Co-packaged optics (CPO) is a packaging change, not a new kind of laser. The photons are the same photons: silicon photonics modulators, continuous-wave lasers, single-mode fibre. What changes is where the conversion happens. In a pluggable module, the switch ASIC pushes an electrical signal across 10–25 cm of PCB and through a connector before a DSP cleans it up and a modulator turns it into light. In a CPO switch, that same conversion happens on the ASIC's package, with millimetres of electrical reach. Everything the industry says about CPO follows from that one geometry change — the watts, the signal integrity, the fibre attach, the service model, and the reason a failed laser no longer means swapping a module.

Why does the optics bill now exceed the switch bill?

Because bandwidth grew faster on the panel than on the die. A 51.2 Tb/s switch is 64 ports of 800G. Each 800G pluggable dissipates roughly 14–16 W (Broadcom's own 60–70% reduction from a 5 W CPO engine implies a 12.5–16.7 W baseline), so the faceplate carries 896 W to 1,024 W of transceivers. Broadcom's product release states the consequence in the plainest language any vendor has used: "Today, pluggable optical transceivers consume approximately 50% of system power and constitute more than 50% of the cost of a traditional switch system." Dell'Oro Group's Sameh Boujelbene, quoted in the same release, puts the trajectory at "exceeding 50% of the switch system power at 51.2 Tbps and beyond." That "and beyond" is doing a lot of work: at 102.4 Tb/s the same arithmetic gives 128 ports and 1,792 W, against an ASIC whose own draw grows by far less per generation. The optics share does not stay at 50%; it climbs.

This is also why the honest framing of CPO is not "greener switches." The switch chip was never the thermal problem in a leaf-spine fabric; the row of hot cages was. An optical engine on the package moves ~576 W out of the airflow path of a single 1RU box, and the benefit shows up as watts per rack unit and degrees of faceplate temperature, not as a smaller electricity bill for the building.

Per 800G portPowerEnergy per bit (our arithmetic)Where the part lives
Pluggable, DSP-based (OSFP/QSFP-DD)≈14 W (range 12.5–16.7)≈17.5 pJ/bitFaceplate cage, field-replaceable, multi-vendor
Co-packaged engine (Broadcom Bailly KPI)5 W including laser6.25 pJ/bitOn the ASIC package, replaced with the switch
51.2 Tb/s switch, all 64 ports pluggable≈896 W—64 cages on one faceplate
102.4 Tb/s switch, all 128 ports pluggable≈1,792 W—128 cages — the density wall
Same two switches with CPO engines320 W / 640 W—Optics inside the package, lasers outside

Note what the table does not show: any reduction in fibre. A 512-port 800G switch still presents 512 optical links at the panel whether the engine is a module or a package. CPO deletes the transceiver, the DSP and the cage; it does not delete the cable plant, the patch panels, or the human who has to clean the connectors. Any TCO model that books CPO as a cabling saving is wrong.

What does 70% lower optics power buy at cluster scale?

Less than the percentage suggests, and it is worth doing the multiplication because the percentage is the number that gets quoted in board decks. Take a 100,000-GPU training campus, one 800G scale-out port per GPU, a non-blocking fabric. The host side does not change: every CPO switch on the market in 2026 still pairs with a pluggable transceiver on the NIC, so 100,000 × 14 W = 1.4 MW of optics sits on the GPU side no matter what the switches do. The switch side does change, and a non-blocking two-tier fabric puts roughly two switch ports on the network for every GPU port — call it 200,000 ports. At 14 W that is 2.8 MW; at 5 W it is 1.0 MW; the delta is ≈1.8 MW.

Now the denominator. A GB200 NVL72 rack draws about 120 kW for 72 GPUs, so a 100,000-GPU campus is in the 150–170 MW range once you add CPUs, power conversion losses and cooling. Our 1.8 MW saving is about 1%. That is not nothing — 1.8 MW is a small hall, or roughly the draw of 1,100 more GPUs — but it is an order of magnitude away from the story the 70% figure implies, and it is smaller than the error bar on most PUE assumptions. Any procurement case that rests on CPO's energy savings is resting on the wrong number.

The number that does hold up is per rack unit. Put 1,792 W of pluggables in a 1RU chassis and you are asking a single slot of a 42U rack to dissipate 40 W per centimetre of height, in a box whose front panel is also the air intake. That is the constraint CPO actually relieves — and it is why Broadcom frames CPO as a prerequisite for the 102.4 Tb/s generation rather than as an efficiency upgrade to the 51.2 Tb/s one.

Why do the lasers still sit outside the package?

Because the laser is the part that fails, and everyone involved knows it. Every commercial CPO design keeps its light source off-package, in an external laser source (ELS) module that plugs into the panel and feeds the engines over fibre. The Optical Internetworking Forum standardised the form factor as ELSFP, with its 02.0 implementation agreement published in January 2025. NVIDIA's description of its own implementation is unusually candid about the reason: the ELS modules sit "in a dedicated thermally controlled environment independent of the main switch chassis," which reduces wavelength drift and "mitigates early failures caused by aging mechanisms," and each module is field-replaceable so an operator can swap it "without disturbing the surrounding switch infrastructure." Concentrating the light source is also what delivers the 4× reduction in total laser count NVIDIA claims, since one high-power laser is shared across many channels instead of one small laser per channel.

