MiniMax's new models, lower inference costs, and open ecosystem jointly support ARR growth and gross margin improvement

Overview

This report summarizes discussions with MiniMax management at Goldman Sachs' 2026 Asia Leaders Conference, focusing primarily on ARR and model launches in the second half of 2026, gross margin improvement from lower inference costs, compute supply constraints, and the role of open weights and Harness products in driving adoption and customer stickiness. Goldman Sachs maintains a positive view and assigns a Buy rating with a 12-month target price of HK$760.

Core views

MiniMax co-founder and President Yeyi Yun attended the Goldman Sachs Asia Leaders Conference in Hong Kong on September 1, 2026, participating in a fireside chat titled "The Next Stage of Frontier AI Models." The report summarizes management's messages into four main themes: second-half ARR and the model pipeline, gross margin improvement driven by cost efficiency, infrastructure optimization amid compute constraints, and the promotion of adoption, monetization, and customer stickiness through the open-weight ecosystem and Harness products. First, model iteration is a direct driver of ARR growth. Management stated that ARR has continued to rise since M3 and multimodal H3 were launched in August. Subsequent products include M3.1, M3 Pro, and H3.1, with M3.1 to be released as soon as possible after its performance and commercial readiness meet internal standards. This means near-term growth depends on whether new models can meet internal quality thresholds and convert technological advances into actual customer usage and recurring revenue. Second, the company expects gross margin to improve sequentially in the second half of 2026, mainly due to ongoing inference optimization and enhanced model efficiency. Management believes that as intelligence capabilities expand, inference-side workloads are becoming increasingly important, including post-training, particularly reinforcement learning. Lower inference costs not only reduce service costs but also accelerate model iteration and improve performance; therefore, the company's cost-effective input-output ratio is viewed as a core competitive advantage. The company's medium- to long-term gross margin target is in the mid-teens for large language models and higher for multimodal products. Third, compute supply remains a significant bottleneck for the foundation model industry. MiniMax is seeking to generate greater output from limited compute resources through infrastructure optimization and improved resource utilization. Management stated that the existing compute capacity is sufficient to support the current model pipeline and remains confident in its ability to secure the compute required to develop future frontier models. This assessment alleviates concerns about near-term disruption to the product pipeline, but as future models scale up, compute access and utilization efficiency will remain critical to execution. Fourth, the open-weight strategy and Harness applications jointly expand the model ecosystem. Open weights help broaden adoption, accelerate innovation across the external ecosystem, and create additional monetization opportunities through revenue sharing under commercial licenses. The report cites H3 Max, launched by fal.ai based on the relevant ecosystem and capable of accelerating video generation, as an example. MiniMax's proprietary Harness products increase user engagement by deeply coordinating and optimizing its internal M-series and H-series models, user workflows, and infrastructure. Products from internet platform customers are also promoting M3 adoption and contributing incremental revenue. The customer mix continues to shift toward enterprises, which currently account for approximately 80% of revenue, with growth coming from new Chinese internet enterprise customers and expanding overseas adoption. Financial forecasts reflect both high growth and persistent losses. The report expects revenue to increase from US$79.0mn in 2025 to US$516.8mn in 2026, US$1,752.9mn in 2027, and US$3,876.8mn in 2028. EBITDA for the same periods is projected at -US$294.7mn, -US$751.7mn, -US$663.6mn, and -US$409.4mn, respectively, while EPS is projected at -US$2.31, -US$2.22, -US$1.66, and -US$1.00, respectively. This shows that revenue is expanding rapidly, but earnings visibility and self-funding capacity still depend on the realization of commercialization and cost efficiencies. Regarding valuation, Goldman Sachs assigns MiniMax a Buy rating and a 12-month target price of HK$760, using a DCF methodology with assumptions of a 12% WACC and a 2% terminal growth rate. The report lists a September 1, 2026 closing price of HK$336.80, an equity market capitalization of HK$111.6bn or US$14.2bn, and an enterprise value of HK$91.6bn or US$11.7bn. The unadjusted target price history shows that the target price declined from HK$1,018 on February 22, 2026, to HK$1,000 on March 3, HK$860 on June 16, HK$800 on August 3, and HK$760 on August 26. However, this report does not explicitly describe the current rating as an upgrade, downgrade, or reiteration. The current valuation thesis relies primarily on model adoption, long-term revenue expansion, and efficiency improvements rather than near-term profitability.

Analysis framework

The report first extracts operating signals from the management meeting, then sequentially connects model launches with ARR growth, inference costs with gross margin, compute supply with R&D execution, and the open ecosystem with customer adoption. Finally, it combines revenue and loss forecasts to derive a 12-month target price using a DCF framework. Its evidence primarily comes from management statements, company data, Goldman Sachs Research forecasts, and FactSet data.

