A link budget looking for a link.

A product without an order

In December 2024, Marvell announced what it called the first DSP built for coherent optics inside a data center campus. Ciena showed a 1.6T coherent-lite pluggable three months earlier. Alphawave sampled its own part in March 2025. Analysts forecast a market of a million units a year or more.

In July 2026, Cignal AI summarized the result. Several operators showed interest at 800G and 1.6T. None placed volume orders. Amazon, which first pushed 1.6T Coherent Lite for campus links, moved on to other options.

This essay asks why. On the way, it answers five practical questions:

1. What is the signal?
2. Does the module need a driver?
3. How much power does it draw?
4. Does it fit an OSFP cage?
5. How large is the market really?

The answer to the main question is short. Coherent Lite sells link budget. Its market is the set of links where the simpler technology runs out of budget.

What the signal is

Inside a data center, almost every optical link uses IM-DD (intensity modulation, direct detection). The transmitter switches light intensity between levels, and a photodiode reads the intensity. Today the levels are PAM4, four intensity levels per symbol.

A coherent link also encodes phase and polarization. The receiver mixes the signal with a local laser, the local oscillator, and a DSP recovers the field. That costs a DSP and a second laser path. It returns two things: more bits per symbol, and a receiver that is much more sensitive.

Coherent Lite is the shortest-reach form of that idea. The standards documents define it this way:

• Format: DP-16QAM, 16 points of amplitude and phase on each of two polarizations. That is 8 bits per symbol.
• Symbol rate: 123.6364 GBd for one 800G carrier, per the January 2024 IEEE 802.3dj baseline for 800GBASE-LR1. The line therefore carries about 989 Gb/s for 800G of payload. FEC and framing use the difference.
• FEC: the host’s RS(544,514) code with a BCH(126,110) inner code. That is lighter than the oFEC that 800ZR uses over long distances.
• Laser: a fixed O-band wavelength near 1311 nm. The baseline allows a 1 MHz linewidth. A tunable C-band laser, the expensive part of a ZR module, is not needed.
• Reach: 10 km in the OIF 800LR agreement. Vendors quote 2 to 20 km.

So the answer to the first question is yes: Coherent Lite is 16QAM. We found no coherent-lite product with DP-QPSK, the simpler four-point format.

The drive question

A coherent modulator needs more drive signals than an IM-DD one. Each carrier needs four: in-phase and quadrature on each polarization. An IM-DD lane needs one.

The count matters at module level:

• A 1.6T DR8 module has eight PAM4 lanes, so eight drive signals.
• Marvell’s 1.6T Aquila layout is four 400G 16QAM carriers, so sixteen.
• Ciena’s layout is two 800G carriers, so eight, at about twice the symbol rate of Marvell’s.

Sixteen linear drivers would be a large cost in a module that must compete with DR8. Marvell’s answer is in its product brief. Aquila supports standard external drivers, but its block diagram shows a silicon photonics modulator with “direct drive” from the DSP. In that design, the DSP output stage is the driver. Ciena and Alphawave do not publish their driver arrangement.

This matches a pattern from the DSP essay. When a DSP integrates the driver, it chooses the modulator for the module. Here the direct-drive path assumes a silicon modulator with a low enough drive voltage.

Watts and the cage

No vendor publishes a coherent-lite module wattage. The best public numbers are estimates from IEEE standards meetings:

At 800G, a 2022 contribution put DR8 or 2xFR4 under 13 to 14 W and LR1 coherent at about 20 to 24 W. At 1.6T, China Mobile estimated about 21 W for LR8 IM-DD and about 40 W for coherent. Both are estimates, and both put coherent at about 1.6 to 1.9 times the IM-DD power.

Ciena disagrees. Its 2024 product blog says that at 1.6T, coherent and IM-DD power start to converge. The OIF 1600CL project targets about 30 W, per Nokia’s OFC 2026 summary. If that target holds, the gap to a 21 W LR8 is smaller, but it does not close.

