AI financing and rising token demand drive ARR growth potential

Overview

This report is the 56th issue of Jefferies' Internet AI Series, focusing on recent AI lab financing, ARR growth potential, cost differences between Chinese and US models, and the launch of StepFun's edge models. Its core view is that rising demand for AI models and agents is driving token consumption growth, while expanded computing resources following financing will increase the commercialization opportunities for AI labs and cloud service providers.

Core views

The report believes that recent financings by DeepSeek, Zhipu, MiniMax, and others will strengthen computing investment and support model demand and commercialization expansion. SemiAnalysis' ARR and profitability forecasts for Anthropic indicate that leading overseas AI labs continue to grow rapidly. Chinese models have clear advantages over US models in API costs, which may drive volume growth in use cases such as Coding, Workspace, and medium- to long-form content. Beneficiaries include Baidu, Alibaba, Tencent, Kingsoft Cloud, AI labs, and Kling.

Analysis framework

The report combines financing events, ARR forecasts, token consumption indices, model API pricing comparisons, and product launch information to assess demand intensity, commercialization elasticity, and potentially benefiting assets across the AI ecosystem.

Methodology notes

• Demand trackingSilicon Data LLM Token Expenditure Index

  Token expenditure index

  The index is weighted by usage concentration, with expenditure rising when demand shifts toward premium models; the report uses it to observe changes in LLM token demand and pricing structure.

• Revenue growth assessmentARR analysis

  Annual recurring revenue

  The report uses Anthropic's ARR milestones, API revenue share, and net dollar retention rate to assess the AI lab's revenue expansion speed and profitability potential.

• Model cost comparisonAPI price comparison

  Comparison of API costs for Chinese and US models

  The report compares blended prices for Chinese and US models and attributes the cost advantage of Chinese models to architectural efficiency from MoE, attention mechanisms, MFU, and other factors.

• Valuation and rating frameworkJefferies Ratings and Valuation Methodology

  12-month expected total return and multi-method valuation

  Jefferies' rating methodology may incorporate market capitalization, maturity, growth/value characteristics, volatility, and expected total return over the next 12 months; target prices may be based on DCF, EV/EBITDA, P/E, FCF, SOTP, NAV, ROE, and other methods.

Asset mapping & comparison

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

• Baidu Inc. / 9888 HK

  Beneficiary of the BAT cloud and AI ecosystem

  Strengths

  Benefits from rising demand for AI models and agents, growing token consumption, and the cost-efficiency advantage of Chinese models.

  Weaknesses

  The report does not provide standalone ARR, gross margin, or new target price analysis for Baidu.

  Comparison

  Part of the CSP beneficiary group highlighted by the report, alongside Alibaba and Tencent.

  Risks

  Fluctuations in model demand, cloud pricing competition, corporate token consumption controls, and macro market volatility.

• Alibaba Group Holding Limited / BABA / 9988 HK

  Beneficiary of the BAT cloud and AI ecosystem

  Strengths

  The report specifically identifies Alibaba as benefiting from AI model and agent demand growth; Qwen is listed as one of the Chinese models with high cost efficiency.

  Weaknesses

  The report does not elaborate on Alibaba's standalone earnings elasticity or target price changes.

  Comparison

  Part of the CSP beneficiary group alongside Baidu and Tencent.

  Risks

  API price competition, computing investment costs, the pace of enterprise demand, and the regulatory environment.

• Tencent Holdings Ltd. / 700 HK

  Beneficiary of the BAT cloud and AI ecosystem

  Strengths

  The report specifically identifies Tencent as benefiting from growing token consumption and rising AI agent demand.

  Weaknesses

  The report does not provide standalone model pricing, ARR, or earnings forecasts for Tencent.

  Comparison

  A beneficiary of the Chinese internet and cloud ecosystem alongside Baidu and Alibaba.

  Risks

  The pace of model commercialization, investment costs, industry competition, and market valuation volatility.

• Kingsoft Cloud / KC

  Beneficiary of cloud services and computing demand

  Strengths

  The report lists Kingsoft Cloud as a beneficiary of growing token consumption and AI model demand.

  Weaknesses

  Its scale and ecosystem coverage may be smaller than those of BAT; the report does not provide standalone financial forecasts.

  Comparison

  More exposed to cloud infrastructure elasticity, differing from BAT's broader internet ecosystem attributes.

  Risks

  Cloud service pricing pressure, capital expenditure requirements, customer concentration, and weaker-than-expected AI demand.

• Kuaishou Technology / 1024 HK

  AI application and content ecosystem-related company

  Strengths

  Listed as BUY on the report's disclosure page, with Kling identified as a beneficiary that may benefit from video generation and AI content demand.

  Weaknesses

  The main body of the report does not provide a detailed breakdown of Kuaishou's or Kling's revenue contribution.

  Comparison

  More focused on AI application and content generation scenarios than CSPs.

  Risks

  Uncertainty surrounding AI product commercialization, content regulation, competition, and the disclosed investment banking relationship between Jefferies and Kuaishou.

• MiniMax Group Inc. / 100 HK

  AI lab and application platform

  Strengths

  Recently raised USD2bn; its enterprise and developer base exceeds 1m, while consumer products have approximately 300m global users; the funds will be invested in AI infrastructure, global commercialization, and product development.

  Weaknesses

  Expanding model parameters, agentic AI applications, and inference workloads will require substantial computing investment.

  Comparison

  A Chinese AI lab financing expansion example alongside DeepSeek and Zhipu.

  Risks

  Computing costs, commercialization execution, model competition, and risks associated with global expansion.

