Goldman Sachs In-depth Report: Who Will Be the Long-term Winner in China’s AI Foundation Model Industry?

```

China's AI large models are standing at a historic turning point. Goldman Sachs believes that the intelligence performance of China's open-source/open-weight large models has approached that of the world's top proprietary models. The adoption scale among domestic companies and global SMEs is rapidly expanding, creating a data flywheel effect that will further drive model iteration and upgrades.

According to Wind Trading Desk, Goldman Sachs' latest report notes, this evolutionary trajectory can be summarized as "from DeepSeek's cost-efficiency moment last year, to Zhipu GLM's intelligence moment this year". The team led by Goldman Sachs analyst Ronald Keung presents a systematic assessment around four core questions in the 50-page report: how China’s AI models achieve high performance at low cost, why open-source is the chosen path and how it is monetized, where the core addressable market is, and who will be the long-term winners.

In terms of competitive landscape, Goldman Sachs introduced a "competitive positioning framework" based on pricing power, cost advantage, and financial strength. Based on this, it determined that in the field of foundational text models, Zhipu (first coverage) and DeepSeek (unlisted) hold the strongest positions; in multimodal field, ByteDance (unlisted) leads. Goldman Sachs also maintains buy ratings for MiniMax and Kuaishou.

Small bets, big returns; efficiency wins

Chinese large models can achieve nearly equivalent performance at far lower costs than their US counterparts, with the core breakthroughs in architecture innovation and parameter efficiency.

The Goldman Sachs report states that parameter scale for China's open-source models is generally between 200 billion to 1.6 trillion, only 2% to 10% of the world's top models, mainly due to limited access to high-end compute power. Meanwhile, innovations like Mixture of Experts (MoE) architecture and sparse attention mechanisms ensure the actual activated parameters are only 3% to 5% of the total, dramatically lowering training and inference costs.

At the model level, DeepSeek V4 Pro has 1.6 trillion parameters, Zhipu GLM5.2 has 0.7 trillion, MiniMax M3 has 0.4 trillion.

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

Meituan's release on June 30 of LongCat 2.0 is viewed by Goldman Sachs as an important milestone for the domestication of China's AI infrastructure—the first 1.6 trillion parameter open-source MoE model entirely trained and deployed on 50,000 domestically produced compute cards. Goldman Sachs believes this proves the feasibility of a localized hardware stack during compute-intensive pre-training, and has profound significance for China's AI models to break dependence on foreign high-end chips.

Market polarization, the strong get stronger

Goldman Sachs describes China's AI model market as forming a "two-layer structure" and identifies two ARR-maximizing quadrants.

In the high-end market, top models such as Zhipu GLM5.2 and Alibaba Qwen3.7 Max are priced at $1 per million tokens, five times that of low-end models, with inference gross margin estimated at 10%-20% (by Goldman Sachs). In comparison, US top models are priced at $4-$8 per million tokens. China's high-end models are only 10%-25% of that price, but maintain positive gross margins thanks to lower parameter activation ratios.

In the low-end market, models for agent tasks are priced as low as $0.06-$0.2 per million tokens, targeting price-sensitive global SMEs and individual users. MiniMax derives 60%-70% of its revenue from overseas. Notably, DeepSeek announced that from mid-July it will introduce peak-valley pricing for the V4 series; peak rates are double those of off-peak, with hybrid pricing at about $0.35 per million tokens (V4 Pro) and $0.12 per million tokens (V4 Flash).

Goldman Sachs predicts API and subscription income for China's AI models will grow from an estimated RMB 35 billion in 2026 to RMB 879 billion in 2030, corresponding to daily token consumption rising from 350 trillion to 4600 trillion—a 25-fold increase.

Open-source strategy: wide penetration, monetization channels to be upgraded

The Goldman Sachs report details the strategic logic and monetization limitations of why China’s AI models generally pursue open-source/open-weight paths.

The core advantages of open-source are deployment flexibility and community ecosystem. The Alibaba Qwen series, DeepSeek, Zhipu GLM, and MiniMax M3 all follow open-source or open-weight approaches. ByteDance's Seed model is a main exception, fully proprietary and closed source. The open-source model allows for flexible deployment inside and outside Mainland China and accelerates iteration via community feedback.

