Reading the OCI MSA, where scale-up optics chose NRZ

Who showed up, and who did not

In March 2026 a new optical consortium announced its founding members: AMD, Broadcom, Meta, Microsoft, NVIDIA and OpenAI. These names rarely sign the same document. The absent names are equally informative. Among the founders there are no module vendors and no connector companies. The companies that buy compute and the companies that make its silicon met to specify light between chips. The industry that now packages that light into boxes was not in the room.

The member list is the first clue to what the OCI MSA (a multi-source agreement) actually is. It is not a module standard. It is the scale-up fabric, the NVLink-class network that connects accelerators into one machine. That fabric now declares that its future is optical, and it writes the rules itself. Scale-up is the traffic that copper is losing as copper’s reach decreases. It is also the most demanding link class in the building, with the shortest distances, the highest bandwidth density, and the least tolerance for latency or failure.

This makes the consortium’s first technical decision look, for a moment, completely backwards.

The modulation went down

The OCI generation-one link is four wavelengths of NRZ at about 50 gigabits each (non-return-to-zero, two signal levels and one bit per symbol). That gives about 200 gigabits per direction per fiber. A second generation reaches 800 gigabits bidirectional per fiber. The stated roadmap scales wavelength counts and rates to 3.2 terabits and beyond.

This is fifty gig NRZ. The electrical world debates PAM6 and PAM8 for 448G. Every pluggable module since 400G uses PAM4 (four signal levels, two bits per symbol). But the most advanced buyers in the industry specified the simplest modulation format there is. The Ethernet mainstream stopped using it years ago. If more complex modulation were progress, this choice would be regression.

Readers of the copper-versus-optics essay will recognize something else. It is the napkin calculation working exactly as printed, one step further than the module industry took it. Do the calculation once more:

• A change from PAM4 back to NRZ refunds the well-known 9.5 dB level penalty.
• At the photodiode, optical dB count double. So the refund is nearly 5 dB of optical budget.
• On these link classes, the full channel allowance is about 3 dB. The refund is bigger than the budget.

What does NRZ cost on fiber? One option is twice the symbol rate for the same bits. That spends bandwidth, which an optical channel has in large quantity. The other option is half the bits per wavelength. That spends wavelengths, and wavelengths on fiber are nearly free. So the trade is simple: buy decibels, the scarce resource, with wavelengths, the cheap one.

The sources show two more advantages that the napkin does not. A 50G NRZ signal has a wide-open eye (the pattern that shows the receiver’s margin). It needs far less equalization and far lighter error correction than PAM4. So the path needs less DSP (digital signal processor) silicon and adds fewer nanoseconds of latency. Latency is critical on a scale-up fabric, where every GPU waits for every other GPU. Here, simplicity is not a cost-saving measure. It is the goal.

From module-centric to silicon-centric

The MSA itself leads with the phrase “shift from a module-centric to a silicon-centric model”. This phrase needs explanation, because it is the structural half of the announcement.

The whole placement ladder above the board assumes the module-centric world. In that world, optics is a product in a box. You buy it, plug it into a cage, and interchange it across vendors at the faceplate. Silicon-centric means that the optical interface becomes a property of the compute package itself. It is an OIO (in-package optical I/O) chiplet beside the die, and the fiber effectively connects to the processor.

Then the thing to standardize is no longer a box. It is the light: wavelengths, rates and formats. With these fixed, one vendor’s in-package optics can talk to another vendor’s optics across the fabric. That is precisely what OCI specifies. It is also why its founders are silicon vendors and buyers, not module companies.

On the map, OCI is the deepest of the three OFC agreements:

• XPO standardizes the box at the faceplate.
• OpenCPX standardizes the engine at the package edge.
• OCI standardizes the light between packages, the layer where boxes no longer exist at all.

What to watch, and what it means here

A consortium announcement is a beginning, not a fabric. An honest reading keeps three questions open:

• Interoperability at these depths needs a demonstration, not a declaration.
• The reliability problem that in-package optics inherits does not disappear because a wavelength plan exists.
• Laser counts and packaging yield, not the spec document, will set the pace of the roadmap from 200G to 3.2T per fiber.

But the direction is unambiguous. Every essay on this site arrives at it from a different starting point:

• The scale-up fabric is going optical.
• Its physics rewards the lowest modulation order and the widest parallelism.
• The optics is moving into the silicon.

The industry is moving its standards to where the photonics actually lives.

What to take away

1. OCI is the scale-up fabric’s own optical standard. Buyers and silicon vendors founded it: AMD, Broadcom, Meta, Microsoft, NVIDIA, OpenAI.
2. Generation one is 4 × 50G-class NRZ per fiber, about 200G per direction, with a roadmap through 800G bidirectional toward 3.2T.
3. A napkin calculation predicts the move back to NRZ. It refunds nearly 5 optical dB on 3 dB budgets and pays in wavelengths, which fiber has to spare. Lighter DSP and lower latency are free additions.
4. Silicon-centric means that the optical port becomes part of the compute package, and the standard governs light, not boxes. That is the OIO rung of the ladder getting its rulebook.
5. Watch demonstrated interoperability and laser reliability, not the press releases. Those two decide when the spec becomes a fabric.