• DeepSeek

  Chinese AI lab

  Strengths

  Recently raised USD7bn, with the report expecting the new capital to strengthen computing resources; it plans to double input/output prices during peak periods upon the release of V4.

  Weaknesses

  Price increases may affect some demand; the report does not disclose detailed revenue or earnings forecasts.

  Comparison

  A recent Chinese AI lab financing case alongside Zhipu and MiniMax.

  Risks

  Model competition, price sensitivity, computing bottlenecks, and demand fluctuations.

• Zhipu

  Chinese AI lab

  Strengths

  Recently raised USD4bn; the report says its strategic focus is AGI, with planned investment in long-horizon task memory architectures, autonomous agents, fully self-learning systems, and safety.

  Weaknesses

  The report notes that Zhipu prioritizes strategic AGI investment over the next 2 years rather than short-term commercialization, so near-term earnings visibility may be limited.

  Comparison

  Its post-financing valuation of USD94bn is higher than the latest valuations of DeepSeek and MiniMax cited in the report.

  Risks

  Long AGI investment payback periods, uncertainty over the technology roadmap, delayed commercialization, and intensifying competition.

• Anthropic

  Growth reference for a leading overseas AI lab

  Strengths

  SemiAnalysis expects July 2026 ARR to exceed USD60bn, API revenue to account for 75%-85%, API gross margin to exceed 80%, and EBIT to reach USD1bn in 3Q26.

  Weaknesses

  High growth depends on continued customer expansion and API demand, while its valuation level is already very high.

  Comparison

  Serves as a benchmark for overseas AI labs and is used to compare the financing and ARR upside potential of Chinese AI labs.

  Risks

  Growth forecasts falling short, computing costs, competition, valuation digestion, and slower customer demand.

• StepFun

  Edge AI model provider

  Strengths

  Launched the Step Edge foundation model, Step Edge Audio, Step Edge GUI, and Step Edge Gen for smartphone and automotive scenarios, emphasizing low latency, data security, and edge-cloud collaboration.

  Weaknesses

  The report does not disclose financing amount, revenue scale, or profitability metrics.

  Comparison

  Compared with cloud-based large model companies, StepFun places greater emphasis on edge-device agent scenarios.

  Risks

  Uncertainty surrounding edge deployment scale, hardware adaptation, model performance, commercialization, and ecosystem partnerships.

Key data

• Recent DeepSeek financingUSD7bnThe report says DeepSeek completed the financing in June 2026.
• Recent Zhipu financingUSD4bnThe report says Zhipu completed the financing in July 2026 and plans strategic AGI investment over the next 2 years.
• Recent MiniMax financingUSD2bnThe report says MiniMax completed the financing in July 2026, to be used for AI infrastructure, global commercialization, and product development.
• Latest DeepSeek valuationUSD50bnThe report describes this as the latest market cap after the financing, as of Friday's close.
• Latest Zhipu valuationUSD94bnThe report describes this as the latest market cap after the financing, as of Friday's close.
• Latest MiniMax valuationUSD10.8bnThe report describes this as the latest market cap after the financing, as of Friday's close.
• Anthropic ARR forecastExceeding USD60bnSemiAnalysis expects Anthropic's ARR to exceed this level in July 2026.
• Anthropic expected EBITUSD1bn in 3Q26SemiAnalysis expects Anthropic to achieve positive EBIT in 3Q26.
• Anthropic API revenue share75%-85%SemiAnalysis estimates that most of Anthropic's revenue comes from API models.
• Anthropic API gross margin forecastExceeding 80%SemiAnalysis expects API model gross margin to exceed 80%.
• MiniMax enterprise and developer baseExceeding 1mIt exceeded 1m in June 2026, compared with 200K in December 2025.
• MiniMax consumer usersApproximately 300m global usersThe report says its consumer-facing products reached approximately 300 million global users.
• Step Edge latencyas little as 0.1 secStepFun's edge models target smartphone and automotive scenarios, emphasizing low latency and edge-cloud collaboration.

Impact & implications

If token consumption continues to grow while Chinese models maintain their cost advantage, cloud service providers, AI labs, AI applications, and the edge AI ecosystem may gain opportunities for revenue growth and valuation re-rating. Financing expansion should help increase computing supply, but corporate token consumption limits, model price changes, and intensifying competition may affect the pace of growth.

Risks

• Employee token consumption limits imposed by enterprises and internet companies may suppress short-term demand.
• Price competition among AI model APIs may affect revenue growth and profit margins.
• Expanding computing resources requires continued capital investment; if financing or supply chains fall short of expectations, model iteration and service capabilities may be affected.
• ARR, EBIT, and valuation forecasts depend on third-party estimates and are subject to assumption bias and execution shortfalls.
• The report contains general investment research disclosures and does not constitute personalized advice for any individual investor.
• Jefferies discloses an investment banking-related relationship with Kuaishou Technology; investors should be aware of potential conflicts of interest.

What to watch

• Whether Anthropic's ARR exceeds USD60bn in July 2026 and whether it achieves positive EBIT in 3Q26.
• The impact of DeepSeek V4's doubling of peak-period input/output prices on demand and revenue after launch.
• Technological progress and commercialization pace of Zhipu's AGI strategic investment over the next 2 years.
• MiniMax's enterprise and developer base, consumer user scale, and global commercialization progress.
• Whether the Silicon Data LLM Token Expenditure Index recovers from its weak range in early July.
• Whether the API cost advantage of Chinese models over US models persists.
• The pace of adoption of StepFun's edge models in smartphone and automotive scenarios.