OCP Global Summit 2026 | Microsoft | Fotini Karinou | Enabling AI Scale-Up: Why Standards Matter

1. Copper Is Out of Runway, but Microsoft Is Solving a Different Problem

Karinou's first message was blunt: reach, bandwidth density and pJ/bit together cap scale-up at the rack boundary. Optics can extend scale-up to multi-rack and multi-row domains, and the system value comes from utilization and resilience, not from the specs of a single link.

She first laid out Microsoft's platform connectivity needs as six use cases:

ID	Use case	Connectivity	Category
A	Switched scale-up fabric	Accelerator ↔ Switch	Scale-up
B	Direct scale-up fabric	Accelerator ↔ Accelerator	Scale-up
C	Host I/O expansion	Accelerator ↔ CPU (PCIe)	PCIe/CXL
D	Memory disaggregation and pooling	CPU/Accelerator ↔ Memory pool	PCIe/CXL
E	Package scale	Accelerator ↔ HBM	Scale-in
F	Chiplet interconnect	Die ↔ Die	Scale-in

The slide's conclusion: the near-term optical opportunities are only A (AI scale-up) and D (general-purpose memory disaggregation), and this talk covered only A. E and F belong to the longer-term scale-in category.

This framing matters. Over the past year, scale-up optics debates have often compared pJ/bit across in-package, in-rack and cross-rack links as if they were one problem; by separating the use cases first, Microsoft is declaring that different use cases need different specs, and it sets up the OCI choice that follows.

2. Breaking Down Three Challenges: Pace, Choices, Interoperability

Karinou grouped the challenges of scale-up specification into three categories.

The first is pace. AI systems evolve faster than traditional Ethernet interface cycles. The slide's timeline reads: 200 Gbps today, 400 Gbps "2027 (?)", 800 Gbps "TBD, >2030 (?)", while OCI Gen1 is marked "today". The question marks are the point: the traditional spec cadence cannot keep up, so a pre-planned multi-generation path is needed.

The second is too many choices. Modulation rate, modulation type, system architecture, component technology and package integration are all open, and every combination fragments the industry further. Karinou's phrase was "align without freezing innovation".

The third is interoperability. A scale-up network only works when the GPU endpoint, the optical interface, the fiber harness and the switch endpoint all interoperate. The slide's diagram shows a GPU/switch connected through an OCI interface and fiber harness to a switch, with light supplied by an ELSFP (external laser small form-factor pluggable).

▲ Microsoft's three scale-up spec challenges: specs can't keep pace, too many technology options, and endpoint interoperability; OCI MSA focuses on a common PHY layer (Image source: Microsoft | 2026 OCP Global Summit OIF session presentation (07:34))

3. The Four Choices of OCI MSA: NRZ, WDM, Bi-Di, Standard Lasers

The OCI MSA was founded in March 2026; the founders shown on the slide are Meta, Microsoft, OpenAI, AMD, Broadcom and NVIDIA, described verbally as "major hyperscalers, GPU and PHY vendors". It does one thing: define the optical link layer for AI scale-up.

Four design choices, each answering a scale-up requirement:

• NRZ modulation: avoids complex modulation that needs heavy FEC, keeping latency and power low.

• WDM: eight wavelengths in total, built on a frequency grid with an established laser supply chain, described as "standard lasers".

• Bi-Di on one fiber: transmit and receive share the same fiber with four wavelengths per direction, halving fiber count while cutting connection points and improving reliability and serviceability.

• Multi-generation path: Gen1 is 50G NRZ; Gen2 raises the per-wavelength rate to 100G; Gen3 adds more wavelengths.

By this math, Gen1 delivers 4 × 50G = 200G per fiber per direction, consistent with the 200 Gbps row matched to "OCI GEN1: today" on the slide; Gen2 at 100G/λ maps to 400G, close to the 2027 row (STT inference; the slide does not state a Gen2 date). We covered the Gen1 line-side details in Meta, Broadcom and AMD Define the OCI 200G Line-Side Spec; this talk adds Microsoft's roadmap logic.

