Everything About AI
NextFin News - MiniMax Group's shares have fallen back below the price at which they were sold to the public, wiping out every gain since the company's January trading debut as investors conclude that China's artificial-intelligence model war is destroying the economics it was supposed to create. The stock dropped as much as 7.3% on Thursday in Hong Kong to HK$159.3, under the HK$165 offer price set at the top of its initial public offering range. Nine months ago, the same shares closed their first day at HK$345, a 109% pop that valued the Shanghai-based lab at about US$13.7 billion.
The reversal is the cleanest signal yet that the market is no longer willing to pay for top-line growth in Chinese AI when the path to profit keeps getting longer. MiniMax is not alone: the probe that Beijing opened into two rivals, the price cuts flooding the model layer, and the lock-up expiries that followed the summer's fundraising have all hit the sector at once. But MiniMax is the first pure-play AI lab to go from IPO darling to below-issue-price, and that makes it the test case for every Chinese model startup still waiting to list.
The Situation: Growth Is Up, the Stock Is Down, and the Market Has Chosen
The numbers on their own look defensible. For the six months ended June 30, 2026, MiniMax reported revenue of US$116.6 million, up 283.1% from US$30.4 million a year earlier — already more than its full-year 2025 revenue of US$79.0 million. Enterprise and open-platform revenue grew 703.1% to US$73.9 million and now makes up 63.4% of the total, up from 30.3%. Gross profit climbed 464.8% to US$20.8 million, lifting the gross margin to 17.9% from 12.1%. Co-founder and CEO Yan Junjie told investors that token consumption in July stood at 20 times its January level.
None of it stopped the shares from sliding under the IPO price. The gap between those two facts is the story. At the IPO, priced at the top of the HK$151–HK$165 range, the company raised HK$4.8 billion (about US$618 million) and sold itself on a valuation of roughly US$6.5 billion; the market then bid that to US$13.7 billion on day one. Today it is worth less than it was offered. In between, management went back to the well: on July 10 the company placed 35.6 million new shares and issued zero-coupon convertible bonds due 2027, raising a combined net HK$15.96 billion, or about US$2.0 billion. The proceeds were earmarked 80% for AI infrastructure. The capital raise arrived just as the lock-up period expired, and the two events together — dilution plus supply overhang — set up the unwind that is now playing out.
The timing matters because the sector's mood has shifted between the IPO and now. In January, the trade was simple: buy the fastest-growing model lab and let the multiple do the work. By October, the trade has inverted. The same investors who cheered 283% revenue growth are now asking what that revenue is worth when the price of a million tokens keeps falling, when open-weight models handle a large share of the world's requests but capture almost none of the money spent on them, and when the company that prints the best benchmark scores is not necessarily the one that keeps the most margin.
The October leg of the selloff has a specific trigger. On October 7, Anthropic cut the price of its small model, Claude Haiku 5.5, to US$0.10 per million input tokens and US$0.50 per million output tokens — one-tenth of the previous price and roughly level with OpenAI's GPT-6 Luna, released the same week. For eighteen months, low price had been the main reason companies outside China adopted Chinese models. When the U.S. frontier labs match that price, the Chinese labs' last structural advantage evaporates overnight. The market did not need a bad earnings print to reassess MiniMax; it needed proof that the price war had reached the point where even the cheapest supplier no longer has an edge. Anthropic supplied that proof.
MiniMax's own CEO framed the tension without resolving it. In the August earnings release, Yan said:
"Intelligence can scale almost without limit; energy and compute cannot. The long-term competition in AI is not just about building more powerful models, but about delivering higher levels of intelligence to more people at lower cost."
That is a compelling mission statement. It is also, read coldly, a description of a commodity market: when the winning move is to deliver more for less, the surplus accrues to the customer, not the supplier.
The Price War Is the Product, Not a Phase
The first question is why 283% revenue growth does not support a US$13.7 billion valuation. The answer is that growth is only valuable if it converts into profit at some point, and the conversion rate in the Chinese model layer is structurally low. MiniMax's gross margin of 17.9% compares with roughly 50% gross margins at leading U.S. model labs, according to IPO-economics analyses published around the listing. The gap is not a temporary efficiency problem; it is the arithmetic of a price war.
