On August 7, 2025, MINIMAX-W (00100.HK) closed up nearly 25 percent in Hong Kong. Zhipu (02513.HK) followed with a 17-plus percent gain. These are the two biggest moves in Hong Kong's newly formed AI large-model listing cohort. The numbers look like a sector breakout. But the only source quoted is Bitget — a cryptocurrency exchange terminal, not an HKEX filing, not a company announcement, not a major financial wire.
There was no earnings release. No product launch. No disclosed contract win. No press conference. The chart simply went vertical. In my years of reading Chinese tech price action, that pattern is not the mark of a fundamental repricing. It is the signature of a capital event — an index inclusion rumor, a southbound flow burst, or a thinly arbitraged float being lifted by a few large orders.
Liquidity didn't wait for the story. It wrote one.
Price precedes narrative in this market. The narrative arrives afterwards, fully packaged, with the eager help of momentum-chasing retail.
The question is not whether the move was real. The transactions settled. The question is what the move means — and what it will look like when the first post-IPO earnings report lands. To answer that, you have to understand the structure underneath the chart.
Context: The Two Tigers That Got Listed First
China's so-called "Six Little Tigers" — the six most prominent private large-model labs — have been burning capital since 2023. The question was always who would exit first. In 2025, the answer came from Hong Kong.
MINIMAX is the consumer-oriented outlier. Its flagship products, Hailuo AI (content and video generation) and Talkie (AI companionship aimed at overseas markets), target individual users. Talkie is the closest thing China has produced to Character.AI. Hailuo competes in the same creative-generation space as Runway and Luma. The company's technical claim to fame: it claims one of the few trillion-parameter MoE (Mixture-of-Experts) models trained domestically, reportedly on Huawei Ascend 910B clusters. That is compute sovereignty as a corporate strategy — a bet that runs far deeper than any single product release.
Zhipu AI is the Tsinghua-lineage institutional player. Its GLM series — from GLM-130B to GLM-4 — maintained a full-stack research pipeline: foundation model, alignment, agents, multimodal. Its revenue base is B-end: government, finance, education, healthcare. Private deployment, API access, and system integration contracts. Slower top line, arguably stickier revenue. The company has been called the closest thing to a "Chinese Anthropic," and the comparison is not entirely off-base: heavy on safety alignment, institutional credibility, and enterprise routing.
Both companies listed under Hong Kong's Chapter 18C, the special chapter allowing pre-revenue technology companies to go public. That single fact matters more than any stock chart: China's frontier AI labs have entered the public market pricing phase. They are no longer private-valuation guesswork. They now have a ticker, a float, and most importantly, a daily mark against which every future round in the Chinese AI ecosystem will be calibrated.
Core: The Divergence Between the Charts and the Businesses
Part One: Two Business Models, One Price Action
The most important structural detail about this dual surge is that MINIMAX and Zhipu are not the same company. They are not even the same species of company.
MINIMAX's trajectory is consumer-led and global-first. Hailuo AI monetizes through subscriptions and usage-based credits. Talkie monetizes through virtual goods and premium features. The economics of C-end AI entertainment are high-variance: user acquisition costs are brutal, retention is uncertain, but when a product hits, the revenue elasticity is enormous. A 25 percent move in this stock means the market is pricing optionality. It is paying for the possibility that Talkie becomes the next breakout consumer AI app — or that Hailuo's video generation reaches a quality threshold that makes Midjourney's engine look legacy.
Zhipu's trajectory is contract-led and domestic-first. Government and enterprise clients pay for API access, private-model deployment, and long-term service. The revenue quality is higher — recurring contracts, signed through procurement processes, with exit barriers once deployed. But the growth curve is slower and heavily dependent on China's public-sector digitization cycle. A 17 percent move in this stock means the market is pricing security: a reliable pipeline of B-end demand and a brand with Tsinghua's institutional credibility attached.
The two companies even diverge on compute. MINIMAX went deep on domestic Ascend hardware for training. Zhipu runs a hybrid of Nvidia GPUs and domestic accelerators. Full-stack sovereignty versus operational flexibility. Global consumer data flows versus China-regulated enterprise data residency.
These are completely different risk profiles. One is a consumer-tech lottery ticket. The other is an infrastructure-tech annuity.
