Rating: Cautiously Bullish | Target Price: HKD 170–195 (base-case fair value range, 12–18 months) | Current Price: HKD 161.3 (close on 2026-10-30) | Margin of Safety: approx. +5% (vs. lower bound of base-case fair range; probability-weighted expected value approx. HKD 185.9, implying +15%)
FX basis: USD 1 = HKD 7.8439 (as of 2026-10-31). The Company reports in USD; amounts without a stated currency in this report are per financial statements (USD); valuations are uniformly converted to HKD.
MiniMax is one of the "dual champions" among China's independent LLM companies (listing around the same time as Zhipu), with H1 2026 revenue of USD 116.6 million (+283.1%), of which B-end API revenue was USD 73.9 million (+703.1%, 63.4% of total); half-year revenue already exceeded FY2025 full-year total, with the revenue surge validated by recognized revenue. However, the "quality" of growth carries three shadows: gross margin of 17.9%, nearly halved sequentially from ~33.7% in H2 2025; adjusted net loss widening 111% YoY to USD 293 million; and management's August ARR figure of "over USD 800 million" being an annualized-results-meeting figure (unaudited, with zero disclosure in company announcements). At the current price of HKD 161.3, market cap is HKD 56.86 billion and EV/ARR approx. 6.1x (Zhipu at 18.8x; Kling at ~36x in the primary market)—the market is effectively paying only "about 1.6x the annualization of recognized revenue"—valuation is reasonably low, but realization of the expectation gap hinges heavily on two milestones: quarterly revenue verification around December 2026 and the January 2027 share unlock (batch size estimates range from 37% to 62%). Cautiously bullish, confidence 0.55, awaiting verification signals for stepwise confirmation.
Event sequence over the next 6–18 months: ① Around December 2026—company voluntarily announces/discloses latest ARR and Q3 revenue via results briefing (first hard checkpoint: whether the year-end USD 1 billion ARR target is supported by recognized revenue); ② January 11, 2027—one-year post-listing share unlock (third-party estimates range from 37.4% to 61.8% of share capital; the October 2026 batch of ~18.65% has already become unlocked), with the actual sell-down rate determining the magnitude of supply shock; ③ H1 2027—disposition of the HKD 6.5 billion zero-coupon convertible bond (conversion price HKD 335, deeply out of the money; maturing 2027-07-14) and progress on the STAR Market A-share offering (CITIC Securities second phase of listing guidance completed); ④ March–April 2027—FY2026 annual report verifying full-year gross margin and the slope of loss narrowing.
Primary stock price driver: conversion rate from annualized ARR to recognized revenue—the single core variable behind the current valuation divergence.
Verifiable expectation gap: the "implied credible ARR" reverse-engineered from the current price at comparable multiples is approx. USD 260–390 million (depending on anchor choice), roughly half of management's August figure and about 1.6x the annualization of H1 recognized revenue; while verified sell-side target price ranges of HKD 533–1,118 (Jefferies lowered to 533 on 2026-10-30; J.P. Morgan Neutral at 260 in August; Morgan Stanley 900 / Goldman Sachs 860) all imply substantial re-rating. The gap cuts both ways: if ARR is verified, upside re-rating; if falsified, the valuation anchor collapses directly to the bear case.
Verification catalysts and falsification conditions: Verification—Q3 recognized revenue ≥ USD 180 million, the company disclosing ARR in announcement form for the first time, gross margin back above 20%; Falsification—Q3 recognized revenue < USD 140 million, FY2026 gross margin < 20%, large-scale sell-down filings after the unlock. Evidence grade for the opportunity is B (key growth metric ARR unaudited, zero disclosure in announcements); grade and real-time odds are computed by the system—this report makes no manual scoring.
Half-year revenue exceeded FY2025 full-year total, and the B-end share rose from 30.3% to 63.4% within a year, completing a revenue mix shift from "C-end companion apps" to "B-end API." The counterevidence must be faced squarely: prepayments and other receivables on the balance sheet grew from USD 16.3 million to USD 266 million in six months (approx. 2.3x half-year revenue, timing aligns with the ramp-up in compute purchases from Alibaba Cloud; composition undisclosed by the company); trade receivables DSO rose from ~41 days to 61 days; API prices were cut 50% one week after M3's launch (Bloomberg; Goldman Sachs lowered its target price accordingly); contract liabilities were only USD 18.3 million—a clear reconciliation gap versus the contract footprint implied by "USD 800 million ARR." This report's adverse scenario: token growth of 20x cannot map linearly into revenue amid price deflation; Q3 recognized revenue is the watershed.
Reverse-solving from Zhipu's current market cap/ARR of 18.8x, the current price implies credible ARR of approx. USD 386 million; but this figure is highly sensitive to the anchor—using EV it's approx. USD 260 million; using J.P. Morgan's 12x forward ARR for MiniMax it's approx. USD 600 million. This report therefore does not treat "implied ARR" as a precise value, only a directional conclusion: the market pays only for recognized revenue plus moderate growth-trust, in a "half-believing" zone. Re-rating requires two conditions, both indispensable: ARR validated by recognized revenue, and peer multiples stabilizing (Zhipu's placement price has fallen from HKD 1,588 to HKD 714; Jefferies cut its cloud business ARR multiple from 50x to 30x—the multiple anchor itself is still compressing).
