US-China AI Intelligence Gap Widens to 13%, Domestic Price Wars Intensify Profitability Concerns

Overview

This research report primarily analyzes the changing gap in model intelligence levels and commercial pricing between the Chinese AI industry and the US. The core conclusion is: with the release of Anthropic's new generation model Fable 5, the US-China AI model intelligence gap has widened from 8% to 13%. Meanwhile, due to US export controls and model anti-distillation technology adoption, the difficulty for Chinese open-source models to catch up increases. More critically, against the backdrop of surging hardware costs, Chinese AI vendors are embroiled in price wars with declining API pricing, constituting a major challenge to the industry's sustainable profitability.

Core views

Intelligence Gap Widening: Anthropic released Fable 5 (including Mythos 5) on June 9, ranking first in global model intelligence scores, 6% higher than Opus 4.8. This caused the intelligence gap between US and China AI models to expand from 8% in May to 13% in June. Institutions believe this gap mainly stems from US labs having access to more powerful AI chips (e.g., NVIDIA Blackwell series). With NVIDIA's next-gen Rubin chips shipping in H2 2026, the compute gap between China and the US may widen further, thereby expanding the intelligence gap. Policy and Technical Barriers: Fable 5 introduced 'distillation protection'; if suspicious distillation activity is detected, the model degrades to Opus 4.8. Additionally, due to cybersecurity vulnerabilities, the US implemented emergency export controls on Fable 5, prohibiting foreign use. Lacking effective identity verification tools, Anthropic suspended Fable 5 and Mythos 5 services globally worldwide. These measures make it harder for open-source models dominated by China to improve performance via distillation; if other closed-source vendors follow suit, it will delay overall progress in the open-source community. Profit Model Concerns: Compared to the intelligence gap, institutions worry more about the commercial profit dilemma of Chinese AI. In the US, driven by hardware costs, average API prices rose 89% from March to June, and Anthropic doubled the Fable 5 API price. Conversely in China, API prices remained flat or declined, down 2% MoM in June, mainly affected by MiniMax's significant 33% price cut. However, Chinese AI vendors face soaring hardware costs like memory and higher cost pressures from adopting Huawei's 'logic folding' technology for chip manufacturing. Under dual pressure of declining revenue-side prices and rising cost-side expenses, institutions believe only by raising API pricing and ensuring sustained growth in Token consumption can acceptable Investment ROI be maintained. Therefore, price competition among domestic AI labs is seen as the current largest risk point. Despite above challenges, institutions still believe current Chinese AI models are sufficient to support application commercialization; Tencent launching A2A services on the WeChat platform is a signal of accelerating C-end AI monetization.

Analysis framework

Institutions mainly conducted analysis by comparing intelligence score indices and API pricing data of frontier AI models in China and the US. First, utilizing third-party data (Artificial Analysis) to quantify the gap in model intelligence levels, combined with hardware supply chain (NVIDIA chip acquisition difficulty) to explain the reasons for the widening gap. Second, by comparing API price trends with underlying hardware cost trends in both countries, revealing risks such as 'revenue increase without profit increase' or 'loss-led market share acquisition' faced by the Chinese AI industry. Finally, combining policy aspects (export controls, anti-distillation technology) and business aspects (Tencent cases), comprehensively evaluating short-term bottlenecks and long-term commercialization potential of the Chinese AI industry.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply-demand framework

  Mismatch between computing power supply and model demand

  The report implies constraints on computing power as a core factor of production. Due to ability to acquire advanced chips (supply abundant), model iteration is fast in US; China limited by export controls (supply constrained), leading to slower model intelligence improvement speed, reflecting decisive influence of upstream resource supply on downstream product competitiveness.

• Company Fundamentals and Financial FrameworkProfitability Quality Analysis

  Impact of Volume-Price Relationship on ROI

  The report points out that when hardware costs soar in China, API prices falling instead is unhealthy. This is based on basic profit logic: When unit cost rises, if unit price falls, unless sales volume explodes, ROI will be compressed. This is a core framework for analyzing sustainability of SaaS or tech service profitability.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Tencent (700 HK)

  Benefit Logic: Report specifically mentions Tencent launched A2A services on WeChat platform, a strong signal of accelerating C-end AI monetization, indicating it is leading in commercialization implementation.

  Strengths

  Huge user base and mature social platforms, rich commercialization scenarios.

  Comparison

  Compared to pure AI model vendors, Tencent possesses stronger traffic monetization capabilities.

• MiniMax

  Damaged/Risk Logic: Report points out its large price cut of 33% was one of the main reasons for the drop in Chinese API prices, reflecting fierce price competition environment.

  Strengths

  Model scores rose significantly recently (+10%), strong technical capability.

  Weaknesses

  Involved in price war, may sacrifice margins for market share.

  Comparison

  More aggressive on pricing strategy, forming stark contrast with high-price maintaining US-funded vendors.

  Risks

  Hardware cost rise eroding profits.

• Alibaba / Xiaomi

  Neutral/Observation: As major players, their model scores remained stable in past 4 weeks, no large fluctuations, but no obvious price or technical breakthrough advantages shown.

Key data

• US-China Model Intelligence Gap13%Expanded from 8% in May to 13% in June, mainly driven by Anthropic Fable 5 release
• Average US API Price Increase+89%During Mar-Jun 2026, partly driven by hardware costs
• China June API Price MoM Change-2%Affected by MiniMax 33% price cut etc., prices continuing to fall
• Anthropic API Price AdjustmentDoubledPrice adjustment compared to Opus 4.8 for Fable 5
• MiniMax Model Score Change+10%Model intelligence score rose from 49.6 to 54.7 in past 4 weeks

Impact & implications

For the Chinese AI industry, simply pursuing improvements in model parameters or scores is becoming increasingly difficult and expensive. Export controls and anti-distillation technologies cut off some low-cost technical catch-up paths. For investors and enterprises, the focus should shift from 'whose model is stronger' to 'who can better achieve commercial monetization'. Companies that can control computing power costs, avoid vicious price wars, and successfully land in vertical application scenarios (such as WeChat ecosystem) will have greater investment value. Conversely, companies trapped in low-price competition and unable to pass on high hardware costs will face profitability pressure.

Risks

• Further expansion of US-China compute gap, deepening model intelligence generation gap
• US export control scope expands, restricting more advanced AI technologies and chips flowing into China
• Intensification of involution in Chinese AI industry, price wars lead to industry-wide profitability deterioration
• Hardware costs (especially memory and domestic chip manufacturing) continue to rise beyond expectations

What to watch

• Shipment status of NVIDIA Rubin chips and acquisition ability of Chinese enterprises
• Whether pricing strategies of major Chinese AI vendors stabilize or rebound
• Commercial monetization progress of various companies in C-end and B-end applications (e.g., growth in Token consumption)
• Whether Anthropic and other closed-source vendors will widely promote 'anti-distillation' technology