ARR is ramping rapidly, but higher R&D investment is widening losses; Goldman Sachs remains optimistic about MiniMax's valuation recovery potential in 2H26

Overview

The report reviews MiniMax's 1H26 results and earnings call. Its core conclusion is that ARR growth was stronger than expected, while model inference efficiency, multimodal capabilities, and commercial adoption continued to improve. However, the company is significantly increasing R&D investment to pursue the frontier of model intelligence and cost efficiency, which will widen near-term losses. Goldman Sachs still believes these investments could create room for a valuation recovery in 2H26 and maintains its Buy rating.

Core views

The report summarizes MiniMax's 1H26 results around three main themes: stronger-than-expected ARR ramp-up, improving gross margin albeit more slowly than expected, and a significant increase in R&D investment. Goldman Sachs believes MiniMax is concentrating resources on pre-training, post-training, and inference infrastructure to improve both model intelligence and unit-cost efficiency. This will widen near-term losses but could strengthen model competitiveness, commercialization scale, and the potential for a subsequent valuation recovery. Revenue and ARR represent the clearest positive changes. ARR exceeded US$800mn in August 2026, up from US$150mn in February and over US$400mn in early June; July Token consumption reached 20 times the January level, while enterprise and developer customers exceeded 2 million, 10 times the level at the end of 2025. Open Platform contributed 63% of 1H26 revenue, up from 30% in 1H25. Management primarily attributed the growth to improved model intelligence and attractive cost efficiency. In particular, usage exhibited step-function rather than linear growth following the launch of M3. Management remains highly confident in its US$1bn year-end ARR target, while Goldman Sachs raised its forecast from US$1bn to US$1.2bn and increased its 2H26/FY26E revenue forecasts from US$373mn/US$493mn to US$400mn/approximately US$517mn. FY26E/FY27E/FY28E revenue forecasts were raised by 5%/38%/29%, respectively, to US$516.8mn/US$1,752.9mn/US$3,876.8mn. Subsequent growth will depend on iterations including M3.1, the 3T-parameter M3 Pro, and H3.1. The second theme is the performance-price Pareto frontier. Management believes inference efficiency not only lowers service costs but also determines the affordable scale of post-training, thereby affecting model capabilities. With architectural innovations such as MiniMax Sparse Attention and infrastructure optimization, M3 achieved API gross margins broadly comparable to or even higher than M2 despite increasing its parameter count from M2's 230B to 430B. MSA 2.0 is expected to further improve the cost efficiency of scaling model intelligence. However, 1H26 gross margin was below expectations due to factors including a revenue mix shift toward lower-margin text models and one-time compensation for Token subscription plans triggered by pricing adjustments after the launch of M3. Goldman Sachs lowered its 2H26/FY26E gross profit forecasts from US$107mn/US$149mn to US$93mn/US$114mn, expecting subsequent improvement to depend on inference optimization, higher cluster utilization, and the pricing power of more intelligent models such as M3 Pro. The third theme is computing capacity and R&D investment. MiniMax uses a diversified computing network comprising self-operated clusters, partnerships with cloud service providers, and Token Factory, which can scale rapidly during demand peaks. It completed cluster, network, and software upgrades in 1H26, bringing effective training utilization to 97%. Overall computing supply growth continues to lag downstream Token demand, but existing capabilities are sufficient to support the launches of M3.1, M3 Pro, and H3.1. Domestic chips could also reduce unit costs and enhance supply resilience in the future. Following its recent financing, the company holds more than US$3bn in cash reserves. Nevertheless, Goldman Sachs raised its 2H26/FY26E R&D expense forecasts from US$331mn/US$531mn to US$512mn/US$809mn. Annualized R&D investment in 2H26 is approximately US$1bn, above the approximately US$600mn annualized level in 1H26. As a result, the 2H26/FY26E adjusted net loss forecasts widened from US$268mn/US$436mn to US$436mn/US$735mn, while FY26E/FY27E/FY28E adjusted net profit forecasts were revised by -69%/-26%/-8%. The report estimates 1H26 free cash outflow at US$521mn, highlighting the clear trade-off between model expansion and the path to profitability. Multimodality is the fourth growth theme. H3 has reached the cost-efficiency frontier among leading video generation models and ranked near the top of the image-to-video Arena leaderboard as of August 15, 2026. Following the model's August launch, its ARR contribution began to accelerate. The open-weight strategy drove global downloads above 24 million, helping expand developer adoption and ecosystem participation. Overseas revenue accounted for 61% of 1H26 revenue, which Goldman Sachs believes provides a distribution advantage for expanding multimodal capabilities into international markets. MiniMax Design uses agent orchestration for content-generation workflows, seeking to lower the barriers to using and the total cost of commercial content production while converting H3's cost efficiency into scalable production use cases. The fifth theme is the agent and application orchestration layer. MiniMax Agent targets general-purpose workspaces, MiniMax Code targets long-horizon complex programming tasks, and MiniMax Design targets end-to-end multimodal commercial content creation. Against the backdrop of major internet platforms such as Tencent WorkBuddy, Alibaba QwenWork, and ByteDance Doubao Work integrating agent entry points, MiniMax faces their integrated advantages across models, products, and distribution, while also benefiting from its role as an important model supplier to these model-neutral workspaces. Goldman Sachs believes the application orchestration layer can provide a more direct commercialization path and accumulate proprietary workflow data to support model and product iteration. However, whether MiniMax can establish sustained retention and pricing power beyond temporary model or product leadership remains critical. On valuation, Goldman Sachs maintains its Buy rating and lowers its 12-month DCF-based target price from HK$800 to HK$760, implying 150.8% upside from the current price of HK$303. The DCF uses a 12% weighted average cost of capital and a 2% terminal growth rate, with bull- and bear-case valuations of HK$1,300 and HK$220, respectively. Based on the report's figures, MiniMax has a market capitalization of approximately HK$100.4bn or US$12.8bn and an enterprise value of approximately HK$80.4bn or US$10.3bn. Its approximately US$13bn market capitalization is below video-generation-model peer Kling's US$18bn post-money valuation and Z.ai/Zhipu's approximately US$61bn market capitalization, while its end-2026 ARR forecast implies a price/ARR multiple of approximately 10 times based on the previous closing price. Goldman Sachs therefore believes the valuation distribution remains skewed to the upside, provided that continued model iterations drive ARR, cost efficiency, and commercialization capabilities. The principal risks explicitly identified in the report include weaker-than-expected model performance amid global foundation-model competition, slower improvement in earnings visibility, weaker commercialization capabilities, intellectual property and content-generation issues, cash consumption and self-financing capacity, and geopolitical risks arising from intensifying China-US technology competition.

