top of page

To Understand the 1.6T Optical Module Shortage, First Understand Two Chips: InP Lasers and DSPs

2 days ago
16 min read

1. Why every 1.6T shortage story ends up at two chips

On its November 12 earnings call, Cisco CFO Mark Patterson was asked which part of the supply chain was most likely to hold back shipments this year. His answer: "The biggest one there is probably optics." On the same call, Cisco raised its FY27 hyperscaler AI infrastructure revenue target from $7.5 billion to $9.0 billion, said the split between optics and systems is roughly 50/50, and announced its first hyperscaler 1.6T optics win for a scale-out use case.

After a line like that, readers usually ask three questions:

  • Is the optical module shortage just a lack of module capacity?

  • Who benefits most: Innolight, Eoptolink, Coherent or AOI?

  • With so many expansion announcements, when does the shortage end?

None of these is wrong, but all of them are asked at the module level. Module makers design, assemble and test. What actually keeps 1.6T modules from shipping is two chips inside the module that cannot be bought in sufficient volume. How many 1.6T modules ship is decided by the InP laser, which handles the light, and the DSP, which handles the electrical signal.

This article does not compare module makers. It takes the two chips apart: what each one does inside the module, why each gets harder to make at higher speeds, and why only a handful of companies can make them. By the end, you should know which three questions to ask about any news on optical module shortages, price hikes, capacity expansions or export bans.

We covered how the individual components work (Driver, TIA, PD, FAU) in Silicon Photonics Transceivers Explained. This article focuses only on the InP laser and the DSP.

2. A different yardstick: it is not module capacity, it is chip allocation

The market tends to estimate 1.6T shipments from how many production lines module makers are adding. But module lines are the easiest part to expand: buy equipment, hire people, lease a building, and it is done in a few quarters. What cannot be expanded quickly are the two chips module makers buy from upstream.

1.6T module shipments = min (laser chip allocation ÷ lasers per module, DSP allocation ÷ DSPs per module, module assembly capacity)

Taking the minimum means shipments stop at whichever of the three numbers is smallest. Module assembly capacity is usually not the smallest one.

  • The optical half (InP lasers): the material is indium phosphide (InP). Substrates, epitaxy and chips are concentrated in a few IDMs, wafers are moving from 3-inch to 6-inch, and capacity is added one fab at a time.

  • The electrical half (DSP): the DSP for 1.6T has to be built on 3nm, volume supply today is concentrated in a few companies, and capacity depends on TSMC's advanced-node allocation. Capacity is added one process node's wafer allocation at a time.

AOI is the clearest example. On its Q3 call, AOI said its 1.6T backlog is about $450 million, but DSPs are the main constraint and it can meet less than 80% of demand. It also cut its year-end capacity target from 650,000 to 500,000 units. The orders are there and the lines are there. What is missing is DSPs.

Simply put: 1.6T shipments are set not by module makers but by the allocation of two chips.

3. The first chip: the InP laser, the "light bulb" that turns electricity into light

3.1 Why an InP laser is needed

The transmit side of an optical module has one job: turn the electrical signal from the DSP into an optical signal that travels through fiber. Silicon does not emit light, so the light source has to be a III-V material. For the 1310nm single-mode light common in data centers, the best material is InP.

There are two common ways to build the light source of a 1.6T module:

  • EML (electro-absorption modulated laser): on a single InP chip, a DFB laser section generates the light and an electro-absorption modulator (EAM) section "writes" the signal onto it. One EML serves one lane.

  • CW laser + silicon photonics chip: the InP laser only emits steady continuous-wave (CW) light, and modulation is done by modulators on the silicon photonics chip. One CW laser can be split across several lanes.

This figure splits a 1.6T DR8 module into its "electrical half" and "optical half." On the left, eight 200G electrical lanes come in from the switch, pass through the DSP, and are handed to the Driver, which drives eight optical lanes. The receive side runs in reverse: the PD receives light, the TIA amplifies it, and the signal returns to the DSP. The blue blocks are where the InP lasers sit; the dark block is the DSP.

The two key chips in a 1.6T module

▲ The two key chips in a 1.6T DR8 module: the DSP in the electrical half, the InP laser in the optical half (Image source: Simple Tech Trend original; data from Broadcom, Marvell, AOI, Credo)

This figure shows which components sit inside a silicon photonics module: where the DSP, Driver, TIA, silicon photonics chip and laser each go. Read it alongside the previous figure to see that the DSP and the laser are two separate, critical chips in the module.