MiniMax's positioning evidence chain is solid: M2 is the first Chinese model on OpenRouter to exceed 50 billion daily tokens; M3 has accumulated 489,000 unique users on OpenCode; enterprise clients/developers grew 10x in a year to 2 million+. But the same third-party forecasts hide a price term—IDC FutureScape forecasts the average delivery cost per agent call falling 87% by 2029 vs. 2025; revenue = volume × price, and the 50% price cut one week after M3's launch proves price erosion has already begun. There is also a structural gate: per MOFCOM/MOST Announcement No. 48 of 2026, model open-sourcing and overseas licensing require technology export registration (the company announced on 2026-10-23 that recent batches have been completed), creating regulatory preconditions for future overseas open-sourcing of M/H series.
Pro forma net cash approx. USD 2.38 billion (including July placement net proceeds of HKD 9.44 billion and CB net proceeds of HKD 6.43 billion, with CB face value counted as debt), approx. HKD 53.1 per share, 33% of current price—this is an asset-value floor, not operating value. At an annualized H1 burn of approx. USD 1.01 billion, the runway is approx. 2.4 years; if the CB requires cash repayment in July 2027 (conversion requires the stock price up +108% from current, low probability), the runway shortens to approx. 1.5 years. Solvency itself is not in question; the real risk is the resonance of "refinancing cadence × unlock supply": the 2026-10 batch (~18.65%) has matured; the 2027-01 batch ranges 37.4%–61.8% per market estimates (61.8% being the cumulative cap of key persons + cornerstone investors), plus the A-share issuance and the updated 20% general mandate—dilution in a severe scenario for existing shareholders is approx. 18.6%.
Beneath the headline of gross margin +5.8pp YoY, the mix is deteriorating: the B-end API share rising to 63.4% (approx. 80% on the August ARR basis) means "the larger the call volume, the higher the inference compute cost"; management gave no quantitative explanation for the sequential decline, and the company-stated "M2.7 inference margin above 40%" is unaudited. The base case requires gross margin climbing to 25%–32% by 2027—this is itself a key assumption, not an established fact; if FY2026 gross margin closes below 20%, the "scale → profitability" narrative will be interrupted, and exit multiples and scenario ranges must be revised down across the board.
| Metric (USD million) | FY2023 | FY2024 | FY2025 | H1 2026 |
|---|---|---|---|---|
| Revenue | 3.5 | 30.5 | 79.0 | 116.6 |
| Revenue YoY | — | +782.2% | +158.9% | +283.1% |
| Gross margin | — | 12.2% | 25.4% | 17.9% |
| Net loss attributable to owners | — | (465.2) | (1,871.6) | (358.0) |
| Adjusted net loss (Non-IFRS) | — | (244.2) | (250.9) | (293.0) |
| R&D expenses | — | 189.0 | 252.8 | 296.9 |
| Selling & distribution expenses | — | 87.0 | 51.9 | 27.0 |
| Net cash flow from operating activities | — | (258.5) | (279.6) | (318.1) |
| Free cash flow (calc: OCF − PP&E purchases) | — | — | (280.6) | (505.7) |
| Cash balance at period end | — | — | 1,050.3 | 1,322.8 |
| Interest-bearing bank borrowings | — | — | 35.5 | 133.6 |
| Gearing ratio | — | — | 343.3%* | 21.5% |
| Pro forma net cash (post-2026-07 financing) | — | — | — | approx. 2,384 |
* Negative net assets due to pre-IPO preferred shares classified as financial liabilities; post-IPO preferred shares converted to equity, with net assets turning positive to approx. USD 1.35 billion as of 2026-06-30. Metrics such as Net Debt/EBITDA and ROIC are not meaningful during the deep-loss period and are not presented.
Drivers of metric changes (YoY ≥ ±20% excerpts, all from company MD&A): ① Revenue +283.1%—"expansion of global client and user base, rapid growth in model inference demand, and improved conversion of model capabilities into products and services"; ② B-end revenue +703.1%—"growth in paying users and enterprise clients, increased API call volumes, and rapid adoption of Token plans"; ③ R&D expenses +138.8%—"continued development of foundation models and multimodal capabilities, increased cloud service expenses related to training"; ④ PP&E purchases rose from USD 0.38 million to USD 187.6 million—per the 8-26 announcement, "sustained and growing demand for high-performance cloud computing resources and related infrastructure for model training and upgrades" (the company thus shifting from rented compute to a self-built hybrid); ⑤ Interest-bearing borrowings +276%—"newly procured borrowings for working capital to support operations"; ⑥ Administrative expenses +103.7%—listing-related expenses (USD 5.7 million included in adjusted add-backs); ⑦ FY2025 attributable loss expansion of 302% mainly due to a USD 1.5899 billion fair-value remeasurement loss on preferred shares (converted to equity post-listing; this item fell to USD 31 million in H1 2026); ⑧ Prepayments and other receivables grew approx. 9x in six months to USD 165.4 million (current portion), trade payables +195% to USD 170.1 million—details undisclosed by the company; this report notes their timing aligns with the ramp-up in related-party compute purchases.
