ARR upside and edge AI models become key focus areas following financing by AI labs in China and overseas

Overview

This report is the 56th installment of Jefferies' China Internet AI series, focusing on recent AI lab financing, ARR growth potential, the cost advantages of Chinese models, and the launch of StepFun's edge models. The report believes that rapidly rising demand for AI models and agents, increased token consumption, and expanding inference workloads will drive demand for AI infrastructure and cloud services, creating potential opportunities for BAT, Kingsoft Cloud, AI labs, and Kling.

Core views

The core views include: first, recent financing by AI labs in China and overseas is expected to strengthen computing resources and support ARR upside amid surging demand; second, Anthropic's commercialization progress is strong, with SemiAnalysis expecting its ARR to exceed USD60bn in July 2026 and positive EBIT in 3Q26; third, Chinese models have lower API costs than US models, while the intelligence gap is narrowing, benefiting use cases such as coding, workspaces, and long-form content; fourth, StepFun's edge model portfolio is bringing agent capabilities to smartphones and automotive on-device scenarios.

Analysis framework

The report primarily combines financing events, ARR forecasts, the token expenditure index, model price comparisons, model architecture efficiency, and product launch information to assess the impact of expanding AI demand on cloud service providers, AI labs, and the on-device AI application ecosystem.

Methodology notes

• Commercialization metricsARR

  Annual recurring revenue

  The report uses ARR to measure the scale and growth rate of AI lab commercialization. For example, Anthropic's ARR grew from USD9bn at the end of 2025 to more than USD60bn expected in July 2026.

• Cost and efficiency analysisSilicon Data LLM Token Expenditure Index

  LLM token expenditure index

  The index is weighted by usage concentration, with expenditure rising as demand shifts toward premium models. The report notes that the index was 1.63 to 1.65 in the week of July 5, 2026, below 2.04 on May 31, 2026.

• Model architectureMoE / GQA / sparse attention / MFU

  Sources of model cost efficiency

  The report believes that Chinese models have lower API costs than US models, partly due to architectural optimizations such as MoE, Group-Query Attention, sparse attention, and Model FLOPs Utilization.

Asset mapping & comparison

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

• Alibaba Group Holding Limited (9988 HK; BABA)

  Beneficiary of cloud services and the Chinese AI ecosystem

  Strengths

  The report lists it among the beneficiary companies and discloses a BUY rating for the related stock.

  Weaknesses

  The report does not provide detailed earnings forecasts or a target price for Alibaba.

  Comparison

  It belongs to the same BAT and cloud-services-related beneficiary chain as Baidu, Tencent, and Kingsoft Cloud.

  Risks

  AI demand or token consumption may fall below expectations, competition may intensify, and returns on capital expenditure remain uncertain.

• Baidu Inc. (BIDU; 9888 HK)

  Beneficiary of the Chinese AI model and cloud-services ecosystem

  Strengths

  The report lists Baidu as one of the BAT beneficiaries and discloses a BUY rating.

  Weaknesses

  The report does not provide a standalone business breakdown or valuation analysis for Baidu.

  Comparison

  Together with Alibaba and Tencent, it represents a major Chinese internet cloud and AI infrastructure platform.

  Risks

  The pace of model commercialization, cost efficiency, and customer demand may affect the extent of the benefits.

• Kingsoft Cloud (KC)

  Cloud service provider beneficiary

  Strengths

  The report explicitly expects rising demand for AI models and agents to benefit Kingsoft Cloud.

  Weaknesses

  The report does not provide standalone financial forecast details.

  Comparison

  Compared with BAT, Kingsoft Cloud has more direct exposure to cloud infrastructure elasticity.

  Risks

  Risks include cloud service price competition, AI computing supply costs, and customer concentration.

• Kuaishou Technology (1024 HK)

  AI application and content ecosystem company

  Strengths

  The report discloses a BUY rating for Kuaishou and mentions Kling as a beneficiary.

  Weaknesses

  The report does not provide detailed operating metrics for Kuaishou or Kling in the main text.

  Comparison

  Compared with other internet platforms, it has greater exposure to AI video and content-generation applications.

  Risks

  Changes in application commercialization, regulation, content costs, and the competitive landscape.

• MiniMax Group Inc. (100 HK)

  AI lab and model application company

  Strengths

  Post-financing funds will be invested in AI infrastructure, global commercialization, and harness product development, while the enterprise and developer base is growing rapidly.

