MiniMax Group Inc.: Platform Revenue Grew 703% to 63.4% of Sales While Gross Margin Is Still Only 17.9%, and HK$361.40 Prices the Optimistic 2030 Case

The quality of the growth is mixed. Gross margin improved from negative 24.7% in 2023 to 12.2% in 2024 and 25.4% in 2025. In the prospectus, Open Platform gross margin reached 69.4% for the first nine months of 2025, while AI-native products were only 4.7%. That made the subsequent shift toward platform revenue look particularly attractive. Yet consolidated H1 2026 gross margin was only 17.9%: better than 12.1% a year earlier but below FY2025's 25.4%. Revenue mix moved toward the historically high-margin platform at the same time that aggregate margin failed to sustain its FY2025 level. That is the single most important unresolved unit-economics question in the accounts.

The cost structure explains why. During the first nine months of 2025, inference cloud resources were 92.7% of cost of sales; cost of sales itself was 76.7% of revenue. MiniMax has shown real technical evidence of pursuing compute efficiency: its M1 release said deep reasoning could consume about 30% of the compute required by DeepSeek R1 under the company's test, and disclosed a reinforcement-learning phase using 512 H800 GPUs for three weeks at a rental cost of US$534,700. The corresponding M1 API was priced at US$0.40 per million input tokens for shorter contexts and US$2.20 per million output tokens. These are company claims and tariff prices, not audited production cost per token. MiniMax still does not disclose group-wide token volume, so an audited cost-per-token calculation is impossible.

The deeper cash problem sits in R&D, much of which is training compute rather than inference cost. H1 2026 R&D expense reached US$296.9 million, 2.55 times revenue and more than the US$252.8 million spent in all of 2025. Gross profit in the half was only US$20.8 million. Selling expense, by contrast, fell 17.9% while sales nearly quadrupled. MiniMax is showing very strong distribution leverage and essentially no evidence yet that gross profit can finance the model-development bill.

The balance sheet buys time. At June 30, MiniMax reported a broad cash balance of US$1.323 billion and US$133.6 million of bank debt. After the reporting date it issued 35.6 million shares at HK$268, receiving about HK$9.444 billion net, and then HK$6.5 billion of zero-coupon 2027 convertible bonds at an initial HK$335 conversion price, receiving HK$6.433 billion net. Treating the convertible cash and principal as broadly offsetting for net-debt purposes, I estimate post-financing net cash around US$2.39 billion before subsequent operating burn. Near-term insolvency is a much smaller risk than dilution and poor returns on that capital.

The geography story has also changed. More than 70% of FY2025 revenue came from outside mainland China, consistent with the prospectus evidence that 73.1% of nine-month 2025 revenue was outside the mainland. By H1 2026, “rest of world” was 60.8% of revenue while mainland China rose to 39.2%. Overseas revenue still grew strongly in absolute dollars; China's platform demand simply grew faster. Describing MiniMax today as a greater-than-70%-overseas-revenue company is no longer correct.

The market's central narrative is now a contest between two interpretations. The bullish one sees a very young company that has turned frontier-model progress into a rapidly expanding commercial platform, has global consumer distribution, has raised enough capital to fight through several model generations and is proving that Chinese frontier models can win workloads on price/performance. The bearish interpretation sees a company operating inside the industry's least settled profit pool: model inference is being commoditised by open weights and hyperscaler subsidies; gross margin remains low; R&D grows faster than gross profit; switching costs for API customers are limited; and a handful of benchmarks or six months of revenue growth cannot establish a durable moat.

The share price itself has traded these interpretations violently. It moved from HK$165 at IPO to HK$345 on the first close, climbed into a March frenzy with a HK$1,330 intraday high, sat at HK$794 by mid-May, then collapsed into the July lock-up and capital raise. A July 9 lock-up expiry made roughly 45% of shares eligible for trading according to Reuters; the stock fell 18.0% that day to HK$297.40. The next day's financing announcement coincided with another 9.7% decline. H3's launch on July 31 was followed by Southbound Stock Connect eligibility on August 6, which produced a 17.1% rise, then heavy mainland buying and the August results. Mainland investors reportedly bought about RMB10.6 billion of MiniMax stock during August.

That trading record is analytically useful. MiniMax's market capitalisation can move tens of billions of Hong Kong dollars because an investor pool becomes eligible, a lock-up ends, a new model ranks well, or a competing model appears. Those are capital-market facts, but they should not be mistaken for evidence that the terminal economics of foundation models have been resolved.

The qualitative portrait is company in transition. MiniMax has moved from model-building start-up, to globally distributed consumer-AI company, to a platform/inference company with consumer applications attached. The transition is commercially impressive and financially unfinished. Its revenue evidence is much stronger than 111 days ago. Its moat evidence is only modestly stronger. Its balance sheet is dramatically stronger. The economics required to justify the stock remain demanding.

Company vertical history and financial review

The founding-date discrepancy in the prior coverage can be resolved cleanly. MiniMax's Cayman holding company was incorporated on June 30, 2021, and important PRC entities were incorporated in November 2021. The prospectus still states that the business was “founded in 2022 by a group of engineers,” and MiniMax's own corporate material describes the operating company as founded in early 2022. Press reports calling it a 2021 founding are best understood as referring to legal incorporation and preparatory entities. For business-history purposes, 2022 is the correct founding year; 2021 is the corporate-formation year.

Founder Yan Junjie had worked at SenseTime before starting MiniMax. That background matters because the company's route was model-first rather than application-first. The first commercial story was not “build an AI companion app and eventually own a model”; MiniMax began developing its own foundation architecture and launched the abab1 text model in April 2022. It obtained its first API customer and launched abab5.5 in May 2023, then added Talkie, Xingye and Speech-01 across the rest of 2023. The early company was already trying to sell the same underlying intelligence in two forms: APIs to developers and products to consumers.

