OCP Global Summit 2026 | OIF | Ian Betty, Steven Hand | Coherent Optics and Management for the AI Era
1. Why 400ZR Still Matters in the AI Era
In this OIF workshop, Ian Betty, senior director at Ciena and OIF board member, covered the evolution of coherent specifications, and Steven Hand, Nokia Fellow and OIF board member, covered the CMIS module management interface.
OIF started its 400G data center interconnect (DCI) project in late 2016 and published its first implementation agreement (IA) in March 2020. 400ZR got two things right: it made coherent optics interoperable across multiple vendors for the first time, and it did so in a client pluggable form factor (QSFP-DD) that plugs directly into switches and routers.
That was the precondition for distributed data center architectures: point-to-point links under 120 km, 64 C-band channels, 25.6T per fiber. In ECOC2026 | 95% of OCS Shipments Are Google's Business we saw AI clusters start to span buildings and campuses; those cross-building links run on this lineage of specs.
400ZR is not an AI spec, but every scale-across coherent module today grew out of its framework.
2. Breaking Down Three Generations: Double the Baud, Double the Tx Power
OIF has published six coherent IAs: three versions of 400ZR (Mar 2020, Nov 2022, Oct 2024), 800ZR (Oct 2024), 800LR (Apr 2025), and the newly published 1600ZR (Sep 2026).
Every generation scales the same way: double the baud rate and stay with DP-16QAM. The catch is that doubling the baud requires doubling the transmit power to keep the same Tx power spectral density.
Spec | Modulation | Baud rate | FEC | Tx output power |
400ZR | DP-16QAM | 59.8 GBd | CFEC | -10 dBm |
800ZR Range A | DP-16QAM | 118 GBd | OFEC | -8 to -1 dBm (amplified, SOA/EDFA) |
800ZR Range B | DP-16QAM | 118 GBd | OFEC | -12 to -6 dBm (low-loss modulators, TFLN) |
800ZR Range C | DP-16QAM | 118 GBd | OFEC | -15 to -12 dBm (no amplification, SiPh) |
1600ZR | DP-16QAM | 236 GBd | OFEC | -5 to -1 dBm (amplification required) |
800ZR was split into three Tx output ranges because meeting customer needs without optical amplification was no longer a given: Range C for unamplified silicon photonics, Range B for thin-film lithium niobate (TFLN) modulators, and Range A for designs with SOA or EDFA amplification.
1600ZR removes that flexibility: in-module optical amplification is now a requirement, so the spec puts much more weight on the transmit OSNR (TOSNR) introduced by the amplifier: 37 dB in-band, 35 dB adjacent channel, 23 dB non-adjacent channel.
▲ Each generation of OIF coherent specs doubles the baud rate and raises Tx power requirements; by 1600ZR, in-module optical amplification becomes mandatory (Image source: Simple Tech Trend original, data from OIF 2026 OCP Global Summit presentation)
3. The 800G Lesson: The Market Rode OIF's Framework, but Not Exactly 800ZR
Betty called the 800G generation "complicated." The project began in November 2020 with no dominant application and no large customer pull, with both DCI 800ZR and campus 800LR in scope. The FEC choice swung between a direct 400ZR overlay and new features such as optical line protection; OFEC won, giving about 2 dB more RSNR than 400ZR.
OpenROADM then built on 800ZR with lookup-table-based probabilistic constellation shaping (LUT-based PCS), defining 800G 8e-DPO (131 GBd) and 600G 6e-DPO (118 GBd) modes.
The line on the slide worth remembering: to date, 800G modules supporting 8e-DPO and 6e-DPO have been a larger scale-across market than 800ZR-only modules, and they ship almost entirely in Range A. In other words, unamplified Range C silicon photonics has almost no share of real 800G shipments; by making amplification mandatory, 1600ZR writes the market's choice into the spec.
