The Bar That Serves AI Can't Serve Profit

In Beijing's Zhongguancun district — the city's tech hub — an AI-themed pub called AGI Bar gives developers free DeepSeek tokens with their drinks. The bar is losing money. The owner says the value of free tokens given away is many times what he earns from sales.

The bar is named after artificial general intelligence, but its registered Chinese name plays on "knowledge distillation" — a term that works for both liquor and the AI technique of transferring knowledge from large models to smaller ones.

One failing business is not an investment thesis. But the bar captures something useful about the gap between Chinese AI culture and Chinese AI economics. The enthusiasm is real. The community is dense. The financial model that sustains either one is not yet visible.

The stock investors are pricing is Zhipu AI (Z.AI) — officially Knowledge Atlas Technology — which listed in Hong Kong in January 2026 as the first pure-play large language model developer to go public anywhere in the world. The company raised $558 million in its IPO. It went on to raise approximately $4 billion in a July accelerated bookbuild of new shares. At its peak, the stock reached a market capitalization near HK$470 billion, roughly $60 billion, far beyond what its revenue could justify through conventional multiples.

But here is the number the narrative has to carry: Zhipu reported approximately $107 million in revenue for fiscal 2025, against a net loss of nearly $500 million. The company loses several dollars for every dollar it earns. (2513.HK) — officially Knowledge Atlas Technology — which listed in Hong Kong in January 2026 as the first pure-play large language model developer to go public anywhere in the world. The company raised $558 million in its IPO. It went on to raise approximately $4 billion in a July accelerated bookbuild of new shares. At its peak, the stock reached a market capitalization near HK$470 billion, roughly $60 billion, far beyond what its revenue could justify through conventional multiples.

But here is the number the narrative has to carry: Zhipu reported approximately $107 million in revenue for fiscal 2025, against a net loss of nearly $500 million. The company loses several dollars for every dollar it earns.

That is the unit economics of a company trading at a market capitalization near $60 billion. Revenue would need to grow 26 times to justify that valuation through conventional multiples. The company's own trajectory and analyst expectations point toward a long road to profitability, with valuations that assume revenue growth orders of magnitude larger than what has been delivered.

At first, the story was about technology. Then it became about a movement. Now it's a question of whether a movement can finance itself.

The open-weight model is the center of the Chinese AI ecosystem, and the center of its economic problem. Open-weight means the trained model parameters — the numerical weights that make an AI system intelligent — are freely downloadable. Anyone with sufficient compute can run them. That is different from open-source software, where the marginal cost of an additional copy approaches zero. Running an AI model requires expensive chips, electricity, and data-center capacity. Every new user increases the infrastructure bill.

Chinese companies have led the world in open-weight distribution. In early 2026, Chinese models accounted for 5.3 trillion of 8.7 trillion tokens served through OpenRouter's top-10 models — roughly 61 percent. Chinese models charge a small fraction of Western prices — some services pricing tokens at roughly one-thirtieth of what U.S. competitors charge for comparable service.

The strategic rationale appears to be geopolitical: turn advanced AI into a commodity, pressure U.S. labs to cut prices, and embed Chinese models into global developer workflows. The strategy is explicit in national policy: treat advanced AI as a commodity, pressure Western labs to cut prices, and embed Chinese models into global developer workflows. The financial payoff for individual companies is secondary to national objectives.

But investors in Zhipu don't hold a thesis about China's industrial policy. They hold a stock. And the stock requires the company to convert cultural influence into revenue fast enough to justify its price.

The business has two segments. About 74 percent of Zhipu's revenue comes from enterprise on-premises deployments — customized AI solutions sold to finance, government, and energy clients. The rest comes from cloud APIs and developer tools. That enterprise focus sounds solid. The margin data suggests otherwise: gross margins on the on-premises segment collapsed to negative 0.4 percent in the first half of 2025, because fulfilling custom projects requires expensive compute that the sales price doesn't cover.

The open-weight strategy compounds the problem. When Zhipu releases a model's parameters for free, other companies download and run it. Cloud providers like Alibaba host it. Corporate customers run it on their own servers for data security. The lab that spent hundreds of millions training the model receives little to no ongoing revenue from the users it created. Only Zhipu's own inference API generates recurring revenue for the company itself, and that API competes against the very model it gave away.

