Copper, and Where It Runs Out

An AI rack contains an unreasonable amount of copper, and the reason is a good one. NVIDIA's GB200 NVL72 carries four NVLink cartridges at the back of the rack holding more than 5,000 active copper cables, meshing 72 GPUs at 1.8 TB/s each for 260 TB/s of AllReduce bandwidth across the domain. The designs were contributed to the Open Compute Project, so this is a published mechanical design rather than a specification sheet.

Copper was chosen over light for a reason that can be stated in watts. Reaching 1.8 TB/s optically would have needed eighteen 800 Gbps pluggable transceivers per GPU — nine on the accelerator, nine on the switch — at roughly 10 to 15 watts each. NVIDIA's chief executive said in 2024 that pluggable optics for the NVL72 would have required an additional 20,000 watts. In a rack that already needs 120 kW of cooling, twenty kilowatts spent turning electricity into light and back is twenty kilowatts not spent on arithmetic. The cheapest interconnect is the one you do not have to power.

And then copper runs out. NVIDIA's own networking lead puts it as copper being the best connectivity if you can use it — the qualification doing the work. At 1.8 TB/s a copper cable degrades within a few feet, which is why the switches sit in the middle of the rack and why the design crams as many GPUs into one rack as it physically can. Both are workarounds for reach, and neither survives contact with a domain larger than a rack. So scale-up goes optical: Vera Rubin NVL576, then Rosa Feynman NVL1152, with co-packaged optics integrated at Feynman. NVIDIA has committed $2 billion each to Coherent and Lumentum for lasers and $2 billion to Marvell for optical I/O.

Two opposite decisions four years apart, both defensible. Copper won while the machine fitted in one rack, and stops winning the moment it does not. Anyone reading this as a reversal has the wrong frame; it is the same trade evaluated at a different size.

The second copper problem is supply rather than signalling, and it is measured in kilograms. At the 54 VDC used inside current racks, NVIDIA states a 1 MW rack needs up to 200 kg of copper busbar, and the racks alone in a 1 GW facility could need up to 200,000 kg. Lower voltage at fixed power means more current, and more current means more metal. The proposed answer is to raise the voltage — 13.8 kV AC converted directly to 800 VDC at the perimeter — with a claimed 85% more power through the same conductor and 45% less copper than a 415 VAC system, in production alongside Kyber racks in 2027.

What is missing is named. Nearly everything here comes from one company: the engineering reasoning, the power figures and the roadmap all originate with NVIDIA or with reporting of NVIDIA statements. The independent technical analysis of the copper-versus-optics decision sits behind a paid newsletter and was not read, and the Open Compute Project whitepaper and solid-state transformer specification behind the 800 VDC material were not retrieved — so a multi-vendor specification is described here through one vendor's account of it. A social-media post describing a cableless midplane redesign for Rubin is listed as a lead and asserted nowhere.