Revenue of USD 116.6 million (+283.1%) beat sell-side consensus (Citi-basis market expectation approx. +272%); Q2 +81.8% QoQ. Structurally, B-end API has become the largest revenue source (63.4%); overseas revenue was 60.8% (73.0% in 2025, domestic B-end ramping faster). Three divergences warrant vigilance: First, gross margin of 17.9% improved YoY but was nearly halved sequentially from ~33.7% in H2 2025; management attributed it to "improving infrastructure efficiency" without explaining the sequential decline; this report judges the mix effect from the jump in B-end share to be dominant; Second, adjusted net loss of USD 293 million widened 111% YoY—revenue grew 3.8x while losses widened, indicating the company remains in a "trading losses for scale" phase, with R&D expense ratio at 254.7%; Third, the interim report itself did not disclose ARR or operating KPIs such as MAU/paying users (the verbal results-meeting figure of August ARR exceeding USD 800 million is annualized data), so external verification can only rely indirectly on contract liabilities (USD 18.3 million, +142% YoY) and receivables/prepayments line items. Subsequent events: July completion of the placement and CB, combined net proceeds approx. HKD 15.88 billion, pro forma cash ammunition of approx. USD 3.35 billion magnitude.
Business model snapshot: A compute-intensive, omni-modal large model platform with a dual engine of "B-end open platform API/Token plans (63.4% of revenue) + C-end AI-native product subscriptions (36.6%)"; point-in-time revenue recognition accounts for 70.8% (API consumption and prepaid top-ups), so recurring attributes are weaker than typical SaaS. Upstream compute is mainly rented from Alibaba Cloud, with self-construction starting from 2026H1 (purchase of PP&E of US$187.6 million; period-end PP&E rose from US$1.6 million to US$74.9 million); downstream, the C-end has product pricing power, while the B-end is billed by usage and industry-wide price hikes have just emerged (2026Q2 China API average prices rose vs. 2025Q1 by +48% input/+80% output, per Morgan Stanley).
Cash content of earnings: During the loss-making period, all three ratios are negative — 2026H1 operating cash flow/net loss of 0.89 (cash outflows broadly in line with book losses), free cash flow/net loss of 1.41 (self-built compute makes cash burn far exceed book losses). OCF outflows have persistently run at 1.05–1.11x adjusted net loss (FY2024 through 2026H1); the company states it "mainly funds cash requirements through proceeds from financing activities": within 9 months of listing, IPO + placement + CB raised an aggregate net amount of approximately HK$21.17 billion. The adjusted figure itself is loose: add-backs consist of SBC (US$28.2 million) and changes in fair value of preference shares, but SBC is a recurring expense every year — after adjustments the company still lost US$293 million. Core operating losses are not window-dressed, but the company is nowhere near breakeven.
Return on capital and capex quality: ROIC is negative (NOPAT < 0, not meaningful). Capex/depreciation jumped from 0.34x in FY2025 to roughly 56x in 2026H1 — a shift from a pure cloud-rental asset-light model to self-built, asset-heavy compute, representing growth investment (training clusters and inference infrastructure); maintenance vs. growth capex is not separately disclosed, so no further determination is possible. The direction of investment intensity is consistent with the "self-controlled inference costs" strategy, but the pace at which depreciation flows into the cost line will weigh on gross margins over the next 2–4 quarters.
Moat and quality red flags: Moat candidates — omni-modal synergy (the company claims roughly 80% of resources are reused across modalities), extreme cost efficiency (cumulative total spend of approximately US$500 million from inception to Sep 2025, versus OpenAI's cumulative US$400–550 billion), high per-capita productivity with a 385-person team of which 73.8% are in R&D, and a globalized revenue mix (60.8% overseas). Red flags — ① cash burn relies entirely on external financing for life support; ② after listing, standard KPIs (MAU/paying users/ARPPU) ceased to be disclosed (the prospectus had period-by-period tables; they disappear from the annual report onward), with only non-comparable metrics such as "cumulatively served 236 million users" disclosed; ③ the top five suppliers account for 57.5% of purchases, with the single largest at 23.3% — Alibaba is simultaneously the second-largest shareholder (~12.65%), the largest compute supplier, and a model competitor — this triple entanglement amplifies bargaining-power and supply-continuity risks; ④ open platform spend per paying customer fell from US$27,020 → 12,454 → 6,167 (2023 → 2024 → first nine months of 2025, per the prospectus), with per-customer revenue continuously diluted and no evidence of pricing power.
Consistency of words and deeds: Verifiable commitments have largely been met — the prospectus projected meeting Chapter 18C commercialized-company revenue requirements (HK$250 million) in 2025; FY2025 revenue of US$79 million (~HK$620 million) confirms this was "achieved"; the prospectus flagged that cost of sales and R&D absolute amounts would continue rising in 2026, and H1 delivered +258%/+139% respectively. There is no formal revenue/profit guidance culture; forward-looking statements are limited to qualitative remarks, with no misses on record so far — assessed as "pragmatic but with minimal expectation management". 80% of the July placement + CB proceeds (~HK$12.7 billion) is committed to AI infrastructure and R&D by end-2027 (implying roughly US$92.6 million per month of additional R&D spending); still within the execution window and yet to be validated.
