Speaking at Fireworks Forge in San Francisco on Tuesday alongside Fireworks.ai CEO Lin Qiao, Nvidia CEO Jensen Huang said Nvidia has no preference in the open versus closed model debate, defending both proprietary and open-weight approaches while arguing that specialized domain intelligence, not general-purpose models, captures most economic value. Huang said "We love closed models" and that there is a solid place for closed frontier models, but argued that adapting open frontier models with proprietary domain knowledge, tools, and evaluation lets companies create what he called "essentially a superintelligence" for their specific needs. Fireworks announced a 25,000-GPU infusion from Nvidia to support its European expansion. Qiao opened the event with ten headlines documenting runaway token costs, saying "Cost is out of control" and that business cannot scale without borrowing a lot of money. Genspark co-founder and CTO Kay Zhu presented Gen-1 Slides, built atop MiniMax M3 and tuned to match Anthropic's Opus 5 quality at $0.30 per million tokens versus Opus 5's $5 per million, a 17x reduction. Huang pointed to agent harnesses as this year's advance and offered two predictions: widespread safe deployment of agents via "Open Shell" and "hardware in the loop."