Co packaged optics is forecast to take over scale up wiring within five years. The decision lives in a link budget that has to close twice: once in decibels, once in dollars.
The fastest way to understand why co-packaged optics has become a five-year story is to start at the top of the demand curve and work down. According to a SemiEngineering analysis published June 10, 2026 by Geoff Tate, Anthropic is running at roughly $47 billion in annual revenue, with OpenAI and Google's Gemini business described as close behind. The source notes that Anthropic has said publicly its growth is constrained by compute capacity, not by customers. That single constraint reframes the entire interconnect question: if the ceiling is compute, then everything that connects compute, every cable, every switch, every photon that lands on a receiver, becomes the rate-limiting part of the business.
The bottleneck is no longer at the datacenter boundary. It is inside a single rack, where a few hundred GPUs or XPUs have to talk to each other fast enough to count as one machine. That kind of "scale-up" wiring is a different problem from "scale-out" networking across racks or buildings, and the SemiEngineering analysis argues the two need to be treated as separate engineering problems with separate economics. Copper, the workhorse of in-rack wiring for two decades, is the technology the source is most directly challenging.
In a copper world, scale-up is bounded by distance, signal integrity, and power. As bandwidth per accelerator climbs and the number of accelerators in a pod grows, the copper traces between chips start losing signal before they reach the next switch. The article, written by industry analyst Geoff Tate, argues that the only path past that physical limit is to move the optical interface as close to the silicon as the package will allow. The destination is co-packaged optics, or CPO, where the laser and the modulator sit next to the switch ASIC inside one package.
The journey is governed by a "link budget" measured in two currencies: decibels, for the optical loss along the path, and dollars, for the cost per bit of bandwidth the system actually delivers. Both budgets have to close for the transition to happen on the timeline the source forecasts.
The decibel side of the budget is unforgiving. A simplified optical chain in a CPO compute rack, according to the same analysis, goes like this: an external laser source module, called an ELSFP, fires light into a polarization-maintaining fiber, which lands at a detachable connector, enters the optical engine transmitter inside the package, exits on single-mode fiber to the face plate, passes through more connectors and possibly an optical circuit switch or shuffle box, and finally lands on the receiver of the next switch. The reverse path mirrors it. Roughly 99% of the laser's power is gone before any of it reaches the receiver, the source says.
The losses break down sharply. From the laser die through the collimating lens, the isolator, the focusing lens, and the MT connector, the ELSFP path loses between 1.5 and 3 decibels, meaning 29% to 50% of the power is shed before the light even leaves the source module. The surviving 50% to 71% has to be enough to drive the rest of the link. ELSFP modules today ship at output powers of 20, 23, or 26 dBm per channel, equal to 100, 200, or 400 milliwatts. Those power levels are what makes the link feasible at all, the source argues. Without them, the budget cannot close.
The cable plant is not generic. The run from the ELSFP to the optical engine is short, typically under a meter, and uses polarization-maintaining fiber, a specialty fiber that the analysis says costs roughly 100 times more per meter than standard single-mode fiber but loses less than 0.001 dB per meter in the O-band around 1310 nanometers. That cost gap is the reason PMF is only used on the inside hop, while the rest of the rack uses ordinary yellow-jacketed single-mode fiber, the "yellow cable" most people picture when they think of datacenter fiber.
The dollar side of the budget is where the real fight is. CPO is not a foregone conclusion. It is an economic claim: that the total component cost of an optical link, broken down across the laser, the fiber, the detachable connector, the optical engine, and the package assembly, will beat a target dollars-per-gigabit threshold that makes it more attractive than the next-best pluggable transceiver.
The author flags detachable connectors as a particular pinch point. Pluggable transceivers are too bulky to land at the package edge in volume, the source says, so the industry needs a new connector category for CPO. Vendors named in the analysis include Senko, Molex (which the article says recently acquired Teramount), Corning, Foci, ACON, Foxconn, and Furukawa. Edge-coupling from those connectors into a photonic integrated circuit inside the optical engine is hard to do at manufacturing scale, and the source flags this as a yield problem, not a solved problem.
The other architectural bet the article highlights is Nvidia's use of TSMC's COUPE platform, which brings the laser fiber in from the top of the package rather than from the edge. TSMC showed COUPE at IEEE ECTC 2025, and the SemiEngineering analysis identifies Nvidia as the first customer. Whether that architecture becomes the default for everyone else, or stays a single-vendor choice, is one of the open questions that will decide whether the five-year transition actually happens on schedule.
There is a reference architecture in the analysis worth holding onto. A representative CPO compute rack connects compute trays to a shuffle box via single-mode fiber, then to a switch tray. The source points to a working demo build with Ayar Labs and Wiwynn. That image is useful because it shows what "scale-up optical" actually means in physical hardware: not a single magic chip, but a chain of components, each with its own dB and dollar cost, that has to add up to a viable system.
The piece the analysis does not resolve is the supply side. A transition from copper to optical inside the rack is also a transition to a more concentrated vendor list. The laser source is a small number of suppliers. The polarization-maintaining fiber is a small number of suppliers. The detachable connectors, if they become standard, will likely consolidate around a handful of vendors. The optical engines themselves require advanced packaging that only a few foundries can deliver. Each of these is a yield, cost, or single-source risk that does not show up in the link budget itself but will determine whether the link budget closes at the volumes the AI buildout needs.
The original analysis frames this as a five-year transition driven by physics, with the dollar crossover as the open variable. Whether that crossover happens on the author's timeline depends on packaging yield, fiber management at scale, the willingness of hyperscalers to commit to a supply chain that does not yet exist in volume, and whether Nvidia's COUPE bet becomes an industry standard or a single-company architecture. The link budget in decibels already points to optical. The link budget in dollars is still being negotiated.