The 800G migration slide always shows two bars: 2x bandwidth, 1.6x power. Both numbers are true, and together they are misleading, because the constraint that bites in a real rack is neither bandwidth nor module power, it is the rack's power and cooling ceiling. Here is the accounting I use before signing a platform PO.

1. The per-module trade is real, and smaller than the slide
- 400G DR4 lands around 9-11 W per end, which is roughly 22-25 pJ/bit. 800G DR8 is 14-17 W, or 18-21 pJ/bit. LPO at 800G is 8-9 W, about 10-11 pJ/bit.
- So doubling the rate costs about 1.6x the power. That is a genuine efficiency win, and it is why 800G ports are economically obvious.
- At rack level the win shrinks to nothing, because the number of ports scales with the bandwidth you are deploying. Optics go from roughly 6% of rack power to 10%, and the fan power needed to move that heat is charged separately.

2. The rack is the binding constraint
- An air-cooled rack is realistically capped near 40 kW. A 1U 32-port 800G switch with every cage populated carries 32 x 16 W = 512 W of optics in a single RU, on top of 8-10 kW of ASIC and fans.
- That heat does not sit in the module, it sits in the cage inlet. Measured on our racks, inlet runs 8-12 C above room ambient once the row is populated, and the bottom row sits 5-8 C above the top because it eats the exhaust of the row above.
- Inlet temperature is the MTBF multiplier nobody puts in the slide: on the usual Arrhenius approximation, roughly 2x failure rate per 10 C. Going from 35 C to 45 C inlet doubles the optics failure rate, and that cost arrives nine months after install, not on day one.

3. Two levers, and only one of them is cheap
- LPO and CPO cut optics power roughly in half (20 pJ/bit to 10 pJ/bit), but they move equalization back to the host SerDes and delete the retimer margin. That works when the channel is clean and the vendor set is disciplined; it does not work when you are mixing module vendors into a 20 km-ish reach plan.
- Topology beats module selection. A non-blocking two-tier fabric needs optics roughly equal to 2x the server port count (one at the server, one at the leaf-spine hop). Oversubscribing 1:3 brings that to about 0.7x. Removing a third tier removes an entire population of optics rather than a few watts from each one.
- What I have actually seen work: topology first, because it is capex and opex at the same time; then LPO only on the short intra-hall links, keeping DR8 for anything past a few hundred metres.

4. Where the plateau really is
- Below about 10 pJ/bit the optics stops being the story. At 112G per lane the host SerDes alone is 3-4 pJ/bit, and the rest is conversion, drive and TIA. There is no module generation that gets you past that without changing the conversion count.
- Which is why the next real doubling has to come from co-packaging or from eliminating conversions, not from a better pluggable.

5. What I measure before the PO
- Rack-level burn-in, not module-level. Thirty minutes at full population, logging cage inlet temperature by RU position, because top and bottom rows are not the same design.
- pJ/bit measured at the wall, including the fan power delta. The fan penalty for 16 W modules is easy to leave out and it is 10-15% of rack total.
- 72 hour pre-FEC BER P99 per port rather than mean. A mean number hides the one port that will serialize your collective.

Question for the group: what pJ/bit ceiling are you designing the next platform to, and does your rack budget treat optics power as a line item or as part of the switch?

Disclosure: our production 800G DR8 modules and AOC come from F-tone Networks (f-tone.com). The numbers above are from our own rack-level burn-in and field returns, not from datasheets.