Welcome to the Halftime Show.
I ran into an engineer at a coffee shop in Nanshan this week who told me he turned down a job at MiniMax 18 months ago. His girlfriend says he choked. This week made it worse.
She may be leaving him.
Let’s lock in.
PLAY OF THE WEEK
The application layer goes public before the frontier layer does.
Zhipu AI and MiniMax (rivals) debuted on the Hong Kong Stock Exchange on January 8 and 9.
Combined raise: $1.1 billion. Both oversubscribed. OpenAI and Anthropic are still laying groundwork for their IPOs. China's AI labs didn't wait. But why?
SCOUT REPORT
→ Zhipu AI raised $560M at IPO. MiniMax raised $539M. Both oversubscribed. Zhipu, founded by Tsinghua professors and backed by Tencent, Meituan, and Ant Group, claims performance matching OpenAI and Anthropic's flagships. MiniMax operates Hailuo - the AI video generator that competes directly with OpenAI's Sora. (Rest of World)
→ Hailuo shipped while Sora was still in preview. MiniMax's video generator launched in the same window as Sora's announcement, found paying users, and iterated in public. By the time Sora was available, Hailuo had a head start measured in millions of generations. (Rest of World)
→ MiniMax generates only 26% of its revenue from China: 20% comes from the US, the rest from Singapore and other markets. MAUs grew from 3.1 million in 2023 to 27.6 million by September 2025. Not just a domestic Chinese champion. It's a global LLM provider that happens to be based in Shanghai. (eWeek)
→ DeepSeek published a technical paper on New Year's Eve. On December 31st. During the global holiday peak. A framework for training AI systems at scale, signalling the next architectural direction. Light weekend reading 🙂 (Link here)
→ Moonshot AI closed a $500 million Series C on December 31st: led by IDG, with Alibaba and Tencent oversubscribing. Kimi's MAUs hit 27.6 million by September 2025. The funding round closed on NYE. Dramatic. Same day as DeepSeek's paper. The Chinese AI ecosystem does not take holidays. (KuCoin)
FILM ROOM
What does that sequencing actually tell you?
OpenAI: $300B+ valuation. Still private. Anthropic: $60B+. Still private. Both laying groundwork for eventual IPOs and neither has pulled the trigger.
Zhipu and MiniMax just did.
The sequencing reveals something important about how the two ecosystems are structured.
US frontier AI labs:
raised enormous private rounds at enormous valuations on the promise of AGI and long term compute dominance
requires patient capital, long horizons, and a public market willing to price a moonshot
none of those conditions are certain right now
Chinese AI labs took a different path. Neither Zhipu nor MiniMax is competing at the frontier:
Zhipu is an enterprise AI company: government contracts, financial institutions, production deployments
MiniMax is a consumer and creator AI company: video generation, voice, music, global users
both are loss making. Both are generating real revenue. Both are pricing the application layer, not the model layer
That distinction matters enormously for public market investors. Buying OpenAI is a bet on whether one company can capture a disproportionate share of a hypothetical future. Buying MiniMax is a bet on whether AI video generation becomes a standard creative tool. One of those is a much easier thesis to underwrite.
MiniMax halting institutional orders a day early is evidence from the streets. Institutional capital in Hong Kong looked at a loss making AI application company and said: we want more of this than you're offering.
STAT OF THE WEEK
$1.1 billion
Combined raised by Zhipu AI and MiniMax’s IPOs. OpenAI's IPO timeline: still TBD. First mover advantage in public markets, as in product markets, compounds.
See you next week,
Jen, live from Shenzhen