Here is a rule that has never failed me. The moment a Goldman analyst, a TechCrunch reporter, a guy on YouTube named Mark who signs off to "the boys," and your own gut all agree on something, somebody is about to lose a fortune, and the smart money says it's the person feeling smuggest. That person, this week, was me. I was feeling extremely smug. I had a scorecard. People with scorecards always lose.

Let me set the table, because the table is packed.

The chorus goes like this. AI is too cheap to be real. The model companies priced intelligence below cost to get everybody hooked, the oldest hustle in the Valley, the free-sample-at-Costco strategy except the sample is your entire engineering department, and now the meter is speeding up exactly as the discount expires. Tokens, the little units they bill you by, fell about 75 percent in a year, and the bills went up anyway, because agents quietly clone themselves into little armies and burn your budget while you sleep, like raccoons, if raccoons had AWS credentials. Uber torched its entire annual AI coding budget by April, which is the corporate equivalent of blowing your whole paycheck before the second of the month and then looking surprised. Somebody out there woke up to a reported half-billion-dollar bill they did not mean to run. Not a typo. Half a billion. By accident. Because nobody was watching the meter, which raises the question of what, exactly, that person was being paid to watch.

And it gets better. Meta told its own employees to maximize token usage as a productivity metric, a sentence that should have ended several careers and possibly a marriage, and then the CTO had to put it in writing that nobody should be using the tools just to use them, which is a thing you only have to say out loud after a great many people have, in fact, been using the tools just to use them. TechCrunch reached for the phrase "crack cocaine." And here's the kicker. It wasn't the reporter who said it. It was a telecom billing lifer, a man who has personally witnessed every pricing crime known to God and the FCC, and he was the one clutching his pearls. When the guy who invented the overage fee thinks the pricing is predatory, pay attention.

And the story has traveled, which is how you know it's no longer news but folklore. When the AI cost panic reaches a marketing YouTuber explaining your own industry back to you on a Saturday, somewhere between a Will Smith reference and a heartfelt goodbye to "the boys," the diagnosis has officially left the building and entered the bloodstream. Even The Economist got around to it, asking whether all these surging tokens were producing any actual profit and dragging out the corpse of the dotcom era to make the point, back when we measured corporate greatness in clicks and eyeballs, two units that also turned out to be worth precisely nothing.

I said all of it. Louder. With footnotes. Four installments. I built a gallows with the kind of care most people reserve for IKEA furniture they actually like. Fifteen red flags. A scorecard. Nine diligence questions engineered to make a founder lunge for the water glass mid-sentence, which, I will admit, I enjoyed designing more than is strictly professional. I split the whole world into wrappers, who rent intelligence from a landlord that is at this very moment measuring their living room for new curtains, and real businesses, who own something that existed before the word "agent" colonized every keynote on earth. I told you the wrapper has four separate ways to die, which is three more than most things need. I told you not to sign the thirty-six-month contract without running the vendor past me first. And I ended on a line I remain insufferably proud of: you are not their customer, you are their burn rate.

It's a good gallows. I stand by the carpentry. The joints are tight. It would hold a much larger man than the one I was hanging.

And then I did the thing every prosecutor eventually does if he's any good at the job, the thing that separates a lawyer from a guy yelling in a parking lot. I went looking for the defense's best witness, fully expecting a clown to honk in on a tiny bicycle, and instead a serious person sat down, adjusted the microphone, and proceeded to ruin my entire week.

His argument has a name. The name is a hundred and sixty years old. And it is the reason I am writing two more parts instead of doing a victory lap with my shirt off. We'll get to the name in Part Two, where I have room to do it the damage it deserves. The short, week-ruining version is this: I might have the whole thing backwards. The exploding bill, the thing I've been treating as a death rattle, might be a birth cry. Cheap might not mean doom. Cheap might mean everywhere. And "everywhere" is not the sound of a company dying. It is the sound of a company eating the world while I stand outside writing its obituary in a notebook.

Here's what poking the argument taught me, before I do the long, serious, genuinely upsetting version below. Three things.

One. Half my best horror stories are governance failures wearing a nice suit. The half-billion-dollar bill is not a law of physics. It is a company that forgot to set a budget and then performed astonishment at the receipt. Gartner ran the math and found that if you actually plan for it, total cost rises only moderately, because the unit price is collapsing nearly as fast as the volume is climbing. Which means a lot of the people screaming about their token bills are the people who left the tap running all night and then published an exposé on the water company. I have, to my horror, been quoting these people.

Two. The bull case is being made by sober adults with spreadsheets, not by guys in Patagonia vests doing breathwork. When Goldman calls the rising bill "margin inflection" instead of "crisis," that is not a press release, it is a thesis, and a thesis is a thing you have to beat, not point at from across the room while making a face.

Three, and this one left a mark. My favorite historical cudgel swings just as hard the other way. I keep invoking 1999. Everybody invoking 1999 thinks they're the genius in the story. But the people who looked at the fiber glut in 1999 and said "this is insane, nobody will ever use this much" were exactly right about the glut and catastrophically, hilariously wrong about everything that mattered. They missed the internet. The whole thing. I would prefer, on balance, not to be those people.

So in Part Two I'm going to do the thing I told you a real journalist has to do, which is build the strongest possible argument that I am an idiot, take it completely seriously, follow it all the way down into the dark, and report back honestly on what's living at the bottom. Fair warning. It's stranger down there than the gallows ever was. The gallows was simple. This is not.

Every AI ad vendor is selling the same line: production got 95% cheaper, so make more, so you win. The first two are true. The third is a lie, and Cheaper, Faster, Quietly Worse proves it with the receipts nobody else bothered to pull.

The cost collapse is real (a $2,000 AI spot that aired during the NBA Finals, built from a 25-to-1 discard pile). But cheaper only matters if the extra output does work. NielsenIQ wired up 2,000 brains and found AI ads score worse on memory even when they look polished, and only beat human ads when nobody can tell they're AI, a disguise that expires a little more every month. So you get a flood of cheap, forgettable, faintly brand-damaging creative, and the only one who reliably gets paid is the platform renting you the engine.

The payoff is a buyer's toolkit: a vendor scorecard, a marked-up "what they say / what punctures it" table, and a map of where AI creative earns its keep versus where it quietly bleeds your brand. It's the thing you bring to the procurement meeting so you don't sign the wrong three-year deal.

Get the full white paper with ADOTAT+

Cheaper, Faster, Quietly Worse is an ADOTAT+ member edition, and the whole toolkit lives behind the membership: the 12-flag vendor scorecard, the marked-up pitch table, the use-it-here/never-here map, and the sourced evidence locker you can quote in a procurement meeting.

ADOTAT+ is where the adtech reporting stops being a hot take and starts being something you can bring to a negotiation. Members get the full Subsidy Cliff series, every field guide and scorecard, and the diligence tools that have saved buyers from signing the wrong three-year deal.

The diagnosis is free. The receipts, the scorecard, and the part you actually use are ADOTAT+.

→ Join ADOTAT+ and read the whole thing.