Paper Analysis | Meta, Broadcom and AMD Define the OCI 200G Line Spec: A Scale-Up Bet on Bidirectional Single Fiber, Micro-Ring DWDM and External Lasers
Meta, Broadcom and AMD have jointly released the OCI (Optical Compute Interconnect) 200G Line Interface Specification v1.0, pushing optical interconnect for AI back-end scale-up toward a concrete spec.
The spec makes four architectural bets: micro-ring (MRR) DWDM for low power and high density; bidirectional single fiber (Group A/B travel in opposite directions on the same fiber, halving fiber count); external laser source (ELS) (via OIF ELSFP, moving the laser out of the engine); and 53.125 Gbaud NRZ × 4λ = 212.5 Gbps per direction (NRZ rather than PAM4).
The rate ladder scales all the way to 1.6T: 200G OCI 1:4 / 400G 2:8 / 800G 4:16 / 1.6T 8:32.
This is a concrete product of the scale-up optical interconnect war: an alliance of a buyer (Meta), a switch/optics vendor (Broadcom) and an XPU vendor (AMD) betting on a "low power, high density, serviceable, simple" engineering philosophy to take on NVIDIA's vertically integrated route. Along the way, it draws a clear spec for the external laser supply chain.
1. Who Published This, and Why It Matters
Start with the byline: this is not a single-vendor spec but the OCI 200G Line Interface Specification v1.0 (March 11, 2026), co-edited by Meta, Broadcom and AMD. The editors span Meta (Amiralizadeh, Alduino, Peng), Broadcom (Ramaswamy, Brosnan, Traverso) and AMD (Streshinsky, Li, Settaluri).
Why does it matter? Because it is buyers and suppliers jointly defining the line interface for scale-up optical interconnect. OCI uses a DWDM wavelength grid plus cascaded micro-ring resonators (MRR) to deliver low-power, high-density optical interconnect for AI back-end networks. After the scale-up MSA battle (which we broke down in What Is XPO? The Loudest New Scale-Up Standard at OFC 2026 (in Chinese); OCI is one of those MSAs), this is the step where the three giants write the line-side physical layer into a concrete spec.
2. The Core: Four Architectural Bets in One Spec
In one sentence: the spec bets AI back-end scale-up optical interconnect on four choices: micro-ring DWDM + bidirectional single fiber + external laser + NRZ.
Let's take them one at a time:
Bidirectional single fiber (saving fiber): to ease fiber connectivity and density pressure in scale-up, each fiber carries both TX and RX, using two CWDM bands (Group A / Group B) traveling in opposite directions. A Type A transceiver transmits Group A and receives Group B; Type B does the reverse. This cuts fiber count in half outright.
53.125 Gbaud NRZ × 4λ: the PMD uses 53.125 Gbaud NRZ modulation, with four wavelengths forming one 212.5 Gbps serial data stream. This choice is worth noting: NRZ rather than PAM4. Density comes from 4λ DWDM and bidirectionality rather than from higher-order modulation's spectral efficiency, a deliberately simple, low-power route.
Rate ladder: the OCI PMA adapts 200/400/800GBASE-R and 1.6TBASE-R as 200G OCI 1:4, 400G 2:8, 800G 4:16 and 1.6T 8:32, where an m:n PMA demuxes m lanes of 212.5 Gbps into n lanes of 53.125 Gbps.
[Image] Source: Meta/Broadcom/AMD, OCI Spec v1.0 - Figure 1-2
3. Key Sections, One by One
3.1 External Laser Source (ELS): The Bet on Moving the Laser Out of the Engine
This is the chapter with the biggest supply-chain implications in the entire spec. OCI mandates an External Laser Source (ELS) compliant with OIF ELSFP (OIF-ELSFP-01.0), with light coupled into polarization-maintaining fiber (PMF) and then routed to the OCI chiplet. The laser is pulled out of the hot package and placed outside the module, trading for reliability and serviceability.
The ELS requirements are very specific: 1 MHz laser linewidth, 30 dB SMSR, 16 dB polarization extinction ratio, −144 dB/Hz RIN, with wavelengths on the CWDM/DWDM grid in Group A (~1308–1315 nm) and Group B (~1328–1335 nm). In effect, this draws a clear external laser spec for CW DFB laser and ELSFP suppliers. We discussed the motivation behind this "laser externalization" route in Polymer Waveguides Survive +20 dBm for Six Hours: The Missing "Optical Redistribution" Piece for ELS-Based CPO (in Chinese).
3.2 Micro-Ring DWDM and the Hidden Cost of Thermal Locking
OCI uses cascaded micro-rings for DWDM. Low power and high density are the upside, but micro-rings require heater locking. One detail in the spec says a lot: when TX squelches (silences modulation), average power must be held constant so the micro-rings at both TX and RX stay thermally locked. That single line exposes the engineering burden of the micro-ring route: thermal management is a problem micro-ring DWDM cannot avoid.