Shareholder friendliness: Assessed as "expansion-financing type, slightly below neutral on shareholder returns" — no dividends (no distributable reserves), no buybacks; three financings within 9 months of listing; the July placement price of HK$268 is 66% above the current price, and the CB conversion price of HK$335 is deep out of the money. If the A-share issuance + 20% general mandate + incentive schemes all land, base-case dilution is roughly 18.6%, with an extreme case of about 30%. Offsetting factors: founder Yan Junjie announced he will forgo compensation until AGI is achieved, and donated 4% of his personal stake for team incentives plus 1% to an open-source fund — the direction of interest alignment is correct.
Risk signals: Yan Junjie concurrently serves as Chairman/CEO/CTO (a voluntary deviation from Corporate Governance Code provision C.2.1); under the WVR structure, he and parties acting in concert collectively control ~78.6% of voting rights; related-party transaction annual caps were raised sharply within a year (Alibaba Cloud procurement from US$115 million → US$300 million); INED resignation and appointment in Aug 2026; core personnel lock-up expires on 2027-01-08. The suspension of KPI disclosure is the governance signal this report weighs most — revenue quality cannot be independently verified externally.
The company reports a single operating segment; the breakdown below is by revenue sub-line (Note 3) and geography:
| Business line | 2026H1 revenue | Share | YoY | One-line business logic |
|---|---|---|---|---|
| Open platform and enterprise services (B-end API/Token plans) | US$73.9 million | 63.4% | +703.1% | Token-metered inference demand; developer ecosystem + enterprise customer ramp |
| AI-native products (C-end: Hailuo AI/Xingye/Talkie/Agent) | US$42.6 million | 36.6% | +100.9% | Subscriptions + in-app purchases; organic growth-driven, sales expense ratio declining with scale |
| Of which: Overseas / Mainland China | US$70.9 / 45.7 million | 60.8% / 39.2% | Overseas share 73.0% → 60.8% | Coverage of 230+ countries and regions; domestic B-end growing faster |
Segment gross profit is not disclosed. Per the prospectus (first nine months of 2025), open platform gross margin was 69.4% — B-end unit economics are significantly better than C-end, explaining why overall gross margin once rose to 33.7% (2025H2) before the mix shift; however, overall gross margin in 2026H1 fell instead to 17.9%, showing that once B-end call volumes scaled exponentially, inference compute costs (including related-party purchases from Alibaba Cloud) ate into the high-margin advantage. "The higher the revenue, the higher the cost" accurately describes the current gross margin bottleneck. Judgment on the profit engine: on the first-nine-months-2025 basis, B-end was the gross profit mainstay; from 2026H1, B-end volume growth has diluted per-call gross margin, and the race between these two curves will determine the profitability inflection point. The gross margin structure gap (B-end modeled margin of 69% vs. actual overall 17.9%) exceeds 10pp — essentially, the "model vendor's ability to pass through compute costs" remains unproven.
Accounting red flags: ① Prepayments, other receivables and other assets (current) rose ~9x in half a year, from US$16.3 million to US$165.4 million; the interim report does not disclose composition or counterparties, and the timing coincides with the ramp-up of related-party compute procurement (severity: medium); ② Disney/Universal/Warner IP infringement litigation (motion to dismiss denied in May 2026, proceeding to substantive trial) with no provision recognized (severity: low — the outcome is indeed not estimable, but the combination of "zero provision + 60.8% overseas revenue" warrants monitoring).
Cross-period consistency: ① Gross margin of 12.2% → 25.4% → 17.9% shows two swings in opposite directions, each attributed to "efficiency improvements" but with explanations that do not close the loop (no quantified attribution for the 7.5pp decline in 2026H1 vs. full-year 2025) — inconsistent; ② MAU/paying users/ARPPU were disclosed period-by-period in the prospectus, then suspended post-listing without explanation — unexplained by the company; ③ Alibaba Cloud related-party procurement of US$3.1 million → 10 million → 75.9 million → 75.6 million (2023 → 2026H1), with consecutive doublings and management's attribution (high-performance compute demand) self-consistent with procurement at roughly 30% of R&D spending — consistent; ④ OCF outflows persistently at 1.05–1.11x adjusted losses, with the financing-funded model consistently described across periods — consistent. The company has been listed for fewer than three full disclosure periods, so the cross-period sample is limited; the above conclusions await verification in the FY2026 annual report.
Current market data (as of close on 2026-10-30): Share price HK$161.3; total shares 352,480,183 (Class A 271.4 million + Class B 81.1 million, excluding 19.4 million potential CB dilution shares); market cap HK$56.86 billion (~US$7.25 billion); EV approximately HK$38.15 billion (~US$4.86 billion). Multiples: PS-TTM 43.9x (TTM revenue US$165.2 million; listed only 9 months, no historical percentile significance — percentile for reference only); PB 5.38x (2nd percentile since listing — net assets are mostly cash, limited reference value); the core of current pricing for this name is EV/ARR and forward EV/Sales: EV/ARR approximately 6.1x (denominator is management's August weekly-annualized figure of US$800 million, unaudited), EV/2027E revenue approximately 6.2x (base of US$750 million). Average daily turnover approximately HK$1.4 billion (HK$1.4 billion on 10-30 alone), ~2.4% of market cap; August southbound net buying of HK$10.05 billion — liquidity is ample, no low-liquidity flag.