Early capital also shaped that strategy. The financing history ran from an angel round in late 2021 through Pre-A, A, A+, Pre-B, Pre-B+ and Pre-B++ rounds before listing. The prospectus identifies sophisticated investors associated with Alibaba, Tencent, miHoYo and IDG, among others. Alibaba became more than a passive shareholder: MiniMax has bought Alibaba Cloud services since 2022, while Alibaba has also purchased MiniMax API services, with post-listing framework arrangements extending through 2028. I found primary prospectus support for Alibaba, Tencent and miHoYo. I did not find equally clear primary-prospectus support for Xiaohongshu as a named equity investor, so I do not treat the widely repeated Xiaohongshu claim as verified in this report.

The company's short life is better divided into four stages than recited launch by launch.

The first stage, 2022–23, was model validation. MiniMax had almost no scale, but it established the pieces of the eventual business: proprietary text models, an API, character-based consumer products and speech. Revenue in 2023 was just US$3.46 million. Open Platform was already 78.1% of that small revenue base, while the consumer products were only beginning to monetise through subscriptions and top-ups. Gross margin was negative 24.7%. The economic meaning is simple: customers existed before efficient inference did.

The second stage, 2024, was consumer-scale validation. Average monthly active users increased from 3.1 million in 2023 to 19.1 million, paying users from about 119,700 to 650,300, and revenue from US$3.46 million to US$30.52 million. Hailuo added video generation; Talkie and Xingye expanded internationally; online-marketing services revenue became 47.8% of 2024 revenue. Gross margin turned positive at 12.2%. The important legacy of 2024 was distribution: MiniMax proved that a Chinese model company could acquire large numbers of global consumers rather than relying only on domestic enterprise contracts.

The third stage, 2025, was architecture and monetisation expansion. MiniMax released Text-01, M1, Hailuo-02, M2 and newer speech/music products while increasing cumulative users above 236 million by year-end. Revenue rose 158.9% to US$79.0 million. AI-native products contributed US$53.1 million and Open Platform US$26.0 million. Gross margin reached 25.4%, while adjusted loss moved only modestly from US$244.2 million to US$250.9 million despite the revenue increase. Selling expense fell sharply. For the first time there was evidence that revenue could scale materially faster than some operating costs.

But 2025 did not establish cash self-sufficiency. The prospectus records net operating cash outflow of US$258.5 million in 2024 and US$209.4 million in the first nine months of 2025; the prospectus's broader “cash burn” measure was US$260.7 million and US$211.3 million respectively. I could not independently reconcile the prior report's approximately US$280 million FY2025 operating-cash-flow figure to a full-year primary cash-flow line in the materials extracted for this re-read. I do not reuse that number as an audited fact. The verified evidence is already sufficient to show heavy cash consumption.

The fourth stage began with the January 2026 listing and has already split into two subplots: operating acceleration and capital-market normalisation. The IPO was priced at HK$165. Reuters reported the base upsized offering of 29.2 million shares at about HK$4.82 billion gross. The prospectus had a 15% over-allotment mechanism, and the subsequent issued-share count of 313.635 million before July is consistent with the fully enlarged offer structure; on that basis, ultimate gross proceeds including the over-allotment were approximately HK$5.54 billion. The public offer was reported more than 1,800 times subscribed.

The IPO valuation narrative was scarcity plus optionality. MiniMax and Zhipu were among the first listed Chinese pure-play foundation-model developers. The stock's HK$345 first close valued MiniMax at roughly HK$106–108 billion, nearly twice Zhipu's first-day market value. Investors initially preferred MiniMax's consumer/global-growth story to Zhipu's more enterprise-heavy profile. By June that ranking had reversed: SCMP reported Zhipu at HK$585.8 billion versus MiniMax at HK$159.3 billion on June 8, showing how rapidly public markets changed their preferred AI narrative.

The operating data then changed even faster. H1 2026 revenue was US$116.6 million versus US$30.4 million a year earlier. Platform and enterprise sales expanded eightfold; consumer-product revenue roughly doubled. MiniMax's revenue mix inverted:

Source: MiniMax H1 2026 interim results.

This is the most important business transition since the IPO. Consumer products are still growing; the platform simply grew far faster. That matters because the prospectus had shown a striking economics gap between the two: in the first nine months of 2025, Open Platform gross margin was 69.4%, while AI-native products managed only 4.7%. If those segment economics persisted, the mix change would be an obvious margin catalyst. H1 2026's 17.9% consolidated margin says investors should not assume they persist.

The full financial arc is short enough to fit in one table:

† Prospectus cash-and-cash-equivalent figures for the earlier periods are not identical to the later company's broader “cash balance” definition; they should not be treated as perfectly comparable.

Sources: prospectus, FY2025 results and H1 2026 results.

Three things jump out. Revenue scaling is exceptional. Gross-margin improvement has been real but non-linear. Losses remain large enough that ordinary profitability ratios are almost meaningless.

The IFRS losses look even worse but require adjustment. FY2025 IFRS net loss was about US$1.87 billion because roughly US$1.59 billion came from fair-value losses on financial liabilities, principally preferred-share instruments that converted on listing. H1 2026 IFRS loss narrowed to US$358.0 million as that accounting effect largely disappeared, but adjusted loss worsened from US$138.7 million to about US$293 million. The cleaner conclusion is that listing cured a distorted liability-accounting presentation, not the underlying operating deficit.

The balance sheet has improved far more than the income statement. June cash and equivalents alone were US$930.9 million; adding financial assets, deposits and restricted cash gave management's US$1.323 billion “cash balance.” Bank debt was US$133.6 million. The July share placement raised HK$9.444 billion net and enlarged issued shares from 313.635 million to 349.235 million, meaning a roughly 11.35% increase in the pre-placement share count and about 10.2% dilution to an existing holder's percentage ownership. The HK$6.5 billion convertible could create another approximately 19.4 million Class A shares at the initial HK$335 conversion price, equivalent to about 5.3% of a pro-forma share count after conversion.