Analysis framework

Goldman Sachs first divides MiniMax's actual 1H26 results and earnings-call information into five themes: ARR and revenue, performance-price efficiency, computing capacity and R&D, multimodality, and the agent application layer. It then uses Token demand, customer count, revenue mix, model parameters, inference efficiency, and computing utilization to explain operating changes. Based on this analysis, it revises revenue, gross profit, R&D expense, and net loss forecasts, and then assesses the target price and potential valuation recovery using DCF, bull/bear scenarios, and peer market-capitalization comparisons.

Methodology notes

• Valuation methodologyDCF discounted cash flow

  DCF valuation

  The report discounts the company's future cash flows to present value and applies a 12% weighted average cost of capital and a 2% terminal growth rate to derive a 12-month target price of HK$760.

• Competition and strategy framework

  Performance-price Pareto frontier

  The report evaluates both model intelligence and inference costs to determine whether architectural innovation and infrastructure optimization can enable MiniMax to expand model capabilities and commercial applications without disproportionately increasing costs.

• Company fundamentals and financial framework

  ARR- and Token-demand-driven revenue forecasting

  Goldman Sachs combines changes in ARR, Token consumption, customer count, Open Platform revenue contribution, and upcoming model launches to derive its 2H26 and FY26-FY28 revenue forecasts.

• Valuation methodology

  Bull-, base-, and bear-case valuations and peer market-capitalization comparisons

  The report presents outcomes of HK$1,300/HK$760/HK$220 under different scenarios and compares MiniMax's market capitalization with Kling and Z.ai/Zhipu to illustrate the potential range and uncertainty of a valuation recovery.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• MiniMax Group (0100.HK)

  The report believes the company benefits from growing demand for foundation models, multimodality, and agents, but its near-term results will face pressure from rising R&D expenses and widening losses.

  Strengths

  Comprehensive multimodal product coverage, relatively strong commercialization capabilities, an advantage in per-Token costs, high organizational efficiency, a relatively high overseas revenue contribution, and more than US$3bn in cash reserves.

