DeepSeek's Three-Move Combo Crashes Agent Costs — LLM Era Shifts to Order War
What This Is
DeepSeek fired three moves in September, driving Agent deployment costs to the floor — V4.1-Flash cuts KV Cache (inference intermediate cache) VRAM requirements to 1/4 and disk requirements to 1/8, while the open-source Harness racked up 235,000 stars in three weeks, with OpenCode and CodeBuddy simultaneously adopting it as their default foundation. The bigger context matters more: a September industry report laid it bare — "2026 AIGC (AI-generated content) stops showing off and starts grabbing orders." China's LLM main storyline is shifting from "scoring one point higher on benchmarks" to "who collects revenue first."
Industry View
The revenue signal is loud: Kuaishou Kling's overseas revenue hits 70% with valuation approaching 100 billion RMB; WPS 365's revenue has grown over 60% for six consecutive quarters with 48 million individual paying users; Doubao started taking commission on hotel orders in August; even a banking-sector Agent project won a 65 million RMB bid. These numbers prove AI is genuinely turning into business. Cold water is also warranted. Kimi K3, despite carrying the "world's largest open-source model" halo, saw monthly active users drop to 7 million — proving model strength alone can't retain users; general-purpose Agent company Manus's 14 billion RMB acquisition collapsed, with large operations and thin revenue making any acquirer feel the pain; a Deloitte survey shows one-third of users have a trust deficit, with pure AI content preference down to just 26%; on the regulatory side, generative synthesis labeling rules have taken effect, and AI companion products have faced multiple regulatory talks. One engineer's line stuck with us: "Don't look at benchmark scores — look at who's willing to pay for mistakes."
Impact on Regular People
- Enterprise IT: Per-inference Agent costs keep falling, the "build vs. buy" scale is rebalancing, and teams who can deliver engineering-grade deployments will see their value rise.- Individual careers: Traditional CRUD (Create, Read, Update, Delete) roles are indeed contracting, but people who can weave AI into business workflows will command higher pay — tool-use proficiency is becoming the new dividing line.- Consumer market: The free-app heyday is over; subscriptions, pay-per-use, and commission models will become mainstream; AI video drops a single clip's cost from tens of thousands to a few dozen yuan, and SMB merchants will be the first to put real money down.