Author: Wall Street News
China's AI large models are standing at a historic inflection point. Goldman Sachs believes that the intelligent performance of China's open-source/open-weight models is approaching that of the world's top proprietary models. Adoption by domestic enterprises and global small and medium-sized enterprises (SMEs) is rapidly expanding, forming a data flywheel effect that will further drive model iteration and upgrade.
According to Chasing Wind Trading Desk, Goldman Sachs' latest report points out that this evolutionary trajectory can be summarized as "from DeepSeek's cost-efficiency moment last year to Zhipu's GLM's model intelligence moment this year." The team led by Goldman Sachs analyst Ronald Keung systematically evaluates four core issues in this 50-page report: how Chinese AI models achieve high performance at low cost, why they choose the open-source route and how to monetize it, where the core addressable market is, and who will become the long-term winners.
In judging the competitive landscape, Goldman Sachs proposes a "competitive positioning framework" based on pricing power, cost advantage, and financial strength. Based on this, it concludes that in the field of foundational text models, Zhipu (initially covered) and DeepSeek (not listed) are positioned as the strongest; in the multimodal field, ByteDance (not listed) leads the pack. Goldman Sachs also maintains its Buy ratings on MiniMax and Kuaishou.
Achieving More with Less, Winning Through Efficiency
Chinese large models are able to achieve performance close to that of their US counterparts at a significantly lower cost, with the core reasons being breakthroughs in architectural innovation and parameter efficiency.
Goldman Sachs' report notes that the parameter scale of China's open-source models generally ranges from 200 billion to 1.6 trillion, which is only 2% to 10% of that of the world's top models, primarily due to limited access to high-end computing power. Meanwhile, innovations such as the Mixture-of-Experts (MoE) architecture and sparse attention mechanisms mean that the actual activated parameters only account for 3% to 5% of the total parameters, significantly reducing training and inference costs.
At the specific model level, 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 leap in programming capabilities of Chinese models to the synergy of factors such as data filtering and reinforcement learning post-training (RLHF/RLTF). On June 27, DeepSeek launched the speculative decoding framework DSpark, which has been deployed in the online services of V4-Flash and V4 Pro. It increases per-user generation speed by 60% to 85% (V4-Flash) and 57% to 78% (V4 Pro) without changing model weights or output quality.
Meituan's release of LongCat 2.0 on June 30 is viewed 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 fully trained and deployed on 50,000 domestic AI accelerator cards. Goldman Sachs believes this demonstrates the feasibility of a localized hardware stack in the compute-intensive pre-training stage, holding profound significance for Chinese AI models to break free from dependence on foreign high-end chips.
Polarizing Market, Stronger Get Stronger
Goldman Sachs describes the Chinese AI model market as forming a "two-tiered structure" and identifies two ARR-maximizing quadrants.
In the high-end market, top models like Zhipu GLM5.2 and Alibaba's Qwen3.7 Max are priced at around $1 per million tokens, five times that of low-end models, with estimated inference gross margins of about 10% to 20% (Goldman Sachs estimate). In comparison, top US models are priced at $4 to $8 per million tokens. Chinese high-end models are only 10% to 25% of that price, but due to their lower parameter activation ratio, they can still maintain positive gross margins.
In the low-end market, models aimed at agent tasks are priced as low as $0.06 to $0.2 per million tokens, tapping into the price-sensitive global SME and individual user market. 60% to 70% of MiniMax's revenue comes from overseas. Notably, DeepSeek has announced the introduction of peak/off-peak pricing for its V4 series starting mid-July, with peak rates being twice the off-peak rates, resulting in a blended price of about $0.35 per million tokens (V4 Pro) and $0.12 (V4 Flash).
Goldman Sachs forecasts that API and subscription revenue for Chinese AI models will grow from an estimated RMB 35 billion in 2026 to RMB 879 billion in 2030, corresponding to daily token consumption increasing from 350 trillion to 4.6 quadrillion, a roughly 25-fold increase.
Open Source Strategy: Widespread Penetration, Monetization Path Pending Upgrade
Goldman Sachs' report details the strategic rationale behind the widespread adoption of open-source/open-weight approaches by Chinese AI models and their monetization limitations.
