NextFin News - Chinese AI lab MiniMax laid out its business strategy at the Goldman Sachs Asia Leaders Conference on September 1, 2026, and the message was blunt: the race in artificial intelligence is no longer won by whoever builds the most powerful model, but by whoever delivers the most intelligence at the lowest cost. The strategy shift is real, the revenue growth is explosive, and yet the market is refusing to cheer. That gap - between a 283% revenue jump and a muted share price - is the story, and it says as much about the commoditization of model inference as it does about MiniMax itself.
The Numbers Behind the Pivot
MiniMax Group (HKEX: 0100) reported first-half 2026 revenue of $116.6 million, up 283.1% from $30.4 million in the same period a year earlier. That six-month figure already exceeds the company's full-year 2025 revenue of $79.0 million. But the headline growth rate masks the more important change: the composition of revenue has flipped.
Revenue from the Open Platform and other AI-based enterprise services jumped 703.1% year over year, to $73.9 million, and now represents 63.4% of total revenue - up from 30.3% in the first half of 2025. AI-native products, which include the Talkie AI-companion app and Hailuo AI video generator, grew 100.9% to $42.6 million. A year ago, consumer products accounted for nearly 70% of revenue. Today, enterprise API and platform services are the larger half. In six months, MiniMax stopped being structurally a consumer-AI company.
The velocity of that shift shows up in annualized revenue. The company said its ARR jumped to $800 million in August 2026, from $150 million in February - more than fivefold growth in six months. The trajectory management outlined runs from $100 million in ARR at the end of December 2025, to $150 million in February 2026, to a stated target of $1 billion by the end of 2026. That is not a straight-line extrapolation; it is an acceleration curve, and it is the single most important number in the bull case.
Gross profit improved 464.8% to $20.8 million, and gross margin expanded from 12.1% to 17.9%, which management attributed to improving infrastructure efficiency. Selling and distribution expenses fell 17.9% to $27.0 million as the company moved to an organic user-growth strategy. But the adjusted net loss widened to $293.0 million from $138.7 million, and research and development expenses rose 138.8% year over year. Cash stood at $1,322.8 million as of June 30, 2026, up from $1,050.3 million at the end of 2025 - a war chest that buys time, but not profitability.
CEO Yan Junjie framed the constraint in a single line in the company's earnings release:
Intelligence can scale almost without limit; energy and compute cannot. By July 2026, token consumption on MiniMax had grown to 20 times its January level. That reinforces a belief we've held since day one: 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.
The implication is uncomfortable for any model lab: demand is exploding, but the scarce inputs - chips and power - are where pricing power concentrates. MiniMax is caught in the middle of that squeeze.
Why the Stock Did Not Celebrate a 283% Revenue Jump
Here is the tension. MiniMax listed in Hong Kong on January 9, 2026, at an offer price of HK$165 and closed that first day at HK$345 - a 109% debut gain that valued the company at about $13.7 billion, easily outpacing fellow Chinese AI "tiger" Zhipu AI's 13% first-day climb. Since then the shares have been volatile, trading in a 52-week range from HK$186.20 to HK$1,330, and the 283% revenue print did not produce a clean breakout. When a hypergrowth earnings number does not move a stock, the market is telling you two things at once: the growth was already priced in, and investors are looking at what the growth cost.
The cost side is the problem. MiniMax is growing revenue faster than it is growing losses in percentage terms - R&D up 138.8% versus revenue up 283.1% - but the absolute loss is still $293 million for a six-month period. At a gross margin of 17.9%, the company retains less than one dollar in five after direct inference costs. For the stock to re-rate, revenue growth must convert into margin expansion, not just top-line momentum.
The market is also asking whether enterprise API revenue is sticky. A sevenfold year-over-year jump can come from promotional pricing, one-off pilots, or genuine recurring contracts. The ARR figure - $800 million and rising - is management's answer to that question. The market will want to see it hold through at least two quarters before treating it as proof rather than promise.
There is also a valuation discipline at work. At a market capitalization that has oscillated between roughly $10 billion and $50 billion depending on the share price, MiniMax trades at a price-to-sales multiple that prices in years of flawless execution. A company burning $293 million a half cannot afford a single missed ARR milestone without a multiple contraction.
The Second-Order Problem: Model Inference Is Becoming a Commodity
The first-order read of MiniMax's results is simple: enterprise demand is strong, so buy the stock. The second-order question is harder, and it is the one most investors are not asking: if every Chinese AI lab can sell API calls, what prevents API pricing from collapsing toward marginal cost?
Yeyi Yun, Co-Founder and President, addressed part of this at the conference. She said multimodality is the ultimate direction for foundation models, and that MiniMax is beginning to merge its M3 language model and H3 video model into a single unified system. She also said the company has designed its models to be chip-agnostic, able to run on both Western and domestic Chinese chips. That chip-agnostic design is not a technical detail - it is a hedge against the export curbs that limit Chinese labs' access to advanced Western semiconductors, and it is a cost-control lever when domestic chips trade at a performance discount. But it is also an admission that no single chip supply is reliable enough to build a moat on.
The mechanism runs deeper than supply chains. In the AI value chain, value is concentrating at the two ends: consumer applications with real willingness to pay, and the cloud providers selling the compute underneath. Model labs sit in the middle, and the middle is where margins get squeezed. MiniMax's international revenue accounted for 60.8% of the first-half total, which diversifies demand but also exposes the company to global price competition from U.S. and Middle Eastern model providers who can undercut on price because their cost of capital is lower.
