OCP Global Summit 2026 | OCP Photonics Panel | Nathan Tracy, Ryan Yu, Andy Bechtolsheim, Rajiv Pancholy | Standards and Future Technologies

1. Why this panel deserves its own read

This year's OCP photonics track grew from a half day to a full day, and submissions still ran at more than twice the available slots. The closing panel covered four distinct positions: a standards body (OIF), an optics and silicon photonics supplier (TeraHop), a system vendor (Arista), and the switch-silicon and CPO pioneer (Broadcom). The moderator said he was from LightCounting.

Bechtolsheim opened bluntly: IEEE and similar bodies have done great work for years, but their pace is far behind AI. Customers want multiple sources and a ramp from zero to a million units in about six months, which changes the whole development process.

2. Breaking down the new job of standards: white papers first, specs second

Tracy said an OIF Implementation Agreement used to take several years; OIF now publishes fast white papers first so everyone agrees on the challenges and vocabulary, which makes the formal spec much faster to write.

He named several projects in flight:

• The EEI working group has published several white papers, including COI (compute optical interconnect), which maps the needs of next-generation energy-efficient optical interconnects and electrical channels. We covered it in OIF Draws the Official Map for AI Interconnect.

• 448G electrical interface specs are under way: "generate the electrons to drive the photons."

• High-density connectivity discussions started at OIF about 18 months to two years ago, and an MSA later delivered XPO quickly.

Yu added that Open CPX followed the same path: announced this March, but debated in OIF meetings throughout the year before. His division of labor: OIF brings breadth, MSAs bring execution speed.

Tracy's most important line: the industry does not have enough engineering bandwidth for every problem that needs solving. There are several NPO variants, plus DPO and CPC, and forums help everyone set priorities. In other words, a standards body's new role is "filter," not "inventor."

3. Disagreement one: product first or standard first, and Broadcom switched playbooks

The moderator put it directly: Broadcom was first to bring CPO to market, when OIF's external laser (ELSFP) was the only available standard, and it won time by shipping to customers for evaluation first. For OCI it reversed course: form the MSA first, then talk products.

Pancholy replied that Bechtolsheim's public criticism after the CPO launch "actually helped us get better." In scale-up, Broadcom wants to "unify the approach first and bring everyone along this time," rather than have one vendor go first and everyone else work out interoperability later.

Broadcom's CPO roadmap is itself the scorecard of the product-first approach:

▲ Broadcom's four-generation CPO roadmap: 25.6T in 2022, 51.2T in 2024, 102.4T (200G/lane) in 2026 and a fourth generation in 2028, with over 1 million cumulative device-hours and zero link flaps (Image source: Broadcom (from Pancholy's earlier talk, not cited in the panel) | 2026 OCP Global Summit presentation)

Broadcom's deck claims CPO is 10x more reliable than pluggables. For third-generation details see OCP APAC Summit 2026 | Broadcom | CPO's Third-Generation Report Card.

The OCI MSA was founded on March 12, 2026, by Meta, Microsoft, OpenAI, AMD, Broadcom and NVIDIA. The Broadcom deck states the goal of "a broader industry participation structure while maintaining pace of innovation," which means six companies set the direction for now and others follow.

4. Disagreement two: comb lasers or four ordinary lasers

Bechtolsheim said that before OCI there were "at least a handful" of slow-and-wide proposals with different wavelength counts and plans, none interoperable. Each required its own investment and tape-out, and from an ROI standpoint nobody invests without an upfront standard.

The key choice in OCI Gen1 is four wavelengths per direction. Comb lasers can produce many colors but are not really in volume production; four wavelengths can be built with four discrete lasers, which removes the comb-laser risk entirely. The panel added that if comb lasers mature to 8 or 16 colors, the same architecture can scale further.

▲ OCI-MSA Gen1 spec: three optical engine placements, and a bidirectional (Bi-Di) link with two groups at 1311 nm/1331 nm and 4 wavelengths per direction per fiber (Image source: Broadcom (from Pancholy's earlier talk, not cited in the panel) | 2026 OCP Global Summit presentation)

This is the panel's methodology: a standard does not pick the strongest technology; it picks what can be built in volume today and leaves an upgrade path. For OCI 200G line-side details, see Meta, Broadcom and AMD Define the OCI 200G Line-Side Spec.

5. Disagreement three: multicore fiber now, or next decade

The only real tug-of-war on the panel was about fiber. Yu framed multicore fiber (MCF) as an alternative path to OCI; another view on stage was that the two stack: multiple wavelengths times multiple cores gives another 4x, 16x in total. Someone also noted that DCI facilities have fiber trays so full they cannot fit more, requiring at least a 2x to 4x reduction.

The moderator poured cold water: the fiber ecosystem is "very, very slow," connectorization and splicing must be in place first, and MCF is "probably a next-decade technology." Someone on stage immediately replied: "They're making good progress on the connector."