That July financing deserves a balanced reading. Raising more than HK$15.8 billion net between equity and convertible debt gives MiniMax enough capital to continue an expensive model race. It also shows why per-share economics matter. The company sold equity at HK$268 shortly after the stock had traded several times higher; investors who owned the scarcity premium saw the company monetise that premium for corporate capital. For the company, that was sensible insurance. For an owner, it was expensive dilution.

Governance amplifies the distinction between company and owner. The prospectus adopted a 10-votes-per-share Class B structure in 2023 versus one vote for Class A on ordinary WVR matters, subject to Listing Rule matters requiring one-share-one-vote. Before the July placement, Class B stood at 81.103 million shares; after the placement it was unchanged while Class A rose to 268.133 million. Prospectus disclosures indicate Yan's founder-controlled interests amounted to about 79.103 million shares and Yun Yeyi's to 7 million. On the disclosed class structure, I estimate Yan now has roughly 22.7% economic ownership but about 69% of WVR votes; Yan and Yun together hold approximately 24.7% economically and around 75.6% of votes. These voting percentages are calculations from the disclosed class counts, not a company-published current-vote percentage.

Minority holders supply capital but cannot realistically displace the founder on ordinary WVR resolutions while that structure remains intact. Yan also combines chairman, CEO and technical leadership roles, and the interim filing explicitly records the board's departure from the normal separation of chairman and CEO functions. Concentration like that can speed technical decision-making. But it merits a governance discount, because the usual external-control mechanism is weak.

Lock-ups are more nuanced than the “six months for controlling shareholders” shorthand suggests. Under the original pre-commercial-company prospectus structure, Yan-related controlling interests and Yun's interest were subject to 24-month restrictions, while miHoYo and IDG sophisticated-investor holdings were described with 12-month restrictions. The prospectus also contemplated shortening the Chapter 18C restrictions if MiniMax satisfied the commercial-company revenue threshold and the Exchange approved the change. MiniMax subsequently dropped the “P” marker and trades as “-W”; Reuters reported that roughly 45% of outstanding shares were due to become unlocked in the week of July 9. The founder's original restriction should not be conflated with the July 45% unlock.

The actual market consequence was unmistakable, even though seller identity was not. MiniMax closed at HK$297.40 on July 9, down 18.0%, on 20.9 million shares of volume, compared with fewer than 1 million shares on several sessions immediately before the expiry. I found no primary disclosure proving that “45% was sold”; Reuters established that 45% became eligible. CCASS movement and turnover cannot identify every beneficial seller. The defensible conclusion is that the expiry changed tradable supply and the market priced that overhang aggressively.

Business model, industry and horizontal peers

MiniMax monetises intelligence in two economically different forms. AI-native products include consumer applications such as Hailuo, Talkie/Xingye, MiniMax and audio products, monetised through subscriptions, top-ups and, historically, advertising. Open Platform sells model access and AI services to enterprises and developers. FY2025 revenue was roughly two-thirds products and one-third platform. Six months later the mix had flipped to 36.6% products and 63.4% platform.

The original consumer model produced reach before it produced compelling unit economics. Average MAU increased from 3.1 million in 2023 to 19.1 million in 2024 and 27.6 million in the first nine months of 2025; cumulative users exceeded 212 million by September 2025 and 236 million by year-end. Paying users rose from about 119,700 in 2023 to 650,300 in 2024 and 1.77 million in the nine months to September 2025.

A crude monetisation calculation is revealing but must not be mistaken for cohort ARPU. AI-native product revenue was approximately US$21.8 million in 2024; divided by the disclosed 19.1 million average MAU, that is only about US$1.14 of annual product revenue per average monthly active user. For nine-month 2025, product revenue of about US$38.0 million divided by 27.6 million average MAU gives roughly US$1.38 per average MAU for nine months. User cohorts, geography and paying periods differ, so these are scale indicators rather than formal ARPU. They still explain why large cumulative-user claims did not translate into large gross profit.

The platform is potentially a better business because every successful external application can generate inference demand without MiniMax paying to build its own consumer distribution. The catch is price competition. Foundation-model APIs increasingly resemble a market where vendors advertise intelligence-per-dollar as aggressively as raw capability. MiniMax itself has made low cost central to product positioning: M1's disclosed API tariff was only US$0.40 per million input tokens for contexts under 200,000 and US$2.20 per million output tokens, while Hailuo-02 was marketed on a claimed 2.5-times architecture-efficiency improvement. H3 was subsequently promoted as costing less than one-third of mainstream video competitors. These statements show management's strategy; they do not establish that MiniMax alone possesses a permanent cost advantage.

The prospectus gives the best available accounting window into inference economics. Cloud resources used for inference were US$26.0 million in 2024 and US$38.0 million for nine-month 2025, representing 96.9% and 92.7% of cost of sales. Inference cloud cost fell from roughly 85% of revenue in 2024 to roughly 71% in nine-month 2025. That is genuine progress. Yet the H1 2026 accounts show total cost of sales back at 82.1% of revenue. Without group token volumes and a current split of inference compute, app-store fees and other COGS, it is impossible to say whether cost per token continued falling fast enough to offset API price cuts.

This distinction is load-bearing for the moat. MiniMax has evidence of technical efficiency: Lightning Attention, long-context design, multimodal architectures and rapid model iteration. M1's disclosed tests claimed 30% of DeepSeek R1's compute for an 80,000-token reasoning workload, and its training disclosures were unusually concrete. Hailuo-02 and later M-series releases continued to emphasise efficient activated parameters and inference. A real moat requires those gains to remain economically valuable after competitors copy architectures, release open weights or cut prices. MiniMax has existed for only about four years. It cannot yet provide that multi-cycle proof.

