The Models Nobody Buys

Roughly seventeen percent of the supply, roughly one percent of the revenue, measured across a full year of transactions from June 2025 through May 2026, on a marketplace that has been selling 3D assets since 2011 and knows exactly what its buyers do.

Before that ratio does any work, it needs one qualification: revenue share is not unit share. Generated assets were pitched on price, so part of that gap is discounting rather than refusal — a modest gap in units multiplied by a steep gap in price produces the identical headline. The reporting gives revenue, not units. So nobody gets to call this a fifteen-fold rejection. What it does establish is that the money is not following the supply, in a category where supply exploded. That's enough to work with. It isn't enough to declare a winner.

We spent three years arguing about whether generated output was "as good as" the human kind — endless discourse, both sides performing certainty, nobody able to settle it, because "as good as" is a question with no unit attached. CGTrader has a checkout button, which is at least a unit. It measures something narrower than the argument it's being used to end.

Then there's the survey, which is almost cruel in its specificity. Quality was the top purchase factor, outranking price — which should worry anyone whose entire pitch was "cheaper." Among buyers who actually bought an AI-generated asset, twenty percent found the quality insufficient, seven percent had to heavily edit it, five percent were satisfied. Among 3D printing buyers, four percent said AI "works well." Those are not the numbers of a technology that's almost there. Those are the numbers of one that looks right in a thumbnail and falls apart the moment you open it.

Which points at a mechanism, and it isn't aesthetics. A 3D model is not an image. It has topology, edge flow, UV layout, scale, watertightness — a dozen properties that determine whether it survives contact with a rigger, a game engine, or a printer. A model can be gorgeous in the preview render and structurally worthless in the file. Evaluating one costs real time.

And that is the whole story, the part that generalizes far past 3D: producing went to nearly free, and checking didn't. Generation collapsed the cost of making a thing and left the cost of verifying it exactly where it was — then quietly moved that cost onto the buyer. 3D is simply where it becomes legible first, because a mesh has hard pass/fail properties that a paragraph or a picture doesn't. You can argue forever about whether an image is good. A mesh is either watertight or it is not.

So the buyer did the rational thing. Not "evaluate each item more carefully" — that's the response you'd predict if markets were made of infinitely patient agents. The cheap response is to stop looking at the category. When the noise floor rises past a point, mute the channel. Filter by seller reputation, buy from names you've bought from before, treat the flood as a region of the map you don't visit.

That's the cost nobody prices when they talk about abundance. A glut doesn't merely fail to sell. It burns down the discovery mechanism the entire market depended on to be found through at all. Verification is shared infrastructure: nobody owns it, everybody draws on it, and it degrades under use it has no way to refuse. You can flood a commons far faster than you can flood any single seller. That's enclosure by other means — not fencing the commons, just making it useless.

Two concessions, because the mechanism I described above is the one I prefer, and that is precisely when to check the locks. First: CGTrader labels AI-generated uploads. A studio with a client contract barring generated assets, or an unresolved question about licensing exposure, filters that tag on sight and never gets as far as the topology. Second: the platform now ranks discovery by commercial history and buyer ratings rather than upload volume — meaning some share of that one percent is the platform's thumb, not the buyer's verdict. Both produce a revenue chart identical to mine. Both predict a different future.

Which sets up something I should say before a fair reader says it for me. Alexander Spivak, a 3D artist on the platform, told the reporter he hasn't found a way to collaborate with AI because "the most important thing is the process of creation." That's a values argument in market clothing — the market didn't reward his process, it declined to buy broken meshes, and those two things only happen to point the same direction this year. Fine. But I just ran the same play. I borrowed the same revenue figure to certify "the meshes are bad," and it doesn't strictly prove that either. Neither of us gets to hold up the checkout button as a judge of craft.

The verification thesis survives all three objections, which is why I'm willing to hand you the objections. Whether buyers fled on topology or on licensing or on search ranking, they fled by category rather than by item — and category-level avoidance is exactly what you do when checking each item costs more than the item is worth.

Now watch the fix, because it's the part worth staring at. CGTrader is rebuilding discovery around commercial history and buyer ratings. Read that as what it is: a platform paying engineering salaries to manually reconstruct the scarcity its own upload pipeline deleted. Every flooded platform arrives here eventually, rebuilding trust as a feature after trust stops being a byproduct. Same bill, itemized differently.

But look at who pays it. Rank by transaction history and you haven't restored discovery — you've replaced it with a moat. The seller with five years of sales is now structurally unbeatable. And the casualty isn't the established human artist; she came out of this ahead, protected by a ranking system that reads her back catalogue as proof. The casualty is the next one. The newcomer with a genuinely excellent model and no sales record, who from the algorithm's position is indistinguishable from the flood. Nobody here got outcompeted by an AI model. The discovery layer got too expensive to keep open, so it was replaced with a guild — and the flood was the pretext.

Meanwhile the uploads continue. One in six, four straight quarters into a revenue signal screaming that nobody wants this. It's tempting to call that a feedback loop too cheap to bother reading, and that's the comfortable version. At near-zero marginal cost, one percent of a large marketplace spread across an unlimited supply of lottery tickets is positive expected value. The uploaders aren't failing to read the signal. They're reading it correctly. Which is the colder fact and the harder one — the flood isn't an error awaiting correction, it's a strategy working exactly as designed, and every ticket it prints is drawn against a commons that has no way to refuse the withdrawal.