3.3 The Deskew State Machine: The Price of Bidirectional Single Fiber
Because the link is bidirectional and 4λ DWDM, there is skew between wavelengths (from dispersion and electrical path differences). The spec defines a deskew state machine (Relink → Detect → Sync → Validate → Mission) that can compensate 0–7 UI of lane-to-lane skew, and requires training patterns to be recognized even at a BER as high as 1E-4. It is a classic trade: using signal processing to absorb optical complexity.
3.4 Optics and Link Budget
Key specs: 53.125 Gbaud NRZ, 2 CWDM groups per fiber, 4 DWDM channels per group; Group A wavelengths 1308/1310.28/1312.58/1314.88 nm, Group B 1327.69/1330.05/1332.41/1334.78 nm; 30 dB SMSR, 6 dBm total launch power per group, 3.5–4.5 dB ER, −6.2 dBm stressed sensitivity, 1E-6 BER floor. The reference link is 500 m of SMF-28 with 2.5 dB insertion loss (mostly from connectors), explicitly defined as "the worst-case link in a data center back-end network." Management runs on CMIS 5.3, including VDM monitoring and a flight data recorder.
4. Technical Highlights: An Engineering Philosophy of Three Savings
String the spec's architectural choices together and you get a very consistent engineering philosophy: save power, save fiber, save heat (by pulling the laser out).
Micro-ring DWDM saves power: rings instead of MZMs/EMLs, betting on low power and high density (at the cost of thermal locking).
Bidirectional single fiber saves fiber: Group A/B run in opposite directions on one fiber, halving fiber count (at the cost of back-reflection / MPI management, which is why the spec includes MPI metrics and ORL tolerance).
External lasers save heat and maintenance: with the laser out of the hot package, reliability and serviceability improve.
Add the simple NRZ (rather than PAM4) route and the message is entirely consistent: scale-up optical interconnect should be low power, high density, serviceable and simple.
5. Industry Context: What the Three-Giant Alliance Means
The biggest industry signal in this spec is who is in the alliance: Meta (hyperscaler, buyer) + Broadcom (switch ASICs and optical IP) + AMD (XPU). This is the demand side and supply side defining the scale-up optical interface together, with an obvious aim of making it a de facto standard to take on NVIDIA's vertically integrated, proprietary scale-up (NVLink plus its own photonics).
For the battle over scale-up's "end-state architecture," see The Great Optical Packaging Transition (Part 2): CPO's Three-Stage Evolution, and Why Scale-Up Is the Endgame.
The actionable take for the supply chain: the spec makes external lasers mandatory and gives concrete requirements (1 MHz linewidth, 30 dB SMSR, PMF coupling, Group A/B wavelength grid). External lasers (CW DFB / ELSFP) go from "optional" to "a clearly specified link in the supply chain". In the shift of scale-up to optics, it is a signal that the laser supplier's position is being institutionalized.
But don't get carried away. The honest reality check: this is v1.0, with many sections marked "Subject to change" (VDM thresholds TBD); micro-ring thermal locking is a real engineering burden; bidirectional single fiber raises the bar on MPI / back-reflection management; and a spec is not volume production. Whether it becomes a de facto standard depends on whether the three companies' ecosystem can pull in other suppliers and buyers while racing XPO and other MSAs.
6. Conclusion
The OCI 200G line spec is not a paper; it is a concrete artifact the three giants want to turn into a de facto standard. With its four choices of micro-ring DWDM + bidirectional single fiber + external laser + NRZ, it takes a clear position: scale-up optical interconnect should follow a "low power, high density, serviceable, simple" route rather than higher-order modulation and vertical integration.
For anyone watching scale-up optical interconnect, the actionable take is this: read OCI as a concrete template of the "three-company alliance" version in the scale-up MSA war. Every spec choice is an architectural stance, and the clause "external laser via ELSFP" directly carves out a spec-defined market for CW DFB laser and ELSFP suppliers. NVIDIA goes its own way, while Meta + Broadcom + AMD want to build an ecosystem around a public spec. That is the real story behind OCI.
References
"200G Optical Compute Interconnect Line Interface Specification," Version 1.0, March 11, 2026. Editors: Siamak Amiralizadeh, Drew Alduino, Yiwei Peng (Meta); Anand Ramaswamy, Mike Brosnan, Matt Traverso (Broadcom); Matt Streshinsky, Mike Li, Krishna Settaluri (AMD).
Referenced specs: OIF CMIS revision 5.3, OIF ELSFP implementation (OIF-ELSFP-01.0), IEEE Draft P802.3dj.
Related Reading
What Is XPO? The Loudest New Scale-Up Standard at OFC 2026, Explained (in Chinese): the full picture of the scale-up MSA battle, of which OCI is one contender.
Polymer Waveguides Survive +20 dBm for Six Hours: The Missing "Optical Redistribution" Piece for ELS-Based CPO (in Chinese): the motivation behind the external laser (ELS) route and the missing piece.
The Great Optical Packaging Transition (Part 2): CPO's Three-Stage Evolution, and Why Scale-Up Is the Endgame: the context for why scale-up is CPO's endgame.
Comments