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.

• Industry analysis frameworkSupply-demand framework

  Compute supply constraint analysis

  The report views compute supply as an industry bottleneck and assesses whether the model pipeline can continue to advance based on existing compute capacity, future access, infrastructure optimization, and resource utilization.

• (Out-of-vocabulary methodology)

  Management meeting takeaways and product roadmap tracking

  Based on management statements at the conference, the report tracks operating signals such as model release timing, commercial readiness, ARR, customer mix, and cost efficiency.

Asset mapping & comparison

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

• MiniMax Group (00100.HK)

  The report believes the company will benefit from the launches of M3, H3, and subsequent models, improved inference efficiency, the open-weight ecosystem, and enterprise customer growth.

  Strengths

  A cost-effective input-output ratio, coordination between internal models and Harness products, an open-weight ecosystem, and enterprise customers accounting for approximately 80% of revenue.

  Weaknesses

  Compute supply is constrained, and Goldman Sachs forecasts that the company will continue to record negative EBITDA and EPS through 2028.

  Comparison

  The report places the company within the global foundation model competitive landscape but does not provide specific quantitative peer comparisons.

  Risks

  Risks related to model performance, commercialization, earnings visibility, intellectual property and content generation, cash burn, and geopolitics arising from US-China technology competition.

Key data

• Enterprise customer revenue contributionApproximately 80%The customer mix continues to shift toward enterprise customers
• Subsequent model pipelineM3.1, M3 Pro, H3.1M3.1 will be released as soon as possible after its performance and commercial readiness meet internal standards
• Medium- to long-term gross margin targetMid-teens for large language models, higher for multimodal productsGross margin is expected to improve sequentially in the second half of 2026
• Revenue forecast2025 US$79.0mn; 2026E US$516.8mn; 2027E US$1,752.9mn; 2028E US$3,876.8mnGoldman Sachs Research forecasts
• EBITDA forecast2025 -US$294.7mn; 2026E -US$751.7mn; 2027E -US$663.6mn; 2028E -US$409.4mnEBITDA remains negative throughout the forecast period, but losses narrow from 2027 onward
• EPS forecast2025 -US$2.31; 2026E -US$2.22; 2027E -US$1.66; 2028E -US$1.00Losses persist throughout the forecast period but gradually improve
• Valuation parametersWACC 12%; terminal growth rate 2%Used for DCF valuation
• Rating and target priceBuy; 12-month target price of HK$760The current price basis is the September 1, 2026 closing price of HK$336.80
• Market capitalization and enterprise valueMarket capitalization HK$111.6bn/US$14.2bn; enterprise value HK$91.6bn/US$11.7bnData from the company overview table in the report
• Target price history (unadjusted)2026-02-22 HK$1,018.00; 2026-03-03 HK$1,000.00; 2026-06-16 HK$860.00; 2026-08-03 HK$800.00; 2026-08-26 HK$760.00The corresponding closing prices were HK$970.00, HK$821.00, HK$416.60, HK$247.20, and HK$303.00, respectively; the report does not explicitly state a rating action

Impact & implications

The report believes model iteration, lower inference costs, and the open ecosystem constitute three mutually reinforcing drivers of MiniMax's growth: new models promote ARR and customer adoption, cost efficiency supports gross margin, and open weights and Harness products increase ecosystem activity and monetization opportunities. However, the high-growth forecasts are still accompanied by persistent losses, and the ultimate outcome depends on the realization of commercialization, cost control, compute availability, and the path to profitability.

Risks

• Amid global competition in foundation models, model performance may fall short of expectations.
• Earnings visibility may take longer than expected to establish.
• Commercialization capabilities may be weaker than expected.
• Intellectual property and content generation may pose risks.
• Cash burn and self-funding capacity may be worse than expected.
• Intensifying US-China technology competition may create geopolitical risks.

What to watch

• Track whether M3 and H3 can continue to drive ARR growth in the second half of 2026.
• Monitor the timing of M3.1's release after it meets internal performance and commercial readiness standards, as well as the progress of M3 Pro and H3.1.
• Verify whether inference optimization and improved model efficiency can deliver sequential gross margin improvement in the second half.
• Monitor how well existing compute resources support the current model pipeline and the company's ability to secure the compute required for future frontier models.
• Observe whether open-weight commercial licensing, revenue sharing, and Harness products can improve adoption and customer stickiness.
• Track customer acquisition among Chinese internet companies, overseas adoption, and changes in the approximately 80% enterprise customer revenue contribution.
• Monitor progress in narrowing EBITDA losses, reducing cash burn, and improving earnings visibility.