Distribution is the second candidate moat. MiniMax had users across more than 200 countries before listing and says it now serves customers, developers and users across more than 230 countries and regions. Few young Chinese model labs have simultaneously built character AI, video, audio, coding and enterprise API businesses with that international reach. The weakness is retention evidence: cumulative users are not active users, and the latest interim result does not provide enough product-level retention, subscription churn or cohort ARPU data to show switching costs.

Capital is a competitive advantage rather than a moat. A post-financing net-cash position around US$2.4 billion can fund multiple expensive training cycles. It substantially reduces the chance that MiniMax has to stop investing when a competitor releases a stronger model. But Alibaba, Tencent, ByteDance and other hyperscalers have much larger cash-generating businesses behind their AI spending. MiniMax's capital must ultimately earn a return from AI; hyperscalers can treat model pricing as a way to sell cloud, advertising or ecosystem services.

That leaves the industry's profit pool unsettled. Upstream cloud and accelerator suppliers get paid whether a model vendor earns money or not. Distribution platforms can monetise AI through existing advertising, commerce, gaming or cloud customers. Model developers occupy the most technologically glamorous layer but face unusually low switching costs: developers can route traffic across several APIs, open-weight models can be self-hosted, and benchmark leadership can change within months. MiniMax is trying to escape that trap by owning both models and applications, while using Open Platform to turn its model investment into external traffic.

Regulation adds friction without destroying the market. China's Interim Measures for Generative AI Services apply to text, image, audio and video services offered to the domestic public and impose obligations around prohibited content, discrimination, intellectual property, personal information, transparency and reliability. The CAC was still running specific clean-up campaigns around AI applications in 2026. For MiniMax, that creates moderation and compliance costs in China; outside China, privacy, IP, app-store rules and payment/distribution constraints form a different regulatory stack.

The competitive landscape is Scenario C: ample competitors, but no listed peer perfectly matches MiniMax.

Z.AI Co., Ltd., listed as Knowledge Atlas Technology until its July 2026 renaming and better known as Zhipu AI, is the closest public pure-play reference. Zhipu came from Tsinghua-linked research, commercialised more heavily into enterprises and government-related workloads and listed one day before MiniMax under 02513.HK. At debut, investors awarded MiniMax almost twice Zhipu's market capitalisation. By early June, Zhipu was worth about HK$586 billion and MiniMax HK$159 billion. The reversal tells us less about stable intrinsic-value differences than about what the market was temporarily rewarding: technical model perception, enterprise monetisation and scarcity.

SenseTime, 00020.HK, is the older warning and opportunity case. It began with computer vision and subsequently pivoted heavily toward generative AI and large-model infrastructure. Its H1 2025 filing reported RMB2.36 billion of group revenue, of which RMB1.82 billion was generative AI, and a 38.5% group gross margin. That is a much larger commercial base and a higher reported margin than MiniMax had then, but SenseTime's long history of losses also shows that scale and respectable gross margin do not automatically pay for frontier-model R&D. The comparison is directionally useful rather than a current multiple anchor because the business mix and reporting period differ.

Alibaba, 09988.HK, sits in a different economic position. It backs MiniMax, sells cloud resources to it and buys API services from it, while Alibaba's own Qwen family competes for the same developers. That relationship captures the industry structure in miniature: the hyperscaler can earn from MiniMax even while competing against MiniMax. Alibaba can also price a foundation model partly to stimulate cloud demand, an option a pure-play model company lacks.

Tencent, 00700.HK, is similarly both capital provider and ecosystem competitor through its own AI efforts. Kuaishou, 01024.HK, is particularly relevant to Hailuo because Kling competes directly in generative video and comes with a large native creator/content distribution network. Reuters described the Chinese AI-video market around H3 as one led by ByteDance and Kuaishou. H3's combination of 2K output, stereo sound, multimodal reference inputs and open weights is MiniMax's attempt to compete on both capability and cost rather than distribution scale.

The strongest answer to why a customer chooses MiniMax today is price/performance across several modalities, plus convenient API access and products that are already used globally. That same customer can leave: APIs can be substituted, open weights reduce dependence on a single vendor, and model rankings change rapidly. MiniMax's competitive niche is a technically credible, global challenger rather than an entrenched platform owner.

The company's strongest evidence of moat is repeated efficient model engineering. Its weakest evidence is customer captivity. I classify the moat as weak today because the first has not yet created the second.

Price path, current fundamentals and valuation

MiniMax has packed what would normally be several market cycles into eight months.

Historical prices are from S&P Global Market Intelligence-derived data and other end-of-day sources; the March figure is the March month-end close, while HK$1,330 was an intraday high.

The first leg, from January through March, was a scarcity and capability trade. The IPO itself was massively oversubscribed. FY2025 results released March 2 then confirmed 159% revenue growth, improving gross margin and greater-than-70% international revenue. M2.7 was unveiled March 18, the same date on which the stock printed its HK$1,330 intraday high. Correlation is not proof that M2.7 alone caused the peak, but the timing shows that model-news sensitivity had become extreme.

At HK$1,330, using the pre-July share count and FY2025 revenue, MiniMax's implied market capitalisation was roughly HK$417 billion and trailing P/S about 677 times by my calculation. No plausible six-month earnings revision generated that valuation. The March peak was overwhelmingly a multiple and narrative event.

The May-to-July collapse had several causes layered together. First, the approaching lock-up expiry transformed theoretical value into potential stock supply. Caixin reported MiniMax had already fallen by about half in the fortnight before the early-July unlock, while Reuters identified roughly 45% of shares becoming eligible. Second, the model sector's pricing environment became more obviously deflationary: around the M3 period, MiniMax cut API pricing while rivals accelerated product releases. Third, the July equity/convertible financing told the market that MiniMax intended to keep spending aggressively and was willing to issue stock.

