A 1.6 terabit-per-second optical module — the part that turns electrical bits into light bouncing through a fiber inside an AI server — dissipates about 20 watts of waste heat into a metal cage roughly the size of a USB stick. Above that threshold, air can’t carry the heat off fast enough. Below it, air is fine. One number, 20 W, bisects the entire cooling industry.
This piece is the path of that 20 watts: what the two cooling architectures actually do, which AI buyer picked which one and why, the supplier list at each layer, and the one materials story that could turn the whole stack upside down by 2H 2027.
Companion piece to What AI Hardware Is Running Out Of — same supply chain, different layer.
The thermal cliff at 20 watts
Optical modules have used air cooling forever. A copper heat sink, a small fan, ambient airflow across the module’s cage — done. It works up to roughly 16 W of dissipation, where the heat sink’s surface area can shed enough heat into the airstream.
Between 16 and 20 W you’re in the gray zone. The module’s still running, but the laser is operating outside its thermal sweet spot, the bit-error rate creeps up, and the effective line rate drops. In a 1.6T-spec module on air, the haircut shows up as roughly 1.4T of useful throughput — a 12.5% efficiency loss you pay for cooling you can’t actually do.
Past 20 W there is no gray zone. The laser drifts off-channel, error rates spike, the link doesn’t come up. So a 1.6T module at full rate must be liquid-cooled.
“Liquid cooling” here isn’t water sloshing around the optics. It’s a precision assembly: a metal cage machined to mate to the module’s hot top surface, a cold plate with internal microchannels, and quick-disconnect couplers routing a dielectric coolant in and out. Same module, full 1.6T rate. That’s the cliff. That’s the entire cooling story.
Who picked what, and why density decided it
Two buyer archetypes, two answers.
- Nvidia — 100% liquid. The Rubin platform shipping through 2026 and into 2027 is built around extreme rack density: Nvidia’s reference design packs 72 GPUs per rack, sometimes more. The optical layer connecting them is correspondingly dense. At that density the rack-level photon-side dissipation is into the kilowatts — air doesn’t even try. Nvidia spec’d liquid end-to-end, and the entire vendor stack lined up because Nvidia is the volume buyer.
- CSPs (Google, Amazon, Meta) — mixed. Hyperscalers run their own rack designs. They cluster around 32 GPUs (or custom accelerator count) per rack — about half Nvidia’s density — so the per-module thermal envelope can sit in the 14–18 W band and a lot of their 1.6T deployments stay on air. Google’s 1.6T liquid-cooling penetration is around 20%: liquid only where they’re pushing density on a specific workload; the rest is air.
The generalization is clean. Whoever owns their own rack design picks à la carte; whoever buys Nvidia gets liquid. And as Nvidia takes a larger share of AI accelerator shipments, the liquid-cooling supply chain drags the broader market with it.
Where the cage gets built — the Amphenol monopoly
The liquid-cooling cage is the part of the assembly that touches the optical module itself — machined metal, internal coolant channels, integrated couplers and seals. It is mechanically critical and patented heavily.
In 2026, Amphenol holds dominant share of the switch-side liquid-cooling cage market — close to monopoly, driven by a portfolio of core patents on the cage geometry and the coupler interface. Nvidia’s reference designs ship Amphenol cages by default. Google’s liquid deployments use a slim shortlist — Amphenol, TE, and Molex — with Amphenol the senior partner.
TE Connectivity is the credible second. It’s been ramping liquid-cage capacity through 2025 and 2026, and the consensus through-2027 projection has it reaching parity with Amphenol in switch-side liquid cooling. That’s a real structural move: a market that today looks like a monopoly resolves into a duopoly inside eighteen months. It is the single most-interesting power shift in this layer of the stack.
Where the parts get stamped — the tier-2 beneficiaries
Behind Amphenol and TE sits a layer of contract manufacturers that actually produce the cold plates, the cage bodies, and the assembled subcomponents. This is where the supply chain quietly redistributes upside.
- Dintech (鼎通科技) is the largest single beneficiary today. It’s the core structural-parts supplier to both Amphenol and TE — meaning whichever side of the looming duopoly wins, Dintech gets paid. Its 1.6T-driven order book has expanded rapidly through 2026.
- Sunway Communication (硕贝德) supplies cold plates and other structural parts to TE. Smaller share than Dintech, but growing into the TE-side build-out.
