A sprint that began in university labs and private research centers has reached the stock market.
On back-to-back days in Hong Kong, two of China’s most talked-about artificial intelligence model builders stepped into the public arena — and the receptions could not have been more different.
Zhipu, formally listed as Knowledge Atlas Technology JSC, completed a roughly HK$4.35 billion ($558 million) IPO and opened above its offer price, nudging the share price up as much as 15% on debut. The Beijing-based large-language-model developer — founded in 2019 by researchers from a top university and singled out by U.S. rivals as a notable competitor — told investors it would spend roughly 70% of the proceeds on R&D. Its prospectus also shows revenue of 312.4 million yuan in 2024, signaling early commercial traction even as the company faces heavy regulatory and technology headwinds.
A day later MiniMax, a Shanghai outfit launched by a former SenseTime executive in 2022, raised about HK$4.8 billion ($620 million) and rocketed in trading — jumping roughly 50% from its HK$165 offer price to trade near HK$248 in early dealings. MiniMax bills its models as multi‑modal, able to process text, audio, images, video and music, and plans to funnel most of its IPO haul into further model development and compute.
What investors are buying — and what they’re nervous about
Why the divergent market reactions? Part of it is narrative and timing. MiniMax arrived with a simpler growth story: a hot founder background, clear capital needs for scaling compute, and investor appetite in Hong Kong for headline-grabbing AI plays. Zhipu’s listing was historic in a different way — it marked the first major Chinese LLM developer to go public — but that milestone came with baggage: close ties to Beijing, a January 2025 placement on the U.S. Commerce Department’s Entity List for alleged military links, and limits on accessing advanced semiconductors and overseas expertise.
That combination helps explain the cautious, sometimes mixed coverage from financial desks. On paper, both companies are chasing the same prize — general-purpose AI models that can be licensed, embedded into cloud services, and sold into enterprise workflows — but the path to profitable scale is capital intensive. Training state-of-the-art models consumes vast compute and chips that are increasingly restricted by export controls, a structural squeeze noted repeatedly by analysts.
China’s broader AI ecosystem is reacting in kind. The IPOs follow a wave of listings by AI chipmakers and other infrastructure players, and they speak to a broader strategy: build homegrown stacks in model architectures, chips and data pipelines to reduce dependence on foreign technology. That ambition is visible beyond the listings themselves — from product experiments to integrations that aim to place AI into everyday tools, echoing moves by global players to embed assistants into software and services. For example, recent efforts to add booking and agentic features into search tools underline how quickly AI features can reshape product expectations (Google’s AI Mode). Image and creative-model competition is also accelerating — Microsoft’s in-house image model rollout is part of that broader push to diversify capabilities and avoid single-vendor chokepoints (Microsoft MAI-Image-1).
The hard math of LLM economics
Investors and executives alike keep circling back to a couple of tensions: how much compute is needed to stay competitive, and how easily those costs can be monetized. Building and fine-tuning large models demands GPUs, fast interconnects and years of engineering — and the revenue models are still evolving. Licenses, cloud APIs, enterprise customization and downstream applications (from search to customer service to creative tools) are all in play, but they remain experimental at scale.
Regulatory risk is another line item. Zhipu’s spot on the Entity List and other geopolitical frictions mean access to certain chips, tools, and partnerships is uncertain. That doesn’t stop demand—Chinese companies and governments are pouring money into homegrown solutions—but it does complicate timelines and raise execution risk for public investors.
Why the listings matter beyond valuations
These IPOs are less about one company’s market cap and more about signalling. They show that capital markets — at least in Hong Kong — are willing to price speculative growth premised on AI leadership. They also make visible a Chinese industrial play to seed an ecosystem of models, chips and application layers that can operate in a more bifurcated global tech landscape.
For traders and technologists the short-term story will be share moves and quarterly metrics. For strategists and policymakers, the launches are a data point in a broader strategic competition over AI infrastructure, talent and supply chains.
If anything is certain: more startups will try the public route, and more investors will test whether high valuations for model makers can translate into sustainable profits. That test will rest on technical execution, access to compute, and the stubbornly hard work of turning clever models into dependable, paying customers — the real litmus test for any AI tiger trying to pounce.