The tape around the lock-up reads cleanly. July 8 closed HK$362.60, up 12.0%, as the market digested reports of an enormous forthcoming model. July 9 closed HK$297.40, down 18.0%, as the major lock-up expired. July 10 closed HK$268.60, another 9.7% lower, as the HK$16 billion financing package came into focus. Volume on July 9 jumped above 20 million shares. That is supply and financing risk being repriced in real time.

The August recovery was just as clearly capital-markets assisted. H3's July 31 release coincided with a 13.2% rise. MiniMax became eligible for Southbound Stock Connect on August 6; the stock rose 17.1% that day and another 9.8% on August 7. HKEX then fast-tracked MiniMax into its Tech 100 family effective August 13. Reported August Southbound purchases of MiniMax reached about RMB10.6 billion, more than the contemporaneous inflow cited for Alibaba or Tencent.

H1 results were released August 26 after the market closed. The next session rose only 3.8%, but the stock reached HK$349 by August 31 after a 16.2% one-day rise. It is reasonable to infer that both the fundamental surprise and the newly available Southbound liquidity contributed to the recovery; the evidence does not allow a clean percentage attribution between them.

This is the share-price decomposition investors should carry forward:

At IPO, the market paid roughly 84 times the subsequently reported FY2025 revenue. By mid-May it paid about 404 times. At September 4 it paid about 205 times FY2025 sales, but only 69 times H1 2026 annualised sales. Adjusting for estimated post-financing net cash brings the latter to roughly 59 times EV/sales. Under my US$290 million FY2026 base revenue assumption, current EV/sales is about 48 times.

The May valuation was driven principally by multiple expansion. The current valuation is supported by much more revenue delivery, but it still assumes that today's exceptional growth survives several years.

The current fundamentals contain four encouraging points and three warnings. The positives are revenue growth, platform mix, selling expense leverage and the balance sheet. The warnings are H1 gross margin, R&D spending and dilution. H1 2026 gross profit was only US$20.8 million against R&D of US$296.9 million. Adjusted loss more than doubled year over year to approximately US$293 million. Trade receivables rose from US$10.7 million at December to US$39.1 million at June, while trade and bills payables rose from US$57.7 million to US$170.1 million. None of those balance-sheet movements is yet evidence of distress, but receivables deserve attention because growth is becoming enterprise-heavy.

Quarterly analysis has a limitation that should be explicit. MiniMax does not provide the four clean audited quarters that a U.S.-style quarterly filer would. The defensible reported periods are FY2025 and H1 2026, with H1 2025 comparatives. I do not manufacture quarterly revenue or margin numbers from incomplete disclosures.

For valuation, P/E, FCF yield and owner earnings are unusable. MiniMax has no positive earnings, no positive free cash flow and no dividend. The five-year operating-cash-flow/net-income ratio requested by the framework is also economically meaningless here: the company has not operated for five full years, IFRS net loss was distorted by preferred-share fair-value accounting before listing, and operating cash flow is negative. Maintenance capex is small relative to cloud/R&D spending and is not separately disclosed well enough to calculate defensible “owner earnings.” The correct method is a revenue-to-margin path with explicit cash burn, dilution and a terminal sales multiple.

So I give the greatest weight to a four-year, 2030 EV/sales framework. It asks how much revenue MiniMax must actually reach, what gross margin it must produce after inference-price competition, how much of its new capital remains, and what multiple a still-growing AI platform might deserve. Current-run-rate EV/sales is a secondary sanity check. Private-lab valuations and listed AI-peer multiples get little weight because both numerator and business models are unstable.

The scenario assumptions below are mine, not management guidance. 2026E starts with the reported US$116.6 million H1. The inference assumption is expressed through consolidated COGS: H1 2026 COGS was 82.1% of revenue, and inference cloud had historically represented more than 90% of COGS in the prospectus period.

The model uses about 369 million shares for base/optimistic cases, including potential conversion of the HK$6.5 billion convertible at HK$335; the conservative case assumes some additional dilution. It does not attempt to predict every employee award. This is valuation-scenario analysis inside a research framework, not investment advice. Inputs on current capital structure and financing come from the interim filing.

The conservative case is not a collapse. Revenue still reaches more than US$1 billion by 2030, gross margin rises materially and MiniMax survives. It generates only HK$60 of present value because a business that still struggles to convert gross profit into operating profit cannot command a large sales multiple once growth normalises.

The base case is demanding in absolute terms. US$1.81 billion in 2030 revenue means MiniMax becomes roughly 23 times its FY2025 size within five years. Gross margin must nearly triple from H1 2026's level, operating margin must turn positive despite frontier-model research, and dilution must remain controlled. Even then my present value is around HK$180.

The optimistic case is a genuine category-winning outcome. Revenue exceeds US$3 billion, gross margin approaches software-platform economics, R&D scales much more slowly than revenue and the market continues paying 12 times sales in 2030. That supports roughly HK$479 of present value and HK$809 as a 2030 terminal share value.

At HK$361.40, the stock already requires something much closer to the optimistic trajectory than the base trajectory.

An expectation-gap calculation makes this explicit. If I use a 10-times 2030 EV/sales multiple, a 15% required return and US$1.2 billion of residual net cash, the current diluted equity value implies about US$2.87 billion of 2030 revenue. Starting from US$290 million in 2026, that requires approximately 77% annual revenue growth for four years. A more ordinary 8-times multiple and 16% discount rate pushes the required CAGR close to 90%. Those are not impossible numbers for a company that just grew 283%; they are unusually unforgiving assumptions to pay for in advance.

Historical-percentile valuation analysis adds little. MiniMax has only eight months of trading, the share count changed in July and the revenue denominator is tripling. A trailing P/S percentile would create false precision. What can be said is that today's valuation is far below March to May mania on a trailing-sales basis and still far above the absolute value produced by a merely very successful base case.

