🎯Nvidia Pushes Specialized AI Models as Fireworks Gets 25,000 GPUs

Nvidia wants your specialized model bill, open or closed

Nvidia doesn't care which models you run. Jensen Huang says he prefers closed models but will happily take your money either way, because he thinks most of the value lives in specialized intelligence, not general-purpose chatbots. That's a convenient position for the company selling the GPUs. If you're shipping AI features, this is a signal about where the money is going. Fireworks, a platform for training and running models tuned to a company's specific needs, just got 25,000 Nvidia GPUs to expand in Europe. Revolut, the British fintech, went self-hosted and slashed its AI coding costs 10x. The pattern is clear: teams are tired of paying frontier-model prices for narrow tasks. The numbers are the story. Gen-1 Slides, a slide-deck generator built on MiniMax M3, was tuned to match Anthropic's Opus 5 on results at 1/17th the cost, priced at $0.30 per million tokens. Huang also argued that open frontier models can be adapted with domain knowledge to reach what he calls superintelligence. Treat that last claim as a sales pitch. The cost data is the part you can actually test.