Goldman Sachs: Who Will Be the Long-Term Winner Among China's AI Large Models
Taylor Wilson
Goldman Sachs published a 50-page deep dive concluding that China's open-source large models now rival the world's best proprietary ones, projecting API and subscription revenue to hit RMB 879 billion by 2030 — a roughly 25× increase from 2026 — and naming Zhipu and DeepSeek as the strongest-positioned foundation-model players.
How did China close the gap with a fraction of the parameters?
China's open-source models range from 200 billion to 1.6 trillion parameters — just 2% to 10% of global frontier models. The core constraint: restricted access to top-end training chips.
Mixture-of-Experts architecture (MoE — a design that activates only a small slice of parameters per query) and sparse attention cut active parameters to 3%–5% of the total. This means → near-frontier intelligence on roughly one-tenth the hardware.
Meituan's LongCat 2.0, released June 30, is a milestone: the first Chinese 1.6-trillion-parameter open-source MoE model trained entirely on 50,000 domestic AI chips, proving local hardware can handle compute-intensive pre-training.
What does the market look like, and where does the money come from?
Goldman describes a "two-tier structure." The premium tier — led by Zhipu's GLM5.2 and Alibaba's Qwen3.7 Max — prices at roughly $1 per million tokens, with inference gross margins of 10%–20%. U.S. frontier models charge $4–8; China's premium tier sits at 10%–25% of that.
The value tier targets agentic tasks at $0.06–0.2 per million tokens, aimed at price-sensitive SMEs and individual users globally. MiniMax draws 60%–70% of its revenue from overseas.
Goldman forecasts China AI model API and subscription revenue will grow from RMB 35 billion in 2026 to RMB 879 billion by 2030, with daily token consumption rising from 3.5 quadrillion to 46 quadrillion. In plain terms = a 25× jump in four years, betting that businesses and consumers treat AI like a utility.
If open-source is free, how do these companies make money?
Alibaba's Qwen, DeepSeek, Zhipu's GLM, and MiniMax's M3 all use open-source or open-weight approaches. ByteDance's Seed models are the main exception — fully proprietary.
Goldman flags that disclosed ARR figures likely significantly understate actual deployment. Zhipu targets $1 billion ARR by end-2026, but Alibaba Cloud's Bailian platform can host the open-source GLM5.2 without paying Zhipu a cent. This means → the model is widely used, but the revenue doesn't all flow back to the model company.
Goldman expects a gradual shift to "open weights + community license" — commercial users must sign revenue-sharing agreements. MiniMax's M-series already does this. In plain terms = moving from "use it for free" to "use it, but share the profits."
How does Goldman rank these companies?
Goldman built a three-dimensional framework: ARR scale × gross-margin advantage + financial strength, examining pricing power, cost advantage, and cash on hand.
In foundation text models: Zhipu (initiating coverage, neutral, target valuation $110 billion) and DeepSeek (unlisted) are positioned strongest. Combined implied valuation for independent AI model companies exceeds $200 billion.
In multimodal / video generation: ByteDance leads with Seedance — reported gross margins of 70% and an ARR run-rate above $2 billion.
Which stocks does Goldman rate Buy?
MiniMax: Buy rating maintained, target price HK$860. The M3 model sits in the high-token-volume, attractive-pricing quadrant, and the current valuation is only 13× end-2026 ARR — a clear discount to Chinese and global peers.
Kuaishou: Buy rating maintained. Goldman expects the stock to benefit in H2 2026 from breakthroughs combining video generation with large language models, plus healthy pricing from tight supply.
This reflects Goldman's deeper thesis: the real differentiator in the next phase is not who offers the cheapest token, but who delivers the most real-world output per unit of cost — one data point shows heavy enterprise AI users consumed 10× the tokens but lifted output by only 2×.
Content is for reference only, not financial advice.