Copper is out of bandwidth. Optics is out of SNR. That asymmetry is the roadmap.

A puzzle worth five minutes

The interconnect industry in 2026 shows something genuinely odd.

The electrical signaling community is in a serious, open debate about a move away from PAM4. The OIF’s 448G framework, released at the end of 2025, keeps PAM6 and PAM8 as live candidates. It also lists a long tail of more exotic formats, and the working groups write careful analyses of each. The optical community looked at the same menu and declined without much discussion. It is often in the same companies, sometimes on the same campus.

Every pluggable generation from 400G through 1.6T runs PAM4 per lane. The optical capacity roadmap adds more lanes and more wavelengths, never more levels.

Same bits. Same decade. Same physics textbooks. Opposite conclusions.

You could explain this as culture, or conservatism, or a bad past experience in optics. The real explanation is better than that, and it is short. The two communities face the same trade at very different prices. One physical difference between their channels sets the prices: where the loss sits in frequency.

A careful look at that difference explains more than the modulation split. It explains most of the roadmap, including things that look unrelated. One example is why co-packaged optics keeps appearing on the same slide as 448G.

Let us calculate it.

What a channel actually cares about

Start with a fact that is easy to say and easy to underestimate. A channel never sees your bit rate. It sees your symbol rate.

A lane advertised as 224G carries 224 gigabits per second. PAM4 packs two bits into each symbol, so the wire changes state only 112 billion times a second. The highest frequency that matters, the Nyquist frequency, is half of that: 56 GHz. If you remember one derived quantity from this article, make it this one. Divide the rate by twice the bits per symbol. The result is the frequency the channel actually has to carry.

Why does this matter? For a copper trace, frequency and loss are almost the same thing. In this range, attenuation grows roughly in proportion to frequency. Intel’s board design guidance puts a good low-loss laminate at roughly 1 dB per inch when the Nyquist is at 14 GHz. At a 56 GHz Nyquist, the same inch costs several times more.

A signal travels across a switch board for ten or so inches. Over that distance the trace alone uses on the order of 35 dB. That is the entire budget a long-reach SerDes (the serializer-deserializer circuit that drives each electrical lane) must survive. This is why 224G host designs focus so closely on board materials. It is also the first half of the asymmetry. On copper, each added hertz costs decibels, and the cost is high.

Keep that fact in mind as we look at the obvious response.

The price of a level

If frequency causes the loss, the escape is clear. Pack more bits into each symbol so that the wire changes state more slowly. Go from PAM4’s two bits to PAM8’s three, and the symbol rate decreases by a third. The Nyquist decreases with it. At a fixed dB per hertz, the loss decreases too.

What does the escape cost? Think about the receiver’s problem. The transmitter has one voltage swing. PAM-L divides it into L levels, which leaves L minus 1 gaps between adjacent levels for the receiver to distinguish. Each gap is 1/(L−1) of the swing. A gap smaller by that factor costs 20·log₁₀(L−1) in signal-to-noise. This is the source of PAM4’s famous 9.5 dB penalty against NRZ. A step from PAM4 to PAM8 costs an additional 7.4 dB.

So the trade for the electrical designer is concrete. You get a third off the symbol rate, and you pay 7.4 dB of SNR. Whether that is a good trade depends entirely on how much loss the lower frequency refunds. Loss scales with frequency. So a symbol rate that is a third lower refunds about a third of whatever the channel cost.

Now the pieces fit. On a 40 dB host channel, a third refunded is 13 dB back, against 7.4 paid. That is a clear win, nearly 6 dB ahead. On a 10 dB channel the refund is 3 dB, and the same 7.4 dB price makes the trade a loss. Somewhere around twenty-something dB of channel loss, the trade changes sign.

That one observation explains why the 448G debate exists. It also explains why the debate stays open: the answer genuinely differs by reach. Long host channels are above the breakeven. Short in-package hops are far below it. A standard has to serve both. The OIF numbers show the two candidate escapes explicitly: 173 GBaud for PAM6 and 149 GBaud for PAM8. That is exactly what the divide-by-bits arithmetic predicts.

