A 1.6 terabit-per-second optical module dissipates about 20 watts of waste heat inside a cage roughly the size of a USB stick. Above that threshold, air can’t carry it off fast enough. One number bisects the entire cooling industry.
↓the cliff
cooling has had to evolve three times — the fourth is coming
can’t keep up
$250 a cage
no fans
to the chip
each generation pulls more
Air cooling works up to 16 W. A copper heat sink, a small fan, ambient airflow — done.
cheap.
Between 16 and 20 W: gray zone. The laser drifts off its sweet spot, error rates creep up — a 1.6T module gives only 1.4T of useful throughput. A 12.5% efficiency hit.
you can’t actually do.
Past 20 W: no gray zone. The laser drifts off-channel, the link doesn’t come up. Full-rate 1.6T must be liquid-cooled.
dielectric coolant.
→two buyers, two answers
Nvidia — 100% liquid. Rubin packs 72 GPUs per rack. At that density rack-level dissipation is kilowatts. Air doesn’t even try.
volume buyer rules.
CSPs — mixed. Google, Amazon, Meta run their own racks at 32 GPUs each — about half Nvidia’s density. Google’s 1.6T liquid penetration is around 20%.
à la carte.
⚡amphenol monopoly
The liquid cage is mechanically critical and patented heavily. In 2026, Amphenol holds dominant share — close to monopoly — driven by core patents on the cage geometry.
by default.
TE Connectivity is the credible second. Consensus has it reaching parity with Amphenol through 2027 — a market that looks like a monopoly resolves into a duopoly in eighteen months.
duopoly.
⌖where the parts get stamped
Dintech (鼎通科技). Largest single beneficiary — supplies structural parts to both Amphenol and TE. Whichever side of the duopoly wins, Dintech gets paid.
sides of the fight.
Sunway Communication (硕贝德). Cold plates & structural parts to TE. Growing into the TE-side build-out.
growing.
Yidong Electronics (奕东电子). Minor TE supplier, expected to be qualified into the broader chain through 2027.
2027 qualification.
※unit economics
A single 2×8 liquid-cooling cage (eight columns wide, two rows deep, sixteen module slots) sells at up to ¥1,800 — roughly $250 at current FX.
raw parts cost less.
⚡the cpo threat
CPO (co-packaged optics) puts the optical engine onto the switch ASIC. No cage, no pluggable transceiver. The whole liquid-cooling stack goes to zero.
cage business dies.
Consensus has CPO as a 2027–2028 story, not a 2026 one. Specific high-density links first, not a wholesale replacement.
erases the stack.
⌖the wildcard
Coherent’s diamond-particle silicon-carbide ceramic. Thermal conductivity ~2× copper. Weight ~60% of copper. Drops chip temperature by about 15°C.
bolts on anywhere.
Independent of the cage stack. Independent of CPO. Revenue begins in 2H 2027 — entirely incremental, owned end-to-end by Coherent.
by CPO.
※the shape
Four layers, one Pareto: buyer (Nvidia = liquid, CSPs = mixed), cage maker (Amphenol monopoly → Amphenol + TE duopoly by 2027), structural-parts contractor (Dintech, Sunway, Yidong), wildcards (Coherent ceramic in 2H 2027; CPO in 2027–2028). The most interesting question isn’t who wins the cage market — it’s whether the cage market still exists in three years.