What expectations the market has already priced in: Back-solving the current price (using Zhipu's market cap/ARR of 18.8x as an external anchor) implies a credible ARR of approximately US$386 million — 48% of management's figure and 1.66x the annualized H1 recognized revenue (~US$230 million). In one sentence: the current price requires MiniMax to prove that roughly half of the US$800 million weekly-annualized ARR can settle into sustained monthly revenue, while management and most sell-side analysts believe monthly-annualized will reach US$1 billion by year-end. This back-solve is anchor-sensitive (EV basis: US$260 million; J.P. Morgan 12x basis: US$600 million); this report only draws the qualitative conclusion of "half-believing, half-doubting". Verified sell-side distribution: median of 6 Buy/Overweight ratings at HK$690 (as of 09-10) is stale — Jefferies cut to 533 on 10-30, J.P. Morgan Neutral 260 on 8-17; sell-side targets overall remain in a rapid downward-revision channel, and this report does not adopt the sell-side median as the market anchor.
Three-tier value (EPV): Asset value (floor) = pro forma net cash of approximately US$2.38 billion ≈ HK$53.1/share (33% of current price), of which HK$1.20 billion comes from the HK$268 placement and HK$830 million corresponds to CB creditors (if cash-settled, the per-share floor drops to approximately HK$29.5). EPV (zero-growth sustainable cash flow): not estimable — current sustainable operating cash flow is deeply negative (R&D at 2.5x revenue, heavy compute capex); any "normalization" requires first assuming growth, which is circular input; this report refuses to fabricate. The current price exceeds the asset floor by approximately HK$108 (67%), which is the growth option the market is paying for; since EPV is not estimable, the growth option share cannot be decomposed on an "EPV + residual" basis, and a scenario distribution is used instead.
Three scenarios and odds (12–18 months):
| Scenario | Probability | Fair value range | vs. current price | Key drivers |
|---|---|---|---|---|
| Bear | 0.28 | HK$58–68 (midpoint 63) | −61% | ARR falsified (Q3 recognized revenue <US$140 million) + gross margin stagnant at 15–18% + CB cash settlement + lock-up expiry selling pressure; 2027E revenue ~US$450 million × 6x PS (anchor: SenseTime 7.9x / 4Paradigm 1.4x, mid-band with further discount) + net cash of ~HK$1.0 billion at the time |
| Base | 0.50 | HK$170–195 (midpoint 182.5) | +5% to +21% | Year-end ARR monthly-annualization partially delivered (FY2026 recognized revenue US$420–490 million), 2027E revenue US$700–800 million (~+65%), gross margin climbing to 25–32%; exit at 9–10x PS (2027E) (anchor: 50–60% discount to Zhipu's current 18.8x market cap/ARR — reflecting compression in its own multiple: placement price HK$1,588 → 714, Jefferies cutting its cloud ARR multiple from 50x → 30x) + net cash of HK$4.0–5.0 billion at the time |
| Bull | 0.22 | HK$330–370 (midpoint 350) | +105% to +129% | M3 Pro/Hailuo 3.0 driving ARR to US$1.8–2.0 billion by end-2027, 2027E revenue US$1.00–1.05 billion, gross margin 35–40%; exit at ~13x PS (2027E) (anchor: ~70% of Zhipu's multiple, lower end of Anthropic's primary-market 15–31x ARR range) |
Probability-weighted expectation of approximately HK$185.9 (+15.2%); margin of safety (base-case floor of 170 vs. current price) +5.4%; distribution is positively skewed but with a deep tail (−61% vs. +117%). The external bearish anchor (J.P. Morgan 2026-08-17 Neutral HK$260, on a 12x expected ARR basis) is significantly above this report's bear case — the difference lies in premises: its bear path is "multiple compression," while this report's bear case is "ARR itself falsified" (numerator and denominator both hit); the two are not directly comparable, and this report does not adopt it as the lower bound. The external bullish anchor (Morgan Stanley 900 / Goldman Sachs 860, Aug 2026) is already below 2.3x this report's bull-case ceiling — sell-side bullish anchors are likewise decaying, and this report's bull case is constructed independently of them.
In-house earnings forecast (results reconciliation anchor): FY2026E revenue US$420–490 million (midpoint US$455 million, implying Q3 ~US$130–160 million, Q4 ~US$180–220 million), adjusted net loss of US$620–680 million; FY2027E revenue US$700–800 million, adjusted net loss narrowing to US$400–550 million. Method: weekly-annualized ARR converted to monthly-annualized at 20–25% (referencing Zhipu's same-month monthly/weekly ratio), with Q3/Q4 sequential growth naturally decelerating; cross-checked against management's guidance (year-end ARR US$1 billion) and Citi's FY2026E (US$577 million) — this report's forecast is below management's path and slightly below Citi's.
