There is a physical limit that has been quietly rearranging the inside of datacenters for the last three years, and it is not about transistors. It is about the fact that an electrical signal travelling down copper attenuates with frequency, and the frequencies have gone up by a factor of sixty-four in a decade.
You can watch the consequences in your own internet bill. Every time a cloud provider announces cheaper egress or a new region, part of the reason is that interconnect costs per gigabit fell — and they fell because the industry moved the hard part of the link from electrons to photons.
Where copper stops
The rule of thumb in high-speed signalling is that as per-lane bandwidth doubles, the usable reach at that rate roughly halves. At 25 Gb/s per lane, a passive copper cable is fine for three metres. At 112 Gb/s, you are down to about a metre. At 224 Gb/s, well under a metre unless you add retimers, and retimers cost power and latency and board area.
Once accelerator racks started drawing tens of kilowatts, the topology changed: fewer, bigger switches with more radix, and cable runs that had to reach across a row of racks. That is exactly the regime where copper stops being a cost saving and starts being a design constraint.
Two curves crossing is the interesting part. Copper port shipments are not collapsing — they are plateauing, because the last metre inside a rack is still cheaper in copper and probably always will be. What is happening is that new ports are almost all optical, because every new cluster is built at a density where copper cannot reach.
The cost crossover that matters
Pluggable optics have dominated for two decades because they are swappable, standardised, and let you mix vendors. Their weakness is that the electrical signal still has to travel from the switch ASIC to the front panel and back, and at 200G+ that path burns more power than the optical conversion itself.
Co-packaged optics puts the optical engine on the same substrate as the switch ASIC, eliminating that electrical detour.
The chart is my own model, and the honest caveat is that CPO is currently more expensive per gigabit — you pay for yield, for the inability to swap a failed module, and for a supply chain that is still one vendor deep. The crossover in my numbers lands around 2028. After that, CPO's advantage compounds because the electrical penalty you avoid grows with every lane-rate doubling.
Three practical consequences:
- Failure modes change shape. A failed pluggable is a field-replaceable part with a ticket. A failed CPO engine is a switch replacement. This is why hyperscalers are deploying CPO in their own custom switches first, where they control the whole stack and can afford to iterate.
- Power budgets shift. Optics moving closer to the ASIC removes retimer power but concentrates heat in a smaller volume, which pushes the thermal problem into the package. If you thought liquid cooling was optional, co-packaged optics makes it structural.
- Vendor lock-in deepens. Pluggables created a competitive market with multiple sources per form factor. CPO ties your optics to your switch vendor. That is a procurement risk, not a technical one, and it is the real reason adoption is slower than the technology merits.
What this means outside the datacenter
Cheaper long-haul bandwidth, eventually. Interconnect is a large fraction of the marginal cost of moving data between regions. When per-gigabit cost falls by half, the floor under cloud egress pricing falls with it. Expect the political pressure on egress fees to be accompanied by an actual cost decline, which is a rare combination.
A different shape of datacenter. Optical links reach kilometres, not metres. That decouples the "compute hall" from the "network hall" and, more interestingly, allows campuses where accelerators are spread across buildings connected by fibre. Land becomes less of a constraint; power and cooling become more of one.
A supply chain chokepoint. Indium phosphide lasers, silicon photonics wafers, and the handful of fabs that can do hybrid integration are all concentrated. Every time the industry doubles optical port volume, it stresses the same three or four suppliers. This is the sort of bottleneck that shows up in a quarterly earnings call as "supply constraints" and in a customer's world as a nine-month lead time.
Your home connection gets better before your office does. Long-haul and metro links are the first to benefit from cheaper optics, because they were already optical. The last hundred metres to a building is a civil engineering problem, not a photonics one, and it will lag by a decade.
How to read the next two years
Three signals:
- Laser count per switch, not port count. CPO's economics hinge on how many lasers you need per switch and whether they are external or integrated. External laser sources (ELS) are the pragmatic near-term answer and the thing to track.
- Retimer revenue. Retimers exist to rescue copper. If retimer revenue flattens while port counts grow, copper is losing — that is the cleanest early indicator.
- Thermal density per rack. Every optical win adds heat where the heat is hardest to remove. Rack density is the number that constrains how fast any of this can actually be deployed.
The engineering summary
Copper is not dead; it is being pushed into a smaller and smaller box, and the box is exactly the last metre. Optics is not free; it is trading a cheap, swappable, multi-vendor component for an expensive, integrated, single-vendor one — and buying a power and reach advantage that compounds.
That trade is worth making when the lane rate crosses the reach limit, which is precisely where the industry is now. The winners in the next cycle will be the teams that understand the exchange rate between reach, power, and serviceability, rather than the ones that pick a side in a format war.
A small ask
If this gave you a usable way to think about interconnect economics, a tip helps keep the analysis coming — the numbers take real effort to assemble and cross-check.
SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Creating this takes real effort — thank you for the support.
Top comments (0)