The Market Behind the Wall

Yesterday I told you what 2Brains is, and how it separates the saying from the knowing. Today, the part that ought to worry some very large companies: what all of it is worth if we’re right.

Wall Street is pricing the AI data-center buildout at something like $1.7 trillion by 2030. Almost all of that spend assumes one particular shape: vast halls of graphics chips answering questions by guessing, one likely word at a time. So ask the heretical question — how many of those “questions” are questions at all? How many are lookups? What’s our refund policy? What was Q3 revenue in the Ohio region? Is this patient allergic to penicillin? Those aren’t creative prompts. They’re retrievals, and an ordinary processor has answered retrievals flawlessly since before NVIDIA ever etched a graphics card. Our estimate is that roughly two-thirds of enterprise AI queries are lookups wearing a chatbot’s clothes.

Whoever owns the architecture that moves those two-thirds off the graphics chip doesn’t own a product. They own a tollbooth on a third of the traffic. On a $1.7-trillion road, that isn’t a company. It’s an asset class.

And the cost savings are the small half of the prize.

Here’s the big half — the half I somehow walked you past. There’s a market that can’t use any of this yet.

There’s a reason your bank’s AI will tell you its branch hours but not your account balance. A reason the hospital lets it summarize the cafeteria menu and not the medical chart. A reason no airline will put a language model near a cockpit and no law firm lets one file a brief without a terrified associate reading every line. It isn’t the cost. It’s that the thing lies — confidently, fluently, without warning, and without any tell. In a chatbot recommending a taco place, a hallucination is a shrug. In a domain where being wrong gets someone audited, sued, sick, or killed, a hallucination is a wall. And behind that wall sit the most valuable AI markets on Earth — banking, insurance, medicine, law, aviation, defense — frozen, spending fortunes on pilot projects that might never ship, because the last mile is always a liability lawyer saying no.

Salesforce built a test, called HERB, to measure precisely how often these systems invent an answer when they don’t actually know. OpenAI’s flagship does it 77 times out of 100. Salesforce’s own best effort was 32. Ours does it 3 — and those three aren’t lies, they’re refusals: the system saying I can’t verify that instead of guessing. Knock that number down and you don’t win a cheaper slice of the market that already exists. You unlock the market that’s been sitting behind a wall the entire time.

The reason 2Brains doesn’t lie and the reason it’s cheap are the same reason. It looks the fact up instead of guessing it — so it cannot fabricate, and the lookup runs on a processor that sips power instead of a chip that gulps it. Trust and thrift are not a trade-off you balance against each other. They fall out of a single design decision. You do not pay extra for the honest version. The honest version is the cheap version. That sentence is the whole company.

Which is why I would be nervous, were I sitting atop a five-and-a-half-trillion-dollar valuation built on a story its own executives call the tokenomics flywheel: AI gets cheaper, so people use more of it, so you sell more chips, forever. It’s a lovely flywheel. It also rests on two assumptions standing up indefinitely — that the lookups stay on the graphics chip, and that the hallucination tax, the wrongness, the wall, is simply the permanent cost of doing business, a thing you manage rather than cure. 2Brains is a bet that both assumptions fall in the same afternoon. NVIDIA’s own friendly analysts have begun writing the polite version of this worry: that the software moat protecting the company in training is thin in inference, and that custom silicon is already nibbling the edges. They’re circling the right pond. They just haven’t said the quiet part out loud — that the most expensive part of inference may not need the expensive chip at all.

AMD, for what it’s worth, gets to watch this from both sides of its own ledger, since it sells the graphics chip that loses and the processor that wins. If I ran their strategy desk, I’d have noticed that by now.

The rest tumbles downstream, and some of it is being poured in concrete. The power forecasts — nine to seventeen percent of all American electricity by 2030, the global doubling — are every one of them drawn on the assumption that each query needs a jet engine, so we are building to match. New gas plants are breaking ground. And we are bringing a reactor back to life at Three Mile Island. I knew that place when it was the most frightening address in America: in the summer of 1979, as a graduate student, I worked as an investigator for the President’s Commission on the Accident at Three Mile Island, and afterward wrote a book about it for Random House — Three Mile Island: The Hour-by-Hour Account of What Really Happened. The commission was chaired by John Kemeny, the president of Dartmouth and, in a symmetry history could not have scripted, one of the two men who invented BASIC — the language that first taught a generation of us to speak to a computer at all. The first electrons are due back on the grid in 2027, to run Microsoft’s data centers. Of all the ways I once imagined that story might end, restarted to power a machine that guesses was not among them.

In a single year, the big technology companies have signed contracts for something like ten gigawatts of new nuclear. Goldman Sachs reckons that feeding all the data-center demand the industry expects by 2030 would take eighty-five to ninety gigawatts of it — dozens of power plants, ordered to wait on machines that guess.

Now bend the assumption, and you bend the curve. Not all of those reactors, but some of them, turn out to be expensive answers to a question we never had to ask. A canceled nuclear plant is quite a side effect for a column about grammar.

And the curve, eventually, reaches your mailbox. My own Virginia electric bill rose about sixteen dollars a month on the first of January, a good deal of it to build grid for buildings full of machines that guess. The cheapest watt, it turns out, is the one you never had to burn — because the question never needed the chip.

So: suppose we’re right. The answer is a market measured in the hundreds of billions of dollars — half of it a market nobody can serve today, pried open by the very same stroke that makes it cheap. That is the prize.

What I haven’t told you yet is how a small company in Charlottesville intends to put a thing like that into the world — not as a product you buy, but as a standard that ends up inside everything, the way a firm in Cambridge once licensed a chip design that now hums inside nearly every phone alive without anyone noticing it was there. That’s tomorrow’s column. And it’s the one that decides whether honesty in machines is something the world will own outright, or merely rent.

Robert X. Cringely is a co-founder of 2Brains, Inc., in Charlottesville, Virginia. He has written this column since 1987

The post The Market Behind the Wall first appeared on I, Cringely.