Conclusion: Reasonable but on the cheap side. Assess quality and price separately: the quality is that of a "growth stock under validation" (genuinely explosive revenue, extremely high per-capita productivity, thick cash — but gross margin of 17.9% is far below Zhipu's 26.4% and SenseTime's 41%, losses not narrowing, and governance transparency discounted); on price, the current price is at a 40% discount to the July placement price of HK$268, 2.2% below the IPO price of HK$165, and down 88% from the March peak of HK$1,330 — the market has already paid a substantial discount for uncertainty, but the discount is not yet sufficient to cover the falsification tail. This is why the stance is cautiously bullish rather than firmly bullish.
Industry definition: MiniMax operates simultaneously in the "foundation model layer + AI-native application layer" — with self-developed omni-modal foundation models (text M series, video Hailuo/H series, speech, music), selling subscriptions to C-end users and API/MaaS to B-end customers. This differs from AI solution deliverers (SenseTime, Fourth Paradigm, Chumen Wenwen) and cloud vendors' in-house models (Alibaba Tongyi, ByteDance Doubao, Tencent Hunyuan — both competitors and upstream compute suppliers).
Industry size: Bloomberg Intelligence estimates the global generative AI market will reach USD 2.3 trillion by 2032 (of which inference accounts for USD 1.3 trillion, CAGR 32%); J.P. Morgan's estimate puts the global AI market at USD 1.4 trillion by 2030; IDC × Inspur estimates the global AI compute market at USD 1.25 trillion by 2030. For China: IDC/CICC estimates the large model market at RMB 14.134 billion in 2023 (+83.9%) → approximately RMB 49.5 billion in 2025; CITIC Securities forecasts China's AI market at RMB 993.0 billion by 2030 (2024–2030 CAGR 35.5%); CIC Consulting estimates China's enterprise-grade AI agent solutions market at RMB 5.6 billion in 2024 → RMB 59.1 billion by 2029 (CAGR 60.2%). Head companies' revenues confirm the explosion: MiniMax's 2024/2025 revenue grew +782%/+159%, Zhipu's 2026H1 grew +399.7%.
Value chain and value distribution: Upstream compute is the largest cost item — NVIDIA's Vera Rubin platform enters production in early 2026, with inference token costs dropping to about 1/10 of the prior generation; cloud vendors' capex arms race (Alibaba Cloud committing at least RMB 380 billion to AI infrastructure over the next three years; Kuaishou approximately RMB 26 billion in 2026). MiniMax has no large self-built compute stock (cumulative training costs of approximately USD 500 million from founding to 2025-09, roughly 1% of OpenAI's cumulative USD 400–550 billion), moving to a hybrid self-built + rented model from 2026H1. Midstream model layer: training costs remain high while inference costs fall rapidly (open-source models disrupting closed-source pricing). Downstream: C-end subscription tiers at RMB 49–700/month; B-end billed by token. Current state of value distribution: gross margin nominally stays with B-end APIs (MiniMax open platform gross margin of 69.4%, per the prospectus 9M2025 figures), but inference compute costs inflate in tandem with call volumes — for the model layer, "the higher the revenue, the higher the compute cost"; benchmarked against Palantir's approximately 80% gross margin, value ultimately settles in the application/platform layer, and model developers must win through cost efficiency — this is precisely the industry logic behind MiniMax's "extreme efficiency" strategy.
Supply-demand and competitive structure: On the demand side, Agentic AI/AI coding has become the primary engine of token consumption — IDC forecasts global active agents growing from 79.4 million in 2026 → 2.216 billion in 2030 (CAGR 129.8%); MiniMax's own July token consumption was 20x that of January. The supply side has a dual structure: compute is in tight balance (IDC estimates global AI compute demand fulfillment rate falling from 79% in 2024 → trough of 71% in 2027, with the absolute gap widening to USD 380.9 billion by 2030); model capability is homogeneously oversupplied — after the brutal 2024–2025 price war (Doubao at RMB 0.0008/thousand tokens; Baidu/iFlytek free), 2026 enters a new phase of tiered pricing and flagship price increases (Morgan Stanley: 2026Q2 China large model API average input price RMB 4.9/million tokens, output RMB 21.9, up +48%/+80% respectively vs 2025Q1; Zhipu's call volume rose 400% after an 83% price increase — demand price elasticity lower than expected). Concentration: global CR5 of the foundation model layer is approximately 58% (third-party estimate, reliability unverified, for cross-reference only); China's C-end AI apps are dominated by big tech (Doubao/DeepSeek/Yuanbao); MiniMax explicitly avoids head-on competition with big tech in general-purpose conversational assistants, focusing instead on companion/video generation/Agent verticals. Barriers to entry: capital threshold starting at the USD 1 billion level, compute access, talent (73.8% of MiniMax's 385 employees are in R&D), data flywheels and developer ecosystems (over 2 million enterprise customers/developers by 2026H1, 10x vs end-2025), and regulatory filing. Substitution threats: open-source weight self-deployment (DeepSeek/Tongyi; Alibaba plans to charge revenue sharing to large commercial customers from 2026 — the era of free open-source models is drawing to a close), and cloud vendors' MaaS cross-subsidization.
