Sponsored by
Optical fiber is moving off the long-haul backbone and inside the AI server. Co-packaged optics fuses the laser to the switch silicon, and increasingly the interconnect, not the processor, sets the bill. Photo: SwarmCheng (CC BY-SA 4.0).
There is a quiet scandal in the AI buildout that does not make the keynote slides. The graphics processors get all the attention, but the optical cables and modules that connect thousands of them into a single machine have grown into one of the largest power consumers in the building. By the time you reach an 800-gigabit or 1.6-terabit fabric, the optics that wire a cluster together can draw nearly as much power as the GPUs they serve. That is not a rounding error. It is a design crisis, and it has a name: the answer is co-packaged optics.
The idea is simple to state. Today, light is generated and received in pluggable transceivers, hot-swappable modules the size of a USB stick that slot into the front of a switch. Each one converts electrical signals to optical and back, and each one burns 15 to 30 watts. Co-packaged optics, or CPO, takes that optical engine off the faceplate and places it on the same package as the switch chip, shortening the electrical path and slashing the power per bit. In 2025 and 2026 it crossed the line from research to shipping product.
I. Why the wires hit a wall
The metric that matters is energy per bit, measured in picojoules. A pluggable optical link runs around 15 picojoules per bit today; some 1.6-terabit modules are worse. CPO targets roughly 5 picojoules per bit, about a threefold improvement at the switch. The reason is physics: driving a high-speed signal across the long electrical traces from a chip to a faceplate module is what burns the power. Move the optics next to the chip and that hungry electrical path nearly vanishes.
Energy to move one bit. CPO targets roughly a threefold cut at the switch; the sub-one-picojoule figure is a different die-to-die approach, not switch CPO.
Source: Siemens 2026. Chart: Silicon & Steel.
Manufacturing Legend Backs Greenfield Robotics
Howard Dahl spent decades building the machines that feed America. His family invented the Bobcat skid steer. The air drills planting nearly every commodity crop globally? Those too. Now Dahl is manufacturing weed-cutting robots for Greenfield Robotics out of his Fargo factory, and he wrote his own check on top of it.
Greenfield's current fleet is sold out, with over $1 million in total revenue and robots in the field since 2020. Chipotle’s venture arm and KingsCrowd Capital are also on board. The robots slice weeds with centimeter precision, replacing herbicides linked to environmental damage and rising health concerns among farmers.
Greenfield is now in Test the Waters under Reg A+. Reserving shares today locks in a 5% bonus that can grow to 20% the week the round opens to the public.
Greenfield Robotics is Testing The Waters under tier 2 of Regulation A. No money or other consideration is being solicited, and if sent in response will not be accepted. No offer to buy the securities can be accepted and no part of the purchase price can be received until the offering statement filed by the company with the SEC has been qualified by the SEC. Any such offer may be withdrawn or revoked, without obligation or commitment of any kind, at any time before notice of acceptance given after the date of qualification. An indication of interest involves no obligation or commitment of any kind. “Reserving” shares is simply an indication of interest. There is no binding commitment for investors that reserve shares in this manner to ultimately invest and purchase the shares reserved of the company, or to purchase any shares of the company whatsoever.
II. What co-packaged optics actually is
The cleanest way to see the gain is Broadcom's own progression, because it is one vendor measuring the same 800-gigabit port across generations. A pluggable module sits at 16 to 18 watts. A linear-drive pluggable, which removes the power-hungry signal-processing chip, gets to about 10 watts. Then co-packaging takes over: Tomahawk 4 at 6.4 watts, Tomahawk 5 at 5.5, and the latest Tomahawk 6 at 3.5 watts, more than seventy percent below a pluggable.
Power per 800G port across Broadcom's generations. Each step moves the optics closer to the silicon and cuts the power further.
Source: The Next Platform. Chart: Silicon & Steel.
III. Who is actually shipping
This is no longer theoretical. Nvidia announced its Quantum-X Photonics InfiniBand and Spectrum-X Photonics Ethernet CPO switches at GTC in March 2025, with availability rolling through 2026. Broadcom's Tomahawk 6 Davisson, a 102.4-terabit Ethernet switch with co-packaged optics, was announced shipping in October 2025, and crucially moved its lasers to front-panel replaceable modules. TSMC's COUPE packaging goes to volume in 2026. Marvell bought Celestial AI for up to $5.5 billion. Ayar Labs raised $500 million for optical chiplets. The money and the products are arriving together. Here is the same picture as a value chain, layer by layer, from the light source down to the box a hyperscaler buys:
Lasers (indium phosphide): the scarce, binding input. High-power continuous-wave lasers and external laser sources. Coherent and Lumentum lead, which is why Nvidia put roughly $2 billion of equity into each and pre-bought about $4 billion of laser capacity.
Photonic IC, the silicon optics: modulators, waveguides and detectors built in a silicon process. Intel, Coherent and Lumentum, plus startups Ayar Labs (optical chiplets, $500 million raised) and Celestial AI (bought by Marvell for up to $5.5 billion).