Marvell silicon photonics module anatomy

▲ Anatomy of a silicon photonics module (Image source: Marvell)

3.2 Examples: the same 1.6T, very different laser counts

AOI put it plainly in an April 2026 article: an 8-channel 1.6T EML module needs eight separate laser chips, each individually aligned and bonded.

Silicon photonics lets several lanes share one laser. The spec sheet for Credo's Oz8 1.6T silicon photonics chip lists "2 integrated lasers, 4 optical lanes per laser, 8 lanes in total." Marvell also explains in a technical blog that a silicon photonics module can let four lanes share one laser, so a full 1.6T module needs only two less expensive CW lasers.

This figure compares how many lasers a discrete (EML) module and a silicon photonics module need for the same bandwidth. It shows directly why silicon photonics saves InP: fewer lasers means less demand for InP chips.

Marvell discrete vs silicon photonics laser count

▲ Silicon photonics modules use fewer, less expensive lasers to deliver the same bandwidth (Image source: Marvell)

But "fewer chips" does not mean "less InP." A shared laser has to feed four lanes, so each CW laser needs much higher power. AOI's CEO gave a rule of thumb on the earnings call: one good InP substrate yields 8,000 to 9,000 100mW lasers, and for high-power lasers you divide that count by 5 to 6.

One line to remember: the EML route uses more chips, the CW route uses more power, and both consume InP.

3.3 The cost: why InP never seems to be enough

The InP bottleneck has three layers.

The first layer is substrates. Mainstream InP wafers are still 3-inch, there are only a few suppliers, and part of the capacity is in China and subject to export approval. On its November 12 Q3 call, IntelliEPI said demand for InP PIN PD epitaxy is rising, but InP substrate shortages are limiting output, and higher substrate costs may lead to price increases. AXT's InP revenue grew from $13.6 million in Q1 and $30.7 million in Q2 to $54.3 million in Q3, with total Q4 revenue guidance of about $90 million.

The second layer is laser fabs. On its FY2026 Q4 call, Lumentum said its EML capacity grew eightfold in two and a half years and is still short of demand, and that CW lasers and EMLs compete for the same InP lines. Coherent's CEO said he sees no InP or EML oversupply before the end of 2028.

Laser fabs are also slow to expand because of the process itself. A CW DFB laser goes through seven process steps: epitaxy, grating, regrowth, front-end, back-end, die attach, and burn-in and test. Output power, single-mode yield and long-term reliability are each set by different steps. In The Most-Patented Steps Are Not the Hardest: Inside the 7 Process Steps of a CW DFB Laser and 13 Vendors' Patent Portfolios, we mapped 123 public patents from 13 vendors to these steps and found that the hardest steps, grating, regrowth and screening, are the ones almost nobody patents; they rely on each company's accumulated process experience. That is one reason InP laser capacity cannot be copied quickly just by buying equipment.

The third layer is the wafer-size transition. Coherent is moving InP wafers from 3-inch to 6-inch at its fabs in Sherman, Texas and Järfälla, Sweden. The company says 6-inch brings 4x production capacity and a greater than 60% reduction in die cost. But 6-inch yields take time to ramp, and this is a two-to-three-year project, not a one-or-two-quarter one.

This figure shows Coherent's 6-inch InP wafers. Compared with 12-inch silicon wafers, InP is still moving from 3-inch to 6-inch, which is one reason InP capacity grows slowly.

Coherent 6-inch InP wafers

▲ Coherent 6-inch InP wafers (Image source: Coherent)

A DIGITIMES report in September estimated 2026 global effective InP capacity at only 600,000 to 750,000 wafers per year, against demand of 2.6 to 3.0 million. Different sources count differently, but the direction is the same: the InP bottleneck is in materials and wafer size, and capacity grows on a timescale of years.

4. The second chip: the DSP, the translator that "cleans up" the signal

4.1 Why a DSP is needed

The electrical signal travels from the switch chip to the optical module through PCB traces and connectors. The higher the speed, the more the signal is attenuated and distorted along the way. At 200G per lane (PAM4, about 106 GBaud), the eye diagram is nearly closed by the time the signal reaches the module.

