Goldman Sachs in-depth report: Who will become the long-term winner in China’s AI large model industry?

China’s AI large models are at a historic turning point. Goldman Sachs believes that the intelligent performance of China’s open-source/open-weight large models has come close to the world’s top proprietary models, while adoption among domestic companies and global small and mid-sized enterprises is expanding rapidly. The resulting data flywheel effect will further drive model iteration and upgrades.

According to the Qichuan Trading Desk, Goldman Sachs’ latest report notes that this evolution track can be summarized as “from last year’s cost-efficiency moment of DeepSeek to this year’s model-intelligence moment of Zhipu GLM.” In a 50-page report led by analyst Ronald Keung, the team conducts a systematic evaluation of four core questions: how Chinese AI models achieve high performance at low cost, why they choose an open-source route, how they monetize, where the core addressable markets are, and who will become the long-term winner.

On the competitive landscape, Goldman Sachs introduces a “competitive positioning framework” based on pricing power, cost advantages, and financial strength, and accordingly concludes that in the base text model category, Zhipu (first coverage) and DeepSeek (not listed) are positioned as the strongest; in the multimodal space, ByteDance (not listed) is leading. Goldman Sachs also maintains Buy ratings for MiniMax and Kuaishou.

Punching above one’s weight, winning with efficiency

Chinese large models can achieve performance close to that of US counterparts at costs far lower than those products, with the core drivers being breakthroughs in both architecture innovation and parameter efficiency.

Goldman Sachs’ report states that China’s open-source models generally have parameter scales ranging from 200 billion to 1.6 trillion, only 2% to 10% of the world’s top models—mainly because access to high-end compute is constrained. At the same time, innovations such as the Mixture of Experts (MoE) architecture and sparse attention mechanisms reduce the ratio of actually activated parameters to total parameters to only 3% to 5%, significantly lowering training and inference costs.

At the level of specific models, DeepSeek V4 Pro has 1.6 trillion parameters, Zhipu GLM5.2 has 0.7 trillion, and MiniMax M3 has 0.4 trillion.

Goldman Sachs attributes the recent jump in Chinese models’ programming capabilities to the combined effects of data filtering and post–reinforcement learning training, among other factors. On June 27, DeepSeek rolled out a speculative decoding framework called DSpark, which has been deployed in the online services of V4-Flash and V4 Pro. Without changing model weights or output quality, it increases each user’s generation speed by 60% to 85% (V4-Flash) and 57% to 78% (V4 Pro).

Meituan’s LongCat 2.0 released on June 30 is viewed by Goldman Sachs as an important milestone for the localization of China’s AI infrastructure—the first completely open-source MoE model with 1.6 trillion parameters trained and deployed based entirely on 50 thousand units of domestically produced compute cards. Goldman Sachs believes this proves the feasibility of a localized hardware stack during compute-intensive pretraining stages, which is of profound significance for Chinese AI models to break away from reliance on foreign high-end chips.

A split market: the strong get stronger

Goldman Sachs describes the China AI model market as forming a “two-tier structure” and identifies two ARR-maximization quadrants.

In the high-end market, top models represented by Zhipu GLM5.2 and Alibaba Qwen3.7 Max are priced at about $1 per one million tokens, which is five times that of low-end models; inference gross margins are about 10% to 20% (estimated by Goldman Sachs). By comparison, top models in the US are priced at $4 to $8 per one million tokens. China’s high-end models are only 10% to 25% of that level, but can still maintain positive gross margins thanks to their lower parameter activation ratio.

In the low-end market, models aimed at agent tasks are priced as low as $0.06 to $0.2 per one million tokens, opening up the price-sensitive global small and mid-sized enterprise and individual user markets. MiniMax has 60% to 70% of its revenue coming from overseas. Notably, DeepSeek has announced that starting from mid-July it will introduce a peak-and-trough pricing mechanism for the V4 series, where peak rates are twice the non-peak rates, with blended pricing of about $0.35 per one million tokens (V4 Pro) and $0.12 per one million tokens (V4 Flash).

Goldman Sachs forecasts that China’s AI model API and subscription revenue will grow from 35 billion RMB estimated for 2026 to 879 billion RMB by 2030—corresponding to daily token consumption rising from 350 trillion to 4,600 trillion, an increase of about 25 times.

Open-source strategy: broad penetration, monetization path still to be upgraded

Goldman Sachs’ report lays out the strategic logic behind China’s widespread adoption of open-source/open-weight routes for AI models, as well as the limitations in monetization.

