Goldman Sachs: Chinese AI Models Are Stealing Token Share from the U.S., Alibaba and Tencent May See Inflection Points in H2
N.R. Finch
Chinese AI models have surged from low-single-digit to nearly 50% of global token share on third-party API platforms, yet razor-thin pricing means they barely register in revenue; Goldman sees conditions for an Alibaba-Tencent stock inflection in the second half.
How did Chinese models grab nearly 50% of token share?
On OpenRouter, a third-party API marketplace, Chinese models now account for nearly 50% of token volume — up from low single digits at end-2024.
The weapon is price: Chinese models charge $0.20–$1.00 per million tokens versus roughly $4.00 for top U.S. models. This means → the cheapest Chinese option costs one-twentieth of the American alternative for the same task.
This week's top three models globally on OpenRouter are all Chinese: DeepSeek V4 Flash, Hy3 preview, and MiniMax M3.
With share that high, why is there almost no money in it?
Chinese models still account for only single-digit percentages of the global AI-model revenue pool. In plain terms = they sell nearly half the "cups" but charge so little that total revenue is a rounding error.
The exception is multimodal — models that handle video and images, not just text. ByteDance's Seedance 2.0 charges $4–$7 per million tokens for video generation, with annualized monthly revenue already at $1.7 billion.
This reflects a stark split: text models have been commoditized; only compute-heavy formats like video still command a premium.
Will Chinese models fully replace American ones?
Goldman's call: not wholesale replacement, but demand stratification.
High-value workloads — large multinationals, complex code generation — will stay on frontier models. Always-on agentic tasks (AI workflows that autonomously execute a chain of operations) will increasingly shift to cheaper Chinese models.
Non-technical headwinds persist: compute-access restrictions, anti-distillation measures (techniques that stop smaller models from copying larger ones), safety regulation, and market-access barriers.
How fierce is the price war inside China?
Five to six independents plus the BATX giants are all competing at the model layer — more fragmented than the U.S. market.
Pricing is sharply polarized: Qwen3.7 Max at $1.40 per million tokens, GLM-5.1 at $0.90, while DeepSeek and MiMo sit at just $0.20. MiniMax made its half-price API promotion permanent at roughly $0.22.
Goldman sees coding capability, multimodal performance, training efficiency, and balance-sheet strength as the long-run dividing line. On an ARR basis: Alibaba's MaaS targets $4.4 billion by year-end (70–80% from Qwen), ByteDance averages $2.2 billion for the full year, and Zhipu and MiniMax each target no less than $1 billion.
Where exactly are Alibaba and Tencent hurting?
Alibaba is down roughly 22% year-to-date, driven mainly by EPS cuts — heavy consumer-AI spending eroded the "other businesses" segment, contributing about 15 percentage points of the decline; P/E compression from 18× to about 16× added another 9 points.
Tencent is down roughly 27%, but the story is the opposite: EPS cuts account for only about 3 percentage points; multiple contraction drove roughly 25 points. This means → the market doesn't think Tencent is earning less — it doubts whether the AI narrative can support the valuation.
Tencent's current 11–12× P/E is the lowest in both Chinese and global peer history.
What could trigger the H2 inflection?
Alibaba's catalysts: accelerating cloud revenue, price increases lifting margins, and quarter-on-quarter narrowing of "other businesses" losses. Goldman believes the EPS-downgrade cycle is near its trough.
Tencent's recovery path: progress on the Hy-series models, faster capex driving cloud-revenue growth, and WeChat AI-agent rollouts in the second half.
Both stocks trade in the lowest tier of global internet-peer valuations, with ample operating cash flow and net cash reserves to fund hyperscale AI investment. In plain terms = compressed this far, a few passing grades on H2 scorecards could be enough to release the spring.
Content is for reference only, not financial advice.