Nvidia is giving its AI models away for free, and it will tell you exactly why.
"The idea is that if more people outside of the top few firms that buy a lot of Nvidia chips can access affordable and pretty good and fine-tunable open source models," the reasoning goes, "they will want more compute and therefore buy more Nvidia GPUs."
Read that twice. This is not a research foundation seeding a field. It is a customer-acquisition memo, stated in the open by a company that no longer needs to disguise its motives because the market has stopped pretending to be surprised.
This is not a research foundation seeding a field. It is a customer-acquisition memo, stated in the open by a company that no longer needs to disguise its motives.
Last week, this column traced Nvidia backstopping its own demand through OpenAI and Groq — a chipmaker financing the buyers of its chips. This week the thread extends in two directions at once: down into a free-model giveaway, and up into a multibillion-dollar equity talk with Perplexity. They are the same machine.
Start with the giveaway, because it is the retail funnel. Free, fine-tunable open models drop the cost of entry for every developer who isn't already a hyperscaler. More developers fine-tuning means more inference, more training, more racks. The models are loss leaders whose payoff arrives as GPU orders. That is the bottom of the conveyor — broad, diffuse pull-through manufactured at zero marginal cost.
The top of the conveyor is where the checks get written, and where the structure stops being honest.
Consider the Poolside deal, announced this month and graded, correctly, as a takeover wearing a subscription's clothing. The package runs roughly $7 billion: a $6 billion non-exclusive license for Poolside's "Model Factory," plus $1 billion of new equity at a $12 billion pre-money valuation, job offers to 109 employees, and, per ainvest.com's analysis, a planned $6 billion investor payout by the end of 2027.
Line those up as you would an unlock calendar. The license fee is the emission; the investor payout is the unlock. By end of 2027, Poolside's existing backers get cashed out to the tune of $6 billion — the same $6 billion Nvidia pays for the technology.
A turnover event is being run through a payment schedule.
Nvidia was careful to insist this was "not an acquisition and it is not an acquihire." The label is legal. The numbers are economic. What Nvidia bought, in order of size: a non-exclusive usage right (Poolside can still sell the same capability to Amazon or a ten-person inference shop), 109 people who know how to build model tooling, and, per ainvest.com's analysis, a minority stake of roughly 7.7% once the post-money settles at $13 billion.
You do not offer jobs to 109 of a startup's staff to license its software. You do it to move the capability without the consolidation — and without the regulatory headache a full acquisition would invite. Microsoft ran the same play on Inflection in 2024: hire the team, license the tech, skip the merger review.
Which brings us to Perplexity, and the tell.
Nvidia is in talks to invest in Perplexity at a valuation north of $30 billion — a jump of more than 50% over the roughly $20 billion the company reportedly carried after its last reported round a year ago. Standard growth-stage financing, on its face. Nvidia is already on the cap table, reportedly having led a Series E extension in July 2025, with Jeff Bezos among the participants in earlier rounds.
But before the equity round, Nvidia reportedly explored something else entirely: paying billions to license Perplexity's technology and acqui-hire its staff. The same target. Two interchangeable structures. It is not clear whether the licensing component is still live or was set aside during negotiations.
That interchangeability is the whole point. When you can approach the same company as either a $6-billion-style license-and-acquihire or a $30-billion equity round, and choose between them the way you'd choose a shipping option, structure has stopped reflecting economics. It has become a costume.
The economics are identical either way: Nvidia deploys capital into a heavy consumer of Nvidia compute, and the compute demand comes back as revenue. What changes is only how the move reads to regulators, to auditors, to a credit analyst squinting at the investment line. License it and it's an expense. Buy equity and it's ordinary venturing. Acqui-hire it and it's a talent story. The demand-capture underneath never moves.
Perplexity is a clean specimen of what Nvidia wants to lock in. Agentic search — its "Perplexity Computer" product automating tasks — burns tokens at a rate that ordinary blue-link search never did. Reuters, citing The Information, reportedly put annualized revenue above $750 million, up from under $250 million at the start of the year, though it's worth flagging that other trackers peg ARR closer to $450–500 million and suspect the $750 million figure is a conflation with a $750 million Azure infrastructure commitment. Perplexity has never disclosed audited figures. Any multiple on a $30 billion valuation is, as one account put it, an exercise in faith.
Faith is the operative word, because the pull-through matters more to Nvidia than the multiple. Higher token consumption is the return. The equity is almost incidental — a way of routing a compute contract through a cap table.
Now widen the lens. Nvidia's equity commitments passed $40 billion in 2026 alone, spread across model developers, neoclouds, and infrastructure firms that almost without exception buy Nvidia silicon with the money. OpenAI, Anthropic, xAI, Poolside, SSI, Groq at $20 billion, Enfabrica at $900 million — and now, maybe, Perplexity.
Here is the conflict that makes all of this combustible. Nvidia is supposed to be the neutral silicon supplier every lab depends on — the arms dealer who sells to all sides. It is now also the equity backer of specific bets among those sides. It sells the picks and it owns some of the miners.
The debt market has already priced the tell. When the company announced roughly $750 billion of AI-related commitments, its own credit default swaps reportedly widened to record levels — a signal that the arrangement looks less like diversification than like a supplier financing its own order book.
That is the phrase to hold: financing its own order book. The free models manufacture demand at the bottom for free. The equity-and-license checks lock in the highest-consumption labs at the top. And in between, the structure of each deal is chosen not to reflect what's actually being bought, but to control how the buying looks.
Nvidia reports next in November. Whether a Perplexity stake lands in the investment line by then will say more about the pace of this cycle than the valuation does. But watch the structure it arrives in — license, equity, or acquihire — because that choice is now the most honest thing Nvidia discloses about what it's really buying.
It is buying its own demand. It has simply gotten good at making the receipt say something else.