Every interconnection in a modern data center will eventually rely on fiber optics. And every fiber optic link requires a laser. The laser provides the carrier wave, which is then modulated and processed by the transmit optical engine, sent through fibers and connectors, and finally received by the receive optical engine. Along the way, each fiber, connector, and photonic device introduces some signal loss.
What is Link Budget?
Link budget is the amount of signal loss that can be tolerated while ensuring sufficient laser power reaches the receiver. A lower link budget means lower power, lower cost, and a lower bit error rate — all critical goals for large-scale AI infrastructure.
In just 60 years, the laser has grown into a $20 billion annual industry.
A Brief History: From Labs to Billions
The laser was independently invented 60 years ago by General Electric, IBM, and MIT Lincoln Laboratory. The basic principle is simple: combining holes and photons releases light. A forward-biased PN junction brings together large numbers of holes and photons, while reflectors provide optical amplification and form a focused beam. Today’s lasers, while still based on this principle, are far more complex.
Laser technology has been used in telecommunications for 30 years — powering transoceanic internet, transcontinental internet, and more recently, fiber-to-the-home broadband.
Lasers in Data Transmission
In data transmission, lasers provide the optical medium for communication. The most basic method uses a single-wavelength laser. Adding more wavelengths increases both cost and complexity but delivers much greater bandwidth. Each laser is modulated via an electrical interface to transmit data. Modulation can be amplitude, frequency, phase, or polarization modulation.
More than thirty years ago, lasers began to appear in data centers with the arrival of pluggable transceivers. These devices convert electrical signals into optical signals and transmit them over laser media such as optical fibers. Data rates have steadily climbed since then. Today, nearly all horizontally scaled data is transmitted using laser-driven pluggable transceivers. The pluggable design means laser failures can be repaired quickly and at low cost.
Lasers are now being developed and initially deployed using co-packaged optics (CPO), replacing copper with fiber optics at scale.
A Booming Market, Driven by AI
This is a massive and fast-growing industry. In 2024, the global laser technology market exceeded $20 billion and is expected to surpass $30 billion by 2030. The rapid growth of the laser market is largely due to AI data centers, which now account for more than half of the market — a share that will grow significantly by 2030. Given the rapid growth in AI data center capital expenditures and the fast transition from copper to fiber over the next five years, the market size will likely far exceed $30 billion by 2030.
Which Laser Dominates AI Interconnects?
There are many types of lasers for large-scale interconnects, but soon the dominant laser for high-speed AI data center interconnects will be the Indium Phosphide (InP) laser — specifically, the continuous-wave (CW) ultra-high power (UHP) laser, a type of distributed feedback (DFB) laser. The DFB laser has a transverse periodic grating structure that forms the laser’s resonant cavity. DFB lasers using InP epitaxial materials generate light in the O-band. Vertical-cavity surface-emitting lasers (VCSELs), by contrast, have a vertical cavity structure and use GaAs epitaxial materials to generate light at wavelengths from 700 to 900nm.
Top Laser Suppliers: A Concentrated Market
The three largest laser suppliers are Coherent, Lumentum, and Sumitomo, which together hold 68% of the market. But competition is fierce, with other suppliers including Broadcom, Mitsubishi, MACOM, Applied Opto, and Landmark.
Each of the top three chipmakers has multiple production sites around the world. However, both Lumentum and Coherent have sold out their production capacity — customers now need to prepay cash to secure supply. Each company has a market capitalization exceeding $60 billion, ten times what it was a year ago.
In March of this year, NVIDIA announced separate $2 billion investments in Lumentum and Coherent to secure supply chain capacity. This happened just before NVIDIA’s GTC conference, where Chief Product Officer Jensen Huang unveiled NVIDIA’s 2028 product roadmap.
At a March OFC briefing, Lumentum demonstrated rapid growth in its InP production capacity — but still not fast enough to meet market demand. (Note: the chip photo in Figure 2 is an InP CW UHP laser chip.)