Peer valuation is equally dangerous. Zhipu's market cap changed by hundreds of billions of Hong Kong dollars within months of listing; using that as “fair value” for MiniMax would substitute one speculative multiple for another. SenseTime has a different legacy business. Alibaba and Tencent are profitable ecosystems rather than model pure plays. Absolute economics must carry the valuation.

The margin-of-safety test is harsh. Current price is roughly six times the HK$60 conservative present value. The margin of safety against that case is zero.

The base case's most fragile assumption is the 58% 2026–30 revenue CAGR. Reducing it to 70% of the original assumption, or about 40.6%, while leaving the 8-times terminal multiple and other framework inputs broadly intact, reduces my base present value to roughly HK$118. That sensitivity is why a high terminal multiple cannot substitute for revenue proof.

The “flat earnings for three years” test is even less favourable because earnings are currently negative. With no dividend and no positive owner earnings, a mechanically unchanged share price would produce a 0% nominal return before considering future dilution. The latest 10-year HKD HKSAR institutional-government-bond auction I could source, dated June 24, 2026, cleared at a 3.371% annualised yield. On that comparison, there is no margin of safety at this buy price.

Margin-of-safety sufficiency verdict: none.

Risks, catalysts and cross-synthesis

The first permanent-loss risk is inference commoditisation. I rate its probability high and its impact high. The observable indicators are consolidated gross margin, API pricing and the cost-of-sales/revenue ratio. MiniMax's own strategy is to make advanced models cheap, and competitors with cloud or platform economics can afford the same strategy. If H1 2026's 17.9% gross margin fails to move above 25–30% while Open Platform becomes most of revenue, the old 69.4% platform-margin evidence will have proved non-durable. The transmission path is direct: lower price per token offsets higher token volume, gross profit fails to cover R&D, cash burn remains high, and the terminal sales multiple contracts.

The second is the training-cost race. Probability is medium-high; impact is high. H1 R&D spending of US$296.9 million was roughly fourteen times gross profit. MiniMax can finance that race today because it raised substantial capital. The risk is that each successful model merely earns the right to fund another expensive model generation. The observable ratio is R&D/revenue: it was about 255% in H1 2026. If it remains above 150–200% through 2027 while gross margin stays below 25%, the financing has extended runway rather than solved the economic problem.

The third is technological displacement. Probability is high; impact is high. MiniMax competes with fast-moving independent labs, hyperscalers and open-weight communities across text, coding, agents and video. H3 arrived into a video market where Reuters specifically identified ByteDance and Kuaishou as major competitors; MiniMax's July share price was also sensitive to rival model launches. A model can remain technically good and still lose commercial share if another model offers similar capability at half the price or integrates directly into a dominant developer/cloud ecosystem.

The fourth is dilution and shareholder-supply risk. Probability is medium; impact is high. July's placement increased the share count by 11.35%; conversion of the 2027 bonds would add roughly 19.4 million more shares at the initial conversion price. The July lock-up episode showed that tradable-supply changes can dominate the stock. Further unlocks, employee awards, conversion or a new raise can cause per-share value to lag enterprise growth. Investors should track issued shares and substantial-shareholder disclosures rather than assuming a fixed denominator.

The fifth is governance. Probability of an actual governance failure is low-to-medium, but impact could be high because the founder's estimated vote control is far greater than his economic stake. Yan's combined chairman/CEO/technical role and WVR control let management pursue long-duration investment even if public holders object. That may be exactly what frontier-model development requires; it also means a minority investor's remedy for disagreeing with capital allocation is principally to sell.

The sixth is regulation and IP. Probability is medium and impact medium-to-high. MiniMax disclosed several intellectual-property infringement disputes at June 30 and had made no provision because the outcomes could not be reliably predicted. Generative video, voice and character products naturally sit close to copyright, likeness and content-moderation boundaries, while the company's international footprint subjects it to several regulatory systems. A material injunction, app-store restriction or content-policy action could hit both user acquisition and international revenue.

Positive catalysts are unusually concrete. Continued platform growth above 200% would show that H1 was not a one-off catch-up. A gross margin sustainably above 25%, and then 30%, would indicate that inference-efficiency gains are beating price deflation. R&D/revenue falling below 150% without a deterioration in benchmark competitiveness would show operating leverage. A large enterprise/API contract outside related-party ecosystems would strengthen the moat case. The strongest long-term catalyst would be positive operating cash flow achieved without slowing model cadence.

Negative catalysts are the mirror image. Platform growth below 100% would puncture the central new thesis, and gross margin below 15% would suggest token economics are worsening. Another large discounted equity raise before the present capital is substantially deployed would call capital discipline into question. And a major model generation that ranks poorly while competitors cut pricing would hit both revenue expectations and the valuation multiple.

The tracking dashboard converts those arguments into observables:

† No official FY2026 results date was available in the sourced IR material as of the research date. March 2027 is an estimate based on FY2025 results having been announced on March 2, 2026; it is not company guidance.

Read against the H1 2026 actuals, this dashboard does not start from a clean slate. Adjusted loss to revenue is already 251% and R&D to revenue about 255%; both are through their alert thresholds rather than approaching them. Gross margin at 17.9%, cost of sales at 82.1% of revenue and trade receivables at 33.5% of half-year revenue all sit outside their working ranges. Only reported revenue growth, Open Platform growth and the share count begin inside objective. The table is a list of things that must improve, not a list of things that are currently fine.

The financial indicators should be read together. Gross margin rising because MiniMax raises prices would mean something different from gross margin rising while API prices keep falling: the latter would be stronger evidence of architectural efficiency. Platform growth should be checked against receivables because revenue quality matters more as enterprise sales increase. Share count is a first-class operating metric for this stock because the company's business can grow while owner participation is diluted.

