NVIDIA CEO Jensen Huang described the practice of using answers from other artificial intelligence (AI) models for training, known as “distillation,” as “competition” rather than theft. This perspective contrasts with U.S. government claims that Chinese AI companies are extracting technology by distilling models from American firms.
In a CNBC interview on the 28th, local time, Huang responded to a question about whether AI distillation constitutes theft by stating, “That’s competition.” He added, “You can test other companies’ products as much as you want,” and noted, “Competitors even dissect our products until only the skeleton remains to understand how they work.”
Huang further remarked, “It would be nice if no one learned from our products and we could continue business comfortably, but honestly, competition makes everything better.” He also suggested that if unwanted use occurs, companies can verify customers and block services.
AI model distillation involves using outputs from high-performing models as training data to improve others. The U.S. government has argued that Chinese firms are leveraging large-scale distillation to replicate advanced models from OpenAI, Anthropic, and others, citing potential sanctions. The Chinese government denies these claims, calling distillation a normal technical practice. Huang’s remarks diverge from the U.S. stance.
Analysts suggest Huang’s position aligns with a business strategy where more companies developing their own models and services benefits NVIDIA. Distillation allows firms to create task-specific models without incurring massive costs to build state-of-the-art models from scratch. As more companies adopt this approach, demand for computing resources—both for model development and ongoing AI service operations—could grow.
NVIDIA has released its open model, NemoTron, to encourage companies to build AI services using its technology. While firms like OpenAI and Anthropic are sensitive to competitors mass-collecting their models’ outputs, NVIDIA benefits from increased AI adoption, as it supplies graphics processing units (GPUs) and related software regardless of who develops the models.
Leave a comment