CSC: AI Application Flywheel Exceeds Expectations, Computing Power Investment Enters Return Verification Phase
Nashnova编辑部
China Securities Construction (CSC) argues the AI investment thesis is shifting from a capex arms race to commercial validation of orders and profits, with the application flywheel closing faster than expected — the market's question is moving from 'who spends the most' to 'who earns it back.'
Big Tech is spending more than it earns — what backs that up?
Google Cloud, Azure, and AWS posted revenue growth of 82%, 43%, and 37% year-on-year respectively. Microsoft's commercial remaining performance obligations hit $678 billion, up 84% YoY. This means → the heavy capex is not blind spending; it is underwritten by a massive backlog of signed-but-undelivered contracts.
Alphabet's Q2 capex reached $44.9 billion — $5.8 billion more than its operating cash flow. Tencent's capex surged 176% YoY to RMB 52.8 billion; Amazon's trailing-twelve-month capex likewise exceeded operating cash flow. In plain terms = these companies are spending more than they generate; the gap is bridged by contract backlogs and confidence in future revenue.
CSC argues the AI cycle's core tension should shift from "how much is being spent" to utilization rates, order conversion speed, and capital payback periods.
Have Chinese models actually caught up? How big is the remaining gap?
GLM-5.3 approaches Claude Opus 4.8 on select coding and agent benchmarks. The gap between Chinese and frontier overseas models has narrowed to roughly one minor version. This means → Chinese models have crossed from "usable" to "competitive" — the gap is no longer generational.
Kimi K3 activates only 16 experts yet improves training scaling efficiency by roughly 2.5× over its predecessor. DeepSeek DSpark lifts per-user inference speed by 60–85% under real production traffic. In plain terms = the same compute investment now delivers more than double the previous output.
Doubao's daily token volume has reached 180 trillion. Alibaba's AI-related products generate annualized revenue exceeding RMB 35.8 billion. Kimi K3's launch day set a company record for the largest single-day ARR — annualized recurring revenue — increase.
Is the Chinese model price war still on?
DeepSeek raised its V4-Pro peak-hour output price to RMB 27 per million tokens from August 17. Zhipu has lifted API prices by over 30% cumulatively this year. Kimi K3's output price reached RMB 100 per million tokens — roughly 3.7× that of K2.6. This means → leading vendors are shifting from "volume through discounts" to testing the market's willingness to pay more. The price-war phase may be nearing its end.
Yet even after the increases, DeepSeek V4-Pro's peak output price is only 33% of Gemini 3.1 Pro, 26% of GPT-5.4, and 16% of Claude Opus 4.8. In plain terms = prices are rising, but Chinese models remain far cheaper than overseas peers — the cost advantage holds, while the gross-margin ceiling has been lifted.
Enterprise AI deployment — how do you tell a product company from an outsourcer?
OpenAI now extends its field deployment engineers (FDEs — technical teams embedded at client sites to operationalize AI) from discovery through production go-live. Salesforce plans to hire 1,000 FDEs, typically in three-person squads serving a single client for months. This reflects that enterprise AI deployment still relies heavily on bespoke delivery — "plug and play" remains distant.
CSC highlights four tracking metrics: ARR supported per FDE, average delivery cycle, feature reuse rate, and post-project subscription revenue share. This means → if every deal requires building from scratch, the company is essentially an outsourcer; only those that codify recurring workflows into standard modules qualify as true product companies.
Whether a company can close the productization loop will be the core validation point separating long-term value in AI application firms over the next one to two quarters.
Content is for reference only, not financial advice.