China’s Free AI Models Won’t Stay Free, Goldman Says. Here’s the Lesson for Anyone Who Builds on Free.

Kimi K3 and GLM-5.2 are near-frontier and used worldwide for nothing — giving Chinese labs ‘reach and status, but not income.’ Goldman predicts ‘paid weights.’ It’s the oldest pattern in tech: free wins adoption, then the economics arrive. Open weights let you fork, but a frozen model in a fast field is a fading asset, not an escape.

Goldman Sachs has a warning for everyone who built on China’s free AI models: the free part was always the marketing, and marketing eventually needs to pay for itself. The mechanism it predicts is a lesson about ‘free’ that extends far beyond AI.

Goldman Sachs has a warning for the millions of developers who built on China’s free AI models: the free part was always the marketing, and marketing eventually needs to pay for itself. Models like Moonshot’s Kimi K3 and Zhipu’s GLM-5.2 are, by Goldman’s own description, “nearly as good as the American frontier” and used heavily worldwide — for nothing. Goldman’s read is that this can’t last, and the mechanism it predicts is worth understanding, because it’s a lesson about free that extends far beyond AI.

The argument

Chinese labs released powerful models under open terms, and the world downloaded and ran them enthusiastically because they were as capable as American frontier models and carried no licensing cost. The labs got something real out of it — “reach and status,” global influence, a seat at the table. What they didn’t get was income.

Goldman’s proposed endgame is “paid weights”: charging cloud platforms commercial licensing fees to host these open models. The reasoning is simply that influence “has to pay for itself” eventually, and the pressure to make it do so is mounting — global AI stocks have sold off sharply (Japan’s Nikkei fell 14% from its June peak, chip stocks slumped over 40%), and Wall Street has flipped from rewarding AI spending to demanding returns on it. A lab burning money to give away a frontier model looks visionary in a bull market and irresponsible in a correction.

The oldest pattern in technology

This is not really a story about China, or even about AI. It’s the free-tier lifecycle that has played out in software for decades, arriving on schedule in a new market.

A new entrant gives the product away to win adoption and undercut incumbents. Free destroys the competition’s pricing power and builds an enormous dependent user base. Then — once the users are committed, their workflows built, their products shipped on top of it — the economics arrive. Rarely a blunt price hike; usually a gentler squeeze: a paid hosting tier, enterprise licensing, premium features, rate limits on the free version. The value was always going to be extracted. The only question was the timing, and the timing is “after you depend on it.”

Anyone who built on a generous free tier that later tightened — and that’s most of the industry — recognises this immediately. Free is a customer-acquisition cost, and acquisition costs are recouped.

The twist that makes AI different

There’s a genuine complication here that Goldman itself flags, and it’s specific to open weights. Unlike a hosted API you can be cut off from, an open-weight model that’s already been released can be forked. If the Chinese labs start charging, developers can keep running the last free version, or switch to whichever rival stays fully open.

That’s a real constraint on the squeeze — but it’s a weaker protection than it sounds, and it’s worth being honest about why. A forked model is frozen: no improvements, no new training, no security and capability updates, no support. In a field moving as fast as AI, running last year’s frozen weights while competitors ship monthly is a slow way to fall behind. “You can fork it” protects you from a price hike the way “you can keep your old phone” protects you from an upgrade — technically true, and quietly untenable over time. The fork is a floor, not a future.

The lesson worth extracting

Strip it to the principle: building your business on someone else’s free tier means building on a decision that isn’t yours to make. The provider decides when free ends, what replaces it, and how much the thing you now depend on will cost. You find out when the email arrives. This is true of free AI models, free APIs, free platform tiers — any dependency where the price is set by someone whose interests will eventually diverge from yours.

The defence isn’t paranoia about ever using a free service; it’s structuring your dependence so that a provider’s pricing decision is a swap, not a crisis. For AI specifically, that means designing so the model is a replaceable component rather than a foundation welded into your product. If pointing your application at a different model — a cheaper one, a self-hosted one, the next open release — is a config change rather than a rewrite, then any single provider’s decision to start charging is an inconvenience you route around, not an existential event.

This is one of the quieter arguments for owning your stack. A self-hosted, source-available platform like VBWD is built with a central LLM connection precisely so the model is a swappable dependency: your AI features point at a connection you configure — a hosted frontier model today, a Chinese open-weight model while it’s free, a locally-run model when the economics or the privacy demand it — and switching is an operational decision rather than a re-architecture. You don’t escape the need to use someone’s model, but you keep the choice of whose, and the leverage that comes with being able to leave. The same logic we’ve applied to consumption pricing applies here: whoever controls the thing you can’t leave controls your costs.

The read

Goldman is very likely right that China’s free frontier models won’t stay free forever — the pressure to monetise “reach and status” is real and rising, and the free-tier-to-paid arc is one of the most reliable patterns in the history of technology. The open-weight fork gives users more protection than a hosted API would, but a frozen model in a fast-moving field is a fading asset, not a permanent escape hatch.

The actionable takeaway isn’t “don’t use free AI models” — they’re genuinely good and genuinely useful right now. It’s “don’t build so that their staying free is load-bearing.” Treat every model as a component you can swap, keep the switching cost low by design, and you get to enjoy free for as long as it lasts without being hostage to the day it ends. Free is a gift with a timer. Build like you can hear it ticking.

Analysis based on Goldman Sachs commentary as reported on 29 July 2026. Market figures and predictions are as reported and are forecasts, not outcomes. Commentary, not investment advice. The VBWD reference illustrates the swappable-dependency approach and is not an endorsement.