Power is not the obstacle. Juniper lists OSFP and OSFP-RHS 1.6T modules from 16 to 43 W, and every estimate above is inside that range. The announced coherent-lite modules use OSFP-XD, the 16-lane member of the family:

• Marvell designs Aquila for 1.6T OSFP-XD.
• Accelink showed a 1.6T coherent-lite OSFP-XD module at OFC 2025: four 400G carriers at 1311 nm, up to 20 km.

So the answer to the OSFP question is yes, with one limit on what we know. The power estimates are inside the range of OSFP-family 1.6T modules, and the announced designs use OSFP-XD. We found no published module power for an OSFP-XD coherent-lite part. The cost is power per bit, not a mechanical limit.

What the extra watts buy

Ciena lists the return. Coherent Lite has a larger loss budget than IM-DD, it tolerates crosstalk better, and it scales with wavelength multiplexing. The loss budget is the important one. A coherent receiver needs less optical power than a photodiode for the same error rate.

That budget has value only where IM-DD runs short. Cignal’s 2025 report gives the reason the 1.6T opportunity closed. Vendors showed that 200G-per-lane PAM4 works at 10 km. A campus link that IM-DD can close has no reason to pay about twice the watts.

The next case is different. Google runs its TPU clusters through optical circuit switches, mirrors that redirect light between fibers. Each switch pass adds optical loss, and the link budget must absorb it. Cignal reports a custom Google project at 2.4T to increase the speed per switch port. It calls this an interim step until 400G-per-lane optics (3.2T) are available in volume.

How large the market is

The forecasts disagree by a factor of ten:

• LightCounting: more than 1 million units a year by 2029, as quoted in Marvell’s December 2024 announcement.
• 650 Group: more than 10 million units within three to five years.
• Cignal AI: more than $1.6 billion by 2029, from its ECOC 2025 keynote.

Against that, Cignal’s July 2026 report says that nobody placed a volume order. Its upside case is that Coherent Lite inside the data center could reach ten times the DCI deployments, DCI meaning links between data centers. Cignal expects more than 200,000 800G coherent DCI modules in 2026. Ten times that is about 2 million a year. That is still only about a fifth of the 1.6T IM-DD volume that Cignal expects for 2026 alone, almost 10 million.

DSP volume follows module volume, because each module carries one DSP. No vendor or analyst publishes coherent-lite DSP shipments. The DSP field is still wide:

• Marvell Aquila: 4 x 400G, sampling since December 2024. Aquila M adds MACsec encryption.
• Ciena WaveLogic 6 Nano: two 800LR paths in one 3 nm DSP, in Ciena’s own module.
• Alphawave Co-Wave AW400-O: DP-16QAM for 800G up to 20 km, sampling since March 2025.
• Celero Communications: a 2 nm coherent DSP and a $275 million round in September 2026. Cignal calls it the frontrunner for Google’s 2.4T project.
• Nokia: a 3.2T-capable coherent lite family, samples in mid-2027.

Five DSP programs and one custom project at one operator: that is the market as of October 2026.

What to take away

1. Coherent Lite is DP-16QAM on a fixed O-band laser, with a lighter FEC than ZR. It is the shortest-reach coherent, not a new format.
2. It needs four drive signals per carrier. Marvell puts the driver in the DSP and drives silicon directly. Ask every other vendor how they do it.
3. The public estimates put it at 1.6 to 1.9 times the IM-DD power at the same rate. Ciena says the gap closes at 1.6T. Neither side published a measured module.
4. Its estimated power is inside the range of OSFP-family 1.6T modules. The announced designs use OSFP-XD.
5. Forecasts span ten times. Orders, as of July 2026, are zero.

Coherent Lite wins only where IM-DD runs out of link budget. At 1.6T, PAM4 closed the 10 km link, and the market disappeared. At 2.4T, an optical switch adds loss, and one operator began a custom project. At 3.2T, the question is whether 400G-per-lane IM-DD keeps its budget. Until it fails, Coherent Lite stays a link budget looking for a link.