How many optical modules does AI actually need?

Four forecasts, one product

This spring, at planning meetings in laser fabs, photonics foundries and module assembly plants, someone put the same four numbers on one slide. Each number is a forecast of 800G module shipments in 2026:

• Goldman Sachs: 33.5 million.
• Citi: 37 million.
• CMB International: 49 million.
• TrendForce: 63 million.

The product is the same and the year is the same, but the highest number is almost two times the lowest. If you must decide whether to double a production line, a range this wide does not tell you what to do.

Do not calculate the average of these numbers. First, find out what each number counts. In this essay, we call that the number’s denominator.

• Goldman and Citi count only 800G modules.
• TrendForce’s 63 million counts “800G and above,” so it includes all 1.6T modules.
• Omdia gives a figure of approximately 20 million units for AI networks. This figure looks too low next to the others. But Omdia counts only the modules in AI clusters, not the usual datacom modules around them.

Many people say that the total module market is approximately 100 million units in 2026. This number is real, but it comes from a calculation, not from a published statement. TrendForce says that 800G-class modules are more than 60 percent of global shipments. TrendForce also says that there are approximately 63 million of them. 63 million divided by 60 percent gives approximately 105 million modules of all speeds, for all buyers. So the five numbers do not really contradict each other. Mostly, each one counts a different part of the same market.

But a real question stays inside one scope. Will 800G in 2026 be nearer to 33 million or to 49 million? Do not select one analyst. It is more useful to count the market again yourself, in three ways. Each way uses data that the forecasts do not share. This is the method of the full essay, and you can use it for any market number that you plan capacity against.

Count the switches

Each module fills a switch port, so the first count starts with the switches. The switch chip sets the number of ports (its radix):

• A 51.2T switch has 64 ports of 800G (51.2 ÷ 0.8 = 64).
• A 102.4T switch has 64 ports of 1.6T (102.4 ÷ 1.6 = 64), or 128 ports of 800G.

These numbers are not design choices. They are division.

650 Group counted approximately 77,000 51.2T switches shipped in 2024. It forecasts 1.8 million in 2028. These two numbers give growth of approximately 2.2 times each year. We interpolated between them, and this step is ours, not 650 Group’s. The result is approximately 370,000 51.2T switches in 2026. This number does not include Tomahawk 6 systems, which started production-volume shipments in March 2026 and add the first 102.4T layer. At 64 ports each, 370,000 switches give approximately 24 million new 800G switch ports in one year.

Ports are not the same as modules. Each optical link has a device at each end. In a leaf-spine network (a network with two tiers of switches), there is a module in each switch port that is in use and in each NIC port. A link that uses copper cable inside the rack needs no module.

• If the full fleet used two-tier networks with optics in all positions, the module count would be approximately one third higher than the switch-port count.
• If copper cable makes the first connection inside the rack, the module count goes back toward the switch-port count.

In both cases, the switch count gives 800G demand in a range from the high 20s to the mid 30s of millions, plus the new 1.6T layer. Dell’Oro gives a check on the cumulative total. 800G passed 20 million switch ports within three years of the first shipment. 400G needed six to seven years to do the same.

Count the DSPs

The second count uses a fact from the bill of materials. A retimed module contains exactly one DSP, the digital signal processor that cleans and retimes the electrical signal. One 8-lane DSP chip retimes the full module, at 800G and at 1.6T. All the shipping parts have eight lanes: Broadcom’s Sian family, and Marvell’s Nova and Ara (now on 3 nm). So merchant DSP shipments count retimed modules one for one.

The correction goes in one direction only:

• Linear-drive modules (LPO) contain no DSP.
• Linear-receive modules (LRO) contain a device for the transmit side only. Google now deploys LRO at 1.6T.

So the DSP count gives fewer modules than the real number, by the linear share. This share is still small.

LightCounting gives two sets of numbers. In 2025, 800G PAM4 chipset shipments increased to almost three times their 2024 level, while module shipments approximately doubled. For 2026, LightCounting forecasts 1.6T chipset sales of more than $2 billion, across “tens of millions of ports.” These two 1.6T statements agree with each other. $2 billion across tens of millions of ports gives a chipset price in the low hundreds of dollars. The industry understands that 200G-per-lane DSP prices are at this level.

Broadcom said that demand for 200G-per-lane modules would pass one million units in the first year of its Sian2 ramp. 2025 passed this minimum easily. The DSP count agrees with the switch count: 800G in the tens of millions, and 1.6T in the millions, moving toward tens of millions.

