Giving away top-tier language models while approaching $1 billion in annual revenue isn’t a contradiction—it’s a sign the market is shedding its old skin. Z.ai, the company behind GLM, is on the verge of becoming the first independent Chinese firm to cross that threshold, according to Bloomberg. The figure is a projection, not yet a closed book, but the trajectory forces a rethink of the rules of the game.

Anyone who thinks open-source or free models are economically unsustainable has to reckon with a startup that has deliberately chosen the opposite path. GLM—open-source models with a large context window and bilingual Chinese-English architecture—are available without licensing costs, yet the company generates substantial revenue. How? By selling enterprise services: cloud APIs, custom fine-tuning, vertical solutions for finance, manufacturing, and public administration. The model isn’t new, but the scale of potential revenue turns it into a case study for the entire industry.

The value shifts from weights to infrastructure

When a powerful LLM is free, the scarce resource is no longer the model itself but the ability to run it reliably in production. Z.ai appears to have understood this earlier than most, building an ecosystem where monetization happens after the download. This has an immediate corollary for those opting for on-premise deployment: the software barrier drops sharply. With GLM freely downloadable, an organization can run inference locally without worrying about recurring per-token or per-call costs, shifting total cost of ownership (TCO) toward hardware and in-house expertise.

The second-order implications concern digital sovereignty. In Europe, where GDPR and data residency requirements are strict constraints, having capable models that run self-hosted without access fees means building AI pipelines that never leave the corporate perimeter. This isn’t a side issue: if a vendor gives away the model and charges only for management and customization, the user remains master of the data. It’s a trust architecture that cloud-only solutions struggle to replicate without complex contractual guarantees.

Who wins and who loses in this scheme

Direct beneficiaries are enterprises with solid technical teams and adequate GPU infrastructure: freed from licensing costs, they can invest in hardware optimization, quantization, and inference acceleration. For Western vendors whose business relies on proprietary APIs, however, the spread of free, high-performing models like GLM is a competitive pressure hard to ignore—it erodes the closed-model moat and reduces willingness to pay a premium in unregulated contexts.

Hardware makers—NVIDIA, but also the new Chinese inference chips—get an indirect but robust boost: the more free models circulate, the more organizations seek on-premise GPUs to exploit them, fueling demand that goes well beyond a few hyperscalers. It’s no coincidence that the Chinese market is pushing domestic accelerators offering good compute density at lower costs, precisely to serve local and regional deployment scenarios.

Of course, caution is warranted: a billion dollars is still an aspiration, and local competitors (from Baidu to SenseTime) aren’t sitting idle. But the structural signal is loud: the pairing of “free model, paid services” is no longer a niche—it becomes a financially ambitious strategy capable of reshaping balances not only in China but wherever data sovereignty and deployment control matter as much as model accuracy.