What if the hardest problem in AI networking isn't speed, but keeping every link running?
In this episode, CEO Martin Lund argues that as AI outgrows a single rack, the interconnect becomes part of the computer itself, and reliability becomes the requirement that cannot be traded away.
Martin started in networking on dial-up modems, and he calls AI scale-up "the most unforgiving, demanding traffic you can put on a network": latency intolerant, loss intolerant, and growing with every generation. His answer is to optimize for the whole AI factory instead of a single component, and to reject fixes that only look good locally, like replaceable lasers that still need a person to swap them. As Martin puts it, a $50 billion data center should not run at 50% utilization because the technician on laser duty was out sick. CScale integrates the laser so failures are handled inside the system, building toward links that need zero touch.
In this conversation, we cover:
- Why CScale optimizes for "the factory," not "the individual rack"
- Why "serviceability is not a solution for continuity" in AI data centers
- How scale-up differs from scale-out, and why tightly coupled accelerators are "actually the machine"
- What "the copper wall" is, and why optics is the way past it
- How designing around the fact that "lasers will fail" lets the scale-up network "behave like copper"
Martin also explains why fewer fibers mean "fewer things to go wrong," why CScale is not "building a company to solve one generation," how the team blends deep domain experts in a collaborative culture, and how better factory utilization can lower the power consumption of AI.
CScale is an optical interconnect company building an integrated light engine for AI scale-up, designed to contain optical failures so thousands of accelerators can keep working as one much larger computer.