Consider what the company is selling into. On OpenRouter, a marketplace where developers send requests to hundreds of models and which publishes its usage figures, the combined share of U.S. models fell from about 70% of tokens in June 2025 to about 30% in June 2026. Chinese labs have won the volume war. Yet on Vercel's AI Gateway, open-weight models — the category Chinese labs lead — handled 29% of tokens in June 2026 but received less than 4% of the spending routed through the platform. MiniMax is a meaningful player in that volume, and it is earning a fraction of the money that volume represents. The company's own accounts show the same gap: it kept only 17.9% of its revenue as gross profit in the first half of 2026.
The second-order effect is what most investors missed at the IPO. A falling token price does two things at once: it expands the addressable market (more developers can afford to build), which is why revenue grew 283%, and it compresses the value of every token already sold, which is why the stock fell. Revenue measures volume; valuation measures the present value of future margin. In a market where the unit price is a race to the bottom, volume growth and multiple compression can run in opposite directions for a long time. MiniMax is living inside that divergence.
There is also a distribution problem hiding inside the revenue mix. Open-platform and enterprise services grew 703% to US$73.9 million, but that segment is the most exposed to API price competition. AI-native products — the consumer subscriptions and in-app purchases for Hailuo AI and Talkie — grew a slower 100.9% to US$42.6 million. The faster-growing bucket is the one with the weakest pricing power; the stickier consumer bucket is growing at half the pace. Investors are not wrong to notice which engine is pulling harder.
The central judgment here is that this is a structural shift, not a cyclical dip. Three pieces of evidence point to a regime change rather than a mean-reverting cycle.
First, the pricing mechanism has changed permanently. Before early 2025, frontier models were mostly closed and owner-priced. Since DeepSeek began releasing widely used open-weight models, most Chinese labs have followed: a model that can be hosted by many providers faces price competition from every host, and the price converges toward the marginal cost of inference. That is a change in market structure, not in sentiment. Market structure does not mean-revert on its own.
Second, the value capture is structurally broken. The Vercel data — 29% of tokens on less than 4% of spend — is not a promotional discount; it is the equilibrium of an open-weight market. Chinese labs process a large share of the world's AI requests and earn a small share of the money spent on AI, because their prices are so low. This explains why a company can grow revenue 283% and still see its equity value fall. Volume is not value when the price per unit is set by the cheapest competitor.
Third, the capital intensity is escalating while the unit price falls. Alibaba, at its Apsara conference, unveiled China's most powerful AI chips and a plan for a 20-gigawatt global data-center network by 2032. When the well-funded conglomerates are committing to that scale of capacity, the pure-play lab must either spend at a comparable pace — which is why MiniMax raised US$2 billion and still runs an adjusted net loss of US$293 million a half — or cede capability. Either way, the margin path narrows.
This is why the cyclical-versus-structural call matters so much. If the selloff were cyclical — a sentiment washout that reverts when sentiment improves — then buying the dip makes sense and the IPO price becomes a support level. If it is structural — a permanent repricing of what a model-layer business is worth — then HK$165 is not a floor; it is a reminder of the multiple the market used to pay for a business model that no longer exists. The evidence above points to structural.
The Counter-Thesis: This Is a Washout, Not a Wound
The strongest case against that read is straightforward: nothing fundamental has broken at MiniMax. Revenue is up 283%, gross margin is improving, token consumption is up 20-fold, and the company sits on US$1.32 billion of cash as of June 30 — enough to fund operations well beyond the next milestone. The July lock-up expiry was a known, one-time supply event; the capital raise bought runway; and the sector-wide selling in October reflected macro pressure and a regulatory overhang that landed on DeepSeek and Moonshot, not MiniMax itself. On this view, the stock is oversold, the overhang is digestible, and the price will mean-revert once the technicals clear.
That argument is correct about the company and wrong about the valuation. A balance sheet can survive a price war; a multiple cannot. Even if the stock bounces on short-covering after the lock-up supply is absorbed, the question is not whether HK$159 is cheap relative to HK$345 — it is whether the market will ever again pay US$13.7 billion for a loss-making model lab growing revenue at triple digits with a 17.9% gross margin. The counter-thesis answers that with faith in the growth curve; the structural read answers it with the unit economics. The two are not reconcilable, which is exactly why the stock sits where it does: the bulls are pricing a recovery in sentiment, and the bears are pricing a permanent compression in the business model.
