Goldman Sachs Deep Research Report: Who Will Become the Long-Term Winner in China’s AI Large Model Industry?

The Chinese AI large model industry is standing at a historic inflection point. Goldman Sachs believes that the intelligent performance of China’s open-source/open-weight large models has already approached that of the world’s top proprietary models, and that adoption scale among domestic companies and global small and medium-sized enterprises is expanding rapidly. The resulting data flywheel effect will further drive iteration and upgrades of the models.

According to Zihu Trading Desk, Goldman Sachs’ latest report said that this evolutionary trajectory can be summarized as “from last year’s DeepSeek cost-efficiency moment to this year’s model intelligence moment of Zhipu GLM.” In a 50-page report led by Goldman analyst Ronald Keung, the team carried out a systematic assessment around 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 assessment, Goldman Sachs introduced a “competitive positioning framework” based on pricing ability, cost advantages, and financial strength. On that basis, it concluded that in the foundational text model space, Zhipu (first coverage) and DeepSeek (not listed) are positioned as the strongest; in the multimodal domain, ByteDance (not listed) is leading. Goldman also maintained “Buy” ratings for MiniMax and Kuaishou.

Punching above your weight through efficiency

Chinese large models can achieve performance close to that of comparable U.S. products at far lower costs, mainly due to a double breakthrough in architectural innovation and parameter efficiency.

Goldman’s report said that the parameter scale of China’s open-source models is generally between 200 billion and 1.6 trillion, only 2% to 10% of global top models, largely because access to high-end computing power is constrained. At the same time, innovations such as the mixture-of-experts architecture (MoE) 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 GLM 5.2 has 0.7 trillion, and MiniMax M3 has 0.4 trillion.

Goldman attributed the recent jump in Chinese models’ programming ability to the combined effects of data filtering and post–reinforcement learning training. On June 27, DeepSeek released 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 per-user generation speed by 60% to 85% (V4-Flash) and 57% to 78% (V4 Pro).

LongCat 2.0 released by Meituan on June 30 was viewed by Goldman Sachs as an important milestone for China’s AI infrastructure to become locally autonomous—China’s first 1.6 trillion-parameter open-source MoE model trained and deployed entirely based on 50k domestically produced compute cards. Goldman believes this demonstrates the feasibility of a localized hardware stack during compute-intensive pretraining, which has far-reaching significance for Chinese AI models to break away from dependence on foreign high-end chips.

Market polarizes, and the strong get stronger

Goldman described China’s AI model market as forming a “two-layer structure,” and identified two ARR-maximizing quadrants.

In the high-end market, top models represented by Zhipu GLM 5.2 and Alibaba’s Qwen 3.7 Max are priced at about $1 per million tokens, which is five times that of low-end models. Inference gross margins are about 10% to 20% (estimated by Goldman). By comparison, top models in the U.S. are priced at $4 to $8 per million tokens; China’s high-end models are only 10% to 25% of that level, but thanks to a lower parameter activation ratio, they can still maintain positive gross margins.

In the low-end market, models for agent tasks are priced as low as $0.06 to $0.2 per million tokens, expanding into price-sensitive global small and medium-sized enterprise and individual user markets. MiniMax has 60% to 70% of its revenue from overseas. Notably, DeepSeek has announced that starting in mid-July it will introduce a peak-and-off-peak pricing mechanism for the V4 series: the peak rate is 2 times the off-peak rate, resulting in a blended price of about $0.35 per million tokens (V4 Pro) and $0.12 per million tokens (V4 Flash).

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

Open-source strategy: broad penetration, monetization path still needs upgrading

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

The core advantages of an open-source strategy lie in deployment flexibility and community ecosystems. Alibaba’s Qwen series, DeepSeek, Zhipu GLM, and MiniMax M3 all adopt open-source or open-weight approaches. ByteDance’s Seed model is the main exception, adopting a fully closed proprietary route. The open-source model allows the model to be deployed flexibly both inside and outside mainland China, and it accelerates iteration through community feedback.

However, Goldman pointed out that the ARR figures disclosed by open-source model companies are likely to seriously underestimate actual deployment scale and revenue potential. Taking Zhipu as an example, its target ARR by the end of 2026 is $1 billion, but the actual global deployment volume of GLM 5.2 will be far higher than the token volume and revenue from Zhipu’s own API channels—Alibaba Cloud’s Bailing MaaS platform can directly host the GLM 5.2 open-source model without Zhipu receiving any fees.

Goldman expects the industry will gradually move from pure open source (MIT license, completely free) to a “open-weight + community license” model—meaning commercial use requires revenue-sharing agreements with the model company. MiniMax’s M series has been the first to adopt this model. Goldman believes this shift will significantly improve unit economics for AI model companies, because they can benefit through revenue-sharing agreements with platforms like AWS Bedrock and Alibaba Cloud’s Bailing, without having to bear inference computing costs themselves.

From “maximizing tokens” to prioritizing ROI

Goldman characterized international market expansion as the most important upside space for China’s AI models, especially in non-U.S. markets.

Goldman’s U.S. research team estimated that by 2030, agent AI will drive a 24x growth in global token consumption to 16k billion tokens per month. Enterprise agents contribute 55x growth, while consumer agents contribute 12x growth. In global markets (outside China), Chinese AI models have already achieved significant token share growth by leveraging performance improvements and pricing advantages.

Goldman’s report said that the global enterprise AI usage paradigm is undergoing a fundamental shift from “maximizing tokens” to “prioritizing ROI.” The former was prevalent from the end of 2025 to the beginning of 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 backend processes, and actual outputs. A Jellyfish AI engineering trend study showed that heavy AI users in enterprises consumed 10x the tokens, but outputs only increased by 2x.

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

Who will be the long-term winner?

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

For pricing ability, it examines listing speed (compared with predecessors and peers), the LMArena arena score (based on user evaluations from large-scale blind tests), and blended pricing levels per million tokens.

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

In foundational text models, Goldman assessed that Zhipu (first coverage, neutral rating, target valuation of $110 billion) and DeepSeek (not listed) are the strongest—both standout in pricing ability and cost advantages. The total implied valuation across independent AI model companies exceeds $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 already exceeded $2 billion. Kuaishou’s Kuaishou Ling and MiniMax Hailuo/its upcoming H3 model are also viewed positively by Goldman. Goldman expects that in the second half of 2026 they will benefit from breakthroughs in functions that merge video generation and LLMs, as well as healthy pricing driven by supply tightness.

Goldman maintains a Buy rating on MiniMax, with a target price of HK$860. The rationale is that its M3 model sits in the ARR-maximizing quadrant of high token volume and attractive pricing, and that its current valuation is only 13x its ARR at end of 2026. Compared with valuation multiples of similar companies in China and globally, there is a clear valuation discount, so the risk-reward profile is skewed upward.