And yet the market moved them in lockstep. Why?
Because the Hong Kong bid right now is not pricing business models. It is pricing a basket. The basket is called "China AI core assets." Within that basket, differentiation disappears. The algorithm priced the ape before the crowd did — the ape being the indexing momentum that hoovers up every Chinese AI name regardless of revenue mix. The crowd, as usual, arrives late.
Value is a consensus, not a contract. The consensus on display here bundles two fundamentally different companies into one price signal.
Part Two: The Thin Book Problem
Here is where my market microstructure experience kicks in.
I spent 2020 stress-testing Uniswap V2 pairs, running 10,000 simulations to map price impact thresholds. The core lesson: in a thin order book, price is not information — it is noise amplified by order flow. A 25 percent move in a newly listed small-cap stock requires far less capital than most retail observers assume.
Both MINIMAX and Zhipu completed their IPOs on a compressed timeline. Chapter 18C listings typically feature small public floats relative to total shares outstanding. Lock-up agreements keep the majority of the register frozen. What remains tradeable — the effective float — can be shockingly small. When the free float is 5 to 10 percent of issuance, a single institutional buyer can move the price by double digits in a single session. And once a daily gain is printed, momentum algorithms and trend-following flows pile in. The chart becomes its own narrative.
This is not to say the moves are fake. They are real transactions at real prices. But a large price change in a small float is a statement about liquidity conditions, not necessarily about business fundamentals.
Liquidity didn't validate the move. It manufactured the move.
My Celsius analysis in 2022 taught me the same lesson in reverse: when a stock or a protocol trades at a premium despite a 15 percent reserve discrepancy, the market is ignoring fundamentals to trade the story. The story can continue until the trading is forced to stop. The key skill is knowing which phase you are in. Looking at the Hong Kong AI names, we are in the story phase. The fundamental verification phase will arrive with the first set of post-IPO financials.
Part Three: The Bitget Signal
The source of this market snapshot deserves more attention than it is getting.
Bitget is a cryptocurrency exchange. Its terminal is the default interface for crypto-native traders. The fact that this price action was captured and broadcast through a crypto terminal tells you which audience is now paying attention to Hong Kong AI listings.
There is a narrative convergence happening. Crypto capital, after the 2022-2023 bear market, learned to look for listed AI plays as a substitute for the equity exposure it cannot always access directly. Chinese AI stocks, listed on HKEX with sufficient trading hours, become the accessible proxy. The capital has a Web3-native temperament: momentum-forward, event-driven, with a short memory for fundamentals. When this capital class enters a small float, it is not buying for 18 months. It is buying for 18 hours, or 18 days.
This matters for interpretation. A 25 percent move powered by crypto-native flows and a 25 percent move powered by institutional rebalancing have entirely different sustainability profiles. The former creates volatility. The latter creates a base. Watching the volume and order composition in the next 10 trading sessions will tell you which one you are watching. The price chart alone cannot.
There is also a policy dimension. If the buying came through the Stock Connect (southbound), it reflects mainland investors reading policy signals — Beijing's continued support for AI industrialization and capital-market revitalization. If it came from global funds, it reflects international confidence in Chinese AI despite sanctions and export controls. These two flows behave differently. Southbound capital is policy-committed but can rotate. International capital is discretionary and can exit in a single session. The exchange's disclosed north-south flow data is not decorative; it is the key diagnostic for the next month.
Part Four: The Compute Strategy Nobody Can See
Here is what the chart does not tell you.
MINIMAX's decision to train its trillion-parameter MoE on Ascend 910B clusters was a national-compute posture before it was a business decision. It bought sovereignty at the price of friction: domestic toolchains, less mature software stacks, and cluster stability issues that Nvidia users simply do not face. If that bet pays off, MINIMAX has a cost and security advantage in the Chinese market that competitors using restricted imports cannot replicate. If it fails — if the chips underperform or supply chains tighten — the model iteration cycle slows precisely when the market is demanding speed. The 25 percent single-day gain will look like a rounding error against a 12-month compute bottleneck.