4. The 1600G Generation: Every Project Has a Hyperscaler Sponsor
The 800G lesson changed how OIF works. Every 1600G project was started by an end customer:
Project | Start | Sponsor | Application | Status |
1600ZR | Q3 2023 | Microsoft | DCI, direct 400ZR overlay | Published Sep 2026 |
1600ZR+ | Q1 2024 | Meta (AT&T support) | Pt-to-pt multi-span amplified, C+L band | Expected within 1-2 quarters |
1600CL | Q4 2024 | O-band campus, high-loss OCS | Key decisions under debate | |
MACsec support | Q2 2026 | Oracle | Coherent module encryption | White paper stage |
1600ZR has a single application code: 16 x 300 GHz channels in C-band, up to 120 km, in OSFP. Channel frequencies were chosen to stay compatible with existing line systems; Microsoft wants to swap out 400ZR without changing the line system.
1600ZR+ brings L-band into an OIF standard for the first time and targets a direct overlay of the widely deployed 800G 8e-DPO (131 GBd) links, with four application codes:
1600ZR+: C-band and L-band, up to 1,000 km
1200ZR+: C-band and L-band, up to 2,000 km
16-QAM with hierarchical tree probabilistic amplitude shaping (PAS) and OFEC, two digital subcarriers, about 252 GBd in total
▲ 1600ZR was published in September 2026: 236 GBd DP-16QAM, 16 × 300 GHz channels, with in-module optical amplification now mandatory (Image source: OIF | 2026 OCP Global Summit presentation (07:14))
Time from start to publication is also shrinking. 400ZR took about 3.3 years, 800ZR about 4, and 1600ZR about 3; if 1600ZR+ publishes within 1-2 quarters as expected, it will also take about 3. Customer-sponsored projects are moving about a year faster than the customer-less 800G generation.
▲ OIF coherent project timelines from start to publication; the customer-initiated 1600G generation is about a year faster than the 800G generation (Image source: Simple Tech Trend original, data from OIF 2026 OCP Global Summit presentation)
5. 1600CL: Coherent Optics Moves Squarely into Campus and OCS
1600CL (Coherent-Lite) is the project to watch this generation, because it takes coherent optics from metro DCI into campus and optical circuit switching (OCS). The slide lists three use cases: 1-20 km campus links, fiber-efficient campus WDM, and high-loss OCS Ethernet front-end and xPU interconnect links.
Two key conditions:
Low latency: high-loss applications require latency below 300 ns
WDM in scope: specific wavelength proposals for campus WDM have already been heard
But major technical decisions are still "being actively debated": multiple concatenated and segmented FEC proposals are in play, DSP framing is open, and OIF is even revisiting its coherent IA optical specification methodology, considering a normative TQM (transmit quality metric) requirement.
▲ 1600CL was initiated by Google in Q4 2024, targeting high-loss campus and OCS applications with latency below 300 ns; FEC and DSP framing are still under debate (Image source: OIF | 2026 OCP Global Summit presentation (08:18))
OCS is Google's home turf. Our view: until 1600CL is settled, high-loss OCS and campus links will remain split between IMDD and proprietary solutions; only after the FEC and latency trade-offs are decided will we know how much of OCS coherent can win.
6. CMIS: The Layer That Makes Millions of Modules Manageable
Hand's half covered CMIS (Common Management Interface Specification), the management interface between host and module. It spans passive copper cables to coherent optics beyond 3,000 km, and is used across 400G, 800G and 1.6T. The slide calls it the highest-volume pluggable management interface ever.
CMIS spares hosts from hard-coding each vendor's module, but as modules grow more complex, two new challenges have emerged:
New features are gated by the host: new module capabilities cannot be used until the host OS is upgraded, which Hand described as a multi-year delay plus an expensive verification phase
No consistent firmware update method: CMIS defines how to download firmware, but in large multi-vendor networks there is no consistent way to get the right firmware onto the right module; Hand said Microsoft and Google asked OIF to tackle this
OIF has two responses:
Extended module advertising (ECOC 2026 demo): CMIS 5.4 modules use CDB 50h/51h commands to report application-dependent properties and interface descriptions, so the host can pass capabilities up to the controller without host software changes. The demo used a Cisco-Acacia 800G ZR OSFP and a Genuine Optics 1.6T OSFP, with host emulation by Wilder Technologies (an Amphenol company).