Revenue is growing rapidly in absolute terms — from $18 million in 2023 to $39 million in 2024 to $107 million in 2025. But losses are growing faster. Each generation of model requires more compute. Each new competitor undercuts prices further. The gap between cultural reach and financial capture widens.

The market has rewarded Zhipu as though it were already the Chinese equivalent of OpenAI — a dominant platform with pricing power and a path to monopoly margins. The IPO priced at a valuation implying that path existed. The $4 billion follow-on offering in July, priced at a 7 to 13 percent discount to the stock's then-current level, raised fresh capital for training and compute expansion while simultaneously diluting existing shareholders.

Then the lock-up expired.

In July 2026, cornerstone investor shares that were locked for six months hit the market — 25.6 million shares, nearly 6 percent of the outstanding float. On July 17, the stock fell 28 percent in a single day. Goldman Sachs estimated $274 billion worth of locked-up shares would be released into the Hong Kong market over the following 12 months, the highest volume ever. Brokers warned that the scale of unlockings created significant liquidity pressure.

A 28 percent single-day decline is not a routine correction. It's a moment when investors who bought at the IPO price — and made returns exceeding 1,000 percent — chose to convert narrative value into cash. The stock has partially recovered since, but the episode revealed something the bull case couldn't address: the people who bought first and held longest are selling.

The ritual says loyalty to the mission. The filing says dilution.

This is where the identity layer matters. The Chinese AI community is not a faceless mob. It's a dense, geographically concentrated ecosystem of developers, researchers, and engineers centered in cities like Beijing, working at labs that compete with each other and collaborate through open-weight releases. AGI Bar hosts parties for AI labs, developer seminars, and university research meetups. The community treats model releases as events. The language shifts from "I built" to "we released."

That identity is consistent with the open-weight philosophy: shared knowledge, collective progress, technology as public utility. The bar owner describes AI as a future utility like water or electricity — more important than the Industrial Revolution. That is a genuine belief, not a marketing slogan.

But the identity also creates a reflexive loop. Developer enthusiasm drives model adoption, which drives community visibility, which attracts investment capital, which raises valuations, which enables the company to spend more on training, which produces better models, which reinforces the community. Every link in that chain is real. None of them generates profit.

The loop works until the community meets a bill. The bill arrives through lock-up expiry, through dilution, through the gap between the $60 billion market cap and the $107 million revenue. The developers can keep building. The shareholders have to decide whether to keep buying, keep holding, or keep selling — and those decisions aren't coordinated.

A useful comparison clarifies the structure. Alibaba makes money from AI not primarily by training models, but by hosting them. Its Volcano Engine holds nearly half of China's public-cloud machine-learning-as-a-service market. Alibaba's advantage is that it captures the infrastructure cost — the compute, the bandwidth, the data-center capacity — while the model labs absorb it.

The model creators do the science. The cloud providers collect the rent.

That doesn't mean Zhipu is worthless. It means the company's value depends on one of two paths: either it becomes the dominant model and converts that dominance into enterprise contracts that cover its compute costs, or it becomes so good that hyperscalers build on top of its models and pay for the privilege. The first path requires pricing power that doesn't currently exist. The second path depends on competitors being willing to pay rather than downloading parameters and running them themselves.

Both paths are possible. Neither one is priced in at the current revenue trajectory.

The falsifiable question is simple: can Zhipu grow revenue by a factor of 10 to 20 before the capital runs out? The company raised $4 billion in July, a signal that institutional investors believe a path to scale exists. But the valuation also implies a very specific outcome: revenue must grow many-fold over the next few years while losses compress. That requires enterprise customers paying enough to cover inference costs, consumer monetization that overcomes a cultural expectation of free services, and third-party licensing from open-weight models that converts into recognized revenue. Recent research suggests the industry is moving toward revenue-sharing contracts and commercial licensing restrictions, which could help. But the mechanics are still being built.

The evidence that would support the bull case: revenue growth exceeding the 5x forecast, gross margins on enterprise contracts turning positive, successful monetization of open-weight licensing, and the company reaching cash-flow breakeven by 2028 as projected.

The evidence that would break it: another failed capital raise, revenue growth slowing below 3x, further margin collapse on enterprise deployments, or a credible competitor from a cash-rich hyperscaler that undercuts Zhipu's on-premises pricing. The stock's 1,200 percent run from IPO leaves very little room for the second scenario.

The bar will keep serving drinks. The models will keep getting better. The question for the stock is whether the people building the technology can build a business around it before the money runs out.