Photonic interconnects will become the standard for AI data centers by 2028, replacing electrical interconnects for GPU-to-GPU communication

Confidence: 6/10

Rescued from the retired optical-computing topic on 2026-08-06. When that topic was deleted on 2026-08-05 this thesis had a live position behind it — LITE (Lumentum), a published BULLISH call — and no rung to sit on. It is the one thesis of the eleven on deleted topics that carried real money, and it is why the hardware scope was corrected to keep optical interconnect in while ruling photonic compute out. With networking now a rung, it has a proper home.

The distinction that keeps it alive: communicating in light is in scope; computing in light is not. Its sibling — optical-computing T2, "photonic compute achieves commercial viability" — was retired with the topic and stays retired.

Supporting evidence: carried over from the retired topic. The evidence base needs rebuilding on this rung: the sources that grounded it (photonic-tensor-processor-pytorch, photonic-backprop-onchip-training) moved here with it, but both are photonic-compute papers, which is the half now out of scope. This thesis currently has a live bet and no in-scope source under it — the clearest ingest target on the rung.

Update 2026-09-24: three in-scope sources now sit on the rung, all vendor-stated: Broadcom's co-packaged-optics switch, NVIDIA's CPO switch announcement (which names Lumentum as a supplier), and NVIDIA's CPO blog (pluggable optics "often 30W per interface" against co-packaged "as low as 9W"). They support co-packaged optics displacing pluggable modules at the switch, which cuts both ways for a thesis about total optical content. None tests the 2028 GPU-to-GPU claim. See frontier and silicon photonics.