The core advantages of the open-source strategy lie in deployment flexibility and community ecosystem. Alibaba's Qwen series, DeepSeek, Zhipu GLM, and MiniMax M3 all adopt open-source or open-weight methods, with ByteDance's Seed model being the major exception, taking a fully closed-source proprietary route. The open-source model allows flexible deployment both inside and outside mainland China and accelerates iteration through community feedback.
However, Goldman Sachs points out that the ARR numbers disclosed by open-source model companies likely significantly underestimate the actual scale of deployment and revenue potential. Taking Zhipu as an example, its ARR target for the end of 2026 is $10 billion, but the actual global deployment volume of GLM5.2 will be far greater than the token volume and revenue from Zhipu's own API channel — Alibaba Cloud's Bailian MaaS platform can directly host the open-source GLM5.2 model without paying any fees to Zhipu.
Goldman Sachs expects the industry to gradually migrate from pure open-source (MIT license, completely free) to an "open-weight + community license" model — where commercial use requires signing a revenue-sharing agreement with the model company. The MiniMax M series has already pioneered this model. Goldman Sachs believes this shift will significantly improve the unit economics for AI model companies, as they can benefit from revenue-sharing agreements with platforms like AWS Bedrock and Alibaba Cloud Bailian without having to bear the inference compute costs themselves.
From "Token Maximization" to ROI Priority
Goldman Sachs characterizes international market expansion as the most significant upside for Chinese AI models, especially in non-US markets.
Estimates from Goldman Sachs' US research team suggest that by 2030, agent AI will drive a 24x increase in global token consumption, reaching 120 quadrillion tokens per month, with enterprise agents contributing a 55x growth and consumer agents contributing a 12x growth. In global (ex-China) markets, Chinese AI models have already achieved significant token share growth leveraging performance improvements and price advantages.
Goldman Sachs' report notes that the AI usage paradigm for global enterprises is undergoing a fundamental shift from "token maximization" to "ROI priority." The former prevailed from late 2025 to early 2026, where companies equated high token consumption with organizational productivity; the latter focuses more on clear task boundaries, daily active agent count, backend process automation, and actual output. Data from a Jellyfish AI engineering trends study shows that heavy AI users in enterprises consume 10x more tokens but only achieve 2x more output.
At the channel level, Alphabet's Gemini Enterprise Agent Platform and Amazon's AWS Bedrock already offer hosting services for Chinese AI models like DeepSeek, MiniMax, Moonshot, GLM, and Qwen. According to The Wall Street Journal, Microsoft's CEO recently stated that Microsoft is considering hosting a version of DeepSeek within Copilot as an optional low-cost model, emphasizing that if hosted, the model would 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, using quantitative metrics to assess each player's probability of long-term success, with the core formula being: ARR scale × gross margin advantage + financial strength.
The Pricing Power dimension examines launch speed (compared to previous generation and similar-tier models), LMArena arena scores (based on large-scale blind user evaluations), and blended price per 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 percentage of total assets, and valuation multiples.
In the field of foundational text models, Goldman Sachs identifies Zhipu (initial coverage, Neutral rating, target valuation $110 billion) and DeepSeek (not listed) as having the strongest positioning, with both excelling in pricing power and cost advantage. The aggregate implied valuation of independent AI model companies exceeds $200 billion.
In the field of multimodal/video generation, ByteDance leads with Seedance. According to LatePost and 36Kr, Seedance has a gross margin as high as 70%, and its ARR run rate has already exceeded $2 billion. Kuaishou's Keling and MiniMax's Hailuo/upcoming H3 models are also viewed favorably by Goldman Sachs, expected to benefit in the second half of 2026 from breakthroughs in the integration of video generation and LLMs, as well as healthy pricing due to tight supply.
Goldman Sachs maintains its Buy rating on MiniMax with a target price of HK$860, citing that its M3 model is positioned in the ARR-maximizing quadrant with high token volume and attractive pricing, and its current valuation is only 13x its projected end-of-2026 ARR, representing a clear discount compared to valuation multiples of Chinese and global peers, resulting in a risk-reward profile skewed to the upside.