Looking vertically across MiniMax's entire life, the capability it has genuinely proven is speed. It went from a 2022 model start-up to tens of millions of active users, US$79 million of annual sales, a public listing and a US$116.6 million half-year within about four years. It repeatedly expanded modalities rather than remaining a text-model vendor. It then converted a consumer-heavy revenue base into a platform-heavy one faster than the IPO numbers implied. Those achievements are too large to dismiss as mere AI-sector beta.

That success also drew heavily on an extraordinary era. Generative-AI demand expanded rapidly, capital markets funded companies long before profitability, cloud providers made massive compute available and users were willing to try new AI products with little incumbency. MiniMax executed well inside those tailwinds. The difficult phase starts when the market asks which model developers can capture economics rather than usage.

Horizontally, MiniMax's advantage is a combination of technical efficiency, multimodal breadth and a surprisingly global consumer footprint, not scale versus Alibaba or Tencent, distribution versus Kuaishou, or enterprise incumbency versus older AI vendors. That combination gives it several routes to monetisation. Its structural weakness is that none of those routes yet creates strong switching costs. A developer can change APIs. A video creator can use several generators. A consumer can download the next compelling AI app. MiniMax has to keep winning repeatedly.

The capital raise buys the time required to attempt that. Post-financing liquidity means the bear case is no longer “cash runs out soon.” The bear case is “the company spends billions proving that foundation-model intelligence is a low-margin input.” That is a subtler and, for equity holders, potentially more important risk.

The market may currently be misjudging two things in opposite directions. Bears who still use FY2025's US$79 million revenue as the business scale are badly behind the evidence; H1 2026 alone was US$116.6 million. Bulls who capitalise the H1 growth rate without demanding gross-margin and R&D leverage are equally vulnerable. The revenue thesis has improved faster than the economic-quality thesis.

For the next year, the critical variable is gross profit, not revenue alone. MiniMax could double revenue again and still disappoint equity holders if pricing keeps consolidated gross margin around 15–20% and R&D remains several times revenue.

For three years, the critical variable is whether Open Platform becomes a repeatable, sufficiently high-margin business that funds a meaningful part of frontier-model development. The company needs both volume and price discipline, or a sufficiently steep inference-cost curve, to make that happen.

For five years, the issue becomes institutional: does MiniMax become one of the small number of independent model platforms with durable developer demand and owned consumer distribution, or does the model layer become a feature supplied by cloud and platform incumbents? The valuation today effectively assumes MiniMax survives on the favourable side of that divide.

Core bull reasons:

• H1 2026 revenue of US$116.6 million exceeded all of FY2025 and grew 283%, giving direct evidence that the revenue base used in the May bearish argument was already stale.
• Open Platform revenue rose 703% and became 63.4% of the company, potentially moving mix toward the part of the business that had shown much higher gross margin in the prospectus.
• The June balance sheet plus July financing provide roughly US$2.4 billion of estimated post-financing net cash, materially reducing near-term financing failure risk.
• MiniMax has repeatedly produced technically competitive, low-cost text, video and audio models while building distribution across more than 230 countries and regions, evidence of execution rather than a single-product accident.

Core bear reasons:

• H1 2026 gross margin was only 17.9% and below FY2025's 25.4%, despite platform becoming the majority of revenue; the expected mix-driven margin benefit has not yet shown up in consolidated economics.
• H1 R&D of US$296.9 million was 2.55 times revenue and about fourteen times gross profit, leaving a very long path to self-funded frontier-model development.
• Current EV/sales is about 59 times annualised H1 revenue by my estimate, so even after the price collapse the stock carries a large premium for several future years of exceptional growth.
• July's 11.35% new-share issuance, an in-the-money HK$335 convertible and the lock-up episode prove that per-share supply can expand sharply even when operating performance improves.
• The core model/API layer has low switching costs and intense competition from independent labs, open-weight models and hyperscalers able to subsidise AI from other profit pools.

The first pre-mortem script is a price war that defeats the margin thesis. Assume 2027–28 brings another wave of comparable open-weight coding and agent models from Zhipu, DeepSeek, Alibaba and others. API prices fall by another 40–50%. MiniMax still grows token volume strongly, but consolidated gross margin remains around 15–20% instead of rising toward the base-case 48% by 2030. R&D stays above US$600–800 million a year as multimodal training expands. Revenue reaches US$1 billion but cash burn forces another equity raise. The market stops paying a platform multiple and values MiniMax at 4–5 times sales. A share price around HK$60–120 becomes plausible even though the company has grown enormously.

The second script is technical and capital-market reflexivity. MiniMax spends heavily on a 2027 flagship generation that arrives behind rival models in coding and video. API growth drops below 80%, while Southbound flows reverse and a major shareholder lock-up creates additional supply. Gross margin falls below 15%, another financing is priced below market, and the diluted share count exceeds 400 million. A simultaneous revenue-estimate cut and multiple compression could remove 70–85% of today's equity value without requiring bankruptcy.

The key difference from the old report is now clear. I would not preserve its analysis unchanged. The prior revenue denominator has been decisively overtaken, the balance sheet is stronger, Open Platform has emerged much faster than expected, and the stock is 54.5% below the mid-May reference price. Those facts justify a less negative research stance than one based on HK$794 and US$79 million of annual sales. The new evidence does not justify treating HK$361 as inexpensive. My base-case economics generate only about HK$180 of present value, while a roughly HK$479 value requires an outcome I deliberately label optimistic.

For a balanced investor, MiniMax is worth following because there is now enough commercial evidence for the business to outrun a static bearish thesis. Ownership at today's price still requires underwriting a very high 2026–30 growth rate, major gross-margin expansion and sustained technical relevance at the same time. I would rather let the company prove one more leg of unit economics than pay today for most of that proof.