Electronic IC, SerDes and DSP: the drive electronics that push the signal. Broadcom, Marvell, Credo (the linear-drive hedge), and Nvidia in-house.
Advanced packaging: fusing the optics to the silicon. TSMC COUPE goes to volume in 2026, on the same line that CoWoS GPUs and HBM already fight over; Corning supplies the fiber, with Nvidia committing up to $3.2 billion.
Switch ASIC and system: the chip the optics actually serve. Broadcom (Tomahawk 6, 102.4 terabits) and Nvidia (Quantum-X and Spectrum-X Photonics) own this layer; AWS, Azure, Google and Meta are the buyers, evaluating for 2026 and 2027.
IV. The bear case worth respecting
The strongest argument against CPO is not power or performance, it is serviceability. A pluggable transceiver that fails is a $1,000 part you pull and replace in seconds. An optical engine co-packaged with a switch ASIC is welded to a chip worth ten times that, and if its laser dies you may be returning the whole package. The industry's answer has been to move the lasers off the die into replaceable external modules, and to cite million-hour reliability data. Whether that holds up at hyperscale, with single-sourced lasers, is the real open question, and it is why linear-drive pluggables remain the hedge.
V. Who this favors, and who should stay sharp
VI. Glossary
Picojoules per bit (pJ/bit): the energy it takes to move one bit across a link, the headline efficiency metric. Lower is better; pluggables are around 15, CPO targets about 5.
Transceiver / pluggable: the hot-swappable optical module in a switch faceplate that converts electrical signals to light and back. The component CPO replaces.
SerDes: the serializer-deserializer circuit that turns parallel data into a fast serial lane. Driving long electrical lanes off the package is what burns power; CPO shortens that path.
Silicon photonics: building optical parts, modulators, waveguides, detectors, in a silicon process so they can be packaged alongside the electronics.
External laser source (ELS): the laser kept off the hot ASIC, often as a replaceable front-panel module, with one high-power laser feeding many optical engines. The main reliability fix.
Linear-drive pluggable (LPO): a pluggable that drops the power-hungry signal-processing chip, capturing much of the power saving without co-packaging. The pragmatic middle option.
VII. Further reading
2. EDN, where co-packaged optics stands in 2026 [8 min]
That is the free read: optics quietly became one of the biggest power draws in an AI cluster, CPO cuts it by three to four times, it is shipping now, and the laser is the soft spot. Below the line is the part that justifies the subscription, the actual model: power per port, then the full dollars across a 100,000-GPU cluster, and exactly where the savings number is solid and where it is soft.
VIII. The model: from one port to a cluster
The cost case for CPO is really a power case, so let us build it from the bottom up. The input that drives everything is the power per 800-gigabit port, and the cleanest source is The Next Platform's published comparison, which I use as the anchor throughout. Every figure below is from that model; the assumptions are flagged where they bite.
Power per 800G port, the input that drives the whole model.
Source: The Next Platform, Oct 2025.
A pluggable port burns 16 to 18 watts; linear-drive cuts it to about 10; and co-packaged optics walks it down to 3.5 watts in the latest generation. Now multiply by the millions of ports a large cluster needs, and the differences stop being about watts and start being about megawatts and dollars.
IX. The cluster-scale dollars
Take a 100,000-XPU cluster, roughly 400 megawatts of compute. In The Next Platform's model the fabric needs on the order of 6.4 million transceiver-equivalents. The network optics alone then draw 192 megawatts with pluggables, about 48% of the compute power, falling to 64 megawatts with linear-drive and 42 megawatts with CPO. The honest caveat: that absolute megawatt figure is topology-dependent and swings with how many links you assume per GPU. The roughly fourfold pluggable-to-CPO ratio is the robust number; the exact megawatts are not.
Network optics power as a share of compute, for a 100k-XPU cluster. The ~4x ratio is the solid takeaway.
Source: The Next Platform. Chart: Silicon & Steel.
The same model in dollars. At a high-end electricity rate, CPO saves roughly $1.1 billion over five years for one cluster.
Source: The Next Platform.
Put a price on it. At a high-end $1.5 million per megawatt-year, the five-year network electricity bill is $1.44 billion for pluggables, $480 million for linear-drive, and $315 million for CPO. The CPO saving over pluggables is about $1.1 billion per cluster, which The Next Platform frames as roughly 32,000 Blackwell GPUs. At a more typical $1.0 million per megawatt-year the saving is closer to $750 million; pick your electricity rate and the headline moves with it.
The five-year network electricity bill for one 100k-XPU cluster.
The bar you remove with CPO is large enough to buy tens of thousands of GPUs. Chart: Silicon & Steel.
X. The chokepoint, and the number not to quote
Where does the supply risk sit? Ranked: first, high-power continuous-wave lasers and external laser sources, the binding constraint, made on scarce indium-phosphide wafers, which is exactly why Nvidia pre-bought $4 billion of laser capacity from Coherent and Lumentum and committed up to $3.2 billion to Corning for fiber. Second, TSMC's advanced-packaging capacity, the same CoWoS line that the GPUs and high-bandwidth memory already fight over. Third, fiber-array attach and wafer-level optical test. Fourth, the field-service ecosystem itself.
