OCP Global Summit 2026 | Oracle | Kannan Raj | Toward NPO Deployments in AI Fabrics

1. Why Oracle Has Standing on NPO: It Already Did This 12 Years Ago

NPO is a new term in most decks, but not for Oracle. Raj opened with an old photo: about 12 years ago, Oracle switches already shipped socketed MBOMs (midboard optical modules) in volume, supplied by TE and Amphenol-FCI at 25 Gb/s per lane, and among the industry's first volume deployments of expanded-beam optics.

That history set the tone of the talk. Raj did not discuss how fast NPO can go, but how socketed optics fail at high volume. He drew one key contrast: back then Oracle deployed "tens of thousands" of InfiniBand switches; now the industry is talking about "millions" of units. Same failure rate, two orders of magnitude more failures.

The lessons of 12 years ago have not expired; the scale is simply 100 times larger.

2. Breaking Down NPO's Promises: Three Checks on Power, Cost and Reliability

Raj first acknowledged that the NPO ecosystem is far broader than CPO's, which is good: electrical connectors, external or internal lasers, polarization-maintaining (PM) fiber, fiber attach units (FAU) and linear drive all bring in new suppliers. He expects NPO to enter scale-up, scale-out and scale-across.

The three checks listed on the slides:

• Power: 50% savings versus fully retimed 200G/lane pluggables (citing NVIDIA).

• Cost: theoretical savings of 30% or more (citing SemiAnalysis).

• Reliability: system FIT expected to be lower than an equivalent pluggable system (citing Meta's OCP 2025 evaluation).

Power targets come in two tiers: slow-and-wide scale-up at below 1-2 pJ/bit, though Raj warned the reverse gearbox eats into part of that benefit; conventional scale-up/scale-out at 5-10 pJ/bit. Another often-overlooked number is electrical channel loss: at 224G, ASIC-to-NPO-engine paths over HDI or flex cable run about 13-18 dB, versus 16-22 dB over conventional PCB.

▲ Oracle's NPO power targets: below 1–2 pJ/bit for slow-and-wide scale-up and 5–10 pJ/bit for conventional scale-up/out; 224G electrical channel loss of 13–18 dB on HDI and 16–22 dB on PCB (Image source: Oracle | 2026 OCP Global Summit presentation (3:34))

The key point: all three checks are written with other companies' numbers. Oracle's own position sits at the bottom of the slide: NPO must prove reliability and power efficiency on par with soldered CPO. In other words, NPO's real competitor is not pluggables but CPO.

3. Connectors: Socketed NPO Bets Reliability on 30 µm and 30 N

The biggest structural difference between NPO and CPO is that the optical engine attaches to the board through a socket rather than solder. Field replaceability is the upside, but Raj listed six connector failure modes, each with a spec number:

• Alignment: both connector types require ±30 µm X/Y alignment, and Type 1 must also align separate RF and MP connectors.

• Signal integrity: at 212.5 Gb/s PAM4, loss, reflections, crosstalk or impedance discontinuities can close the margin.

• Retention: both types target ≥30 N, and Type 1 must maintain RF performance under X/Y pull loads.

• Thermo-mechanical distortion: uneven cold-plate loading or warpage can stress contacts and solder joints; Type 1 may need support-foot features.

The trade-off between the two is clear: Type 1 splits high-speed RF from management and power (MP), with up to 64 differential lanes; Type 2 combines them into one pin field with up to 72 differential lanes, but tighter planarity demands. Raj stressed that a connector that passes factory test may behave very differently after shipping vibration on its way to the data hall. Multiply by hundreds of thousands to millions of connectors, and any small probability becomes a weekly repair ticket.

4. Lasers: ELS vs. ILS Is a Trade-off in Blast Radius

Laser specs are anchored on OIF ELSFP-02.0-1:

Spec	VHP (Very High Power)	UHP (Ultra High Power)
Laser power	100 mW	200-400 mW
Cooling	Uncooled	Cooled
Optical launch power	20 dBm	23-26 dBm

The efficiency target is 20%+ wall-plug efficiency (WPE) and below 1 pJ/bit. Raj cited ECOC data showing CW laser FIT below 0.1, but said plainly it is not apples to apples: those are OSFP-class laser powers, while NPO and CPO need UHP lasers whose FIT is expected to be higher, and there is no field data on UHP lasers yet.

Lay out the math from the slide and the problem becomes clear:

▲ At 20 FIT, about 18 of 100,000 UHP lasers fail per year; at one million lasers that becomes about 175, so FIT must fall to 2 to hold failures at the original count (Image source: Simple Tech Trend; data from Oracle's 2026 OCP Global Summit presentation)

Laser FIT is far lower than module FIT, but one problem Raj said "is not going away": mode hopping. Whether ELS or ILS, the only fix is screening at manufacturing, step by step from die to chip-on-carrier to optical engine or ELS module.

As for choosing an external laser source (ELS) or internal laser source (ILS), Oracle's comparison centers on operations:

• ELS: more power options, low blast radius and field upgradability; the cost is PM fiber, more interfaces and insertion loss, and higher system cost. Raj said PM fiber stress and epoxy are already pain points Oracle is starting to see.

• ILS: better integration, lower BOM and system cost, no PM fiber; the cost is that when the laser fails, the whole NPO box goes back for field return, and thermal management relies entirely on the liquid-cooling loop.

We covered the power logic of ILS in CPO lasers don't have to hit 400 mW: NPO integrated lasers; Oracle's table adds the other side: the cost advantage of ILS is paid for in field-return probability.

