Nvidia CEO Jensen Huang declared that the majority of economic value in artificial intelligence will come from specialized models trained on specific domains of knowledge, not general-purpose systems, according to remarks made at the Fireworks Forge event in San Francisco on Tuesday. Speaking with Lin Qiao, CEO and founder of Fireworks.ai, Huang said the chip giant remains neutral on whether open or closed AI models dominate the market, as long as its hardware powers the infrastructure. The event highlighted growing industry focus on reducing the runaway costs of deploying AI systems.

Huang emphasized that companies can achieve what he called "artificial superintelligence" by taking open frontier models and adapting them with domain-specific knowledge and specialized tools. Fireworks.ai, which offers a platform for businesses to train and deploy models customized to their particular needs, announced it had received 25,000 Nvidia GPUs to support its European expansion. Meanwhile, Qiao presented a series of ten headlines showcasing the extraordinary expense of uncontrolled token spending, warning that businesses can't scale without borrowing significant capital. Genspark, an AI workspace company, demonstrated this cost disparity with its Gen-1 Slides model, which delivers quality comparable to Anthropic's Opus 5 at one-seventeenth the price—$0.30 per million tokens versus $5 per million.

"We love closed models," Huang stated, adding that there's a solid place in the world for proprietary frontier models. However, he also argued that the notion of everyone relying on general intelligence isn't logical, given the world's diversity of good ideas. Qiao told attendees that "cost is out of control" and warned that business models can't scale without substantial borrowing, underscoring the platform's emphasis on affordable, open-model alternatives.

Huang's position reflects Nvidia's strategic calculation that specialized intelligence represents the real commercial opportunity, rather than betting exclusively on general-purpose AI systems. The CEO has previously defended open weight models against pressure from leaders at Anthropic and OpenAI, recognizing the economic appeal of AI services that offer greater affordability and less vendor lock-in. Looking ahead, Huang predicted that this year's breakthrough—agent harnesses that convert AI models into agents—will be followed next year by widespread adoption of what he called "Open Shell" for deploying agents safely, along with hardware-in-the-loop systems for robotics. The enterprise AI landscape appears poised to fragment into countless specialized applications rather than consolidate around a handful of general-purpose models, a shift that favors platforms enabling rapid customization over monolithic closed systems. For companies navigating this transition, the choice between open and proprietary approaches may ultimately matter less than the ability to translate domain expertise into purpose-built intelligence that solves specific problems efficiently.