Forward or Backwards

Topics › Computers › Chips

Silicon Photonics: Why AI Chips Need Light

Published · 21 min

Watch on YouTube

Why can an AI chip be ready to calculate and still wait? This video follows the connections between chips: copper signal losses, silicon waveguides, hybrid lasers, Intel’s research milestones, pluggable transceivers and co-packaged optics.

NVIDIA’s 2026 production claims are distinguished from independent operational evidence. Intel’s energy comparison is labelled as a vendor demonstration. We examine heat, manufacturing yield and repair, then ask whether light removes a real bottleneck or relocates it.

Educational documentary. Not financial or investment advice.

In these topics

Chapters

  1. The bottleneck nobody designed for
  2. Copper's wall
  3. Light on silicon: the idea
  4. Why silicon resisted light
  5. Intel’s lab breakthroughs: 2004–2010
  6. Pluggable optics: the stepping stone
  7. Co-packaged optics: moving light inside the package
  8. The 2026 production moment
  9. The promise: power, heat and uptime
  10. The catch: yield, repair and serviceability
  11. Who is building this, and who still relies on copper
  12. Does light actually remove the limit?

More from Forward or Backwards on YouTube

Sources and credits

Photo credits (Wikimedia Commons)

Primary sources

  • Intel, Introducing Intel’s Advances in Silicon Photonics, February 2004.
  • Intel Labs, The 50G Silicon Photonics Link, July 2010.
  • Miller, Integrated Optics: An Introduction, September 1969.
  • Soref and Lorenzo, silicon integrated optics papers, 1985 and 1986.
  • Intel, Optical I/O Chiplet, 26 June 2024.
  • NVIDIA, Vera Rubin Full Production, 31 May 2026.
  • NVIDIA, Q2 Fiscal 2027 results, 26 August 2026.
  • TSMC technology symposium release, 24 April 2024.
  • IBM Research, simulation-based CPO analysis, 8 January 2024.
  • Yi and Wilkerson, arXiv:2603.21313, 22 March 2026; preprint.

Not regulated financial advice.