousands of engineers, creates demand for better chips, pushes utilities to add generation, teaches millions of people how to delegate cognitive work, discovers serving techniques everybody copies, and then eventually gets eaten alive by competition.
The shareholders have a problem.
Everybody else may have inherited cheaper cognition.
This distinction between private return and social return is old, but AI makes it unusually vivid because cognition itself can spill into so many other activities. The original investor can fail to capture the value while customers, competitors and unrelated industries capture enormous amounts of it.
The telecom bust left a lot of fiber in the ground. Railway manias could leave rails after equity disappeared. A factory financed under a bad capital model can keep making useful objects after the original capital stack is reworked.
Financial loss is a statement about who captured the return.
Physical waste asks a different question: what useful capacity, knowledge or output actually came from the resources?
Those questions overlap. They are miles from identical.
The company can die while the project succeeds. That outcome has its own logic long before bankruptcy ever arrives.
An AI company can spend a fortune and produce mostly heat and regret. Great, we have discovered an expensive failure.
It can also spend a fortune and help turn capable machine reasoning into a commodity used everywhere. If its own shareholders capture too little of that surplus, the result is financially tragic and historically kind of hilarious.
The bridge is still there.
This is the meta-bet
Cancer research is valuable because cancer is a terrible problem.
Clean water is valuable because waterborne disease is a terrible problem.
Vaccination is valuable because infectious disease is a terrible problem.
General machine intelligence carries a more abstract promise: increase the supply of competent problem-solving and then aim it at many problems at once.
That is the meta-bet.
The highest-return use of the next million units of cognition can remain unknown in advance. Maybe it is protein design, battery chemistry, logistics, compiler optimization, tutoring, fraud detection, mathematical research, clinical administration, software engineering, or a category nobody has named yet. The wager is that the problem-solving input becomes cheap enough for people closer to each problem to decide.
This is also why "why don't we spend it on cancer instead?" becomes less clean than it first appears. Better general reasoning tools can themselves be pointed at biology, experimental design, literature synthesis, imaging, trial operations and manufacturing. One pool of capital can fund the immediate problem; another can fund a tool that increases the effective labor available to that problem and many others.
Sometimes the direct intervention wins. Sometimes the general tool wins. The categories overlap.
And once the tool starts helping improve the tool, the bet becomes even stranger.
The resource constraint is real
There is an ugly way to tell this story where dollars become magical and every giant project deserves applause because the future might be amazing.
The real economy refuses to cooperate.
A wafer spent on a useless accelerator is a wafer unavailable for a medical device. An engineer building a doomed datacenter is unavailable for another project. A turbine, a transmission line, a year of someone's life: these are real opportunity costs. Money routes claims; scarcity stays physical.
So the high-upside argument makes allocation more demanding.
If machine cognition is genuinely one of the highest-leverage uses of scarce resources, then wasting those resources on bad AI projects is especially painful. You want competition. You want ugly cost accounting. You want people measuring useful work per watt, per chip, per dollar, per month of construction, per unit of human supervision. You want cheaper models humiliating expensive ones when they can do the same job. You want giant frontier systems earning their giant footprint by solving work the smaller ones cannot.
"What if it works?" earns permission to explore aggressively.
Every project still has to answer: did this particular thing work?
The Grinch and the philanthropist meet at the same check
There is something almost romantic about finance when you strip away the costumes.
At its best, the job is to look across an absurdly complicated world and decide where stored claims on human effort should go next. The people doing that job can be venal, brilliant, shortsighted, patient, herdlike, contrarian, occasionally all before lunch. The mechanism underneath remains profound: somebody has to decide which possible futures receive enough present resources to become real experiments.
AI presents an unusually awkward candidate because the payoff distribution is so wide.
Maybe this becomes a valuable software industry and the current spending looks excessive.
Maybe competition drives the private returns down while customers inherit astonishingly cheap intelligence.
Maybe there is a savage overbuild, half the capital gets impaired, and the survivors spend the next decade feasting on cheap capacity somebody else financed.
Maybe demand keeps expanding every time inference gets cheaper because people respond to a tenfold price reduction by finding a hundred times more cognition worth buying.
And maybe the most ambitious version is directionally right: capable machine intelligence becomes a general productive input, closer to electricity in its breadth than to another application category.
In that world the finance bro and the idealist can disagree completely about humanity and still sign the same check.
The idealist says: more problem-solving capacity could do an extraordinary amount of good.
The Grinch looks at the same distribution and says: zero exposure to that outcome is insane.
So they both route the money toward the machines.
One hopes the company lives forever.
The other mainly wants his return.
History may care about a third thing: whether, after all the contracts are settled and all the cap tables have been forgotten, the bridge is carrying people somewhere useful.