- Yidong Electronics (奕东电子) is currently a minor TE supplier, expected to be qualified into the broader supply chain through 2027.
The thing to notice: the tier-1 connector makers own the IP and the customer relationship; the tier-2 contract manufacturers own the actual capacity. When demand spikes — as it did when 1.6T deployments started cliff-jumping into liquid cooling in 2026 — it’s the tier-2 manufacturers whose revenue moves first and biggest, because the tier-1s can’t physically expand their own production fast enough.
Unit economics
A single 2 × 8 liquid-cooling cage assembly — eight columns wide, two rows deep, sixteen module slots — sells at up to ¥1,800 (roughly $250 at current FX). That’s the assembly price; the raw structural parts going into it cost a fraction of that, which is where Amphenol and TE keep their gross-margin headroom.
The per-rack arithmetic isn’t small. Multiply by the number of cage assemblies in a Nvidia Rubin-class rack and by the number of racks compounding through the Rubin generation — and that’s before any of the higher-priced components above the cage (CDUs, manifolds, rack-level plumbing). The cooling stack’s revenue line moves with rack count, and rack count is exactly what is compounding right now.
The CPO threat — what could erase this whole market
The clean threat to everything above is CPO (co-packaged optics). The pluggable module disappears; the optical engine moves onto the same substrate as the switch ASIC. No cage, no faceplate, no swappable transceiver — the optical layer becomes part of the switch chip.
If CPO ships at scale, the entire pluggable-module liquid-cooling stack — Amphenol’s near-monopoly, TE’s catch-up, Dintech’s order book — goes to zero. The optical heat moves onto the ASIC and gets cooled by the same cold plate that’s already there for the switch silicon. Everyone above suddenly has no line item.
The current consensus is that CPO is a 2027–2028 commercial story rather than a 2026 one. Nvidia’s roadmap has CPO surfacing on specific high-density links before the end of the decade, not as a wholesale replacement yet. But it is the one shift that erases rather than reshuffles the cooling-stack winners — which makes it the single most-watched variable in this market.
The wildcard — Coherent's diamond-particle ceramic
A very different lever sits at Coherent — the materials company best known on this site for its near-monopoly on Faraday rotators in the optical isolator market (covered in What AI Hardware Is Running Out Of). Its quieter Engineered Materials division has developed a diamond-particle composite silicon-carbide ceramic with two stand-out numbers:
- Thermal conductivity is roughly 2× that of copper.
- Weight is around 60% of copper.
Used as a heat spreader, baseplate, or custom cold plate, it drops the chip’s operating temperature by about 15°C at fixed dissipation. That is a huge thermal headroom in an industry that prices throughput in single-digit-celsius margins.
It is independent of the cage / cold-plate stack above. It is independent of CPO. It bolts onto whichever cooling architecture wins. Coherent expects meaningful revenue contribution to begin in the second half of 2027 — entirely incremental, no displacement risk, owned end-to-end by Coherent.
It is worth tracking specifically because it is one of very few stories in cooling-adjacent hardware that doesn’t get unmade by the CPO transition. Whoever ends up cooling the chip will want better materials; this is the better material.
The shape
Pull back, and the cooling stack today is a four-layer Pareto:
- The buyer. Air or liquid by rack density. Nvidia: liquid. CSPs: mixed.
- The cage maker. Near-monopoly in 2026 (Amphenol), credibly turning into a duopoly by 2027 (Amphenol + TE).
- The structural-parts contractor. Sees demand first and captures upside hardest (Dintech, then Sunway, then Yidong).
- The wildcards. Coherent’s diamond ceramic (additive, 2H 2027) and CPO (subtractive, 2027–2028).
The most interesting question isn’t who wins the cage market — it’s whether the cage market still exists in three years. If CPO ships at scale, Amphenol, TE, and Dintech all lose their best line item simultaneously. If it doesn’t, the current Pareto holds for another hardware cycle and the tier-2 contractors keep compounding. Either way, the people paying $250 a cage today aren’t pricing in the binary risk above them — which is why the story is worth tracking on its own.
Figures here are drawn from current supply-chain research (industry Q&A notes and sell-side reports through mid-2026) and stated as the thesis, not independently re-derived.
Adjacent reading: What AI Hardware Is Running Out Of (the broader scarcity map — Coherent’s Faraday rotators and the 200G EML cliff are in there); Nokia / DCI (the inter-data-center link layer that feeds these modules).