A second cost keeps the debate honest. Narrower eyes need stronger error correction (FEC). Preliminary OIF analyses for 448G flagged that the RS(544,514) code the ecosystem relies on looks risky under PAM6 and PAM8. Stronger FEC costs latency, logic and watts, in the host silicon where all three are scarce. PAM6 adds its own problem: log₂6 is 2.585, so bits do not map cleanly onto symbols. The dB arithmetic shows where higher orders can win. Implementation decides whether they do.

You can move all of these levers yourself here, and see the trade change:

Fiber refuses the deal

Now take the same offer to the optics floor. Both sides of it disappear.

The refund side first. Standard single-mode fiber attenuates about 0.4 dB per kilometer at 1310 nm. Most important, that figure does not change with your symbol rate. The 500-meter run of a DR link costs 0.2 dB whether the wire changes state at 50 GBaud or 200. A lower symbol rate is the whole point of higher PAM orders on copper. On an optical link, it refunds nothing at all.

The price side is more interesting. A photodiode produces current in proportion to the optical power on it. Electrical power is proportional to the square of current. In logarithms, the result is a factor of two: every optical dB counts as two electrical dB at the receiver. So the conversion halves the penalty. The 7.4 electrical dB that PAM8 costs becomes 3.7 dB of optical budget.

Is 3.7 dB a lot? One comparison answers the question. The IEEE gives 400GBASE-DR4, the standard 500-meter link, a maximum channel insertion loss of 3.0 dB. An early 802.3dj baseline proposal for 200G-per-lane DR keeps the same 3 dB channel inside a 6.5 dB total power budget. So the PAM8 step costs more than the entire channel allowance, on a link that gets nothing back for it. The optics engineers are not conservative. They can subtract.

Fiber has one thing that copper cannot buy at any price: room to multiply. A second fiber is cheap. A second wavelength on the same fiber is almost free, because photons pass through one another without interaction. Stacked carriers pay no crosstalk cost, but stacked voltage levels do. So optical capacity grows along axes that copper does not have:

• DR4 becomes DR8 when the fibers double.
• FR4 becomes FR8 when the wavelengths double.
• Every generation stays at PAM4 and multiplies. Every module roadmap from 400G to 1.6T shows exactly this pattern.

The one real escape from this logic is coherent detection. It recovers phase as well as amplitude, so dense constellations become possible again. Long-haul links use it to run 16QAM and beyond. Inside the data hall, coherent so far loses on cost and power, not on physics. The IM-DD (intensity modulation, direct detection) approach with parallel lanes continues to win. Coherent is worth watching, but not yet worth betting the hall on.

Reach, on a schedule

Now go back to copper with both halves of the asymmetry. Together they give the roadmap’s most consequential number.

Engineers design a SerDes to survive a roughly fixed loss budget. They build long-reach parts to a 35 to 40 dB class. Loss per meter follows the Nyquist frequency. Reach is just the budget divided by the loss per meter. So every doubling of the lane rate roughly halves the distance copper can go.

The published figures agree. They are in fact slightly worse than a clean halving, because connectors and skin effect do not scale linearly:

• about 2.5 meters of passive cable at 112G
• about 1 meter at 224G
• under half a meter projected at 448G

Before you trust those figures, check them in reverse once against a source we already used. The budget divided by the reach says a 112G cable loses about 15 dB per meter. Is that plausible? Intel’s board guidance of roughly 1 dB per inch at a 14 GHz Nyquist gives 39 dB per meter of PCB trace. Twinax cable has several times less loss than board laminate at the same frequency.

So mid-teens per meter for cable, at twice the frequency of that board number, is exactly where the two sources should agree. They do. Two independent measurements that agree with each other make a number trustworthy. You can do the same division again when the next generation’s reach is announced.

With real furniture, at real scale, the decrease looks like this:

At 112G, a cable crosses four racks, and copper easily wires a scale-up row. At 224G it barely reaches the next rack. While reach halved, the system to wire grew. A GB200 rack draws about 120 to 132 kW, and the roadmap points at megawatt racks. So a fixed amount of compute spreads across more cabinets, not fewer. The distance needed increased while the distance available decreased. For rack-to-rack links the two lines crossed at 224G. That is why scale-up fabrics go optical now and not at some later date.

The next fact makes this a schedule and not a trend. Lane rates do not double when the industry feels ready. They double when the switch does. A switch ASIC carried 512 SerDes lanes for three generations. Die edge and package limit that number. So far, its bandwidth and its lane rate were the same fact:

• 512 × 100G was the 51.2T generation.
• 512 × 200G is the 102.4T Tomahawk 6, now shipping in production volume.
• 512 × 400G is one route to the 204.8T generation. The other route is twice the lanes at 200G, the route that the XPO module serves.