【Company-profile scores】

• Fundamental quality: medium
• Growth: high
• Moat: weak
• Financial soundness: medium
• Management credibility: medium
• Valuation attractiveness: low
• Risk level: high
• Suitable investor type: high-risk speculation

【Investment rating】

• Rating: Watch
• One-line thesis: Revenue has outrun the old thesis, but HK$361 still prices a 2030 outcome much closer to bull than base.
• Acceptable hold price: HK$155–205, corresponding to a band inside ±15% of the approximately HK$180 base-case present value.
• Clearly overvalued price: HK$530–600, beginning above 110% of the approximately HK$479 optimistic-case present value.
• Current-price classification: outside the three bands; above the acceptable-hold range but below the clearly-overvalued threshold.
• Whether to wait for a better price: yes. The price trigger is the ideal-buy band below, together with evidence that gross margin is at least 20% and rising and that platform revenue is still growing above 100%. Waiting risks missing a genuine model/platform breakout; paying HK$361 instead risks owning a base-case negative-return asset.
• Target holding horizon: 3–5 years.
• Expected annualized return: conservative approximately -24.6%; base approximately -2.5%; optimistic approximately +22.3%, measured to the 2030 scenario values.
• Max-loss risk: roughly 70–85% in a combined price-war, margin-stagnation and renewed-dilution scenario; the clearest trigger would be platform growth below 80–100% while gross margin remains below 15–20%.
• Reassessment-trigger signals: sustained gross margin above 30%; platform growth above 150% after the H1 2026 comparison becomes harder; R&D/revenue below 100%; positive operating cash flow without slowing model cadence; or, negatively, diluted shares above 400 million before the existing capital base has produced a margin inflection.

【Ideal Buy Price】40–48 HKD

Basis: the top of the range is 20% below the approximately HK$60 conservative present value; the lower end provides roughly one-third discount. A price decline into the band caused by deteriorating fundamentals would invalidate the mechanical buy signal.

【Valuation Range】

• current: 361.40 (close as of 2026-09-04)
• bear (conservative · ideal buy zone): [40, 48]
• base (fair · acceptable hold zone): [155, 205]
• bull (optimistic · above the clearly-overvalued line): [530, 600]

Sources and research uncertainties

Primary evidence carries the greatest weight in this report. The central sources are the December 31, 2025 HKEX prospectus; MiniMax's FY2025 results; the August 26, 2026 H1 results announcement; the July placement and convertible-bond disclosures embedded in that interim filing; and HKEX/share-price data. The prospectus does the most work here because it contains the corporate incorporation history, pre-IPO financing, WVR mechanics, revenue and user cohorts, geography, segment gross margins, inference-cloud costs and original Chapter 18C lock-ups.

MiniMax's FY2025 and H1 2026 disclosures provide the two critical post-IPO operating reads. FY2025 established US$79.0 million revenue, 25.4% gross margin, US$250.9 million adjusted loss and the greater-than-70% international mix. H1 2026 established US$116.6 million revenue, the platform mix inversion, 17.9% gross margin, US$296.9 million R&D expense, the June liquidity position and the July financing.

For the market narrative I use end-of-day historical-price data, contemporaneous reporting from SCMP, Reuters and Caixin, and official HKEX communications on Stock Connect/index access. Those sources establish the correct HK$345 first close, the lock-up overhang, July price/volume moves and August access-driven flows.

For technical-cost evidence I use MiniMax's own model releases because architecture, training-compute and API-tariff claims originate with the company. They should be treated as company claims unless reproduced independently. M1's compute comparison and tariff, Hailuo's architecture-efficiency claims and H3's commercial positioning are evidence of MiniMax's strategy, not proof of a durable moat.

For regulation, the primary source is the Cyberspace Administration of China's Interim Measures for Generative AI Services. For the risk-free comparison, I use the June 24, 2026 HKMA 10-year HKD HKSAR institutional-government-bond auction yield of 3.371%, the latest directly sourced 10-year HKD result in this research set rather than pretending it is a September 4 spot yield.

Five uncertainties remain material.

First, MiniMax does not disclose group-wide token throughput and audited inference cost per token. API tariff data cannot substitute for production cost. This is the largest unit-economics blind spot.

Second, H1 2026 does not provide the same product/platform gross-margin split that the prospectus provided for nine-month 2025. The fall in consolidated margin therefore cannot be cleanly decomposed between pricing, product mix, inference utilisation and other COGS.

Third, the roughly 45% July unlock is verified as shares becoming eligible for sale, not as 45% actually sold. I found no authoritative public tally identifying every beneficial seller after expiry.

Fourth, I did not independently recover a full-year FY2025 operating-cash-flow line sufficient to validate the prior report's approximately US$280 million figure. The verified prospectus figures are US$258.5 million outflow for FY2024 and US$209.4 million for nine-month 2025. I do not use the unverified FY2025 number in valuation.

Fifth, fully diluted shares will depend on convertible exercise, employee awards and future capital raising. The valuation uses approximately 369 million shares in its main cases to incorporate the current convertible mechanically, but it is not a forecast of the ultimate 2030 share count.

Other tickers mentioned

• 02513.HK: Z.AI Co., Ltd., formerly Knowledge Atlas Technology and better known as Zhipu AI, is the closest listed Chinese pure-play foundation-model peer and a key capital-market comparison.
• 00020.HK: SenseTime is an older listed AI-platform reference showing the scale and profitability challenge of large-model investment.
• 09988.HK: Alibaba is simultaneously a MiniMax investor, cloud supplier, API customer and foundation-model competitor.
• 00700.HK: Tencent is a pre-IPO backer and represents the hyperscaler-backed competitive model that can subsidise AI from larger ecosystem profits.
• 01024.HK: Kuaishou's Kling competes directly with Hailuo in generative video and benefits from an established creator/content distribution network.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.