The DSP (digital signal processor) takes in this distorted signal, digitizes it, equalizes it, retimes it, and sends a clean signal to the Driver to drive the laser. The receive side works the same way: the PD's optical signal is amplified by the TIA, and the DSP restores it into an electrical signal the switch can read.

The value of the DSP is that the electrical side and the optical side can each be designed, and each have their own impairments, with the DSP aligning the two in between. The cost is power and money: Juniper's technical documentation notes that the DSP accounts for over one-fourth of the cost of a typical transceiver module.

This figure shows the difference in signal path between a retimed module with DSPs and an LPO module without them. On the left, both transmit and receive pass through a retimer; on the right, only the linear Driver and TIA remain. The figure also lists 800G module power: about 15–16W for a single-mode DSP module and about 7–8W for an LPO module. The block that gets removed is the subject of this section.

Semtech DSP retimed module vs LPO module

▲ DSP retimed module (left) vs LPO linear module (right), with 800G module power (Image source: Semtech)

4.2 Two examples: Broadcom's and Marvell's 3nm DSPs

Today the 1.6T generation runs mainly on two 3nm DSPs.

  • Broadcom Sian3: a 3nm 200G/lane PAM4 DSP announced in March 2025. Broadcom says it enables sub-13W 800G and sub-23W 1.6T modules, more than 20% less power than the previous Sian2, and supports both EML and silicon photonics modules. A second chip, the 5nm Sian2M, targets sub-25W 1.6T SR8 (multimode) modules.

  • Marvell Ara: announced in December 2024 as the industry's first 3nm 1.6T PAM4 DSP, with eight 200G lanes on both the host and optical sides, cutting 1.6T module power by more than 20%. At its October investor day, Marvell said its 3nm 1.6T DSP is in volume production and that it has demonstrated a 2nm 3.2T DSP.

This figure shows the basic architecture of Ara: eight 200G electrical lanes on one side, eight 200G optical lanes on the other, 1.6T in total. It shows why a 1.6T module usually needs only one DSP, but that one DSP has to be 3nm.

Marvell Ara DSP interfaces

▲ The Ara DSP supports eight 200G electrical interfaces and eight 200G optical interfaces (Image source: Electronic Design)

Even surrounding chips are chasing the 3nm DSP's power budget. On November 9, MaxLinear announced a multiphase power management IC for 1.6T DSPs, claiming up to 0.5W of savings per DSP, with sampling in February 2027. The fact that the DSP draws enough power to justify its own power chip says a lot about its weight in the module.

4.3 Where the industry stands: 3nm volume sits with two suppliers today, and the key term is "retiming"

A brokerage supplier table shows that volume supply of 3nm DSPs for 1.6T today comes mainly from Marvell and Broadcom. For 800G, which uses 7nm/5nm DSPs, Credo, MaxLinear and Chinese vendors also supply. That list will not stay this short: the 1.6T DSP power management IC MaxLinear announced on November 9 is designed around its own Rushmore DSP, and newcomers are catching up.

That is the nature of the DSP bottleneck: it is not that nobody can make it, it is that few companies ship in volume today, and they all queue for TSMC's 3nm capacity. One brokerage estimates 2027 demand at about 90 to 100 million units each for 800G and 1.6T, but says DSP shortages will limit 1.6T shipments to 60 to 70 million.

Meanwhile, module makers are trying to use fewer DSPs. This is where one key term comes in: retiming. Juniper defines three types:

  • Fully retimed (FRO): both transmit and receive pass through DSPs. The most robust option, and the most power-hungry.

  • LRO (Linear Receive Optics): a single DSP retimes the transmit side; the receive side is linear.

  • LPO (Linear Pluggable Optics): no DSP in the module at all; signal processing is handed to the switch chip's SerDes.

Nokia's ICE-D 1.6T DR8 LPO module draws less than 10W, which the company says is about 75% lower than a fully retimed pluggable.

This figure uses a single 200G lane to compare the three approaches: fully retimed, LRO and LPO. It shows where the DSPs sit, how many there are, and the corresponding module power. Fewer DSPs mean lower power, but responsibility for signal quality shifts toward the switch chip.