CMIS Firmware Package IA: a .tgz package containing a YAML metadata file (vendor name, part number, hardware revision, target firmware version, file name, SHA-512 checksum) paired with the original firmware load, which needs no change and is backward compatible with CMIS 5.x; OIF also liaised with OpenConfig on the northbound interface (NBI), the first time OIF has defined work in step with OpenConfig.
▲ The CMIS firmware package pairs YAML metadata with the original firmware file, is backward compatible with CMIS 5.x, and aligns with OpenConfig to enable automated updates of multi-vendor modules (Image source: OIF | 2026 OCP Global Summit presentation (16:14))
Next on the roadmap: an I3C management interface, 1.6T modules for IEEE 802.3dj (APSU/ILT link training), the CMIS-sec secure channel, MACsec engine management, and management of CPO, NPO and new form factors such as XPO and Open CPX. We covered Open CPX in ECOC2026 | Ciena Turns the 6.4T Optical Engine Back into a Shelf Part with Open CPX; for optical engines to become shelf parts, their management interface has to be standardized too.
7. Risks and Counterarguments
Spec is not market: what actually scaled at 800G was OpenROADM's 8e-DPO mode, not pure 800ZR; whether 1600ZR repeats that depends on shipments.
1600CL timing is the most uncertain: FEC, DSP framing and specification methodology are all still debated; on the roughly three-year cadence of earlier generations, a decision before end-2027 would already be optimistic (STT estimate).
Power and cost once amplification is mandatory: the 1600ZR slides list no module power. Fitting an amplifier and a 236 GBd DSP into OSFP is a thermal challenge; 400ZR was in the 15-18 W class (stated verbally), and no 1.6T figure was given.
New CMIS features still need host support: hosts and controllers must first support CDB 50h/51h and the new OpenConfig models.
8. What It Means for Taiwan's Supply Chain
The talk named no Taiwanese companies; the ECOC demo modules came from Cisco-Acacia and Genuine Optics, so what follows is inferred from the spec content only.
Coherent module value concentrates in the DSP and amplifier. With amplification mandatory in 1600ZR, the 236 GBd DSP, the amplifier and TOSNR control become the entry barrier; unamplified silicon photonics no longer has a matching class in the spec. Taiwanese companies aiming at 1.6T coherent modules must answer on amplifier integration and thermal design, not just packaging capacity.
White-box switch makers need to prepare for new CMIS features. Module advertising and firmware packages both require host-side support; Taiwan's switch ODMs are where host software lands, and their support for CDB 50h/51h and the OpenConfig firmware model will affect how fast they win hyperscaler orders.
DCI demand is certain and the product cadence is clear: 1600ZR is published and 1600ZR+ should follow around 1H 2027 (based on the 1-2 quarter estimate), consistent with the DCI momentum we saw in Earnings Highlights: Fabrinet (FN) | Q4 FY2026, and a useful time anchor for coherent component demand.
9. Conclusion
The workshop had no new products, but it drew the coherent roadmap clearly. On specs, baud doubles every generation and amplification has gone from optional to mandatory; on organization, OIF has moved from a customer-less 800G generation to every project having a hyperscaler sponsor, cutting timelines by about a year; on management, CMIS is moving from defining registers to letting modules advertise their own capabilities and automating firmware updates.
Separate the timelines. What is certain is 1600ZR (published Sep 2026) and 1600ZR+ (expected within 1-2 quarters), covering DCI and long-haul scale-across; 1600CL for campus and OCS is still under debate. Three things to watch over the next 12 months: first 1600ZR interop tests and module power, the 1600ZR+ IA publication date, and when 1600CL settles its FEC and latency trade-offs.
Coherent optics has moved from a metro networking business to infrastructure for AI clusters spanning buildings and regions; the next battleground is campus and OCS, and the rules are not yet written.
This article is for technology and industry trend analysis only and does not constitute investment advice.
Comments