CSC October AI Monthly: Accelerating Model Iteration, Agent Commercialization Gaining Momentum
nashnova research
CICC's October AI report finds model iteration uninterrupted by safety debates, with large-scale Chinese funding sustaining a capability catch-up and Agents driving a new ARR growth cycle by raising per-user task density.
How fast are overseas models iterating?
OpenAI and Anthropic released GPT-6 Sol/Luna, Opus 5.5 and GPT-6.1 Sol in quick succession. GPT-6.1 Sol approaches GPT-6 Astra in capability at one-fifth the cost.
This means → iteration is compressing cost and raising the capability ceiling at the same time. "Stronger and cheaper" is no longer aspirational.
As of August, Claude-supervised work accounted for 26% of Anthropic's R&D output. In plain terms = AI is now helping build AI — R&D automation is feeding back into iteration speed itself.
Where does overseas monetisation stand?
Anthropic deepened partnerships with Salesforce and Bain, pushing Agents — AI programs that execute multi-step tasks autonomously — into enterprise workflows. ChatGPT's ad-revenue run-rate has reached $1 billion annualised.
Muse is expanding the personal-task entry point, unlocking consumer-side commercial value. This means → monetisation no longer relies on enterprise clients alone; consumer willingness to pay is being validated too.
Responsible Scaling Policy (RSI) — a self-regulatory framework for managing AI risk — is emerging as a consensus standard among OpenAI, Anthropic and peers.
What is fuelling China's catch-up?
Alibaba completed a ~HK$80 billion share placement; Zhipu closed ~$5 billion in equity-and-debt financing. Tencent ramped up overseas compute procurement. This reflects Chinese AI firms trading large-scale capital for a time window.
On the commercial side, MiniMax's July token consumption hit 20× its January level; Zhipu raised its year-end ARR target from $2.4 billion to $3 billion.
The report argues the next growth phase hinges on effective task volume, retention and paid conversion. In plain terms = user numbers alone are not enough — what matters is whether users keep using the product and keep paying.
How has the market performed over two months?
From July 31 close to September 30: IGV rose 12.6%, the Philadelphia Semiconductor Index rose 11.6%. A-share telecom and electronics gained 4.7% and 3.2%; computing and media fell 5.2% and 3.2%.
On valuation, IGV's composite PS-TTM — price-to-sales based on trailing twelve months — stood at 8.59×. Shenwan computing and media traded at 2.80× and 2.68×, down 10.1%, 21.6% and 18.6% respectively from end-2025.
This means → overseas software is seeing valuation recovery backed by AI-driven payment and data-call growth, while Chinese application-layer performance remains split — no unified rally yet.
What does CICC recommend watching?
The report highlights three threads: compute services, domestic chips and super-nodes, plus B-side AI application vendors with proven scenario, data and enterprise-delivery capabilities.
CICC stresses that future pricing will weigh customer procurement, sustained usage and profit contribution more heavily. In plain terms = the market is no longer asking "who is doing AI" — it is asking "whose AI is actually being bought, and bought again."
Whether companies can sustain ARR growth delivery will be the key checkpoint for further valuation recovery.
市场有风险，内容仅供研究参考，不构成投资建议。