Research|Comment on Hock Tan’s View on CPO: Copper isn’t Dead Yet— But Optics is Inevitable
Since Hock's comment at CPO, many subscribers have come to us with questions. We will address them all here and try to explain the underlying technical principles and reasoning as clearly as possible. Hock's comment does not affect the CPO industry — GTC is what matters most.
Hock Tan’s recent comments on CPO are not really a denial of the technology. The main point he is making is about timing. More specifically, he is arguing that optics for intra-rack scale-up will not become mainstream by 2027. His thesis is straightforward: before roughly 2028, most scale-up or rack-level interconnects will continue to rely on DAC/copper. At short distances, copper still has advantages in latency, power, and cost. Broadcom also believes its SerDes technology can keep extending the reach of 200G/400G copper links, so customers have little incentive to switch to optics too early. As he put it in the call, “beyond 2028 we see optics coming in dramatically,” while today, “every single time they always pick copper.”
This view is actually not inconsistent with what we see in the current system architectures, and it broadly aligns with what we discussed in our earlier GTC preview.
What Hock is really describing is chip-to-chip or XPU-to-XPU connectivity within the current rack form factors, essentially something like Vera Rubin NVL72. In that architecture, scale-up within the rack continues to use copper, and optics are used only when the connection crosses a larger cluster boundary. The discussion about optical scale-up at a much larger domain has never been about NVL72, but rather systems like NVL576 and beyond. Even there, the model is still copper inside the rack and optics between racks, which has long been the industry’s baseline evolution path. Fully optical intra-rack connectivity is likely not available until a later generation, potentially around the Feynman timeframe. In other words, Hock is saying that this generation of short-reach scale-up will stay copper, not that optical scale-up will never happen.
