Zhipu AI Raises $560 Million in Hong Kong as First AI Company to Go Public Despite Losing Money 15 Times Faster Than It Earns Revenue

China's First AI Unicorn Goes Public—And Reveals the Industry's Profitability Problem

Zhipu AI's blockbuster Hong Kong debut exposes the fundamental tension shaping the global AI race: soaring adoption paired with economics that don't yet work

When Zhipu AI priced its Hong Kong IPO at HK$116.20, drawing 1,159-times oversubscription in the public tranche, the milestone was more than just another frothy tech listing. As the first large language model company to go public globally, Zhipu's debut crystallizes a paradox now defining artificial intelligence: investors are desperate for exposure to a transformative technology whose business model remains alarmingly unproven.

The company, operating as Knowledge Atlas Technology , will commence trading Wednesday with a market capitalization near HK$51 billion—despite losing RMB 2.36 billion in the first half of 2025 on just RMB 161.78 million in revenue. That's a burn rate outpacing revenue generation by roughly 15-to-1, yet investors lined up anyway. The oversubscription wasn't retail mania alone: international institutional demand reached 15 times the available shares, signaling sophisticated capital believes something fundamental is shifting.

The Geopolitical Wedge Creating Commercial Opportunity

What sophisticated investors see is less about Zhipu's technical prowess than its strategic positioning. When Anthropic moved to restrict Chinese entities from accessing Claude's API following U.S. export control tightening, Zhipu moved swiftly, offering developers a migration path to its GLM-4.7 model through simple API endpoint swaps. This wasn't innovation—it was availability arbitrage.

The timing proves instructive. Zhipu was added to the U.S. Entity List in January 2025, cutting off American capital markets and semiconductor access. By anchoring in Hong Kong and pitching "sovereign AI" infrastructure, the company transformed regulatory constraint into a moat. For Chinese enterprises navigating compliance requirements and procurement committees wary of geopolitical dependencies, GLM-4.7 becomes the default not because it leads benchmarks, but because it's the compliant option that's actually accessible.

This explains why Zhipu allocated 70% of its HK$4.17 billion net proceeds to general-purpose model R&D through 2028 rather than pursuing profitability. The company is funding a race to become infrastructure—the AI equivalent of running fiber-optic cables while the internet was still dial-up.

Unit Economics That Make Venture Capitalists Wince

Yet Zhipu's prospectus reveals troubling details obscured by IPO euphoria. Cloud-based services—theoretically the highest-margin, most scalable segment—generated negative gross profit in the first half of 2025. Total gross margin compressed from 56.3% in 2024 to 50% in the first half of 2025, even as revenue doubled year-over-year.

This isn't execution failure; it's the fingerprint of China's AI price war. Token pricing has collapsed as competitors from Alibaba to Baidu subsidize inference to capture developer mindshare. Zhipu's strategy of undercutting on cloud inference to build distribution makes strategic sense—if margins can eventually recover through premium agent products or enterprise workflow bundles. But the prospectus offers no evidence that inflection point is near.

The business that actually generates cash today is on-premise deployment—essentially bespoke enterprise projects that scale like consulting, not software. On-prem revenue reached RMB 161.78 million in the first half of 2025, but this model trades margin for revenue visibility. It's the opposite of the "API flywheel" narrative that commands software-like multiples.

What This Debut Actually Predicts

Zhipu's public offering won't be the last time markets grapple with this tension. MiniMax debuted Thursday with similar dynamics; U.S. giants OpenAI and Anthropic are expected to test public markets in 2026 despite losses. What Zhipu's pricing reveals is that investors will pay for distribution and strategic positioning even absent profitability—but only in companies credibly positioned to eventually control workflow outcomes, not just serve tokens.

The real test arrives in late 2026, when investors assess whether GLM-4.7's aggressive developer adoption translated into paid enterprise agents and workflow automation. Until then, Zhipu's share price will trade on narrative, not numbers—a fitting metaphor for an industry that has generated transformative technology but not yet proven it can generate sustainable profits at scale.

NOT INVESTMENT ADVICE