5. Linear Drive Has a Track Record, but the DSP's Job Cannot Disappear With It

This is the only section where Oracle showed its own deployment data. Raj said Oracle has deployed "the largest LPO of any fabric anywhere," with these 800G results:

• Share of links with at least one down transition: 3.244% for LPO-to-LPO versus 5.681% for FRO-to-FRO, meaning FRO links see about 1.75x more transitions than LPO.

• LPO greatly outperforms FRO in p99 pre-FEC BER.

• At 1.6T, signal integrity no longer supports fully linear links, so Oracle is moving to LRO (half-retimed), supporting DR4+ (4 dB) and FR4, with interop demonstrated across multiple suppliers and multi-vendor 3 nm DSPs; power should fall further as LRO DSPs move to 2 nm.

▲ Oracle 800G deployment data: 3.244% of LPO links saw down transitions versus 5.681% for FRO; 1.6T moves to LRO with multi-vendor interoperability completed (Image source: Oracle | 2026 OCP Global Summit presentation (12:18))

The conclusion is that "linear drive correlates with a stable, resilient network," which is good news for NPO. But Raj immediately raised Oracle's current challenge: once the fabric is handed to a customer, Oracle cannot see what happens on the host side. Pluggables still allow remote telemetry through the OCP reliability workstream; once NPO removes or halves the DSP, five functions (retiming, equalization, FEC/BER counting, black-box logging and an attributable fault boundary) must move to the switch side. The DSP may disappear, but its operational responsibilities cannot.

6. Link Flaps Do Not Disappear Just Because Pluggables Do

The final table is the core of the talk. Raj split link flaps into two types: persistent (pointing to a catastrophic failure) and transient ("mysterious flaps that pop up and vanish"). Comparing pluggable and NPO systems item by item:

• Reduced: dust and contamination at the optical engine, DSP loss of lock (still present in LRO-type NPO), and thermal and environmental swings (better controlled by liquid-cooling loops).

• Unchanged: laser mode hopping, and latent defects from SMT micro-cracks and ESD damage (Raj said these often surface three to four months after deployment).

• Worse: socketed connector failures and SI deviations, higher intra-rack fiber density and panel-side contamination in scale-up, and most critically repeat offenders: a pluggable can simply be swapped, but an NPO engine fault means sending the whole box back; only ELSFPs remain swappable.

▲ From pluggable to NPO, some link-flap sources go down, some stay the same and some are new; the cost of a failure shifts from swapping one module to returning the whole system (Image source: Simple Tech Trend; data from Oracle's 2026 OCP Global Summit presentation)

▲ Oracle compares link-flap sources item by item for pluggable and NPO systems, concluding that link-flap risk remains with NPO (Image source: Oracle | 2026 OCP Global Summit presentation (19:42))

Raj's summary was blunt: failure modes and FIT rates have not changed much; they have just been pushed deeper into the box. So Oracle wants lower FIT across the board for connectors, lasers and FAUs, and to "ensure field returns are a low-probability event."

7. Risks and Counterarguments

• No timeline: the talk gave no date or volume for Oracle's NPO deployment, only acceptance criteria. Reading it as "OCI is about to adopt NPO" is an overreach.

• None of the three benefits are Oracle's numbers: the 50% power, 30% cost and lower system FIT figures come from NVIDIA, SemiAnalysis and Meta respectively; Oracle presented no NPO validation results of its own.

• LPO data does not transfer directly to NPO: the 1.75x figure compares pluggable LPO with FRO; NPO adds sockets and FAUs, so link stability must be measured again.

• No verdict on ILS vs. ELS: Oracle listed the trade-offs without taking a side. If hyperscalers lean toward ELS, part of NPO's cost advantage will be eaten by PM fiber and extra interfaces.

• Out of step with Huawei's scoring: Huawei gave NPO a 4:2 edge over CPO at ECOC (see ECOC 2026: Huawei scored CPO vs. NPO, and NPO leads 4:2), while Oracle's view is that NPO's blast radius is high; the two are measuring different dimensions.

8. What It Means for Taiwan's Supply Chain

• Socketed connectors are the new spec battleground. The ±30 µm, ≥30 N, 212.5 Gb/s PAM4 and cold-plate warpage items Raj listed are exactly the validation data connector makers must deliver. The Open CPX MSA is where this connector spec comes from, and the Open CPX member list shown by Ciena at the same summit includes two connector makers, LOTES and FIT (Foxconn Interconnect), plus Taiwanese members such as Accton and BizLink (see ECOC 2026: Ciena turns 6.4T optical engines back into off-the-shelf parts with Open CPX). Oracle's failure-mode table is effectively the exam they now have to pass.

• Switch ODMs must absorb the operational functions the DSP leaves behind. Telemetry, black-box logging and fault boundaries moving to the switch side means new firmware and system-integration work for white-box switch ODMs, and a chance to differentiate.

• Laser screening capability is worth more than laser power. Oracle said mode hopping can only be screened out at manufacturing starting from the die, and with no field data on UHP lasers, the laser and optical-engine suppliers that can show screening flows and accelerated-aging data are the ones that will reach hyperscaler qualification lists. The talk named no Taiwanese company on this point.

9. Conclusion

The most valuable takeaway is not any single number but how a hyperscaler looks at NPO: not by pJ/bit, but by how big a piece has to be replaced when something fails. NPO removes the largest failure source, human handling and contamination, but pushes the remaining failure modes into the box, and the blast radius grows from one module to one switch.

So what Oracle gave was not a timeline but conditions: NPO reliability on par with CPO, FIT falling as fleets reach millions of units, and DSP telemetry moved to the switch. Three things to watch over the next 12 months: the first field FIT data for UHP lasers, volume reliability reports for Open CPX connectors, and progress on switch-side telemetry specs in the OCP reliability workstream.

NPO will be won or lost not on lab pJ/bit, but on how many switches come back from the data hall each year.

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