There is also a peer signal that cuts against the simple "sector is broken" story. Zhipu AI, listed as Z.ai on January 8, 2026 — one day before MiniMax, at an offer price of HK$116.20 — has posted a year-to-date return of more than 300% as of October 21, driven by its GLM-5.2 release and a July follow-on placement. Brokerage coverage has begun to diverge: Bernstein started coverage of China's AI labs in August and favored Z.ai over MiniMax, while JPMorgan cut MiniMax's price target for a second time as Zhipu pulled ahead. The market is not rejecting Chinese AI; it is picking winners within it. MiniMax's problem is not that it is a model lab. It is that it is not the model lab the market currently prefers.
The regulatory backdrop adds a second layer of dispersion. On September 10, Anthropic published a 154-page threat-intelligence report accusing seven China-based labs — Alibaba, Moonshot AI, DeepSeek, Zhipu, MiniMax, Xiaomi, and SenseTime — of routing customer requests through its Claude models to harvest outputs for training, a practice Anthropic calls illicit distillation, covering roughly 190 million exchanges between December 2025 and August 2026. China's Cyberspace Administration summoned representatives from all seven companies and narrowed its formal probe to DeepSeek and Moonshot, with no penalty decided and neither company commenting publicly. MiniMax was named but not charged. The asymmetry matters: in a market searching for reasons to discriminate between labs, "named but not probed" is not the same as "cleared," and it keeps the discount on the whole peer group.
Conclusion: Who Benefits, Who Is Exposed, and What Would Prove This Wrong
Cash in the mechanism and the implications split cleanly by time horizon.
In the short term, the move is liquidity-driven: lock-up supply, a crowded long-only trade, and October sector selling. That leg can reverse quickly — a 7.3% intraday drop is tradable, and oversold bounces are common after washouts. In the medium term, the driver is fundamentals: whether MiniMax can grow revenue fast enough to outrun margin compression, and whether the consumer product business can become the larger engine. In the long term, the driver is structural: whether the open-weight model layer is a moat or a commodity, and whether the market will ever pay a growth multiple for a business whose unit price is set by its cheapest competitor.
The beneficiaries are the buyers of compute and intelligence — the developers and enterprises paying one-tenth to one-fortieth of Western frontier rates — and the well-capitalized platforms that can subsidize models to win distribution. The exposed are the pure-play labs that must fundraise to keep pace: Moonshot AI, which has filed confidentially for a Hong Kong listing, DeepSeek, and StepFun, all of which are now pricing their IPOs into a market that has just learned to distrust triple-digit revenue growth without a margin path.
Three scenarios frame the next six months. The base case: MiniMax trades in a wide range around the IPO price as lock-up supply is absorbed and the market waits for the next earnings print to test whether gross margin can keep expanding. The upside case: the company shows two consecutive quarters of gross-margin expansion toward 25% while revenue growth stays above 50% year over year — proof that scale is finally bending the cost curve — and the stock rerates back toward the July placement price of HK$268. The downside case: token prices fall another leg lower as a frontier lab resets the floor again, revenue growth decelerates below 40%, and the market retests the post-fundraising lows.
The falsifying signal is specific: if MiniMax reports gross margin at or above 30% for two consecutive quarters while maintaining revenue growth of 50% or more year over year, the commoditization thesis is wrong — it would mean the lab has found a margin path through the price war, and the structural call should be abandoned. A weaker version of the same test: if the shares reclaim and hold HK$268 — the price at which insiders and institutions were willing to buy fresh equity in July — for 20 consecutive trading days, the market is telling you the dip was cyclical after all.
MiniMax's IPO was sold as the moment China's AI champions joined the public markets. Nine months later, the public markets have delivered their answer: they will fund the infrastructure, they will use the models, but they are no longer convinced that the model layer itself is where the profit will sit. The company that taught investors to pay for intelligence may end up teaching them the harder lesson — that cheap intelligence is a customer's windfall, not a supplier's moat.
Explore more exclusive insights at nextfin.ai.