▲ OCI MSA multi-generation path: Gen1 is 4λ × 50G NRZ at 200G per fiber per direction, Gen2 moves to 100G per wavelength, Gen3 adds wavelengths; BiDi halves the fiber count (Image source: Simple Tech Trend original; data from Microsoft 2026 OCP Global Summit presentation)

Put simply: OCI does not chase per-lane speed; it scales bandwidth with three degrees of freedom: wavelength count × per-wavelength rate × bidirectional transmission on one fiber. That is a completely different scaling logic from the fast-and-narrow 448G SerDes path.

▲ OCI MSA optical implementation: NRZ, WDM standard lasers and single-fiber BiDi, moving from GPU/switch endpoints through a common PHY to a multi-vendor ecosystem (Image source: Microsoft | 2026 OCP Global Summit OIF session presentation (09:58))

4. Breaking Down the Optical Interconnect Stack: Only One of 12 Layers Is Settled

Karinou immediately acknowledged that OCI is only a first step. She listed the optically interconnected system stack as 12 layers: materials; modulators, photodetectors and lasers; optical interfaces; electrical interfaces; optical engines and lasers; form factors; packaging; server and switch platform integration; connectors; fiber shuffles; rack and pod integration; and finally performance, reliability and interop testing.

Many bodies can address these layers: MSAs, OIF, OCP, APC, IEEE and ITU-T. The slide named only two "starting points":

• OIF COI white paper: published last year, the first time industry partners jointly wrote down scale-up requirements.

• OCI MSA: the NRZ, WDM, Bi-Di optical interface.

The remaining 11 layers still need ecosystem alignment. That line is the thesis of the talk: the bottleneck in optics has shifted from component performance to "who defines the other 11 layers".

5. The Division of Labor: What OIF, OCP and APC Each Own

Microsoft's proposal is "complementary, not competing" ownership, with the OCI MSA at the center and three organizations each filling a part:

Organization	Example focus areas
OIF	ELSFP for OCI, CMIS improvements, link reliability, link initialization, interoperability
OCP	Physical optical trays and rack designs, form factors, system reliability, connectors and interfaces, ecosystem requirements
APC	Wafer-to-system standards, SiP manufacturing specifications, CPO and SiP test and measurement, OSAT-level work

OIF's ELSFP specification already exists and is being extended to cover OCI laser requirements; APC (Advanced Photonics Coalition) defines manufacturing and test from wafer to system.

▲ Microsoft's proposed division of labor around OCI MSA: OIF owns ELSFP and CMIS, OCP owns optical trays and racks, APC owns silicon photonics manufacturing and CPO test (Image source: Microsoft | 2026 OCP Global Summit OIF session presentation (13:42))

Mapping the 12 layers onto these three organizations reveals two gaps: electrical interfaces and materials are not named. The electrical interface gap is especially important, and the Q&A in the next section hits it directly.

▲ STT maps Microsoft's 12-layer optical interconnect stack against the ownership table: the optical interface is defined by OCI MSA and most layers have an owner, but the electrical interface and materials are unassigned (Image source: Simple Tech Trend original; data from Microsoft 2026 OCP Global Summit presentation)

6. How This Compares With the ECOC Spec Sheet

In ECOC 2026: Microsoft Writes Down the Optical I/O Spec Sheet we covered Microsoft's long-term targets for wide-and-slow optical I/O. The Q&A at OCP drew a clearer line between near term and long term:

• Non-laser sources such as micro-LEDs: Karinou said scale-up must guarantee multi-rack reach, which OCI provides; micro-LEDs, VCSELs and nanolasers "still have a road to mature", and reach-limited technologies may land first in in-package applications.