Three ways to build the same 200G lane

▲ Three ways to build the same 200G lane: fully retimed, LRO and LPO; fewer DSPs, lower module power (Image source: Simple Tech Trend original; data from Juniper, Broadcom, Nokia)

One line to remember: the DSP buys signal quality and interoperability; removing it saves power and 3nm allocation.

5. One table to break down InP lasers vs DSPs


InP laser

DSP

Role in the module

Emits light (an EML also modulates)

Equalization, retiming, error correction

Material / process

III-V InP, moving from 3-inch to 6-inch

Silicon, 3nm (2nm next)

Units per 1.6T module

8 EMLs, or 2 high-power CW lasers with silicon photonics

Usually 1 (0 for LPO)

Main suppliers

A few IDMs: Lumentum, Coherent, Broadcom, Mitsubishi Electric, Sumitomo Electric

At 3nm, mainly Marvell and Broadcom today, with newcomers catching up

Unit of capacity

Fabs and substrates, measured in years

TSMC advanced-node allocation, measured in process nodes

Alternatives

Silicon photonics + shared CW lasers, VCSEL (short-reach multimode)

LRO, LPO

Price signals come from

Substrate and epitaxy makers (AXT, IntelliEPI)

DSP lead times and allocation

Geopolitics

Part of substrate capacity is in China and needs export approval

Designed in the US, made in Taiwan

What it decides for 1.6T

How many lanes of light can be made

How many modules can be made

One line to remember: the InP bottleneck is in materials, the DSP bottleneck is in process; one expands slowly, the other has few volume suppliers today.

6. Back to 1.6T: every design is doing the same thing

Lay out the 1.6T modules on the market and every design is doing the same thing: deciding whether to use less InP or less DSP.

Route

Laser

DSP

Example

What it uses less of

EML + DSP (DR8/FR8)

8 InP EMLs

One 3nm

Broadcom Sian3 with EMLs, 1.6T under 23W

Neither; the most mature

Silicon photonics + CW + DSP

2 high-power InP CW lasers

One 3nm

Credo Oz8 (2 lasers × 4 lanes)

Fewer InP chips

Silicon photonics + LPO

CW laser

None

Nokia ICE-D 1.6T DR8 LPO, under 10W

Less DSP

VCSEL + DSP (SR8)

GaAs VCSEL

One 5nm

Broadcom Sian2M, 1.6T SR8 under 25W

No InP, but short reach only

This figure places each route on one chart: the horizontal axis is InP laser demand per module, the vertical axis is DSP demand and node, and disclosed module power is labeled next to each point. The lower-left corner, which saves on both chips, is the ideal zone. No route sits there today.

Position map of 1.6T routes

▲ Where each 1.6T route sits on InP laser demand and DSP demand; positions are STT's reading (Image source: Simple Tech Trend original; data from Broadcom, Nokia, Credo, AOI)

6.1 EML + DSP: the most mature, and the heaviest user of both chips

This is the mainstream approach carried over from the 800G generation: one EML per lane across eight lanes, plus one 3nm DSP. Suppliers are familiar and yields are stable, but it uses the full amount of both InP chips and DSPs. InP IDMs such as Lumentum and Coherent are at the core of this route.

6.2 Silicon photonics + CW + DSP: saving InP chips by splitting light

Lasers drop from eight to two, and laser alignment steps drop from eight to two. The cost is higher laser power, so far fewer lasers come out of each InP substrate. As this route grows, InP demand shifts from "chip count" to "high-power specs."

6.3 Silicon photonics + LPO: removing the DSP

Nokia's 1.6T LPO module draws under 10W, compared with Broadcom's sub-23W for a 1.6T DSP module, a saving of more than half. LPO requires a strong enough SerDes on the switch chip, and the whole link has to be validated together. It suits short links where one vendor controls both ends, such as within a rack or a single pod.

This figure shows Nokia's 1.6T DR8 LPO module. It looks like any other 1.6T module; the difference is the missing DSP inside, which is why its power can drop below 10W.

Nokia 1.6T DR8 OSFP LPO module

▲ Nokia 1.6T DR8 OSFP LPO transceiver (Image source: Nokia)

6.4 VCSEL + DSP: no InP, but short reach only

VCSELs use GaAs substrates and bypass InP entirely, but multimode fiber reach is short, so they only suit links within a rack or between adjacent racks. They relieve pressure on InP, but still need a DSP.

One line to remember: every 1.6T route picks one of the two chips to use less of, and that choice decides where the route gets stuck.

