Overview
- DeepSeek on Tuesday open-sourced its production inference stack DeepInfer to help operators and developers lower the cost and energy of large-model serving by improving distributed inference scheduling and runtimes.
- At GTC Berlin on Wednesday NVIDIA proposed using telco base stations as distributed AI nodes and unveiled the DSX POD modular AI-factory reference design, which pre-integrates power, networking and liquid cooling so sites can go from ready infra to production in as little as 14 days.
- Google’s CEO said the company is reallocating cloud compute toward frontier AI research and that internal 24/7 AI agents are already used to monitor work, a stance that signals tighter near-term capacity and makes multi-year capacity commitments harder for customers.
- Investors responded favorably to recent product and earnings news for several AI firms, lifting tech stocks, but market participants remain wary of financing strains in the sector after debt stress at some cloud and infrastructure players and heavy leverage at major backers.
- Specialized embodied-AI firms and chipmakers are securing capital and reporting progress — for example, Huixi announced sizeable equity raises to scale embodied-AI SoCs and Yushù said G1 robot volumes and GPT-6 tests show early zero-shot control gains — a sign that physical-AI commercialization is accelerating.