Reports indicate that OpenAI and SoftBank Group have announced a joint $1 billion investment in SB Energy ($500 million each) to accelerate the construction of large-scale AI data centers and power OpenAI's "Stargate" project. SB Energy has shifted from renewable energy to data center development and operation, planning to launch two gigawatt-scale facilities this year; the Texas project will handle the new data center construction needs for Stargate, expanding computing power and electricity supply.

CNBC cites TrendForce and manufacturer statements, indicating that AI training/inference is driving HBM shortages and squeezing traditional DRAM capacity. TrendForce forecasts a 50%–55% increase in average DRAM prices in Q1 2026 compared to Q4 2025. Micron stated that HBM manufacturing occupies about three times the capacity of traditional memory, and its 2026 capacity is already sold out; the proportion of memory in overall system costs has risen from 10%–18% to about 20%, potentially pushing up prices for end devices like PCs.

An Apple research team proposed the GSM-Symbolic benchmark, which keeps the reasoning structure of problems unchanged, only replacing variables and numbers and adding information irrelevant to solving them, to test whether models rely on pattern matching. Testing over 20 models showed that variable changes and "distractor statements" significantly lower accuracy; among them, OpenAI o1 Preview accuracy dropped by 17.5%, and Microsoft Phi 3 dropped by up to 65% in relevant settings, exposing model vulnerability to symbolic changes and relevance judgment.

The Cyberspace Administration of China has drafted the "Regulations on Personal Information Collection and Use by Internet Applications (Exposure Draft)" and is soliciting public comments, with feedback due by February 9, 2026. The document aims to regulate personal information collection and use activities by internet apps and software development kits (SDKs), advancing based on higher-level laws such as the "Cybersecurity Law," "Personal Information Protection Law," and "Network Data Security Management Regulations." It seeks to strengthen personal information protection and clarify compliance boundaries, providing a policy basis for personal information compliance audits of internet products and AI applications.

InfoQ reports that MongoDB Server has a "MongoBleed" vulnerability CVE-2025-14847, where attackers can trigger a decompression length handling error by crafting malformed zlib compressed packets without authentication, reading adjacent sensitive data and credentials in heap memory, with active exploitation already reported. Wiz estimates 42% of cloud environments have vulnerable configurations, and Censys identified approximately 87,000 exposed servers. MongoDB Atlas has been updated; self-hosted users need to upgrade promptly or disable compression and tighten network exposure.

Multiple reports indicate that Chinese regulators are reviewing Meta's $2 billion acquisition of AI Agent company Manus. Manus is a Singapore-based company with Chinese backgrounds, and the review focus is interpreted as risks of AI technology and talent outflow and export control compliance. If required to submit additional materials or adjust the deal structure, it could affect Meta's acquisition pace in agents and platform-level AI and introduce greater uncertainty and longer transaction cycles for cross-border "team acquisitions/talent acquisitions."

Meshy announced Meshy AI Creative Lab at CES 2026, targeting non-CAD users to quickly convert generative 3D models into directly 3D-printable full-color files and physical fulfillment orders. The workflow includes: generating models with text-to-3D or image-to-3D in the Workspace; automatic mesh repair and geometry adjustment to meet physical manufacturing constraints; material, surface finish, and slicing parameter recommendations by use case; and integration with manufacturing partners for real-time pricing and fulfillment options. The official announcement states the features will roll out globally in batches starting January 2026.

THE DECODER compiled data released by Epoch AI, stating that global AI computing power has exceeded the scale of 15 million H100 equivalents, with related hardware's total power consumption exceeding 10 gigawatts. This metric quantifies the scale of training and inference hardware using an "H100 equivalent" standard, indirectly reflecting that the large model competition has spilled over from individual model capability to constraints in electricity, facilities, and the supply chain. For cloud providers and data center operators, electricity acquisition, energy efficiency, and deployment speed will directly impact the upper limit of available compute supply and unit costs.