Lasers Are The Heartbeat Of The Optical AI Data Center

Last month we discussed how all interconnects will be optical in the data center in five years, but that’s only part of the story. Every optical interconnect needs a laser.

The laser provides the carrier, which is modulated and manipulated by the transmitter optical engine through fibers and connectors to the receiver optical engine. Each fiber, connector, and photonics device that the laser light passes through results in some signal loss.

The link budget is the amount of signal loss that can be tolerated to have sufficient laser power at the receiver. A leaner link budget means lower power, lower cost, and lower error rates.

Lasers have grown to a $20 billion/year business in 60 years
Lasers were independently invented 60 years ago by GE, IBM, and MIT Lincoln Labs. The basic concept is simple — combining holes and photons releases light, a forward-biased PN junction brings lots of holes and photons together, and reflectors provide optical amplification and a focused beam. The lasers used today are built on this approach, but they are much more complex.

Lasers have been used in telecommunications for 30 years for trans-oceanic internet, then trans-continental, and recently “fiber to the home” internet.

For transmitting data, the laser is the carrier that provides the light medium to communicate. The most basic approach uses a single wavelength. Adding more wavelengths increases cost and complexity, but also delivers more bandwidth. Each laser is modulated by electrical interfaces to transmit data. Modulation can be of the amplitude, frequency, phase, or polarization.

Lasers came into use in the data center more than three decades ago with pluggable transceivers, converting electrical signals to optical signals carried on a laser medium over optical fibers, with steadily increasing data rates. Today, essentially all scale-out data is transmitted using laser-powered pluggable transceivers. The pluggable format allows a laser failure to be quickly fixed at low cost.

Lasers are now in development and initial deployment with co-packaged optics (CPO), which replaces copper with fiber for scale-up.

This is big business, and it’s growing fast. The global laser technology market was over $20 billion in 2024. It’s projected to exceed $30 billion in 2030. The rapid growth of the laser market is due to AI data centers, which already account for more than half of the market, and will represent a much larger share by 2030. The market will likely be much bigger than $30 billion in 2030 given the rapid growth in AI data center CapEx and the rapid transition of scale-up from copper to optical over the next 5 years.on

Lasers for scale up interconnect
There are many kinds of lasers, but the one that is coming to dominate high-speed AI data center interconnects is indium phosphide (InP), specifically continuous wave (CW) ultra-high power (UHP), which is a type of distributed feedback (DFB) laser. DFB lasers have a lateral periodic grating structure that forms the laser’s resonant cavity. DFB lasers with InP epi material generate O-band wavelengths. Vertical-cavity surface-emitting lasers (VCSELs) have a vertical cavity structure and use GaAs epi material to generate 700 to 900nm wavelengths.

Fig. 1: InP Lasers are the workhorse for AI data center interconnects. Source: Lumentum

The three largest laser suppliers are Coherent, Lumentum, and Sumitomo, which together have 68% share (source: Fact.MR). But the field is crowded with other suppliers, including Broadcom, Mitsubishi, MACOM, Applied Opto, and Landmark.

Each of the big three has its own manufacturing facilities in multiple locations, but both Lumentum and Coherent are sold out and require up-front cash to obtain capacity. Each has a market cap of more than $60 billion, which is 10x what they were just a year ago.

In March, Nvidia announced a $2 billion investment in each of Lumentum and Coherent to secure supply chain capacity. This happened shortly before Nvidia’s GTC, where Jensen Huang showed CPO starting on the Nvidia roadmap from 2028.

At its OFC briefing in March, Lumentum showed a rapid growth in InP capacity, but it’s still not growing fast enough to keep up with demand. (Note the die photo in Figure 2 is a InP CW UHP laser die).

Fig. 2: Lumentum InP capacity is growing fast, but demand is increasing faster. Source: Lumentum

Coherent also did an investor briefing at OFC, showing its InP output capacity doubled in 2026 and will more than double in 2027 — and continue to grow from there. Notably, Coherent is the first to move to six-inch wafers for InP.

The primary driver of this explosion in demand is tied to the rapid growth in AI data center CapEx and the shift over the next five years in scale-up interconnects from all-copper to predominantly CPO, with some NPO (near-packaged optics), an intermediate step between copper and CPO, and some VCSEL in slow and wide configurations.

Lasers are quoted in mWatt power, but sometimes also in dBm, which is a log scale:
0 dBm = 1mW;
10 dBm = 10mW;
20 dBm = 100mW;
30 dBm = 1W.

InP CW UHP lasers are typically 300 to 400 mW, with some reaching 600mW. Anyone doing CMOS finFET chips might find it odd that ultra-high power means 300 to 600mW. From a laser point of view, this is high power, because <50mW was typical just a few years ago. The rapid march upwards in bandwidth is due to a combination of growing bandwidth combined with using one laser to power now 4, 8, or 16 fibers.

The rated power of an InP laser is the laser output power. The heat dissipation is another 3X to 4X.

InP CW UHP lasers today are primarily 1,310nm. This is the middle of the O-band, which is infrared and runs from 1,260nm to 1,360nm wavelength. The O-band is chosen for CPO for having the lowest chromatic dispersion — where different wavelengths travel at slightly different speeds, causing optical pulses to spread in time as they travel — and relatively low signal loss per meter.