Cycle and regulation: The industry is in a "commercialization validation + consolidation phase" — the IPOs of Hong Kong's two large model champions provide public pricing anchors, API prices are turning from decline to rise, leading companies' loss ratios are narrowing (SenseTime's EBITDA turned positive in 2025H2), and late-stage fundraising is receding; meanwhile there is a primary-secondary valuation inversion (primary markets price on scarcity, secondary on revenue). Leading indicators: token call volumes (DeepSeek's daily peak of 8 trillion topping OpenCode), ARR (Zhipu USD 1.6 billion in August, Kling approximately USD 500 million in 2026-03, MiniMax over USD 800 million in August), big tech AI capex, API price month-on-month changes, compute gaps, and model leaderboard cadence. Regulation: domestically, the dual-filing system under the Interim Measures for the Management of Generative AI Services, the Measures for Labeling AI-Generated Synthetic Content (effective 2025-09), and the "AI+" action opinions (demand-side positive); overseas, the FTC's 2025-09 6(b) investigation into AI companion products at seven companies (excluding MiniMax/Talkie, but regulatory tightening across the same track) and the EU AI Act's GPAI obligations effective 2025-08; MOFCOM/MOST Announcement No. 48 (2026) requiring technical export registration for model open-sourcing and overseas licensing (the company announced on 10-23 that recent batches have been registered) — this is a compliance gate, not a ban, for MiniMax's H3 open-source overseas strategy.
Quantified chain of the demand inflection point (this report judges the industry to be at a structural demand inflection): driver trajectory (global active agents 2026–2030 CAGR 129.8%) × unit usage (an Agent consumes tens to hundreds of times more tokens per task than a single-turn conversation; MiniMax M2.5 reduces the cost of "4 Agents working continuously for one year" to USD 10,000 — falling costs unlock usage) × penetration ramp (enterprise customers from 214,000 → 2 million+, roughly 10x in a year) × company share (M2 was the first Chinese model to consume 50 billion tokens daily on OpenRouter; M3 accumulated 489,000 unique users/8.1 trillion tokens on OpenCode — share trend rising but absolute share lags DeepSeek). Two-way test: the upward volume is verified by both third-party and company figures; the downward price is equally real — IDC FutureScape predicts per-call cost falling 87% by 2029 vs 2025, M3 cut prices 50% within a week of launch, and GPT-6 API prices halved — the growth implied at current prices (approximately 1.6x annualized recognized revenue) falls within the supportable range of the net effect of "volume up × price collapse" and is not overextended; however, the sell-side median of HKD 690 implying "volume and price both rising" lacks evidence on the price side.
Peer benchmarking table (unified framework of this report):
| Company | Revenue Scale | Revenue Growth | Gross Margin | Market Cap/Valuation | Core Differences vs MiniMax |
|---|---|---|---|---|---|
| MiniMax (00100.HK) | 2026H1 USD 116.6 million | +283.1% | 17.9% | HKD 56.86 billion (10-30 close) | Reference base: overseas revenue 60.8%, B-end 63.4%, market cap/ARR approximately 9.1x |
| Zhipu (2513.HK) | 2026H1 approximately USD 133 million | +399.7% | 26.4% | HKD 235.8 billion (10-30) | Most direct comparable: mainly domestic B/G-end, market cap 4.1x MiniMax's, ARR 2x, market cap/ARR 18.8x; but October placement price cut from HKD 1,588 to 714 — its own valuation is deflating |
| SenseTime (0020.HK) | 2025 RMB 5.015 billion | +32.9% | 41.0% | HKD 46.69 billion | Heavy-asset self-built compute + government/enterprise delivery, representing the mature state of "having crossed breakeven" (PS approximately 7.9x) |
| Fourth Paradigm (6682.HK) | 2025 approximately RMB 7.1 billion | around +40% | 37.7% | HKD 12.05 billion | Near-profitable enterprise AI platform, PS approximately 1.4x — a valuation floor reference after growth deceleration |
| Kling AI (Kuaishou segment, primary market) | 2026Q1 single quarter over RMB 650 million | +300%+ | Not disclosed | USD 18 billion post-money in independent financing | Global leader in vertical video generation, ARR approximately USD 500 million, approximately 36x ARR — primary-market single-round figures, no liquidity adjustment, directional reference only |
| Palantir (PLTR.US) | FY2025 guidance approximately USD 4.4 billion | +53% | approximately 80% | approximately USD 455 billion | Profitable AI platform as an endgame reference (PS approximately 108x) — 3–5 years ahead of model companies' current state; not directly comparable but indicates where value settles |
Company positioning in the industry: one of China's two independent large model "champions" and the chaser behind Zhipu — 2026H1 revenue scale is comparable to Zhipu's (USD 116.6 million vs approximately USD 133 million) and Q2 QoQ +81.8% is accelerating, but market cap (HKD 56.86 billion vs 235.8 billion) and ARR (USD 800 million vs 1.6 billion) clearly lag. Share trend rising. Side-by-side caliber note: the company claims to be "one of only four global omni-modal first-tier players"; the third-party aggregated caliber (Artificial Analysis Intelligence Index, 2026-09, low reliability unverified) shows its flagship M3 index of 29.2 lagging Tongyi Qwen3.8 Max (45.4) and Zhipu GLM-5.3 (44.8); an intermediate caliber (Artificial Analysis open-source model first tier; M2.5's SWE-Bench Verified record per company figures) indicates its text models are first-tier competitive in coding/Agent scenarios, though not first in China on general intelligence leaderboards. Sources of moat: omni-modal synergy, cost efficiency, high per-capita productivity, globalized product matrix, and developer ecosystem; what remains unproven is pricing power (spending per paying customer continuously declining, M3 limited-time and permanent discounts).
