Goldman Sachs sees a turning point for China’s AI model market
Goldman Sachs says China’s large model industry has reached a pivotal stage in its development. In the bank’s view, open-source and open-weight Chinese models are now approaching the intelligence level of the world’s top proprietary systems, while adoption by domestic enterprises and global small and medium-sized businesses is expanding quickly. That, the report says, is creating a data flywheel that can feed the next round of model upgrades.
According to Trading Alpha, Goldman described the current trajectory as moving from “DeepSeek’s cost-efficiency moment last year to Zhipu GLM’s intelligence moment this year.” The 50-page report, led by analyst Ronald Keung, focuses on four questions: how Chinese models are achieving strong performance at low cost, why developers are choosing open routes and how they may monetize them, where the core addressable market sits, and which companies are most likely to win over time.
Lower cost, competitive performance
Goldman says Chinese open models typically range from 200 billion to 1.6 trillion parameters, or just 2% to 10% of the size of the world’s leading models, largely because access to high-end computing remains constrained. Even so, design choices such as mixture-of-experts, or MoE, and sparse attention have pushed active parameters down to only 3% to 5% of total parameters, cutting both training and inference costs.
At the 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 links the recent jump in coding performance among Chinese models to a combination of data filtering and reinforcement-learning-based post-training.
The report also points to DeepSeek’s speculative decoding framework, DSpark, which launched on June 27 and has already been deployed in the online services of V4-Flash and V4 Pro. Without changing model weights or output quality, DSpark lifted generation speed per user by 60% to 85% for V4-Flash and 57% to 78% for V4 Pro.
Meituan’s LongCat 2.0, released on June 30, is described by Goldman as a major milestone for domestic AI infrastructure. The bank says it is China’s first 1.6 trillion-parameter open-source MoE model trained and deployed entirely on 50,000 domestically produced compute cards. In Goldman’s assessment, that shows a local hardware stack can work even in compute-heavy pretraining, an important step for reducing dependence on foreign high-end chips.
A two-tier market is taking shape
Goldman describes China’s AI model market as evolving into a two-layer structure and identifies two quadrants where annual recurring revenue, or ARR, can be maximized.
At the high end, top models such as Zhipu GLM5.2 and Alibaba’s Qwen3.7 Max are priced at about $1 per million tokens, roughly five times the level of lower-end models. Goldman estimates inference gross margins for this tier at about 10% to 20%. By comparison, top U.S. models charge $4 to $8 per million tokens. Chinese high-end models are priced at only 10% to 25% of that level, yet can still preserve positive gross margins because their active-parameter ratios are lower.
At the low end, models aimed at agent tasks are priced as low as $0.06 to $0.2 per million tokens and are targeting price-sensitive global SMEs and individual users. Goldman says 60% to 70% of MiniMax revenue comes from overseas. DeepSeek has also announced peak and off-peak pricing for the V4 series starting in mid-July, with peak rates set at twice off-peak rates. Its blended pricing is about $0.35 per million tokens for V4 Pro and $0.12 for V4 Flash.
Goldman forecasts China’s AI model API and subscription revenue will rise from an estimated RMB 35 billion in 2026 to RMB 879 billion in 2030. Over the same period, daily token consumption is projected to grow from 350 trillion to 4.6 quadrillion, an increase of about 25 times.
Open deployment drives adoption, but monetization is still evolving
Goldman says the appeal of the open-source and open-weight path in China comes from flexible deployment and community reach. Alibaba’s Qwen series, DeepSeek, Zhipu GLM and MiniMax M3 all follow open-source or open-weight approaches. ByteDance’s Seed model is the main exception, remaining fully closed and proprietary.
That model makes it easier to deploy systems inside and outside mainland China while speeding up iteration through community feedback. Still, Goldman says ARR figures disclosed by open-model companies may significantly understate real deployment scale and revenue potential. Zhipu is one example. The company has set a target of $1 billion in ARR by the end of 2026, but Goldman says the actual global deployment volume of GLM5.2 is likely to be much larger than the token volume and revenue flowing through Zhipu’s own API channel, because Alibaba Cloud’s Bailian MaaS platform can host the open model directly without paying Zhipu.
The bank expects the industry to move away from purely open MIT licensing toward an “open weights plus community license” approach, where commercial use requires revenue-sharing agreements with model developers. MiniMax’s M series has already adopted that structure. Goldman says the shift could sharply improve unit economics because model providers would benefit from sharing arrangements with platforms such as AWS Bedrock and Alibaba Cloud Bailian without carrying inference compute costs themselves.
Goldman sees overseas expansion as the main upside
Goldman identifies international markets, especially those outside the U.S., as the largest upside for Chinese AI models. Its U.S. research team estimates agentic AI will lift global token consumption by 24 times to 120,000 quadrillion tokens per month by 2030, with enterprise agents contributing 55-fold growth and consumer agents 12-fold growth.
The report says Chinese AI models have already gained token share outside China thanks to better performance and lower pricing. It also argues that enterprises are changing how they evaluate AI use. The market is moving from “maximizing tokens” to “ROI first.” The earlier approach was common from late 2025 to early 2026, when companies treated high token consumption as a sign of organizational productivity. The newer approach puts more weight on clear task boundaries, daily active agents, backend process automation and measurable output.
Data from a Jellyfish AI engineering trends study showed heavy AI users in enterprises consumed 10 times more tokens but produced only twice as much output. On distribution, Alphabet’s Gemini Enterprise Agent Platform and Amazon’s AWS Bedrock already offer hosted access to Chinese models including DeepSeek, MiniMax, Moonshot, GLM and Qwen. The Wall Street Journal also reported that Microsoft CEO recently said the company is considering hosting a version of DeepSeek in Copilot as an optional low-cost model. If that happens, he said, the model would run inside Microsoft’s cloud ecosystem so customer data remains within Azure.
Who does Goldman see as the long-term winner?
Goldman built a three-dimensional framework to judge long-term positioning across players. Its core formula is ARR scale multiplied by gross-margin advantage, then combined with financial strength.
Pricing power is measured by launch speed relative to prior and peer models, LMArena scores based on large-scale blind user testing, and blended pricing per million tokens. Cost advantage looks at throughput, cache hit rates, active-parameter ratios and inference gross margins. Financial strength covers cash on hand, net cash as a share of total assets and valuation multiples.
In foundational text models, Goldman names Zhipu and DeepSeek as the strongest positions. Zhipu is initiated with a Neutral rating and a target valuation of $110 billion, while DeepSeek remains private. Goldman says both stand out on pricing power and cost advantage. It also says the combined implied valuation of independent AI model companies now exceeds $200 billion.
In multimodal and video generation, ByteDance leads with Seedance. Citing LatePost and 36Kr, Goldman says Seedance has a 70% gross margin and an ARR run rate above $2 billion. The bank also remains constructive on Kuaishou Kling and MiniMax Hailuo, along with the upcoming H3 model, saying they could benefit in the second half of 2026 from feature breakthroughs as video generation converges with LLMs and from healthier pricing under tight supply.
Goldman keeps a Buy rating on MiniMax with a target price of HK$860. Its case is that the M3 model sits in an ARR-maximizing quadrant defined by high token volume and attractive pricing, while the company trades at just 13 times estimated end-2026 ARR, a clear discount to peers in China and globally.