The core advantages of an open-source strategy lie in deployment flexibility and the community ecosystem. Alibaba’s Qwen series, DeepSeek, Zhipu GLM, and MiniMax M3 all adopt open-source or open-weight approaches. ByteDance’s Seed models are the main exception, taking a fully closed proprietary route. The open-source model allows flexible deployment both within and outside mainland China, and iteration is accelerated via community feedback.

However, Goldman Sachs points out that the ARR figures disclosed by open-source model companies are likely to seriously underestimate actual deployment scale and revenue potential. For example, Zhipu’s target ARR by end-2026 is $1 billion, but the actual global deployment volume of GLM5.2 will be far higher than the token volume and revenue generated through Zhipu’s own API channels—Alibaba Cloud’s Bailian MaaS platform can directly host the open-source GLM5.2 model, without requiring Zhipu to be paid any fees.

Goldman Sachs expects the industry will gradually move from pure open-source (MIT license, completely free) toward a “open-weight + community license” model—meaning commercial use requires revenue-sharing agreements with the model company. MiniMax’s M series has already taken the lead in adopting this model. Goldman Sachs believes this shift will significantly improve AI model companies’ unit economics: model companies can benefit from revenue-sharing agreements with platforms such as AWS Bedrock and Alibaba Cloud Bailian, without having to bear inference compute costs themselves.

From “token maximization” to prioritizing ROI

Goldman Sachs characterizes international market expansion as the most important upside space for China’s AI models, especially in non-US markets.

Goldman Sachs’ US research team estimates that by 2030, agent AI will drive a 24-fold increase in global token consumption, reaching 16k billion tokens per month; enterprise agents contribute 55-fold growth, while consumer agents contribute 12-fold growth. In global markets (outside China), China’s AI models have already achieved significant gains in token share thanks to performance improvements and pricing advantages.

Goldman Sachs’ report states that the global enterprise AI usage paradigm is undergoing a fundamental shift from “token maximization” to “ROI-first.” The former was prevalent from late 2025 to early 2026, when enterprises equated high token consumption with organizational productivity. The latter focuses more on clear task boundaries, the number of daily active agents, automation of back-end processes, and actual output. A Jellyfish AI engineering trend research shows that heavy AI users in enterprises consume 10 times the tokens, but output increases only by 2 times.

At the channel level, Alphabet’s Gemini Enterprise Agent Platform and Amazon’s AWS Bedrock have already provided hosting services for Chinese AI models such as DeepSeek, MiniMax, Moonshot, GLM, and Qwen. As reported by The Wall Street Journal, Microsoft CEO recently said Microsoft is considering hosting a DeepSeek version in Copilot as an optional low-cost model, and emphasized that if DeepSeek is hosted, the model will run within Microsoft’s cloud ecosystem, ensuring customer data remains within Azure.

Who will be the long-term winner?

Goldman Sachs has built a three-dimensional competitive positioning framework to quantitatively assess each player’s probability of long-term outperformance. The core formula is: ARR scale × gross margin advantage + financial strength.

The pricing power dimension examines上市速度 (the comparison with previous generations and peer models), the LMArena arena score (based on user evaluations from large-scale blind tests), and the blended pricing level per one million tokens.

The cost advantage dimension examines throughput (tokens per second), cache hit rate, parameter activation ratio, and inference gross margin. The financial strength dimension examines cash on hand, net cash as a proportion of total assets, and valuation multiples.

In the base text model space, Goldman Sachs finds that Zhipu (first coverage, neutral rating, target valuation $110 billion) and DeepSeek (not listed) are positioned as the strongest. Both stand out in pricing power and cost advantages. The overall implied valuations of independent AI model companies total more than $200 billion.

In the multimodal/video generation space, ByteDance leads with Seedance. According to LatePost and 36Kr, Seedance’s gross margin is as high as 70%, and its ARR run rate has exceeded $2 billion. Kuaishou’s build and MiniMax Hailuo/its upcoming H3 model are also seen favorably by Goldman Sachs. It expects that in the second half of 2026 they will benefit from functional breakthroughs in the fusion of video generation and LLMs, as well as healthy pricing driven by supply tightness.

Goldman Sachs maintains a Buy rating on MiniMax, with a target price of HK$860. The rationale is that its M3 model sits in the ARR-maximization quadrant with both high token volume and attractive pricing. Also, its current valuation is only 13 times the ARR by end-2026. Compared with valuation multiples of comparable companies in China and globally, there is a clear discount, so the risk-reward profile is skewed to the upside.

Risk warning and disclaimer

        The market is risky; investment requires caution. This article does not constitute personal investment advice, and it does not consider the special investment objectives, financial conditions, or needs of any individual user. Users should consider whether any opinions, viewpoints, or conclusions in this article align with their specific circumstances. Investing based on this is at your own risk.