The generic financial framework for large model companies misses four decisive dimensions; this chapter addresses them separately:
① ARR caliber discipline. The company-reported August ARR of "over USD 800 million" is a weekly-annualized figure disclosed at the earnings call (a single week's revenue × 52, confirmed by management post-call), unaudited and with zero disclosure in company announcements (including the 10-23 business update); it corresponds to a monthly average of approximately USD 66.7 million, 2.66x Q2's monthly average (approximately USD 25.1 million). Contract liabilities on the balance sheet are only USD 18.3 million (+142% YoY — a positive leading indicator but small in absolute terms), a significant reconciliation gap against the "USD 800 million annual recurring revenue." This report uniformly applies a 20–25% haircut for monthly conversion and sets validation thresholds accordingly: Q3 recognized revenue ≥ USD 180 million = validated; < USD 140 million = falsified.
② Token volume-price scissors gap. Volume: July token consumption was 20x January's; the M2 series' February daily average tokens were 6x December 2025's (Coding Plan +10x). Price: M3 cut prices 50% within a week of launch (blended pricing approximately USD 0.22/million tokens); DeepSeek's price cut and GPT-6 API halving in September 2026. The true slope of revenue = net volume-price effect; the industry API average price turning up QoQ in 2026Q2 (+48%/+80%) is a positive signal, but MiniMax's own "spending per paying customer" has declined for three consecutive years, and the company has yet to prove pricing power.
③ Open-source funnel and export gate. H3 achieved 24 million+ downloads in three weeks of open-sourcing, 300+ derivative models, adaptation across 16 chip platforms, and NVIDIA launching a 4K video enhancement solution based on H3 — open-source customer acquisition → API/subscription conversion is a primary engine of the ARR ladder (after the M3 launch, ARR jumped from USD 430 million in April to 800 million+ in August). If the 24 million downloads convert at 0.5–1% with enterprise ARPU of USD 10,000–50,000/year, that implies annualized revenue potential of USD 100–500 million (extrapolated). Gate: technical export registration under MOFCOM Announcement No. 48 has been completed for recent batches; subsequent overseas open-sourcing of models still requires batch-by-batch registration.
④ Binding of compute costs to financing cadence. Related-party purchases from Alibaba Cloud rose from USD 3.1 million → 10 million → 75.9 million → 75.6 million over 2023–2026H1, with the 2026 annual cap raised to USD 300 million (USD 400/500 million for 2027/2028, USD 1.2 billion over three years); self-built compute capex over the same period was USD 187.6 million. Compute commitments (rigid spending) + R&D (planned monthly increase of USD 92.6 million) mean cash burn can only rise, not fall — the financing cadence (A-share listing, new placements, CB disposal) is an endogenous valuation variable, not exogenous noise; each round's discount and dilution directly rewrites fair value per share.
Stance: Cautiously bullish, confidence 0.55, time horizon 12–18 months. Valuation judged "reasonable to slightly undervalued," baseline fair value HKD 170–195, probability-weighted expectation approximately HKD 185.9 (+15.2%), margin of safety +5.4% (vs baseline lower bound) — a positively skewed distribution (bear −61%/baseline +5%~+21%/bull +105%~+129%) but with a deep falsification tail, plus two high-volatility events (December validation, January lock-up expiry) concentrated within the next three months — hence a cautious rather than resolute bullish posture.
Strategy statement: This report's positioning logic is in two steps — before validation, express the "implied expectations too low" view with a small position, adding after Q3 recognized revenue ≥ USD 140 million and if actual lock-up expiry selling proves moderate; if Q3 falsifies or large disposal filings emerge after the lock-up expiry, exit and await repricing near the asset floor (approximately HKD 53, or approximately HKD 29.5 in the extreme case after further CB repayment deduction). Upside trigger: the share price entering the baseline fair value range (above HKD 170) confirms repair has begun; above HKD 195 the margin of safety is exhausted — evaluate taking profit.
Risk and catalyst alerts (written summary): Risks — ARR weekly-annualized caliber falsification, gross margin stagnating below 20%, resonance of CB repayment (2027-07) with lock-up supply, industry price war reigniting, IP litigation injunction tail risk (no provision accrued), Alibaba-related triple-binding supply and governance risks, transparency discount from suspended KPI disclosure. Catalysts — Q3/ARR disclosure (2026-12), M3 Pro launch and open-sourcing (2026Q4–2027Q1), H3 ecosystem commercialization (MiniMax Design subscription), A-share listing tutoring acceptance and filing (2027H1), Alibaba API order ramp (2027 related revenue cap USD 22 million, +193% vs 2026), CB-guided conversion or low-cost refinancing. The machine-readable monitoring checklist for the above conditions has been saved alongside this report; re-review upon any trigger.
This report is based on publicly disclosed documents and public news; the price anchor is the 2026-10-30 close of HKD 161.3 (frozen caliber); earnings-call figures such as ARR are unaudited — please refer to subsequent company announcements.