Count the accelerators

The third count starts from the machines that the network serves. Operators build scale-out networks per accelerator. So module demand follows GPU shipments multiplied by an attach ratio, which is the number of modules for each GPU. The tier structure sets this ratio:

• Two tiers, one 800G NIC port per GPU, and copper for the first connection in the rack: approximately 1.5 modules per GPU.
• Three tiers, rail-optimized: approximately 2.5.
• Goldman’s model for the step between generations: approximately 3 800G-equivalents per H100-class GPU, and 4.5 per B300-class GPU.
• ConnectX-8 designs with 800G per GPU: toward 6.

The accelerator counts for this year are not certain, and the reason is useful. Rubin had thermal and memory delays. After these delays, KeyBanc expects 5.5 to 6 million Blackwell and 1.7 to 1.8 million Rubin in 2026. Before, JP Morgan modeled a year with more Rubin. But both mixes give a total of approximately 7.2 to 7.8 million high-end NVIDIA GPUs, and the module count depends mostly on the total.

Then add the other accelerators. For Google, TrendForce expects more than 6 million 800G-class modules for approximately 4 million TPUs. AWS and AMD add more volume. The result is approximately 12 million accelerators. Multiply by 2.5 to 4.5, and the accelerator count gives approximately 30 to 54 million 800G-equivalents.

The three counts have no shared inputs, and they give one range:

• 800G: from the low 30s to approximately 50 million.
• 1.6T: between 7 and 20 million, depending on how fast qualification goes.
• Total market: approximately 100 million modules.

LightCounting says that demand is approximately 30 percent more than supply this year. This is the demand that it means. The four forecasts on the slide were not really two times apart. Each one counted a different scope.

The suppliers and the buyers are more concentrated than the unit count shows. On the supply side, LightCounting’s 2025 ranking puts Innolight first, with $5.3 billion of revenue. Eoptolink is second, after it passed Coherent. Trade reports put Innolight alone at approximately 23 percent of global shipments. Together, the two Chinese leaders supply an estimated 60 percent of NVIDIA’s pluggable volume and, per TrendForce, approximately 80 percent of Google’s 2026 orders.

On the buy side, supply-chain tallies give the numbers below. Their sources are weaker, so we mark them as such.

• Meta: 10 to 12 million 800G modules in 2026.
• Google and Microsoft: approximately 12 million together.
• Amazon: approximately 5.5 million.
• NVIDIA: the largest buyer at 1.6T.

At the top, a market of 100 million units is a discussion between approximately six buyers and three sellers.

The scoreboard, and a better source

After a year of counting, we can compare the counts with the results. Cignal AI is the only firm that publishes unit counts each quarter. Its 2025 total was more than 42 million modules at 400G and above, an increase of more than 80 percent. 800G alone was more than 20 million.

In the first quarter of 2026, shipments were almost 20 million in one quarter, an increase of 145 percent. At this rate, a full year is near 80 million. The range from the three counts is still correct. TrendForce’s 63 million for 800G and above was the top of the first range, and it no longer looks unusually high.

A better source than these forecasts is now available. In July 2026, a market model became part of a regulated filing. Zhongji InnoLight’s Hong Kong listing document gives datacom revenue by lane generation from 2021 to 2030. The document attributes this model to CIC, using LightCounting data. The issuer paid for that chapter, so we do not trust it because of who wrote it. We trust it because we can check it from the inside:

• Each column sums to the total that the document prints.
• The growth that its rows give agrees with the 33.8 percent that it states.

Both checks run as tests on this site.

• 3.2T
• 1.6T
• 800G
• 400G
• 200G and below

This chart changes the shape of the answer. The market does not grow because 800G gets larger. 800G revenue increases to $18.0 billion in 2028, then decreases to $16.2 billion by 2030. From 2026 to 2030, it grows by only 2.1 percent each year, while the total more than triples. The growth comes from below, as 1.6T and then 3.2T appear under the generation that you counted. A generation does not stop suddenly. The newer generations grow larger than it, and that makes a fab harder to plan.

The filing also gives a new example of the first lesson of this essay. Its 2025 revenue for 400G and above is $17.0 billion. At the $280 industry price that the filing also gives, this is approximately 61 million units. Cignal counted 42 million. Both numbers are correct:

• InnoLight’s audited figures are 14.6 million high-speed modules at RMB 2,333. With its published revenue share of 28.1 percent, these figures give the $17.0 billion again.
• InnoLight’s own price is higher than the industry average. So its unit share must be lower than its revenue share, and this puts the industry near 61 million.
• The larger group of products includes AOCs and lower-cost form factors. The 400G-and-above count does not include them.

