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

Author: Wall Street Journal

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 approached that of the world's top proprietary models, and the adoption scale by domestic companies and global small and medium enterprises is rapidly expanding. The resulting data flywheel effect will further drive model iteration and upgrades.

According to the Wind Trading Desk, Goldman Sachs' latest report states, "This evolutionary trajectory can be summarized as 'from DeepSeek's cost efficiency moment last year to Zhiyu GLM's model intelligence moment this year.'" The team led by Goldman Sachs analyst Ronald Keung conducts a systematic assessment in this 50-page report around four core issues: how Chinese AI models achieve high performance at low cost, why they choose the open-source route and how to monetize, where the core addressable market is, and who will become the long-term winners.

In terms of competitive landscape assessment, Goldman Sachs has introduced a "competitive positioning framework" based on pricing capability, cost advantage, and financial strength, and based on this, it identifies that in the field of foundational text models, Zhiyu (first coverage) and DeepSeek (unlisted) are positioned the strongest; in the multimodal field, ByteDance (unlisted) is leading. Goldman Sachs also maintains a buy rating on MiniMax and Kuaishou.

Small investments yield big returns, efficiency wins

Chinese large models can achieve comparable performance to similar products in the United States at a significantly lower cost, primarily due to breakthroughs in architecture innovation and parameter efficiency.

Goldman Sachs' report points out that the parameter scale of China's open-source models generally ranges from 200 billion to 1.6 trillion, only 2% to 10% of the world's top models, mainly due to limitations in acquiring high-end computing power. At the same time, innovations such as the mixture of experts architecture (MoE) and sparse attention mechanisms mean that the actual activated parameters account for only 3% to 5% of the total parameters, significantly reducing training and inference costs.

In terms of specific models, DeepSeek V4 Pro has 1.6 trillion parameters, Zhiyu GLM5.2 has 0.7 trillion, and MiniMax M3 has 0.4 trillion.

Goldman Sachs attributes the recent leap in programming capability of Chinese models to the synergistic effects of data filtering, reinforcement learning post-training, and other factors. On June 27, DeepSeek launched the speculative decoding framework DSpark, which has been deployed in the online services of V4-Flash and V4 Pro, increasing user generation speed by 60% to 85% (V4-Flash) and 57% to 78% (V4 Pro) without changing model weights or output quality.

Meituan's LongCat 2.0, released on June 30, is seen by Goldman Sachs as an important milestone in the localization of China's AI infrastructure—this is China's first open-source MoE model with 1.6 trillion parameters trained and deployed entirely based on 50,000 domestically produced computing cards. Goldman Sachs believes this demonstrates the feasibility of a localized hardware stack during the compute-intensive pre-training phase, which has far-reaching significance for Chinese AI models to break free from dependence on foreign high-end chips.

Market polarization, the strong get stronger

Goldman Sachs describes the Chinese AI model market as forming a "dual-layer structure" and identifies two quadrants for maximizing ARR.

In the high-end market, top models represented by Zhiyu GLM5.2 and Alibaba Qwen3.7 Max are priced at about $1 per million tokens, five times that of low-end models, with an estimated gross margin of about 10% to 20% (according to Goldman Sachs). In contrast, top models in the U.S. are priced at $4 to $8 per million tokens, while Chinese high-end models are only 10% to 25% of that, but can still maintain positive gross margins due to lower parameter activation ratios.

In the low-end market, models aimed at agent tasks are priced as low as $0.06 to $0.20 per million tokens, targeting price-sensitive global small and medium enterprises and individual users. MiniMax derives 60% to 70% of its revenue from overseas. Notably, DeepSeek has announced that starting from mid-July, it will introduce a peak and off-peak pricing mechanism for the V4 series, with peak rates being twice that of off-peak rates, and mixed pricing of about $0.35 per million tokens (V4 Pro) and $0.12 (V4 Flash).

Goldman Sachs predicts that the API and subscription revenue of Chinese AI models will grow from an estimated 35 billion RMB in 2026 to 879 billion RMB by 2030, corresponding to a daily token consumption increase from 350 trillion to 4,600 trillion, an increase of about 25 times.

Open-source strategy: widespread penetration, monetization paths to be upgraded

Goldman Sachs' report details the strategic logic behind the widespread adoption of open-source/open-weight routes by Chinese AI models and their monetization limitations.

