Goldman names DeepSeek, ByteDance, Zhipu as China AI winners

Goldman Sachs analyst Ronald Keung put out a note on July 12 titled "Who Will Be the Long-Term Winner in China's AI Large Model Industry?", and the interesting part isn't the picks themselves so much as the fact that a bulge-bracket bank is now writing price targets on Chinese frontier labs at all. CNBC reported that Keung's team named DeepSeek, ByteDance and Zhipu as its long-term winners, with ByteDance singled out as best-in-class on AI video generation.

The specific calls are what a China-facing investor actually has to work with. Goldman initiated coverage on Zhipu with a HK$1,880 price target and a Neutral rating, pegging the company at roughly $110 billion, and put Buy ratings on MiniMax and Kuaishou. DeepSeek and ByteDance are privately held, so the bank can flag them as preferred without a rating attached. On the model itself, Keung wrote that Zhipu's GLM5.2 has "almost reached frontier performance," with adoption by domestic enterprises and global SMEs rising fast enough to sustain a high frequency of upgrades.

What makes this a global story rather than a China-only one is the pricing gap. Goldman puts top Chinese models like GLM5.2 and Alibaba's Qwen3.7 Max at about $1 per million tokens, roughly a fifth to an eighth of the $4-8 charged by US equivalents, with an estimated gross margin of only 10-20% at that level. The bank sees Chinese AI model API and subscription revenue growing from about 35 billion RMB in 2026 to 879 billion RMB by 2030, on daily token consumption rising roughly 25-fold over four years.

The honest caveats are worth stating. A Neutral rating on Zhipu after the Hong Kong vehicle has already run about 70% in sixty business days is Goldman telling you a lot of the good news is priced in, and MiniMax's own equity is down more than 70% over the same window despite the Buy call. The report also doesn't get into how sustained compute access under US export controls affects the next GLM revision, or what path a public-market investor has to DeepSeek and ByteDance at all.

The part worth watching is enterprise procurement over the next couple of quarters, because if global SMEs really are switching to the roughly $1 tier at scale, the US inference-margin story starts to look a lot more fragile than the model leaderboards suggest.