OCP Global Summit 2026 | Linktel | Grace Zhang | Dispersion Managed Transceiver for Next-Gen AI Fabrics

1. Why 10 km Is the New Battleground

AI-cluster optics used to be about 500 m and 2 km links inside and between racks. Linktel's talk targets the next ring out: scale-out expansion, campus-wide scale-across GPU pooling across buildings, and OCS-based all-optical SuperPOD fabrics, with reach requirements of 6, 10 and 20 km.

CWDM is the workhorse at these distances. The slides give three reasons: one single-mode fiber pair carries 4 to 8 wavelengths; fewer fiber end faces reduce contamination risk and simplify connector handling; and Linktel cites LightCounting's forecast that CWDM FR4/FR8 will lead 4x200G and 8x200G shipments through 2030. We covered what OCS demands from optics in ECOC2026 | 95% of OCS Shipments Are One Customer: Google, which is exactly where the FR-versus-DR debate plays out.

The problem is that CWDM runs into a physical bottleneck: fiber chromatic dispersion. The talk abstract is blunt: standard 1.6T CWDM modules only support links under 3 km, and mainstream 800G CWDM variants cap out at 10 km.

Bandwidth can be added with faster SerDes, but dispersion eats reach at a compounding rate.

2. Breaking Down Dispersion: Double the Rate, Quarter the Reach

Dispersion-induced group delay is Δt = D(λ)·L·Δλ. Grace Zhang's derivation is simple: doubling the data rate doubles the signal's spectral width Δλ and therefore the group delay, while the symbol period halves, so the tolerable Δt also halves. Combine the two and maximum reach drops to one quarter.

The numbers on the slide, without optical dispersion management:

• 100G/lane: 10 km

• 200G/lane: down to 2.5 km

• 400G/lane: only 600 m

▲ Each time the lane rate doubles, unmanaged-dispersion CWDM reach drops to a quarter; DMD reaches 20 km, 10 km and 3 km at 100G, 200G and 400G, respectively (Image source: Simple Tech Trend original, data from Linktel's 2026 OCP Global Summit presentation)

A second chart shows why CWDM is hit especially hard: 10 km of fiber has near-zero dispersion around 1311 nm, but about -40 ps/nm at 1271 nm and about +18 ps/nm at 1331 nm. Of the four wavelengths, the two outer lanes take the worst dispersion, with opposite signs. That is Linktel's entry point: there is no need to compensate every lane, only to rescue the outer ones.

3. What the DMD Is: A Thermally Tuned Silicon GT Etalon

The DMD itself is a single-cavity silicon Gires-Tournois (GT) etalon with an integrated micro-heater and thermistor, placed in the receiver optical path:

• Near-zero insertion loss: the back HR coating reflects >99.9%, so all light comes back out; the power spectrum is unchanged and only the phase spectrum shifts, with interference inside the cavity creating dispersion.

• Compensates either sign: the DMD's dispersion curve oscillates with wavelength and can be positive or negative. Silicon's refractive index is highly temperature-sensitive, so tuning temperature slides the curve left or right onto the signal. At 1331 nm, for example, the fiber adds positive dispersion, so the DMD is tuned into its negative region to cancel it.

• DSP closed-loop self-locking: the module reads SNR from the DSP and adjusts DMD temperature to the optimum, adapting to fiber lengths of 0, 2 or 10 km and to transmitter wavelength drift over long-term operation.

▲ DMD is placed only on the outer wavelengths at the receiver: 1271 nm and 1331 nm see the largest dispersion with opposite signs, and DMD compensates each through temperature tuning (Image source: Simple Tech Trend original, data from Linktel's 2026 OCP Global Summit presentation)

The block diagrams of the two 800G modules show DMDs only on the outer 1271 and 1331 lanes; 800G FR4 uses 2, while the dual-group 800G 2LR4+ and 1.6T 2FR4 use 4. It is a design that solves the biggest problem with the fewest parts.

4. The Scorecard for Three Products

Item	800G 2LR4+ 20km	800G FR4 10km	1.6T 2FR4 10km
Transmitter	EML, 100G/lane	EML, 200G/lane	Silicon photonics, 200G/lane
DMD channels	4	2	4
BER after fiber	20 km: 1e-9 to 1e-10	10 km: 1e-10 to 1e-11 (212G)	10 km: 1e-11 to 1e-12 (212G)
Link budget	>10 dB on all 8 lanes	Not disclosed	7.2 dB
Stage	Beta	Beta	Alpha

▲ Linktel's dispersion-managed transceiver roadmap: 800G 2LR4+ 20 km, 800G FR4 10 km and 1.6T 2FR4 10 km; the first two won Lightwave Innovation Awards (Image source: Linktel | 2026 OCP Global Summit presentation (07:29))

Details worth noting:

• Eye recovery: for 800G FR4 at 212 Gb over 10 km, TDECQ at 1271 nm and 1331 nm could not be measured without the DMD and came in at 0.76 dB and 0.83 dB with it; for 800G 2LR4+ over 20 km the figures were 1.5 dB and 2.5 dB.

• Switch-level validation: 36 800G 2LR4+ modules ran in a TH5 switch through 10 thermal cycles from 10°C to 70°C, holding pre-FEC BER below 1e-7 with FEC BIN no higher than 4. 800G FR4 was tested with 10 modules over 4 cycles.