Read the denominator before the number. This rule also applies to good sources.

The filing counts dollars. A fab counts units, and no filing publishes units by generation. The most complete public unit series comes from Goldman Sachs, which increased its forecast again on 7 September 2026. Together with Cignal’s 2025 count, it shows the same market in units.

• 3.2T
• 1.6T
• 800G
• 400G
• Combined, split not published

Two unrelated sources agree on one date. In units, 1.6T passes 800G in 2027. In the filed revenue model, 1.6T also passes 800G in 2027.

The two sources do not agree about the next generation. Goldman forecasts 23 million 3.2T modules in 2027. The filed model shows no 3.2T revenue until 2028. This difference is not rounding. It is a full year of the next ramp, and it is the number to watch.

The chart also shows how fast the forecasts change. In the spring, Goldman’s 2026 forecast for 1.6T was 7 million. By September, it was 33 million, which is above the top of the range from the three counts. 400G is still approximately 22 million of Cignal’s 2025 count, but neither series counts it separately after 2025. Buyers did not stop buying it: in the filed model, 400G revenue continues to increase until 2028. The forecasts stopped counting it.

The year silicon crosses half

Inside these unit counts, a technology change is almost complete. LightCounting says that 2026 is the first year in which transceivers with silicon photonics modulators are more than half of the market, by revenue. That market is approximately $40 billion. CMB International gives the split by generation:

• 800G: approximately 50/50 between EML-based designs and silicon photonics designs. An EML is a laser with a modulator on the same chip.
• 1.6T: approximately 30/70.

Innolight, the largest vendor, gave public data in the same direction. Silicon photonics is already more than half of its 800G and 1.6T shipments, and the 1.6T share is the higher of the two.

You do not need an opinion about technology to see why. You can count it:

• An 800G DR8 module with EMLs needs eight modulated lasers, one for each lane.
• The silicon photonics version needs two CW (continuous-wave) lasers. They supply light to a PIC (photonic integrated circuit), and the PIC does the modulation. The shipping Carmel8 PIC has exactly two.
• The 1.6T bill of materials follows the same pattern: eight 200G EMLs, or two CW lasers and one PIC. Both versions use one DSP.

Now multiply by the market. At 40 million 800G modules, the EML half uses approximately 160 million EMLs. The silicon half uses approximately 40 million CW lasers. Per TrendForce, combined EML and CW-DFB laser capacity is approximately 50.7 million units each month this year, approximately double the 2025 level. That total does not separate the two laser types, so it cannot show which type is short. The EML evidence is a different TrendForce report. In December 2025, TrendForce said that NVIDIA had secured capacity at the key EML suppliers, with lead times beyond 2027. It named CW lasers with silicon photonics as the main alternative route. The market did not choose silicon photonics because it liked it. EML lasers were not available, and a silicon photonics module needs one quarter as many lasers.

For readers of the module BOM essay, this is the same supply calculation, seen from one level higher. The laser line was the first bottleneck, and the technology mix changes around it now.

From modules to wafers

Each silicon photonics module now contains a photonic die, and dies come from wafers. So we can take the count one level lower, to the place where companies decide 2027 capacity now. The calculation has three steps:

1. An 8-inch wafer has approximately 28,000 mm² of usable area.
2. A shipping 800G DR8 PIC is 52.5 mm² (Credo’s Carmel8, 7 by 7.5 mm). The 648-per-wafer figure reported for TSMC’s PIC line gives a die of approximately 90 mm², if those are 300 mm wafers. Our heavy case assumes 110 mm², for a larger die.
3. Per PhotonDelta, photonic yields are 30 to 70 percent.

Divide the area by the die size and apply the yield. One wafer then gives approximately 115 to 340 known-good modules, and the result depends fully on your assumptions.

That dependence is the finding. With the lean assumptions, 30 million silicon photonics modules need approximately 90,000 wafers each year. With the heavy assumptions, they need approximately 260,000. At realistic fab utilization, the capacity that must exist is approximately 100,000 to 370,000 8-inch wafers each year. The model below gives you three controls: die area, yield and utilization. Move them for thirty seconds to see how wide the honest range is. The three controls change the answer by almost 4 times, but the module count changes it only linearly.

The wafer arithmetic, knobs exposed

Pick how many silicon photonics modules 2026 has to build, how much photonic silicon each one carries, and how well the fab runs. The output is annual 8-inch wafer starts, and the capacity that has to exist to make them.

