CITIC Securities: Domestic Large Models Enter Top Tier Across the Board; Focus on Computing Power and Cloud Industry in H2
Miles Bennett
A CSC research report concludes that China's domestic large models have entered the global first tier across the board; surging model capability is driving exponential Token consumption growth, making compute infrastructure and cloud the clearest investment theme for H2.
What paths are the three overseas giants taking?
OpenAI is pivoting to enterprise monetisation. GPT-5.5 hit 82.7% accuracy on Terminal-Bench 2.0 — up over 7 percentage points — and supports up to 1M tokens of context (how much text the model can process in one go).
This means → OpenAI's focus has shifted from "build the strongest model" to "get enterprises actually using it." Ultra-long context is built for complex corporate document workflows.
Anthropic locked down finance, healthcare, and other high-barrier enterprise verticals early through code and compliance strengths. Its unreleased flagship Mythos may reach 10T parameters, priced at $25 / $125 per million tokens input / output — five times the Opus line.
Google is covering every lane: Gemini 3.1 Pro leads ARC-AGI-2 at 77.1% accuracy, yet its API stays at a flat $2 / $12 per million tokens. In plain terms = Google is trading price for scale — the same playbook as giving Android away for free.
What backs the claim that Chinese models have reached the top tier?
CSC notes that Alibaba, DeepSeek, Kimi, MiniMax, Xiaomi, Zhipu, and other leaders have sharply accelerated iteration this year, closing gaps in multimodal input, long context, complex-task execution, and coding.
The key lever is architectural innovation at the base layer — advances in compression and attention mechanisms (the core component that decides which parts of the input the model focuses on) have let domestic models overtake overseas rivals in select multimodal scenarios.
This means → the gap is no longer "can or can't" but "who does a given scenario better." Capabilities are converging at a high level; competition has moved to real-world deployment.
Token consumption is surging — who benefits first?
Stronger models → higher adoption → exponential Token consumption growth → steep rise in compute demand. This causal chain is the core of CSC's H2 investment thesis.
This reflects a critical inflection: large models are moving from "tech demo" to "scaled consumption," and compute shifts from "powerful enough?" to "plentiful enough?"
In plain terms = the better models get, the more calls they field, and the tighter the servers and chips behind them become — compute infrastructure and cloud are the most direct beneficiaries.
How should investors position for H2?
AI track: CSC recommends following demand, with three priorities — compute facing price hikes and shortages, cloud infrastructure driving efficiency, and select high-momentum applications.
This means → the key validation point is whether relevant companies can lock in capacity during the supply-tight window. Securing capacity is the prerequisite for earnings upside.
Non-AI track: policy-driven digital-currency 2.0, autonomous driving, and commercial spaceflight — CSC flags these as structural opportunities independent of the AI cycle.
Content is for reference only, not financial advice.