  Weaknesses

  A rising contribution from text models weighs on gross margin, large-scale pre-training, post-training, and inference investment widens losses, and free cash flow remains negative.

  Comparison

  Its approximately US$13bn market capitalization is below Kling's US$18bn post-money valuation and Z.ai/Zhipu's approximately US$61bn market capitalization; it also faces competition from the integrated model, product, and distribution ecosystems of major internet companies.

  Risks

  Model performance or commercialization may fall short of expectations, the path to profitability may be delayed, intellectual property and content-generation risks may arise, cash consumption and self-financing pressure may increase, and geopolitical risks related to China-US technology competition may intensify.

Key data

• August 2026 ARROver US$800mnUp from US$150mn in February 2026 and over US$400mn in early June
• End-2026 ARR forecastUS$1.2bnGoldman Sachs previously forecast US$1bn; management's target is US$1bn
• Token consumptionJuly was 20 times the January levelReflects a significant acceleration in model usage
• Enterprise and developer customersOver 2 million10 times the level at the end of 2025
• Open Platform revenue contribution63% in 1H2630% in 1H25
• FY26E/FY27E/FY28E revenueUS$516.8mn/US$1,752.9mn/US$3,876.8mnPreviously US$493.1mn/US$1,271.1mn/US$3,004.8mn, raised by 5%/38%/29%, respectively
• 2H26/FY26E gross profitUS$93mn/US$114mnPreviously US$107mn/US$149mn
• 2H26/FY26E R&D expensesUS$512mn/US$809mnPreviously US$331mn/US$531mn
• 2H26/FY26E adjusted net lossUS$436mn/US$735mnPreviously US$268mn/US$436mn, primarily due to increased R&D investment
• FY26E/FY27E/FY28E earnings per shareUS$-2.22/US$-1.66/US$-1.00Basic earnings per share before exceptional items; previously US$-1.32/US$-1.32/US$-0.92
• 1H26 free cash outflowUS$521mnGoldman Sachs estimate
• Cash reservesOver US$3bnFollowing recent financing, available to support model and computing-capacity investment
• Effective training utilization97%Level following infrastructure upgrades in 1H26
• M3 and M2 parameter counts430B/230BM3's parameter count nearly doubled, but its API gross margin remained broadly comparable or even higher
• 1H26 overseas revenue contribution61%Supports international distribution of multimodal products
• H3 global downloadsOver 24 millionAdoption scale driven by the open-weight strategy
• 12-month target priceHK$760Previously HK$800; the current price of HK$303 implies 150.8% upside
• Key DCF assumptionsWACC 12%, terminal growth rate 2%Used for the base-case target price valuation
• Bull- and bear-case valuationsHK$1,300/HK$220The report believes the valuation outcome remains skewed to the upside

Impact & implications

The report believes MiniMax is trading higher R&D spending and cash consumption for improvements in model intelligence, inference efficiency, and computing scale. If M3.1, M3 Pro, and H3.1 continue to drive Token demand, customer adoption, and pricing power, revenue growth and gross-margin improvement could support a valuation recovery in 2H26. Conversely, persistently expanding R&D expenditure would delay earnings visibility and increase demands on cash reserves and self-financing capacity.

Risks

• Global competition in foundation models is intense, and MiniMax's model performance may be weaker than expected.
• R&D investment is increasing rapidly, and improvement in earnings visibility may be slower than expected.
• Commercialization capabilities may fall short of expectations, and ARR growth may not fully translate into revenue and profit.
• Issues related to intellectual property and generated content may affect product adoption and commercial expansion.
• Continued cash consumption may weaken the company's self-financing capacity.
• Intensifying China-US technology competition may create geopolitical and computing-supply risks.

What to watch

• Track M3.1's stability, post-training scale, and inference-efficiency improvements, as well as their impact on ARR.
• Monitor the progress of updates to the 3T-scale M3 Pro and H3.1 in 4Q26 and their commercial adoption.
• Assess whether computing supply can keep pace with Token demand and how cluster utilization and domestic chips affect unit costs.
• Track how the revenue contribution from text models, inference efficiency, and pricing power jointly affect gross margin.
• Monitor the H3 open-weight ecosystem, global downloads, overseas revenue, and the commercial conversion of MiniMax Design.
• Assess whether MiniMax Agent, Code, and Design can establish sustained user retention and pricing power rather than relying solely on temporary model leadership.
• Track changes in R&D investment, free cash outflow, cash reserves, and earnings visibility.