Nvidia's CEO, Jensen Huang, has stated that a promising new chip technology, co-packaged optics, is not yet reliable enough for use in the company's flagship GPUs.

The technology, which uses laser beams to transfer data via fiber optic cables instead of traditional copper, is more energy-efficient and faster.

However, Huang emphasized that copper connections remain 'orders of magnitude' more reliable than today's optical alternatives, making them the preferred choice for now.

Speaking at Nvidia's annual developer conference in San Jose, Huang announced that the company will use co-packaged optics in two upcoming networking chips designed for server switches, increasing their energy efficiency by three and a half times.

These switch chips will be released later this year and into 2026, marking a gradual technological step forward. However, Huang clarified that Nvidia currently has no plans to implement optical connections between GPUs, as reliability remains a key priority for its AI-focused customers like OpenAI and Oracle.

Silicon Valley startups such as Ayar Labs, Lightmatter, and Celestial AI have invested heavily in co-packaged optics, seeing it as essential for building more powerful AI systems. Nvidia itself has backed some of these ventures, despite Huang's cautious approach.

While optical connections could eventually help AI models process complex tasks more efficiently, Nvidia is prioritizing proven technology for its near-term roadmap, ensuring stability in an industry preparing to invest hundreds of billions in AI infrastructure.