Consider the pricing arithmetic. MiniMax's API-vlm video model carries a published pay-as-you-go price of $0.01 per request, effective July 22, 2026, and the company has layered a Token Plan subscription on top to lock in recurring usage. That is a rational response to commoditization - subscriptions create stickiness that per-call pricing cannot - but it also caps the price per unit of intelligence. Every competitor can copy a price cut in a day; copying a switching-cost architecture takes quarters.
The structural question, then, is whether MiniMax's moat is technological or organizational. A model advantage compounds for months, not years - competitors can replicate architecture, and open-weight releases from rivals force the whole industry's price curve down. Organizational speed - the ability to ship, price, and integrate faster than rivals - is harder to copy. That is the bet the company is making, and it is why management talks about agility more often than it talks about parameter counts.
The Competitive Field Is Not Standing Still
MiniMax is not pivoting in a vacuum. The same commercialization pressure is visible across China's AI lab cohort. Moonshot AI, Zhipu AI, and ByteDance's model unit are all pushing frontier performance while racing to monetize. The rise of DeepSeek earlier in the cycle reset the baseline for what a Chinese lab can achieve at what cost, and it trained investors to expect capability parity to arrive faster and cheaper than Western incumbents priced in.
The difference for MiniMax is timing and mix. Zhipu AI, listed a day before MiniMax, leaned into enterprise and government contracts - stable, but perceived as less exciting in a hype-driven market. MiniMax leaned into consumer apps first, then pivoted to enterprise API. The market rewarded the consumer story at the IPO; it is now demanding proof of the enterprise story. That is a harder test, because enterprise contracts are scrutinized on retention and gross margin, not download rankings.
There is also a platform risk that hangs over every Chinese AI consumer app. Company disclosures flag that consumer apps face potential restrictions on major app stores and social platforms, while enterprise customers may prefer to work directly with cloud providers such as Alibaba or Tencent, which offer integrated AI services alongside core infrastructure. That is the bear case in one sentence: MiniMax's consumer channel can be throttled by platform gatekeepers, and its enterprise channel can be bypassed by the infrastructure owners. The pivot to API is, in part, a hedge against both.
The Bull Case, and What Would Break It
The strongest counter-thesis comes from the sell side. Goldman Sachs maintains a Buy rating on MiniMax with a 12-month target price of HK$860 - a figure the bank reiterated through the summer after previously setting it near HK$1,000 and trimming it. At a share price around HK$357, that target implies roughly 141% upside. The bank's model has MiniMax's revenue surging from $79 million in 2025 to $300 million in 2026, then to $880.1 million in 2027 and more than $2.4 billion in 2028. The bull case rests on three pillars: the $1 billion ARR target, organizational agility as a defensible moat, and deep integration with domestically produced chips as a supply-chain advantage.
A broader consensus is also turning more constructive. After the interim results, 21 analysts covering the stock lifted their 2026 revenue forecast to $451.9 million - up from $385.1 million before the print - with an average price target of HK$628. The sentiment shift is clear: the market is revising up what MiniMax can sell, even as it waits to see what MiniMax can keep.
Here is the falsifying signal. The bull case breaks if ARR fails to reach $1 billion by the end of 2026, or if gross margin stalls near 17.9% while the adjusted loss continues to widen. A second warning sign would be enterprise customers migrating to in-house models or to integrated AI offerings bundled by cloud providers such as Alibaba or Tencent - a risk flagged in company disclosures, since those providers can price infrastructure and models together in ways a standalone lab cannot match. A third: if token-consumption growth decelerates from the 20-times-January pace while per-token pricing keeps falling, revenue growth and margin move in opposite directions, and the unit-economics story unravels.
What Comes Next: Three Horizons
Short term - sentiment and liquidity. The shares will track AI-hype cycles, mainland-Chinese investor-access flows into Hong Kong tech, and the stabilization window around lock-up expirations. Volatility is the baseline, not the exception, for a stock that has traded between HK$186 and HK$1,330 in twelve months. The Goldman target of HK$860 and the analyst consensus of HK$628 create a visible tug-of-war that will show up in daily trading.
Medium term - fundamentals. Two numbers matter: the $1 billion ARR run rate by year-end 2026, and the gross-margin trajectory from 17.9%. If ARR hits target and margin expands toward the mid-20s, the current valuation starts to look cheap. If ARR stalls and margin compresses, the multiple contracts regardless of revenue growth. The interim report due cycle and any quarterly ARR updates are the catalysts that will force a decision.
Long term - structure. The winner in Chinese AI will not be the lab with the biggest model. It will be the one that delivers the most intelligence per dollar, with a defensible route to positive unit economics. MiniMax has proved it can sell AI at scale. What it has not yet proved is that selling AI pays.
The base case: ARR approaches $1 billion by the end of 2026, revenue roughly triples, and losses narrow gradually as infrastructure efficiency improves. The upside case: the unified M3-plus-H3 model creates a durable enterprise moat, gross margin expands beyond 25%, and the Goldman target becomes a floor rather than a ceiling. The downside case: API price competition compresses margins, ARR growth decelerates below the $1 billion target, and the market re-rates MiniMax as a low-margin inference utility rather than a platform business.
MiniMax has spent two years proving it can build models that rival the Americans. The next two years will decide whether it can build a business that rivals them too. The conference message was confident. The numbers are impressive. The market, characteristically, is waiting for the third thing: evidence that the pivot converts growth into profit.
Explore more exclusive insights at nextfin.ai.
