China's AI Surge Is Closing the U.S. Gap-But 2.7T Models Meet a Hong Kong Sell Wall

Moonshot's 2.8T Kimi K3 Narrowed the Perceived Gap

Moonshot's new Kimi K3 is a 2.8 trillion-parameter model that the company says approaches the performance of leading U.S. frontier systems. Reuters also said independent benchmarks pointed to strong capabilities, reinforcing the idea that China's gap at the top end is narrowing. Moonshot added that Kimi K3 performed competitively with Fable 5 and substantially outperformed other leading U.S. models on GPU kernel optimization.

Add MiniMax's reported 2.7-trillion-parameter language model, possibly due this quarter, and the pattern is clear: China's leading labs are making large scaling jumps rather than small steps.

Why this matters now

This is not just a benchmark story. The quicker question is whether better model output, lower-cost options, and open-weight distribution can start to affect commercial adoption.

MiniMax said H3 would extend the open-weight approach into video generation, a sign that Chinese developers are trying to widen access and customisation. If that strategy keeps working, investors may have to reassess how durable U.S. pricing power really is.

What the evidence still does not prove

Reported benchmark wins do not automatically mean durable frontier parity. More importantly, MiniMax's 2.7T model is still only reported, not confirmed. That keeps this a race on release cadence, liquidity, and adoption rather than a settled verdict on long-term leadership.

Hong Kong Capital Markets Are Becoming Part of the Race

Model launches are no longer feeding only leaderboards. They are also shaping financing opportunities.

Zhipu AI's fundraising shows the capital signal

Zhipu AI is trying to raise around $4 billion in Hong Kong, with proceeds earmarked for R&D, talent training, and compute deployment. That would create a direct link between market interest and model capacity.

That matters because labs are moving toward the trillion-parameter threshold, where analysts say systems may get better at autonomous multi-step reasoning. Once that happens, the edge may depend less on a single benchmark run and more on who can keep reinvesting.

MiniMax shows how quickly model news can turn into cash

MiniMax's Hong Kong IPO settled at the top of the pricing range after multiple over-subscriptions, raising HK$4.19 billion, or about $538 million. That does not prove durable economics, but it does show that Hong Kong investors are willing to fund Chinese AI ambition.

Watch three flows from here: - Whether Zhipu can close its accelerated bookbuild - Whether MiniMax can turn listing momentum into its expected third-quarter model release - Whether more labs keep turning to Hong Kong for capital

If those channels stay open, catch-up can accelerate faster than current fundamentals alone would suggest.

Policy Controls and Hong Kong Supply Overhangs Can Still Dampen the Story

Beijing may still turn a strong model pipeline into a commercially constrained one. Authorities have held talks with firms over restricting overseas access to China's most advanced AI models, discussed new measures to restrict who can fund domestic AI startups, and considered making leaks or theft of proprietary AI a national security offence. Add reports that regulators are consulting on tightening export controls on AI and semiconductor technologies, and the risk is obvious: benchmark strength matters less if commercialization cannot cross borders cleanly.

That is why this remains a trade as much as a thesis. The upside case works only if better models can convert into foreign users, paid API demand, or open-weight adoption outside China. The risk is not that the progress is fake. It is that policy can narrow the addressable market at exactly the moment investors are trying to price scale.

Hong Kong's lock-up wave adds another layer of risk

Hong Kong also has its own technical overhang problem. Knowledge Atlas is freeing 25.6 million shares, nearly 6% of its outstanding shares, while MiniMax has 45% of its outstanding shares set to unlock. Reuters cited Goldman Sachs estimates of $274 billion of locked-up shares hitting the market over the next 12 months and noted that prices have historically dipped after releases. Morgan Stanley also warned that these events can create liquidity headwinds even when fundamentals remain intact.

If policy stays manageable and float pressure is absorbed, the sector can still rerate. If either breaks, supply and regulation can overwhelm model progress quickly.