• 50G NRZ versus 448G SerDes: an audience member asked how 50G NRZ optics connect to XPU SerDes climbing to 224G and 448G. Her answer was that the long-term goal is to "get rid of the SerDes" and drive optics close to the accelerator with wide parallel interfaces, where power and bandwidth density are best.

• Packaging path: the first step will not be co-packaged with the accelerator but an off-package solution, moving to tight integration eventually.

• Retimed or linear: when asked by Broadcom, she said only that the default is the most power-efficient option, that linear or half-retimed is "hopefully" possible, and that it is still being evaluated.

What these answers reveal: near-term OCI optics need a conversion layer to high-speed SerDes, and no one owns that layer. That is the practical meaning of the electrical interface gap in the previous section. For the fast-and-narrow versus slow-and-wide debate, see ECOC 2026: Three Hours of Fast-Narrow vs Slow-Wide, and the Dividing Line Is 448G.

7. Risks and Counterarguments

• Side effects of specs set by a few hyperscalers: the OCI MSA was founded by six large companies, which speeds alignment, but it may not be optimal for other customers and could create a different kind of split with the IEEE Ethernet path.

• NRZ's scaling ceiling: beyond 50G and 100G NRZ, raising the per-wavelength rate would require more complex modulation, bringing back the FEC and latency penalties Microsoft wants to avoid, so Gen3 can only add wavelengths.

• Practical challenges of Bi-Di: bidirectional transmission on one fiber is more sensitive to reflections and crosstalk; Karinou also acknowledged that using Bi-Di for data center interconnect would require replacing existing fiber, which may not be cost-effective.

• The division of labor is only an "example": the slide says "Example Focus Areas", and the talk made no commitment on whether the three bodies will actually split the work this way or on what timeline.

• Very few hard numbers: the talk gave no pJ/bit, reach or cost targets, and Gen2 and Gen3 timelines were not disclosed.

8. What It Means for Taiwan's Supply Chain

• Laser source specs are being defined at OIF. OCI uses WDM standard lasers supplied by ELSFP external laser modules, and ELSFP for OCI is being extended at OIF. For Taiwan's InP laser epitaxy and chip makers, this is a rare window where the spec is not yet fixed and participation is possible; the eight-wavelength frequency plan and power requirements are the details to track next.

• Halving fiber count changes harness and connector volume. Bi-Di carries the same bandwidth on half the fibers, so demand for fiber harnesses, connectors and fiber shuffles becomes "fewer units, higher spec per unit"; Taiwan's cable and connector makers should focus on value per fiber rather than fiber count.

• OCP optical tray and rack design is home turf for Taiwan's ODMs. As scale-up moves from in-rack to multi-rack, if tray and rack optical integration specs are defined at OCP, rack integrators need to join the workstreams early.

• APC's OSAT-level work maps directly to Taiwan's packaging and test houses. Once SiP manufacturing specs and CPO test and measurement are standardized, OSAT process and test capabilities become comparable specs rather than per-customer customization.

9. Conclusion

The value of Microsoft's talk is that it changed the question from "which optics is best" to "who defines the remaining 11 layers". The OCI MSA gives scale-up a manufacturable optical interface with NRZ, WDM and Bi-Di, plus a multi-generation path of "per-wavelength rate × wavelength count"; but laser supply, trays and racks, packaging and test, and above all the electrical interface conversion are still waiting for standards bodies to take ownership.

The timeline needs to be read separately. OCI Gen1 has a spec today, Gen2 at 100G/λ lands around 2027 by STT's inference, but the full ecosystem division of labor is still only an example. Over the next 12 months, watch three things: when the OCI Gen2 spec is published, when OIF finalizes ELSFP for OCI, and whether OCP forms a corresponding optical rack workstream.

Optics beating copper is no longer a physics problem but a coordination problem; whoever writes the remaining 11 layers into specs first decides what the scale-up supply chain looks like.

This article is for technology and industry trend analysis only and does not constitute investment advice.