• 1.6T 2FR4: with inner FEC enabled, BER stays below 1e-12 at 226 Gbps, and the DMD long-term locking penalty is 0.5 dB.

▲ 1.6T 2FR4 over 10 km: BER below 1e-11 at 212 Gbps and a 7.2 dB link budget, with the DMD temperature window scan in the middle (Image source: Linktel | 2026 OCP Global Summit presentation (16:09))

The most common objection is thermal control. Linktel scanned the DMD temperature window over 10 km of fiber plus a variable attenuator: with 1 dB of extra loss, the speaker said the window exceeds 3°C; at 4 dB it is still more than 1°C. Grace Zhang's point is that ±1°C control is "very easy" for the hardware.

5. 400G/Lane: A 3 km Proof of Concept

Linktel also took a step into the next generation, using a thin-film lithium niobate (TFLN) modulator at 190 GBaud PAM4 (380 Gbps) on 1271 nm, the worst-case wavelength:

• Without dispersion management, the eye is completely closed after 3 km

• With the DMD, a clean four-level PAM4 eye is restored, with 3.3 dB ER and -0.6 dBm average power (3.4 dB ER at 0 km)

This echoes the conclusion of ECOC2026 | Fast-Narrow vs Slow-Wide: The Dividing Line Is 448G: 400G/lane is the make-or-break point for IM-DD. The DMD at least shows CWDM need not hand 3.2T straight to coherent. But be clear: this is a single-lane, AWG-driven result measured on a lab scope, not a module, and with no BER reported.

6. Competitive Comparison and Open Strategy

Linktel put DMD-based CWDM side by side with two alternatives:

Metric	DMD CWDM	LAN-WDM	Coherent-Lite
Readiness for 200G/L 10km	Good	Fair	Good
Potential for 400G/L	Good	Fair	Good
Supply chain (laser, DSP)	Good	Poor	Fair
Interop with existing FR4	Good	Poor	Poor
Cost	Good	Fair	Poor
Power consumption	Good	Good	Fair

▲ Linktel compares DMD CWDM, LAN-WDM and Coherent-Lite on six metrics, and rates its own DMD approach ahead on all of them (Image source: Linktel | 2026 OCP Global Summit presentation (17:17))

The key is the middle two rows: LAN-WDM needs a different set of laser wavelengths and Coherent-Lite needs a different DSP and module architecture, and neither interoperates with existing FR4. The DMD approach uses exactly the commercial CWDM wavelength grid, and that is its strongest argument.

Linktel also announced it is opening access to core DMD optical designs and listed three US patent applications, 19/026,490, 19/295,055 and 19/295,081, inviting DSP vendors and cloud operators to co-develop. Opening the design while holding patents signals a plan to make DMD a spec option first, then monetize through components and licensing.

7. Risks and Counterarguments

• The self-assessment has no numbers: the comparison table uses smiley faces only, with no actual cost, power or BER comparison; it is a vendor view.

• "Negligible power" comes without watts: the temperature scan sits between 95°C and 98°C, meaning the DMD is heated to around 96°C; actual heater power and margin at high ambient temperature were not disclosed.

• Still early: both 800G products are in Beta and 1.6T is in Alpha, with no volume date; the 1.6T link budget of 7.2 dB is tighter than the 10 dB of 800G 2LR4+.

• The LightCounting chart does not fully support "lead": on the same chart, 4x200G DR4 volumes are actually larger than FR4, so CWDM dominance in 2030 is not settled.

• 400G/lane is eye diagrams only: single lane, 3 km, no BER, and a 3.2T module is still separated by DSP, lasers and packaging.

8. What It Means for Taiwan's Supply Chain

• The slides name no Taiwanese companies; Linktel's invitation goes to DSP vendors and cloud operators, and no confirmed Taiwanese role is visible yet.

• Neutral to positive for FR4 component suppliers. The DMD approach reuses FR4 lasers, PDs, TIAs and DSPs, folding 10/20 km demand back into the CWDM 1271/1291/1311/1331 nm part numbers, without the narrow-spaced LAN-WDM wavelengths. Makers of CWDM EMLs and PDs do not need a separate product line for longer reach.

• The silicon-photonics 1.6T 2FR4 is another signal. Linktel's 1.6T transmitter is silicon photonics, meaning the 10 km class is also moving to SiPh with an external light source, so CW laser demand will not stop at 500 m DR.

• The thing to track is who adopts first. If a hyperscaler writes DMD into a 10 km spec, the receiver gains a temperature-controlled silicon device plus supporting optics, a new process step for precision optical assemblers, but for now it is only a potential opportunity.

9. Conclusion

The value of Linktel's talk is that it does not treat dispersion as a problem that requires a new architecture. One heated silicon etalon, fitted only on the two outer wavelengths, with everything else kept as FR4, pulls 200G/lane CWDM from 2.5 km back to 10 km and pushes 100G/lane to 20 km.

Keep the timeline separate. With 800G in Beta and 1.6T in Alpha, what we will see before end-2026 is sampling and customer qualification, not volume production; 400G/lane only shows IM-DD still has a path in the 3.2T generation. Three things to watch over the next 12 months: when 1.6T 2FR4 enters Beta, actual heater power and high-temperature reliability data, and whether a cloud operator writes the DMD approach into a 10 km spec.

Dispersion is not a reason to switch to coherent; it is a reason to add one more chip in the receiver. Whoever turns that chip into a standard part wins 10 km.

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