The core advantage of the open-source strategy lies in deployment flexibility and community ecology. The Alibaba Qwen series, DeepSeek, Zhiyu GLM, and MiniMax M3 all adopt open-source or open-weight methods, while ByteDance's Seed model is a major exception, using a completely closed-source proprietary route. The open-source model allows for flexible deployment of models both within and outside mainland China and accelerates iteration through community feedback.

However, Goldman Sachs points out that the ARR figures disclosed by open-source model companies are likely to severely underestimate the actual deployment scale and revenue potential. For example, Zhiyu aims for an ARR target of $1 billion by the end of 2026, but the actual deployment of GLM5.2 globally will far exceed the token volume and revenue from Zhiyu's own API channels—Alibaba Cloud's Bailian MaaS platform can directly host the GLM5.2 open-source model without any fees paid to Zhiyu.

Goldman Sachs expects the industry to gradually shift from pure open-source (MIT license, completely free) to an "open-weight + community license" model—where commercial use must sign a revenue-sharing agreement with the model company. The MiniMax M series has already adopted this model. Goldman Sachs believes this shift will significantly improve the unit economics of AI model companies, as they can benefit from revenue-sharing agreements with platforms like AWS Bedrock and Alibaba Cloud Bailian without bearing the inference computing costs themselves.

From "token maximization" to ROI priority

Goldman Sachs characterizes the international market expansion as the most important upward space for Chinese AI models, especially in non-U.S. markets.

Goldman Sachs' U.S. research team estimates that by 2030, agent AI will drive global token consumption to grow 24 times, reaching 120 trillion tokens per month, with enterprise agents contributing a 55-fold increase and consumer agents a 12-fold increase. In global markets (outside of China), Chinese AI models have achieved significant token share growth due to performance improvements and price advantages.

Goldman Sachs' report indicates that the AI usage paradigm among global enterprises is undergoing a fundamental shift from "token maximization" to "ROI priority." The former prevailed from late 2025 to early 2026, where high token consumption was equated with organizational productivity; the latter focuses more on clear task boundaries, daily active agent counts, backend process automation, and actual output. A Jellyfish AI engineering trend study shows that heavy AI users in enterprises consumed 10 times the tokens but only doubled their output.

On the channel level, Alphabet's Gemini Enterprise Agent Platform and Amazon's AWS Bedrock have both provided hosting services for Chinese AI models such as DeepSeek, MiniMax, Moonshot, GLM, and Qwen. According to the Wall Street Journal, Microsoft CEO recently stated that Microsoft is considering hosting a version of DeepSeek on Copilot as a low-cost model option, emphasizing that if DeepSeek is hosted, the model will run within Microsoft's cloud ecosystem, ensuring customer data remains within Azure.

Who are the long-term winners?

Goldman Sachs has constructed a three-dimensional competitive positioning framework to quantitatively assess the long-term winning probabilities of various players, with the core formula being: ARR scale × gross margin advantage + financial strength.

The pricing capability dimension examines the speed of market entry (compared to previous and peer models), LMArena arena scores (based on large-scale blind user evaluations), and the mixed pricing level per million tokens.

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

In the foundational text model field, Goldman Sachs identifies Zhiyu (first coverage, neutral rating, target valuation of $110 billion) and DeepSeek (unlisted) as the strongest positioned, both excelling in pricing capability and cost advantage. The overall implied valuation of independent AI model companies exceeds $200 billion.

In the multimodal/video generation field, ByteDance leads with Seedance, with reports from LatePost and 36Kr indicating that Seedance has a gross margin as high as 70%, and its ARR run rate has exceeded $2 billion. Kuaishou's Keling and MiniMax Hailuo/upcoming H3 model are also viewed positively by Goldman Sachs, expected to benefit in the second half of 2026 from functional breakthroughs in video generation and LLM integration, as well as healthy pricing due to supply constraints.

Goldman Sachs maintains a buy rating on MiniMax, with a target price of HKD 860, reasoning that its M3 model is in the ARR maximizing quadrant of high token volume and attractive pricing, and its current valuation is only 13 times the ARR at the end of 2026, indicating a significant discount compared to the valuation multiples of similar companies in China and globally, with a risk-reward ratio skewed upward.