Goldman Sachs In-depth: Five Key Debates on Chinese AI

The pace of releases for China's AI large models has significantly accelerated, but the investment debate around their narrative has deepened and become more controversial. Goldman Sachs has published its latest in-depth report, systematically outlining five core investment debates, covering model performance between China and the US, domestic competitive landscape, token growth drivers, the transition to domestic chips, and trends in consumer-level AI Agents.

Goldman Sachs states that the annualized recurring revenue (ARR) of China's AI large models has continually exceeded expectations this year, with explosive token demand driving cloud infrastructure into a high-prosperity cycle. The quarterly performance of US cloud giants (Google Cloud up 63% year-on-year, Azure up 40%, AWS up 29%) has provided strong evidence of industry prosperity.

Analysts upgraded MiniMax's rating from "Neutral" to "Buy," maintaining Alibaba's "Buy" rating. Analysts expect Alibaba Cloud's revenue growth in the March quarter to reach 40%, higher than last quarter's 36%. Meanwhile, China's major cloud service providers still have room to expand capital expenditures, which provides multi-year prosperity support for the cloud and data center sector, making it the preferred sub-industry within China's internet sector.

Debate One: How do the Chinese and US models really perform?

Goldman Sachs believes that, based on multiple benchmark tests, Chinese models excel in pricing competitiveness, inference speed, and agent task completion rate, with certain dimensions narrowing the gap with leading US models (GPT/Opus/Gemini series).

Analysts indicate, As the gap in performance of China's foundational models narrows, pricing power and ARR have improved. Complex programming and multi-modality will become key competitive dimensions in the next stage.

Very tight computational constraints are pushing Chinese AI companies onto a unique development path—focusing on training and inference efficiency, data quality, and post-training optimization, constructing efficient architectures with fewer chips and memory.

As access restrictions for US SOTA models to Chinese users tighten, usage rates for Chinese AI models in high-value scenarios are expected to increase, forming a positive data flywheel, especially in the programming domain. In multi-modal direction, ByteDance's Seedance 2.0 and Alibaba's Happy Horse have achieved industry-leading levels; MiniMax’s upcoming Hailuo 3 is seen as the next major catalyst.

Debate Two: Fragmented competition landscape, differentiation is key

DeepSeek V4 was released on April 23, supporting a 1 million token context window while requiring only 7%-10% of the KV Cache memory compared to its previous version V3.2, and offering price discounts to developers. This has once again sparked discussion on the competitive landscape of Chinese AI models.

Goldman Sachs believes that the pullback of MiniMax and Zhipu's stock prices since the March highs is mainly due to concentrated releases of competing models such as DeepSeek V4, Tencent Hy3.0, Xiaomi MiMo V2.5, Alibaba Qwen3.6, etc.

Investors' core doubts regarding the competitive landscape focus on three points:

First, decreased entry barriers and market fragmentation—Xiaomi entering the field, Tencent Hy3’s training cycle was less than three months, both indicate that model development costs for mid-sized manufacturers are rapidly decreasing;

Second, competition in models between 20 billion and 300 billion parameters is becoming increasingly intense; this range is the main battleground for agent applications. While MiniMax M2.7 has built recognition and received platform endorsement, DeepSeek V4 Flash, Hy3.0, and MiMo V2.5 all pose competitive pressure;

Third, pricing war risk—Hy3's free preview version once topped OpenRouter’s usage ranking, but Goldman points out such free periods tend to be short, and the global tight supply-demand for computational power will continue to support improved pricing power of Chinese AI models.

In landscape assessment, Goldman Sachs believes the core advantage of independent AI model companies lies in efficient organizational structure and fast decision-making; large Internet giants, leveraging strong operating cash flow from core businesses, are more capable of seizing AI infrastructure and cloud computing opportunities. However, they need to set up independent incentive mechanisms for their AI chip and model teams to counter talent competition from independent AI-native companies. Programming capabilities, multi-modality, and task completion rates will drive future pricing power, which may evolve from charging per token to charging per successful task.

Debate Three: Token growth is sustainable, cloud vendors have pricing upside potential

The continuously rapid expansion of token consumption is the key variable supporting growth in AI cloud services.

