AI Daily Brief — 03 January 2025
Friday delivered the first big number of 2025: Microsoft’s Vice Chair and President Brad Smith pledged roughly $80 billion in AI-datacenter capex for the firm’s fiscal year, with more than half landing in the United States. The post doubled as Microsoft’s opening lobbying salvo to the incoming Trump administration. arXiv added three notable papers, including Tencent’s omni-modal VITA-1.5 and RUC’s open attempt to reproduce o1-style slow thinking for multimodal models.
Top stories
- Microsoft commits ~$80B to AI-enabled data centers in FY2025. Brad Smith published “The Golden Opportunity for American AI” on the corporate blog, sketching a three-part national strategy — infrastructure investment, AI skilling, and exports — and disclosing that “more than half” of the $80B FY25 spend will be in the US. The single largest hyperscaler AI-capex anchor announced to date. via Microsoft
- Markets and trade press amplify the $80B number. CNBC, Data Center Dynamics, and Network World all carried the disclosure on Jan 3, framing it as the early signal that 2025 hyperscaler capex would dwarf 2024 — and as pre-positioning for the AI Diffusion Rule expected from the outgoing Biden team within ten days. via CNBC
- VITA-1.5 (Tencent / NJU) targets GPT-4o-level real-time vision and speech interaction. A 16-author technical report on end-to-end training and serving of a multimodal LLM combining vision, speech input, and real-time speech output. via arXiv
- Virgo (Renmin University / BAAI) probes whether multimodal LLMs can be given o1-style slow thinking. Fine-tunes a VLM on textual reasoning traces and finds that long-form chain-of-thought transfers into the vision setting — an early open attempt to reproduce o1 in multimodal land. via arXiv
- Auto-RT (ISCAS) automates jailbreak-strategy search for red-teaming LLMs. Removes the human-crafted-attack bottleneck and searches the strategy space directly. via arXiv
Money, infra & hardware
Microsoft’s $80B capex figure dominated the day’s wire, but it was a Friday in a pre-CES week — chip vendors held back. NVIDIA, AMD, and Intel kept their announcements for the CES week starting Jan 6. Baidu’s Kunlunxin filing from Jan 2 continued reverberating in Asia. ExecutiveBiz read Smith’s post as Microsoft’s bid to shape the new administration’s AI export framework before Biden’s exit.
Who shipped
Frontier labs: still quiet on the products front. OpenAI, Anthropic, Google DeepMind, Meta, xAI: no dated releases. Sam Altman’s “Reflections” essay would arrive Sunday Jan 5.
Open-source pulse
The open-weights newswire stayed dark. Chinese labs were holding releases for the pre-CNY window; the first big drop — MiniMax-01 — was twelve days away.
By the numbers
- $80B — Microsoft FY2025 AI-datacenter capex pledge
- 3 notable arXiv papers (VITA-1.5, Virgo, Auto-RT)
- Most-mentioned company: Microsoft
- Notable absences: NVIDIA, AMD, Intel (CES embargo)
Compiled by AI Feed’s editor from verified web sources for 3 January 2025.