  Weaknesses

  Expansion of inference workloads and model parameters may create ongoing capital and computing pressure.

  Comparison

  It is another Chinese AI lab financing case alongside DeepSeek and Zhipu.

  Risks

  Risks include commercialization, computing costs, model competition, and global execution.

• Tencent Holdings Ltd. (700 HK)

  Beneficiary of the cloud-services and internet AI ecosystem

  Strengths

  The report lists Tencent among the BAT beneficiaries and discloses a BUY rating.

  Weaknesses

  The report does not provide a standalone estimate of Tencent's AI revenue.

  Comparison

  It belongs to the same major internet platform beneficiary chain as Alibaba and Baidu.

  Risks

  Changes in AI capital expenditure returns, model competition, and the regulatory environment.

• StepFun

  Edge AI model developer

  Strengths

  It released the Step Edge foundation model, Step Edge Audio, Step Edge GUI, and Step Edge Gen for smartphone and automotive on-device agent scenarios.

  Weaknesses

  The report does not disclose StepFun's financing amount, valuation, or revenue data.

  Comparison

  Compared with cloud-based large models, StepFun emphasizes low latency, on-device data security, multimodality, and cloud-device collaboration.

  Risks

  On-device deployment performance, ecosystem compatibility, business model, and competitive pressure.

Key data

• Recent DeepSeek financingUSD7bnFinancing in June 2026, which the report expects to strengthen computing resources.
• Recent Zhipu financingUSD4bnFinancing in July 2026, emphasizing strategic AGI investment over the next two years.
• Recent MiniMax financingUSD2bnFinancing in July 2026, to be used for AI infrastructure, global commercialization, and product development.
• DeepSeek post-financing market capitalizationUSD50bnThe latest market capitalization after Friday's close, as disclosed in the report.
• Zhipu post-financing market capitalizationUSD94bnThe latest market capitalization after Friday's close, as disclosed in the report.
• MiniMax post-financing market capitalizationUSD10.8bnThe latest market capitalization after Friday's close, as disclosed in the report.
• Anthropic expected ARR超过USD60bnSemiAnalysis expects it to exceed this level in July 2026.
• Anthropic expected EBITUSD1bnSemiAnalysis expects this to be achieved in 3Q26.
• MiniMax enterprise and developer base超过1mExceeded 1m in June 2026, compared with approximately 200K in December 2025.
• MiniMax consumer product users约300m全球用户The global user base for consumer-oriented products, as disclosed in the report.

Impact & implications

If demand for AI models and agents continues to rise, token consumption, inference workloads, and AI infrastructure demand will expand in tandem, potentially benefiting cloud service providers and computing-related platforms. The cost-efficiency advantage of Chinese models may also increase adoption by developers and enterprises, driving commercialization in applications such as coding, office work, and long-form content. StepFun's edge models further indicate that AI capabilities are extending from the cloud to on-device environments, potentially improving real-time interaction, security, and multimodal experiences in smartphone and automotive scenarios.

Risks

• The token expenditure index remained weak in early July 2026, and some technology and internet companies reportedly imposed caps on employee token consumption, which may suppress near-term usage.
• If the efficiency of converting computing investment into revenue after AI lab financing falls short of expectations, ARR upside may be below the report's expectations.
• Intensifying price and capability competition between Chinese and US models may lead to lower API prices and compress profit margins.
• The deployment of on-device AI models in smartphone and automotive scenarios still faces challenges related to performance, power consumption, privacy and security, and ecosystem integration.
• The report discloses that Jefferies has investment banking relationships with some covered companies; investors should be alert to potential conflicts of interest.

What to watch

• Market acceptance of higher input and output prices during peak periods following the release of DeepSeek V4 on July 15, 2026.
• Whether the Silicon Data LLM Token Expenditure Index rebounds from the 1.63 to 1.65 range seen in early July 2026.
• Whether Anthropic's ARR exceeds USD60bn in July 2026 and achieves positive EBIT in 3Q26, as expected by SemiAnalysis.
• Progress in Zhipu's investment over the next two years in AGI, long-horizon task memory architectures, autonomous agents, self-learning, and safety.
• MiniMax's enterprise and developer base, global consumer user scale, and global commercialization progress.
• Actual deployment, latency performance, and cloud-device collaboration of StepFun Edge models in smartphone and automotive scenarios.