Adding lanes needs a larger package, so most capacity doublings so far were lane-rate doublings. By the arithmetic above, each one is also a halving of copper’s reach. It comes on the silicon schedule, whether the cabling plan is ready or not.

This makes the modulation debate of two sections ago less important. Even where PAM8 wins its 6 dB, six dB gives back only centimeters at these loss slopes. Modulation improves life inside the budget. It cannot move the wall.

Only geometry moves the wall: make the electrical path shorter. That is the entire content of the shift from front-panel pluggables toward near-package and co-packaged optics (NPO and CPO). It is also the reason those acronyms share every roadmap slide with 448G. They are not on the same slide by coincidence. They answer the same problem.

The module, caught in between

One more consequence follows, and it affects the pluggable module.

Follow the logic of each side to its end. The host silicon has 30 to 40 dB channels, so it has every reason to go up to PAM6 or PAM8 at 448G. The optical line, for the reasons above, will still run PAM4. When those two formats differ, the module between them is no longer a re-timer that cleans a signal. It is a translator that converts between number systems.

Translation is DSP work. So the optical DSP (the digital signal processor in the module) gets a bigger job. It is already in volume production on 3nm and already competes for that capacity. This happens exactly when the industry hoped that module electronics would get smaller.

The same fork explains the future of linear-drive optics (LPO) better than the usual power-based explanation. LPO removes the DSP and lets the host SerDes drive the optics directly. That is possible only while both sides of the module use the same format. At 800G today, they do. Where the host goes up the PAM ladder and the line cannot follow, that condition fails.

So linear drive does not lose so much as move. It does well exactly where the electrical path is short enough for PAM4 from end to end, which means the NPO and CPO geometries. DSP-based modules keep the long, lossy paths that made higher-order modulation attractive at the start. Both architectures survive. The asymmetry gives each one its own territory.

The asymmetry, seen whole

From far enough away, the whole landscape is one picture.

One physical difference prices the same modulation trade oppositely on the two media: copper pays for hertz and fiber does not. That price split has these results:

• It settles the PAM debate differently by reach.
• It keeps the optical line at PAM4 and pushes host silicon to higher orders.
• It halves copper’s reach on the switch’s schedule.
• It makes fabrics optical from the outside in.
• It turns the module into a translator.
• It sorts DSP and linear-drive architectures into their own territories.

None of these were separate decisions. They are one asymmetry, which moves through the stack.

This gives a broader lesson about how to read industry debates, and it is the one lesson worth carrying beyond interconnect. Sometimes two competent communities study the same problem and disagree for a long time. Then the useful step is rarely to choose a side. Find the constraint that differs between them, because that constraint usually does all the work. When you find it, the disagreement stops being a controversy and becomes a map.

Under this particular map is a physical division of labor older than any roadmap. Electrons interact strongly, and this makes switching and gain possible. So electrons were always going to do the computing. Photons ignore one another, and this makes massive parallel carriage cheap. So photons were always going to do the connecting.

The industry does not choose between copper and optics. It finds, one budget line at a time, where the handoff between those two kinds of particle belongs. Each lane-rate doubling moves the handoff closer to the silicon. That, in one sentence, is the roadmap.

What to take away

1. Divide the lane rate by twice the bits per symbol, and the result is the Nyquist frequency. On copper, that frequency almost equals loss. This one-line calculation is behind everything else here.
2. Higher PAM orders refund about a third of channel loss and cost 7.4 dB. So they pay only on long, lossy channels. That is why 448G modulation is genuinely contested and why the answer differs by reach.
3. On fiber the refund is zero, and the square law doubles every optical dB. So the PAM8 step costs 3.7 dB against a DR budget of about 3. Optics multiplies lanes and wavelengths instead, at PAM4.
4. Copper reach halves each time the switch generation doubles the lane rate. Modulation cannot move that wall. A shorter electrical path can, and NPO and CPO are exactly that.
5. If host silicon and the optical line use different formats, the module becomes a translator and the DSP’s job grows. Linear drive concentrates where paths are short enough to stay PAM4 throughout.