According to the National Data Bureau, China's average daily token usage in March exceeded 140 trillion, up more than a thousandfold from 100 billion at the start of 2024, and about 40% higher than at the end of 2025. On the enterprise side, ByteDance’s Doubao model saw daily token usage exceed 120 trillion, doubling over three months; the number of enterprise clients with daily token consumption exceeding 1 trillion on Volcano Engine increased from 100 at the end of 2025 to 140.

On the third-party API aggregator platform OpenRouter, Chinese AI models have consistently gained market share from American AI models over the past two months, mainly benefiting from comprehensive advantages in performance, long-context, agent capabilities, and pricing of newly released Chinese models.

Goldman Sachs notes that the peak daily token consumption on OpenRouter occurred in the last week of March, with subsequent weekly averages at about 79% of the peak; but as AI Agents take on more 24/7 tasks, and enterprise incentives for using AI tools persist, Goldman expects token demand to continue sequential growth.

On the capital expenditure side, Goldman estimates that US hyperscale cloud vendors’ combined capex will exceed $700 billion in 2026, with China at more than $70 billion (Goldman forecasts).

Chinese cloud vendors’ capex as a share of operating cash flow is about 60%, significantly lower than America’s at about 90%, providing financial space for further expansion from the second half of 2026 through 2028.

Goldman expects upside potential in token pricing, partly driven by improved model performance, partly by pass-through of cloud service price increases due to tight supply and rising costs. According to Gartner, Alibaba Cloud’s share of the Asia-Pacific IaaS market will rise to 22.5% in 2025 (from 20.8% in 2024), and its global IaaS market share will rise to 7.7% in 2025 (from 7.2% in 2024).

Debate Four: Accelerating shift toward domestic chips

Multiple factors are catalyzing the accelerated shift of China’s AI industry toward domestic chips.

Main directions include Huawei Ascend 910C and 950 series (expected mass production expansion starting in the second half of 2026), and self-developed chips by internet giants (such as Alibaba’s Pingtouge series).

Goldman Sachs believes the migration to domestic chips will accelerate from 2026 to 2028, but short-term supply bottlenecks still exist. Chinese AI model training will increasingly rely on highly optimized computational efficiency architectures. Additionally, rising memory costs will spur short-term capex increases by China’s hyperscale cloud vendors, echoing similar comments from US cloud vendors.

Debate Five: OS-level Agent could become a paradigm shift, super apps’ moat faces test

In the consumer AI assistant field, Doubao continues to lead.

According to QuestMobile, Doubao’s DAU in March reached about 150 million, accounting for about 63% of China’s AIGC to-C application total usage time, with March usage up 78% month-on-month. Alibaba Qwen, leveraging its transactional functionality, ranked second in user growth.

On April 24, Doubao embedded the “Doubao helps you choose” feature in its app navigation bar, officially entering the e-commerce track. Goldman believes task execution, seamless transaction capability, and social push mechanisms will be key to user retention and scale expansion.

The current core debate in the market is: Will OS-level Agents represented by Doubao’s phone assistant, or in-app Agents represented by the soon-to-be-launched WeChat AI Agent, dominate the next phase of consumer AI entry points?

The potential expansion of OS-level Agents triggers deeper structural discussions: Super apps in China (e.g., WeChat), operated in closed ecosystems, will face whether OS-level Agents cannibalize their traffic, or whether they block agents on grounds of data privacy and security, thereby fostering new competitive dynamics in instant messaging.

Goldman Sachs defines OS-level Agentic AI as a profound paradigm shift, believing it may, in the medium term, occupy major traffic entry points, relegating independent apps to backend tool providers, weakening their user engagement and data accumulation advantages.

Goldman expects fierce competition around interoperability, data permissions, and ecosystem control. Internet giants (like Tencent and Alibaba) deeply integrating payment, logistics, and social relationship chains may defend their moats by deepening commercialization of in-app Agents, while hardware makers and independent AI players will actively push for broader

~~~~~~~~~~~~~~~~~~~~~~~~

The above highlights are from Chasing Wind Trading Desk.

For more detailed analysis including real-time interpretation and frontline research, please join [Chasing Wind Trading Desk ▪ Annual Membership]

Risk Warning and DisclaimerThe market has risks, investment needs to be cautious. This article does not constitute personal investment advice, nor does it take into account specific users’ particular investment goals, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investing accordingly is at your own risk.