Scenario

LeanBaseHeavyCustom

SiPh modules, 2026

30M

Photonic silicon per module

80 mm²

Dies per module

MonolithicTX + RX

Known-good-die yield

60%

Fab utilization

75%

Modules per wafer

191

318 gross die

Wafer starts per year

157k

200 mm, shipped

Capacity that must exist

210k

17k per month

In 300 mm equivalents

70k

area-converted starts

Mid-size 200 mm fabs

0.7

at ~25k wafers per month each

• This scenario, capacityour estimate210k
• TSMC capacity, mid 2026estimate via TrendForce270k
• TSMC capacity, exit 2026estimate via TrendForce405k
• One dedicated 200 mm fabassumption, ~25k wafers/mo300k

Anchors: 52.5 mm² is the die area of a shipping 800G DR8 PIC (Credo Carmel8, 7 × 7.5 mm). The 648 die per wafer reported for TSMC's PIC line gives about 90 mm² on a 300 mm wafer. Yields span the 30 to 70 percent band PhotonDelta reports for photonic chips. The 110 mm² heavy die, utilization and the fab-size denominator are assumptions. All bars are 200 mm equivalents.

Now compare the public data points with that range, but first remove one of them. QY Research counted approximately 63,700 8-inch-equivalent silicon photonics wafers shipped in all of 2025. This essay first read that count as agreement with the bottom of the range. The count does not pass inspection:

• The same publisher gives a processed-wafer price of $7,620 and a 2025 market of approximately $485 million.
• 63,700 × $7,620 = $485 million. So the count and the price are one estimate, not two observations.
• Tower alone holds contracts for $1.3 billion of 2027 silicon photonics revenue. This is more than the publisher’s calculation leaves for the full industry in that year.

What remains is the capacity that companies are building, and that signal was stronger anyway. TrendForce relays institutional estimates for TSMC’s PIC line, in wafers each month:

• Approximately 500 at the start.
• 10,000 by mid-2026.
• 15,000 by the end of 2026.
• At least 25,000 by 2028. At the top, this is approximately 675,000 8-inch equivalents each year from one line.

Tower received $290 million in prepayments and is multiplying its output. STMicroelectronics says that its photonics line output quadruples by 2027. Nomura still expects capacity to be 5 to 15 percent less than demand this year. Our estimate, one to a few hundred thousand 8-inch wafers each year, is the scale that the industry builds toward. The top of the range is more than one full mid-size 200 mm fab for photonics only.

One caveat stops this from being a comfort. The early customers named for the TSMC line are NVIDIA, Broadcom and AMD. These are co-packaging programs, not pluggable module makers. Capacity that exists is not the same as capacity that this market can buy. The largest new line in photonics points at a different product.

One level lower is the substrate. Almost all of these wafers start as photonics-grade SOI (silicon on insulator), a specialty product with effectively one volume supplier. Soitec’s Photonics-SOI revenue passed $100 million in its fiscal 2026. Soitec guides it to more than double this year, and it recently qualified a 300 mm fab in Singapore for it.

Substrate demand is not a separate estimate. It is the same wafer count: one starting substrate for each wafer start, plus test and qualification overhead. So a wafer count in the low hundreds of thousands is also an SOI count. A supply chain this concentrated needs the same attention that the laser line needed in 2024.

What to take away

1. Read the denominator before the number. 33.5M, 49M, 63M and “100M” count different scopes: 800G only, 800G and above, and the full market. They do not describe different markets.
2. Three independent counts give the 2026 range: approximately 30 to 50 million 800G and 7 to 20 million 1.6T modules. The counts are switch radix times shipments, one DSP per retimed module, and accelerators times attach ratio. The 2025 result (more than 42 million at 400G and above) and the first quarter of 2026 (near 20 million) are both in the range. Goldman’s September forecast for 1.6T, 33 million, is above it.
3. Concentration is the real shape of the market. Approximately six buyers and three sellers control most of the volume, and Innolight has approximately one quarter of global shipments.
4. Silicon photonics becomes more than half of transceiver revenue in 2026 for a countable reason. Two CW lasers replace eight EMLs in each module, and EML lasers were short first.
5. 30 million silicon photonics modules need approximately 100,000 to 370,000 8-inch wafers of capacity each year. At the top, that is more than one full mid-size 200 mm fab for photonics, on an SOI substrate with a single source. The one published wafer count that seemed to agree with this is circular, and this essay withdraws it. The real data point is the capacity that companies are building, and the largest new line is for co-packaging, not for pluggables.

This lesson applies outside optics too. Before you build capacity against a market number, you can count it again from a related count: the chips that pair with the product, the systems that hold it, or the machines that use it. When three unrelated counts keep giving the same range, you can stop the argument about forecasts and decide what to build.