However, Goldman Sachs notes, open-source model companies' disclosed ARR numbers very likely severely underestimate actual deployment scale and revenue potential. Taking Zhipu as an example, its 2026 year-end ARR target is $1 billion, but global deployments of GLM5.2 far exceed the token volume and income managed by Zhipu's own API channel—Alibaba Cloud's Bailian MaaS platform can host the GLM5.2 open-source model directly without any payment to Zhipu.

Goldman Sachs expects the industry will gradually shift from pure open-source (MIT license, fully free) toward "open weight + community license"—meaning commercial use requires a revenue sharing agreement with the model company. MiniMax M series has pioneered this model. Goldman Sachs believes this transition will significantly improve unit economics for AI model companies, as they can benefit from revenue sharing agreements with platforms like AWS Bedrock and Alibaba Bailian without bearing the inference compute cost themselves.

From "token maximization" to ROI priority

Goldman Sachs identifies global market expansion as the most important upside for Chinese AI models, especially outside the US.

Goldman Sachs' US research team estimates that by 2030, agent AI will drive global token consumption up 24-fold to 1.2 quadrillion tokens per month, with enterprise agents contributing a 55-fold increase, consumer agents a 12-fold increase. In global markets (outside China), China's AI models have already gained significant token market share thanks to improved performance and price advantages.

Goldman Sachs notes global enterprises are undergoing a fundamental shift in AI usage from "token maximization" to "ROI priority". The former prevails in late 2025 to early 2026, equating high token usage with organizational productivity; the latter focuses on clear task boundaries, daily active agents, back-end process automation, and actual output. A Jellyfish AI engineering trends study showed heavy enterprise AI users consumed 10 times more tokens but only raised output twofold.

At the channel level, Alphabet’s Gemini Enterprise Agent Platform and Amazon AWS Bedrock both host Chinese AI models including DeepSeek, MiniMax, Moonshot, GLM, and Qwen. According to The Wall Street Journal, Microsoft’s CEO recently said Microsoft is considering hosting versions of DeepSeek in Copilot as an optional low-cost model, emphasizing that if hosted, DeepSeek would operate within Microsoft’s cloud ecosystem, ensuring customer data remains on Azure.

Who is the long-term winner?

Goldman Sachs has developed a three-dimensional competitive positioning framework to quantitatively evaluate each player’s probability of long-term success, with the core formula: ARR scale × gross margin advantage + financial strength.

Pricing ability measures listing speed (vs previous and peer models), LMArena leaderboard scores (based on large-scale blind user reviews), and hybrid price per million tokens.

Cost advantage considers throughput (tokens/sec), cache hit rate, parameter activation ratio, and inference gross margin. Financial strength looks at cash on hand, net cash as share of total assets, and valuation multiples.

In the foundational text model field, Goldman Sachs identifies Zhipu (first coverage, neutral rating, target valuation $110 billion) and DeepSeek (unlisted) as strongest, both excelling in pricing power and cost advantage. Total implied valuation for independent AI model companies exceeds $200 billion.

In the multimodal/video generation field, ByteDance leads with Seedance. According to LatePost and 36Kr, Seedance’s gross margin is as high as 70%, with an ARR run rate over $2 billion. Kuaishou Koling and MiniMax Hailuo/about-to-be-launched H3 models are also favored by Goldman Sachs, expected to benefit in the second half of 2026 from breakthroughs in video generation and LLM fusion plus healthy pricing due to supply constraints.

Goldman Sachs maintains a buy rating on MiniMax, target price HK$860, reasoning that its M3 model sits in the high-token attractive pricing ARR maximization quadrant, with current valuation only 13 times its 2026 year-end ARR—significantly discounted compared to peers in China and globally, with an upside-leaning risk-return profile.

Risk warning and disclaimerMarkets are risky, investment should be cautious. This article does not constitute personal investment advice and does not take into account individual users' special investment objectives, financial status, or needs. Users should consider whether any opinions, views, or conclusions herein are suitable for their specific situation. Investment according to this is at your own risk. ```