The physics behind that decision is not subtle. Lasers degrade with temperature and with thermal cycling, and a package carrying a several-hundred-watt switch ASIC is the hottest place in the machine. Vendors who build the module say the same thing in different words: the laser is among the least reliable active components in the optical path, which is why it has always been socketed and swappable. Co-packaging a laser with an ASIC converts a five-minute field replacement into an RMA.

So the architecture that ships is a hybrid, and it is worth being precise about it: modulators and photodetectors go on the package; the laser stays a pluggable consumable on the panel. CPO does not eliminate the wear-out part. It moves everything except the wear-out part.

What does CPO take away?

The last commodity in the rack. A pluggable transceiver is bought from at least three qualified vendors, priced annually, and improved on somebody else's roadmap: 400G modules became 800G modules became 1.6T modules, and each generation arrived as a purchase order rather than a forklift. CPO folds that line item into the switch. When your optics improve next, you buy a switch. When one fails, you replace a chassis. When you want a second source, you change platforms. That is a normal thing to do for an ASIC and a strange thing to do for the one component that used to obey a commodity price curve — and it is the reason the optics industry's own roadmaps keep a pluggable tier alive (400G/800G DR and FR modules remain the host-side standard for every CPO platform shipping today).

The failure granularity, and the vendor says so. Broadcom's own engineering blog, discussing its joint fibre-reliability white paper with Corning, warns that "improper fiber quality control and cable assembly procedures can lead to fiber breakage and have significant operational consequences such as costly checkpoint restarts and full appliance replacements." Read that sentence twice: the failure mode of a CPO switch is not a dead port, it is a full appliance replacement, and in a training cluster the cost is not the appliance but the checkpoint restart — the same restart whose cost we put at 6.5 TB written per event in our checkpoint post. A thousand-plus fibres bonded to one package is a manufacturing achievement and a serviceability liability, and the mitigation is procedural: fibre routing discipline "similar to PCB routing," in Broadcom's words.

The measurement you can trust. Every power figure in this article is a vendor KPI measured on the optical interconnect, not an independent teardown of a named hyperscaler SKU. Broadcom's 70% is "optical interconnect power," which is not the same claim as "the switch uses 70% less power" — a Tomahawk 5 still switches 51.2 Tb/s and still draws what it draws. NVIDIA's 3.5× is a platform claim against "traditional methods." Treat both as directional, and treat the shape of the argument as the durable part: optics were half the box, that share was rising, and the only way to keep building bigger boxes was to stop treating the transceiver as a separate product.

Where does it actually stand in November 2026?

Further along than the sceptics said and much less far than the slide implies. Three platforms define the market, and only two of them are shipping at volume.

PlatformFabric / capacityStatus (Nov 2026)What it tells you
Broadcom Bailly (Tomahawk 5)Ethernet, 51.2 Tb/s, 8 × 6.4 Tb/s enginesIn deployment / volume production through system partnersThe reference design the industry quotes; 5 W per 800G is its published KPI
Broadcom Davisson (Tomahawk 6)Ethernet, 102.4 Tb/s CPOSampling since late 2025, no volume-production statement as of September 2026The generation where CPO stops being optional — 128 cages is the density wall
NVIDIA Quantum-X PhotonicsInfiniBand, 115.2 Tb/s, 144 × 800G, liquid-cooledCustomer delivery from June 2026, production deploymentsFirst CPO switch in real training fabrics; four-ASIC design with a detachable optical sub-assembly
NVIDIA Spectrum-X PhotonicsEthernet, up to 409.6 Tb/s, 512 × 800GVolume production from August 2026The Ethernet answer, and the one that competes with Bailly/Davisson head-on

Two observations that matter more than the table. First, CPO still has a host side: the NIC in the server uses a pluggable, so the change is confined to the switch tier and the pluggable market does not disappear — it relocates. Second, the conservative camp is not just waiting. Several switch vendors still decline to bond a thousand fibres inside a package and prefer linear pluggable optics for the next generation, which is a rational bet if your customers care more about field service than about watts per rack unit. The deployment split that follows — hyperscalers with their own spares depots on CPO, everyone else on LPO and pluggables — is the most likely 2027 shape, and it is a procurement question before it is a technology question.

For anyone specifying a fabric this quarter, the practical reading is unglamorous. Ask what the switch's optical interconnect power per 800G is, not what its efficiency multiplier is. Ask what the field-replaceable unit is when an engine fails, and whether the answer is a module, a sub-assembly or the box. Ask whether the lasers are inside or outside the package, and who stocks the spares. And price the whole thing against a pluggable build of the same bandwidth, because the honest case for CPO is not that it is cheaper today — it is that at 102.4 Tb/s the pluggable build may not be buildable at all. That is a different argument, and it is the one that survives scrutiny.

There is a symmetry worth noting with the rest of this series. The fabric was the third bill after compute and memory; the optics line inside it was the part everyone treated as a rounding error and then discovered was half the switch. The pattern repeats across the stack — memory, packaging, cooling — the component that was supposed to be a commodity becomes the constraint, and the industry's answer is always to integrate it and lose the commodity. CPO is that move applied to light.