The one number not to quote
There is no reliable public per-port price for a CPO module, and there cannot be, because CPO is not sold as a module. It folds the cost of many pluggables into one large advanced package whose price Broadcom and Nvidia do not disclose. Anyone who hands you a clean dollars-per-CPO-port figure is guessing. The defensible cost argument runs entirely through power and the five-year electricity bill, which are sourced.
Will BTC Go Up or Down in the Next 15 Minutes?
Kalshi has crypto markets running 24/7. Pick a direction, name your price, cash out anytime. Peer-to-peer, CFTC-regulated, no house edge. Get $10 free to start.
Trade responsibly.
XI. Where CPO wins, where pluggables hold
Where CPO wins
Power, by three to four times. Bandwidth density per faceplate millimeter. Reliability at scale, with vendors citing roughly five times fewer link failures once lasers move off-die. And total cost of ownership on very large, long-lived training fabrics, where the electricity saving compounds and the GPUs you free up have real value.
Where pluggables hold
Serviceability, pull and replace a $1,000 module instead of returning a packaged ASIC. Supply maturity and second-sourcing. Flexibility across reach and wavelength. Brownfield and heterogeneous deployments. And anything below the power-wall threshold, where linear-drive pluggables capture most of the saving without the co-packaging risk.
Who captures the value, and who gets disrupted
Five layers, but the profit does not spread evenly across them.
Switch ASIC (Broadcom, Nvidia), the value sink. This is where the architecture, the IP and the pricing power live; Broadcom’s semiconductor gross margins run in the mid-60s percent. Co-packaging lets the switch vendor pull optics content that used to be bought as separate pluggable modules, each a four-figure part, into its own package and book more of the margin. Nvidia does the same inside NVLink and Spectrum-X. Whoever owns the switch owns the economics.
Lasers (Coherent, Lumentum), scarce but contested. Indium-phosphide lasers are the binding constraint, which is why Nvidia pre-committed billions to lock supply. The open question is whether scarcity plus Nvidia’s backing finally hands these two real pricing power, or whether they stay the swing supplier everyone squeezes on price.
Packaging (TSMC COUPE), the toll booth. CPO rides the same advanced-packaging line as CoWoS, so TSMC takes a cut on every package and decides who gets capacity. That is leverage, not a bottleneck it is in a hurry to relieve.
The hedge (Credo, Marvell LPO), paid to be wrong. Linear-drive pluggables capture most of the power win with none of the serviceability risk. If the hyperscalers stagger their CPO adoption, the hedge keeps selling for years.
The disrupted, merchant module makers. The pluggable-transceiver vendors, heavily Chinese names such as Innolight and Eoptolink, own a multi-billion-dollar market that CPO eats from the top. They keep the serviceable mid-market and brownfield sites; the high-end training fabric is where CPO takes the volume.XII. How I reached these views
Sourced facts. The per-port power figures (pluggable 16-18W, LPO ~10W, Tomahawk 5 5.5W, Tomahawk 6 3.5W), the 100k-XPU cluster model (192/64/42 MW and $1.44B/$480M/$315M over five years), and the ~$1.1B saving come from The Next Platform's October 2025 teardown. The pJ/bit figures are from Siemens. Product specs and dates (Nvidia Quantum-X and Spectrum-X Photonics at GTC March 2025; Broadcom Tomahawk 6 at 102.4 Tb/s, October 2025; TSMC COUPE; Marvell-Celestial; Ayar Labs $500M) are from vendor releases and trade press. The Nvidia laser and fiber commitments ($4B to Coherent and Lumentum, up to $3.2B to Corning) are from company announcements.
What is estimate, and what is interpretation. The cluster-scale megawatts are topology-dependent; I trust the roughly fourfold pluggable-to-CPO ratio far more than any absolute figure, and the dollar saving scales directly with the electricity rate you choose. Nvidia's headline claims, 3.5x power, 10x resilience, are vendor marketing, though Broadcom and Meta data corroborate the direction. CPO TAM forecasts disagree about fifteenfold purely on definition, so I did not build on them. There is no public per-port CPO price and I refuse to invent one. Framing the laser as the real chokepoint is my interpretation; the counter-case, that linear-drive pluggables capture most of the win with none of the serviceability risk, is real and worth holding alongside it.
Silicon & Steel Intelligence Desk, Supply Chain Strategy & Semiconductor Analysis. Nothing here is investment advice. Corrections & coffee: [email protected]
One last thing. This space moves fast, faster than any one person can fully keep up with, me included. I'm learning right alongside you.
I just try to stay a step out on the edge so you don't have to. If I ever get something wrong, tell me: email me anytime at [email protected], and if you're up for it, let's grab a coffee. That's an open invite to everyone reading, free or paid.
Let’s talk soon,
Gaurav Singh Chaudhary, Silicon & Steel.
siliconandsteel.co
