GTC: AI factories & AttnRes
Nvidia’s GTC pushed the ‘AI factory’ narrative and the idea of ‘physical AI,’ while analyst recaps flagged a surveillance paradox on event portals — lots of openness on models, little openness on infrastructure. (youtube.com) (youtube.com) (x.com)
Jensen Huang used the GTC keynote to boost Nvidia’s market projection, raising the company’s demand signal to $1 trillion through 2027. (atlan.com) Nvidia unveiled what industry coverage described as a five‑layer AI platform and introduced the Vera Rubin compute system, which the company says delivers roughly 10× inference performance per watt versus prior generations. (forbes.com; atlan.com) The company published a “Physical AI Data Factory Blueprint” for collecting and curating real and simulated robot and vehicle data and listed partners including FieldAI, Hexagon Robotics, Linker Vision, Milestone Systems, Skild AI, Uber, and Teradyne Robotics in the reference design. (automate.org) On the show floor Nvidia demonstrated more than 110 robots and announced RoboTaxi collaborations with BYD, Hyundai, Nissan and Geely as examples of companies that will use its new data‑collection and inference tooling. (atlan.com) Multiple analyst write‑ups warned that open‑weight model releases at GTC come paired with heavy CUDA and stack optimizations that effectively tie peak performance to Nvidia’s infrastructure, creating what commentators called an “open models, closed infrastructure” lock‑in dynamic. (byteiota.com; siliconangle.com) Separately, the Attention Residuals (AttnRes) method appeared on arXiv this week as a proposed drop‑in replacement for fixed residual aggregation that uses softmax attention over prior layer outputs, and the MoonshotAI GitHub repo provides reference code for Full and Block AttnRes implementations. (arxiv.org; github.com)