7. The real bottleneck: both chips sit with very few suppliers

InP and DSPs look like two different problems, but the bottleneck is the same in nature: both are concentrated in a very small number of suppliers, and both take years to expand.

  • DSP: 3nm volume today comes mainly from Marvell and Broadcom. AOI can meet less than 80% of its orders, while Innolight says its locked-in 1.6T DSP supply is close to its order demand. Whoever secures allocation ships.

  • InP: EML capacity grew eightfold in two and a half years and is still short; there are only a few substrate makers; and IntelliEPI says even its PD epitaxy is limited by substrates.

This figure compares two points in time. On the left is the 1.6T generation of 2026–2027, where the DSP is the tighter chip. On the right is the 3.2T generation from 2028 (400G per lane), where the DSP roadmap is visible but the laser and modulator technology has not converged, so pressure may shift back to InP.

Bottleneck today vs the 3.2T generation

▲ The bottleneck shifts: DSPs are tighter in the 1.6T generation, while pressure may move back to the optical half at 3.2T (Image source: Simple Tech Trend original; data from AOI, Lumentum, Marvell, Coherent; right column is STT's reading)

Our view: the ceiling on 2027 1.6T shipments is set by DSPs, not InP. There are three reasons. First, AOI itself says InP will not be the binding constraint through 2027, and DSPs are. Second, the silicon photonics + CW route reduces the number of InP chips by sharing lasers, while DSPs have no equivalent substitute; LPO only fits some use cases. Third, few companies ship 3nm DSPs in volume today, so how their allocation is split directly decides each module maker's shipments.

Three conditions could flip this:

  1. LPO/LRO adoption at scale in 1.6T: if hyperscalers are willing to use LPO in scale-out networks too, DSP pressure would ease significantly and the bottleneck would move back to InP.

  2. More 3nm DSP suppliers reaching volume: if newcomers such as MaxLinear qualify for 1.6T and ramp, DSP allocation will no longer be concentrated in a few companies, and the bottleneck will ease.

  3. The arrival of 3.2T: at 400G per lane, Marvell has already demonstrated a 2nm 3.2T DSP, but whether lasers and modulators will be EML, silicon photonics or thin-film lithium niobate is still open. We think that at 3.2T, the bottleneck will move from the DSP back to the optical half.

One threshold being discussed for the proposed FCC rule on China-made optical modules is "65% US BOM content." The 3nm DSPs for 1.6T today come mainly from two US companies, and one DSP alone is over a quarter of module cost; high-end EMLs are also concentrated in US and Japanese InP IDMs. How the BOM share works out depends on who these two chips are bought from, so the threshold really acts on the component layer, not on module makers.

8. Next time you see 1.6T optical module news, ask these three questions

  1. Is the shortage in the optical half or the electrical half? "InP substrates," "EML," "CW lasers" and "epitaxy" are the optical half; "DSP," "3nm" and "lead times" are the electrical half. Separate them, and you know which group of companies is affected.

  2. What kind of laser does this company use for 1.6T? EML needs eight, silicon photonics needs two high-power CW lasers, and VCSEL needs no InP. For the same one million units shipped, InP demand can differ several times over.

  3. Where do its DSPs come from, how many are locked in, and does it have LPO? Few companies ship 3nm DSPs in volume today, and the allocation a company secures is the ceiling on next year's shipments. Companies with LPO or LRO products depend less on DSPs.

Summary

Cisco says optics is the biggest supply constraint, but optics is not a single link in the chain. Broken down, how many 1.6T modules ship depends on two chips inside the module: the InP laser that emits light and the DSP that cleans up the signal.

The InP bottleneck is in materials and wafer size, and capacity grows over years. The DSP bottleneck is that few companies ship at 3nm in volume today, and capacity depends on TSMC's allocation. Every 1.6T design is really choosing which of the two chips to use less of: silicon photonics saves InP chips by sharing lasers, LPO removes the DSP, and VCSEL bypasses InP.

The signals most worth watching are Marvell's and Broadcom's 3nm DSP shipments, when new DSP suppliers reach volume, and the share of hyperscaler scale-out networks that adopt LPO. If DSP allocation loosens or LPO ramps, the 1.6T bottleneck will move from the electrical half back to the optical half.

References

Related reading

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page