Morning Brief - 4 November 2026

AI infrastructure costs are reshaping enterprise budgets while memory shortages bite consumer hardware, and China's semiconductor and AI ecosystems hit new milestones overnight.

Enterprise AI budgets buckle as agentic workflows burn tokens 10x faster than chat

Uber exhausted its annual AI allocation in four months as Claude Code and similar agentic tools drove token consumption 10x to 50x beyond conversational queries. GitHub has moved from unlimited seat licenses to metered credits, and Gartner now forecasts token expenses for AI-assisted coding will exceed typical developer salaries by 2028 in many cases. Read more →

DRAM prices double as memory makers prioritize AI data centres over consumer PCs

DRAM pricing has climbed to approximately twice its December 2025 baseline as Samsung, SK Hynix, and Micron reallocate fabrication capacity toward high-bandwidth memory and server-grade DRAM for AI infrastructure. The shift is not a supply shock but a calculated move driven by hyperscale data centre demand from US cloud providers. Read more →

Western startups strip censorship from Chinese models, but few buyers see the need

Firms like Hirundo and Lazarus AI are editing weights in DeepSeek, Qwen, and other Chinese open models to remove embedded censorship and geopolitical bias for Western enterprises. The pitch trades compelling performance at lower cost against alignment shaped by Beijing's content controls, but most commercial users see little reason to modify the models. Read more →

Trajectory raises $45M to let enterprises train specialized AI through feedback loops

The Stanford-founded startup backed by Sequoia and Nvidia wants law firms, financial analysts, and healthcare providers to fine-tune models for tasks like contract review without hiring expensive consultants. The 18-person team raised the round roughly a year after founding, aiming to close the gap between general capabilities and domain-specific work. Read more →

ChangXin Memory samples 4F² DRAM, diverging from the 6F² path rivals adopted

The Hefei chipmaker plans to ship samples of server-grade DRAM built on a 4F² cell architecture before year-end, a design Micron, Samsung, and SK hynix abandoned years ago. The choice places CXMT on a different technical trajectory, driven as much by US export controls and fab tool access as by engineering considerations. Read more →

Google bans Videostrong after 1.3 million Android devices turned into criminal botnet

The Shenzhen manufacturer deliberately embedded backdoors in Android TV boxes, telehealth equipment, karaoke machines, and digital signage over multiple years, operating them as nodes for ad fraud, data theft, and unauthorized network routing. Google has terminated the partnership and dismantled the infrastructure after discovering the systematic integration. Read more →

China's inference workloads now exceed training volumes as deployment scales

Inference generated by deployed AI systems across China has surpassed the volume of data used to train those models, marking a shift in how compute resources are allocated. The milestone comes from the China Centre for Information Industry Development, an arm of the Ministry of Industry and Information Technology, and signals that the centre of gravity has moved downstream. Read more →

Empyrean acquires Easy-Logic for $33M to shore up chip verification capabilities

China's largest electronic design automation vendor will acquire a 50.38 percent stake in the domestic verification specialist for 218.8 million yuan, addressing a structural weakness in the country's chip-design stack. The move comes as the semiconductor ecosystem rallies around Huawei's Tau Scaling roadmap and consolidation pressure mounts. Read more →

Compiled by OTW Briefing Desk from Opentechwire reporting.