It is possible to build InP CW UHP lasers in other frequencies with narrow distributions of light. There are many ways to do this. In manufacturing, the masks can be differentiated for changes in DFB grating design by controlling epitaxial growth or by using e-beam lithography to modify die-by-die after the wafer is produced. The more wavelengths and the more closely packed they are, the more complicated it is to make the lasers. InP laser wavelengths can be spaced as close as 1.5nm apart, so at some future date a lot could fit within the 100nm-wide O-band. In operation, laser frequency can be further tuned. Varying temperature will change frequency, and changing input power will change frequency.

For CPO, lasers are kept off the GPU/XPU package for reliability/replaceability, but also due to heat sensitivity. To maintain a specific frequency, the temperature of the laser must be within narrow ranges, which is much easier to do with an external laser than one packaged with a 1,000+ watt GPU. The external laser is typically mated with a thermoelectric cooling device (TEC), which keeps the laser junction temperature within a very narrow range: +/-20C. Lumentum, Coherent, and others manufacture TECs.

ELSFP: External laser small form-factor pluggable for scale up
The actual die for an InP laser is very small. Each InP laser is different, but they don’t output circular beams typically. For example, in one type of InP laser the output of the laser comes from a horizontal slit such that the beam is of an oval shape, about 2x to 3x wider than tall. The fiber is round. By shining the laser directly at the fiber, at least 2/3 of the power would miss the fiber (among other issues).

Fig. 3: ELSFP optical path. Source: Gemini, from prompt

What is needed is an optical path. As shown above, the laser beam goes through a collimating lens, which reshapes the beam from a non-circular shape to a circular one. Next, the light passes through an isolator, which prevents light from reflecting back into the laser. The isolator is made of yttrium iron garnet and is magneto-optical. When a magnet has voltage applied, it controls polarization to block light from bouncing back into the laser. (Coherent claims it makes the majority of isolators today.) Then, a second lens focuses the light into a fiber, which is cut at an angle to maximize the light that enters the fiber and at optimal angles.

This is complex, so customers typically buy ELSFPs with what they need. The ELSFP is complex, with microcontrollers and dozens of components. Coherent claims it is the only supplier that makes all of the components of an ELSFP. Because of the extra components, and because of losses in the internal optical path, the wall plug efficiencies of ELSFP are 10% to 15%. So 10% to 15% of the wall-plug power is converted to light power, the rest to heat.

Coherent’s ELS (external laser source) is shown below. It is designed to be pluggable, as are all ELSFPs today.

Fig. 4: External laser source for CPO. Source: Coherent

Lasers for scale-up CWDM and DWDM
CWDM = Coarse wavelength division multiplexing.

DWDM = Dense wavelength division multiplexing.

Initially, lasers for CPO are shipping at 1,310nm: a single wavelength. A huge amount of bandwidth can be transmitted with this. But more bandwidth is always better. Multiple wavelengths can be transmitted on a fiber and in both directions. The limit is the width of the O-band (1,260 to 1,360nm = 100nm wide) and how precisely lasers can be tuned and centered so they don’t overlap (2nm going to 1nm over time). Having more wavelengths increases bandwidth, but it also increases complexity and cost.

The recent OCI-MSA is an Optical Computer Interconnect Multi-Source Agreement, published by AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI. It is an open, interoperable optical interconnect spec for AI scale up. They propose 8 wavelengths, 4 in each direction.

Fig. 5: 4 wavelengths for each direction for 8 total. Source: OCI MSA Version 1.0 Spec

One group is dense, centered around 1,311 nm with ~2.3nm spacing, the other around 1,331nm. The exact spacing is in GHz, which varies slightly with the wavelength. Each wavelength can have a spread between min and max of +/-0.2nm. So there is a distinct gap between each wavelength, even when densely distributed.

To build an ELSFP for this, there will need to be 8 InP lasers, each tuned to the specific wavelength.

InP wavelength drift is .1nm/°C, which seems small, but over 20 °C, that is 2nm — which may exceed the spacing to the next wavelength and cause crosstalk if the adjacent lasers vary significantly in operating temperature. To avoid this, all of the lasers in an ELSFP contact with the same TEC, which controls temperature within 20 °C so that all lasers are at about the same operating temperature. Laser frequencies can be further tuned by small changes to operating power, which will shift frequency.

At OFC, Lumentum demonstrated 16 DWDM channels with 200 GHz spacing centered on 1310nm. This isn’t yet a product but a demonstration of capability.

Fig. 6: 16 DWDM channel demonstration at OFC March 2026. Source: Lumentum

This is not a first. At last year’s OFC 2025, Ayar Labs demonstrated a 16-lambda laser at 200GHz spacing centered at 1,300nm.

Also at OFC, Scintil Photonics demonstrated an interesting alternate, denser method of building a DWDM laser source for 8 or 16 channels with 100GHz spacing. The company assembled InP die on a silicon photonics die, with the SiPho die able to more precisely control the wavelength. The tradeoff may be efficiency and output power. But if precise and tight wavelength control is the goal, this may be an optimum solution. Scintil is qualifying now for low-volume production in 2027.

Lasers and the CPO end-to-end link budget
Why do CPO lasers need to be 400mW or more? It’s because the laser is shared across 8 to 16 links, the ELSFP has significant internal losses, and there are cumulative losses in an end-to-end link from XPU/switch to switch/XPU.

Next month, we’ll dig into this in more detail. There are a surprising number of steps where losses occur. Minimizing the end-to-end link budget is key to lower laser power and lower bit error rates. But reliability and flexibility considerations add connections that increase the losses. More on these tradeoffs next month.