Computer Industry Weekly: AI Investment Refocuses on the Intelligence Frontier of Large Models
Summaries are public. PDF access requires an active membership; all members have the same access. Sign in
Guolian Minsheng Securities maintains a 'Recommended' rating on the computer sector, arguing that the dense run of model updates since August 2026 refocuses AI competition on the intelligence frontier. GPT-6 Astra lifted Terminal-Bench Science 0.1 from 22.4% (GPT-5.6 Sol) to 64.6%, and Claude Fable 5.1 raised Humanity's Last Exam from 55.5% to 59.1%, suggesting frontier capability has not hit a clear plateau. Scaling continues: GLM-5.3, built on a base identical to GLM-5.2, gained over 50% on internal Z.ai Code Bench purely through post-training; Kimi K3 reaches 2.8 trillion parameters; GPT-6 Astra was pretrained on more than 100,000 GPUs at Stargate. Domestic 'Chinese models plus Chinese chips' is the main line: Zhipu runs low-cost inference on 100,000-card domestic clusters, with GLM-5.3-Flash drawing over 62 trillion tokens in six days of anonymous testing, per-token inference cost down 80% since the start of the year, and its 'compute multiplier' up about 14x YoY; first-half revenue reached RMB954m, up nearly 400%, with API revenue of RMB825m, up 2,736%. The MIIT's new AI application service-provider program targets 2,000+ providers by end-2026 and 3,000+ by end-2027. DeepSeek's price increase opens a domestic volume-and-price commercialization window, favoring open-source model firms and compute-leasing vendors. Risks: AI policy shortfall, intensifying competition.
Key exhibits
Exhibit 1
Exhibit 2
Exhibit 3
Exhibit 4
Exhibit 5
Exhibit 6
Exhibit 7
Exhibit 8