Why fiber matters becomes clear from the 1,024-XPU cluster in the Broadcom deck:

▲ A 1,024-XPU scale-up cluster needs 131,072 fibers, or 32,768 per switch rack; with 4-core MCF (STT assumption) that drops to 8,192 per rack (Image source: Simple Tech Trend original; data from Broadcom 2026 OCP Global Summit presentation)

• Each GPU has 25.6 Tbps over 128 Bi-Di fibers; STT calculates 200 Gbps per fiber per direction, exactly four wavelengths at 50G.

• GPU side 1,024 × 128 and switch side 128 × 1,024 both equal 131,072 fibers (STT calc).

• Each switch rack carries 32,768 fibers; each GPU rack 8,192.

TeraHop has published a 1.0 spec with fiber vendors, and Tracy said OIF's ZR interoperability demos at OFC and ECOC this year included multicore fiber. Both aim at the same thing: get customers trialing first. As Yu put it, without trials it simply won't happen.

6. The hardest line: optics reaches the GPU last

The final audience question hit the key point: why are there many CPO demos on switches and few on GPUs?

The answer on stage was economics: a GPU going down costs far more than a scale-out switch going down, so reliability is proven on switches first, then brought to the GPU. The follow-up was even more direct (speaker not labeled in the transcript): if you commit a native OCI interface (50G NRZ) to your next GPU now, that GPU ships about two years later; if the optics fail, the GPU cannot ship, "you're dead." So the technology must be 100% proven, reliability included.

▲ The gates for optics entering scale-up: first build reliability in switches, then pass known-good-engine testing, and only then GPU-native OCI (Image source: Simple Tech Trend original; data from Broadcom 2026 OCP Global Summit presentation and the photonics panel)

Testing is the second gate. An audience member assumed NPO might arrive in the second half of 2027 and asked how EIC and PIC can be tested together at volume. Pancholy said Broadcom first ensures a known good engine, then places a known-good switch die; another speaker described an ECOC vendor cutting trenches into wafers to do wafer-level testing of edge-coupled devices with micro-probes. Yu said TeraHop went from 4 and 8 channels to 16 and 32, has shipped tens of millions of transceivers, and that gradual steps are manageable.

7. The shared anxiety: lead times went from 3 months to a year

The closing question was "what keeps you up at night." Put the four answers together and you get the real 2027 bottleneck:

• Bechtolsheim: optics lead times used to be 3 months and are now 9 months to a year, partly due to laser shortages; order millions more today and delivery is next summer.

• Yu: short on practically everything, with shortages appearing in unexpected places; growth needs to be managed, since growing too fast plants the next problem. InP makers are pushing 6-inch wafers and GaAs is being evaluated, but that is two to three years out.

• Tracy: the pace of change, with new products scaling to millions overnight; "there are still only 24 hours in a day."

• Pancholy: whether the industry hits the adoption window, and that a single scale-up node grows from 72 or 144 today to roughly a thousand once optics arrive.

8. Risks and counterarguments

• "Standards first" could become "six-company standards." Six large companies set the direction of the OCI MSA; if broader participation comes too slowly, other slow-and-wide approaches may go their own way and repeat the interoperability problem.

• Switch reliability is not GPU reliability. More than 1M device-hours with zero link flaps is switch data; GPU packages have different thermal and yield conditions and cannot be extrapolated directly.

• The cost of discrete lasers is not disclosed. Four discrete lasers lower technical risk but raise laser count and packaging cost, and the panel gave no numbers.

• NPO in 2H 2027 was only an audience member's assumption, not a commitment from any panelist.

9. What it means for Taiwan's supply chain

• Packaging is already on the list. On Broadcom's TH6-Davisson CPO partner slide, the packaging column is TSMC and SPIL, and the deck notes EIC-to-PIC attach uses TSMC COUPE. If GPUs get optics two years later, as the panel's sequence implies, packaging houses' optical revenue will lean on switches in the near term.

• Rack and system partners include Delta and QCT, listed alongside Celestica, Jabil, HPE and Supermicro as TH6-Davisson rack partners.

• The ELSFP list includes Foxconn group's FIT, alongside Eoptolink, O-Net and Source Photonics.

• Test and yield are the next challenge. The panel repeatedly stressed known-good engines and wafer-level test but named no Taiwanese test company, so this remains a watch item.

10. Conclusion

The panel broke "when does optics enter scale-up" into three gates: agree the standard (OCI is set), prove reliability on switches (Broadcom is on its third generation), and wait for 100% validation on the GPU (a GPU design cycle of about two years). All three share one logic: at a pace of a million units in six months, nobody will bet a GPU on a technology not yet in volume.

So STT's call is: through 2027 the main optical battleground in scale-up remains switch-side and rack-to-rack NPO/CPO, and GPU-native optical I/O will not reach volume before 2028. If a GPU with native OCI ships within 2027, this call is wrong. Over the next 12 months, watch three things: whether the OCI MSA opens to more members, whether optics lead times fall back from 9-12 months, and when the first XPU with a native optical interface tapes out.

The job of a standard is no longer to pick the strongest technology, but to make a million units buildable in six months.

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