Zhipu's architecture straddles both worlds: Nvidia GPUs and domestic accelerators, training on a hybrid stack, with inference optimized for cost. This makes Zhipu more operationally flexible but less strategically distinctive. The company's moat is the GLM ecosystem's grip on B-end contracts, not compute singularity. The same hybrid flexibility, however, could prove to be an entirely adequate posture if export controls loosen or if inference economics become the dominant competitive variable.
The public market cannot price this nuance in a single session. It only sees the ticker. But long-term holders should understand that compute choice is a durational bet embedded in both companies. The stock moved 25 percent in a day. The compute strategy will take 18 to 36 months to prove or disprove.
Structure is not a cage; it is a launchpad. But launchpads only matter if the rocket has fuel. The fuel schedule is the next two earnings cycles.
Part Five: The Primary Market Echo
There is a second-order consequence of these two listings that most coverage misses.
The Hong Kong prices now serve as the valuation anchor for the four unlisted "Six Little Tigers": Moonshot AI, Baichuan, 01.AI, and StepFun. When a listed peer trades at a market cap implying 20 to 40 times revenue, the unlisted peers' boards and investors will demand comparable valuations in their next private rounds. This sounds bullish. It is, in fact, a double-edged sword.
Higher private valuations mean bigger checks required, longer diligence cycles, and a higher bar for the next round's return. If the public market corrects in the next six months — say, the first earnings reports disappoint — the unlisted companies will face a "valuation inversion." Their last private round price will exceed what the public market is willing to pay for their revenue. That gap freezes follow-on fundraising and forces down-rounds across the entire cohort.
The echo effect of a 25 percent surge is not pure optimism. It sets an expectations trap for the entire Chinese AI private ecosystem.
Part Six: The Compliance Premium Nobody Quantifies
There is another invisible variable: how these companies behave under Hong Kong disclosure rules.
Listing on HKEX subjects both firms to stricter corporate governance, ESG disclosure, and — for AI companies operating in China — ongoing compliance with the country's generative AI regulations and algorithm-filing requirements. Zhipu, with its Tsinghua lineage, has invested heavily in safety alignment and value alignment. That is a genuine differentiator in enterprise procurement, but it is nearly impossible for the market to quantize in a single quarter. It shows up in contract wins, not in press releases.
MINIMAX's Talkie business, by contrast, carries a different tail risk. AI companionship products face regulatory scrutiny over minor protection and content safety, especially in overseas markets. A single content-safety incident abroad could trigger a compliance review that hits both the user base and the stock price. The market is currently paying zero attention to this risk. It is too busy watching the candle.
Contrarian: What You Are Not Being Told
Let me state it plainly: the surge tells you almost nothing about the quality of these businesses. It tells you about the capital available to chase the story.
Three things are being ignored in the enthusiastic coverage.
First, no fundamental data has been disclosed yet. Neither company has published a full post-IPO quarterly financial report. The market priced these moves without verified revenue, margin, or cash-burn numbers. That is not investing; that is guessing with leverage. I built a sentiment index during the Bitcoin ETF race in 2024 to separate signal from narrative noise. The signal here is minimal. The noise is enormous.
Second, the direction of the bid matters more than the size. As noted, the two dominant hypotheses — southbound policy-driven flows versus international discretionary capital — imply completely different next moves. The price chart cannot distinguish between them. Only the flow data can.
Third, wash-trading is not a mechanism invented in Western NFT markets. My BAYC floor-price analysis in 2021 identified whale wash activity as the mechanism behind a 30 percent floor collapse within hours. Hong Kong's regulators are stricter than the NFT market ever was, but thin floats remain vulnerable to engineered volume, matched orders, and cross-book wash patterns. The price move is real. The book behind it needs verification before any conclusions are drawn.
The valuation story being told today — that Chinese AI has reached public-market maturation — is partly true. But the timing of this surge, the source of the flow, and the absence of financial disclosure all suggest something more fragile: a liquidity event wearing a fundamentals costume.
Takeaway
Watch the volume. Watch the southbound ledger. Watch the next five to ten sessions. A real breakout holds at the highs; a liquidity illusion fades on declining volume.
And most importantly: wait for the first financial report. The market just priced the story. The contract comes when the numbers print. Until then, treat the candle as what it is — an algorithmically amplified signal in a thin book, not a judgment on China's AI future.
The structure is the signal. The candle is just noise.
