We won the compute race, and the prize was a thermodynamics problem. GPUs made performance abundant and, in doing so, relocated the constraint from the chip to the infrastructure around it. Rack densities have climbed from 10–20 kW to 100+ kW in under five years, and auxiliary systems already consume 30–40% of facility energy before useful work begins.
In Part Two of Week 24, I make the case that the next bottleneck after GPUs isn’t a faster GPU, it’s whether you can power and cool what you’ve already committed to build. Photonic co-processors (like q.ant’s Native Processing Server, in production at the Leibniz Supercomputing Centre) attack all three post-GPU constraints at once: lower power draw, far less cooling to build, and much higher density per rack, because the heaviest math no longer arrives as heat.
The catch: photonic efficiency only becomes infrastructure relief when it’s designed in, captured at the rack and the substation, not in a procurement line.
Inside the full brief:
The three infrastructure trajectories, GPU-only, hybrid, and a build redesigned around light, and what each does to power, cooling and density.
The post-GPU readiness checklist for your next build.
The one capex-reframing question: if 20–40% of suitable workloads moved to photonics, how many megawatts and how much cooling would you never have to build?
Keywords: post-GPU bottleneck, AI power and cooling crisis, photonic co-processor, q.ant Native Processing Server, rack power density 100kW, data center thermodynamics, hybrid photonic infrastructure, sustainable multi-GW AI campus, thin-film lithium niobate, Energy Dominance
Full article with the three infrastructure trajectories and the post-GPU readiness checklist: