The next AI infrastructure bottleneck isn't compute — it's moving data at energy costs transistors can't sustain
We talk constantly about scaling laws and model capabilities. The constraint nobody's discussing enough: getting data across chips, nodes, and racks at the power density modern datacenters can't keep up with.
A single H100 SXM draws 700W. Eight of them in a server = 5–6kW, just for GPUs. Scale to 1,000 servers and you need dedicated power infrastructure most cities don't have.
I wrote a deep dive on why photonic computing — using light instead of electrons to move data — is the infrastructure shift that everyone building at scale will eventually have to reckon with.
Covers: why interconnect is the real bottleneck, how photonic chips work differently from silicon, which companies are building in this space, and realistic timelines.
https://pawankjha.substack.com/p/from-gpus-to-photons-the-quiet-revolution