At the OFC show, Coherent also held an investor briefing, showing that its InP capacity would double in 2026, more than double again in 2027, and continue growing thereafter. Notably, Coherent was the first company to produce InP devices on 6-inch wafers.
The main driver behind this surge in demand is the rapid growth of AI data center capex, combined with the industry-wide transition from all-copper interconnects to a mix dominated by CPO, with some NPO (near-package optics) as an intermediate step between copper and CPO, and some VCSELs for slower, wider configurations.
Power: Milliwatts Matter, but Context is Everything
Laser power is typically measured in milliwatts (mW), but sometimes in dBm, a logarithmic scale:
- 0 dBm = 1 mW
- 10 dBm = 10 mW
- 20 dBm = 100 mW
- 30 dBm = 1 W
InP CW UHP lasers typically deliver 300 to 400 mW, with some reaching 600 mW. Engineers working on CMOS FinFET chips might find “ultra-high power” of only 300–600 mW surprisingly low. But from a laser perspective, this is already high power — just a few years ago, power levels below 50 mW were common. The rapid increase in bandwidth comes from both higher bandwidth per laser and the fact that one laser can drive 4, 8, or even 16 fibers simultaneously.
The rated power of an InP laser refers to optical output power. The heat dissipated is typically 3 to 4 times that amount.
Why 1310nm? The O-Band Advantage
Currently, InP CW UHP lasers operate mainly at 1310nm. This wavelength sits in the middle of the O-band, an infrared band ranging from 1260nm to 1360nm. The O-band is chosen for CPO because it offers the lowest dispersion — different wavelengths of light travel at slightly different speeds, causing optical pulses to spread over time — and relatively low signal loss per meter.
Using InP materials, manufacturers can create CW UHP lasers at other frequencies with narrowband distributions. There are many ways to achieve this. During manufacturing, by controlling epitaxial growth or using electron beam lithography to modify masks for different DFB grating designs chip by chip after wafer production. The more wavelengths and the tighter the spacing, the more complex the laser manufacturing. InP lasers can have wavelength spacing as small as 1.5nm, so in the future, many wavelengths can be accommodated within the 100nm-wide O-band. During operation, the laser frequency can be further tuned — changing temperature changes frequency, and changing input power also changes frequency.
Thermal Management: Keeping Lasers Cool and Stable
For CPO, to improve reliability and replaceability, and to avoid the laser’s sensitivity to heat, lasers are typically packaged separately from the GPU/XPU. To maintain a specific frequency, the laser’s temperature must be controlled within a very narrow range. Using an external laser is much easier than packaging the laser inside a 1000-watt GPU. External lasers are typically used with a thermoelectric cooler (TEC), which keeps the laser junction temperature within a very tight range: ±20°C. Companies like Lumentum and Coherent manufacture TECs.
ELSFP: External Laser, Small Form-factor, Pluggable
Actual InP laser chips are very small. While every InP laser design is different, they generally do not output a circular beam. For example, in one type of InP laser, the output comes from a horizontal slit, so the beam is elliptical, with a width about 2 to 3 times its height. Optical fibers are circular. Shining a laser directly into a fiber would lose at least two-thirds of the power (among other problems).
What we need is an optical path. The laser beam first passes through a collimating lens to convert the non-circular beam into a circular one. Next, the beam passes through an isolator to prevent light from reflecting back into the laser. This isolator is made of yttrium iron garnet, which has magneto-optical properties. When voltage is applied to a magnet, it can control polarization and thus block light from reflecting back into the laser. (Coherent claims to produce the majority of isolators currently on the market.) Then, a second lens focuses the light into an optical fiber. This fiber is cut at an angle to maximize the amount of light entering the fiber and ensure the light strikes at the optimal angle.
Due to the complexity of the ELSFP, customers typically purchase the complete assembly according to their needs. The ELSFP itself is complex, containing a microcontroller and dozens of other components. Coherent claims to be the only supplier capable of producing all components of the ELSFP. Due to the additional components and internal optical path losses, the power conversion efficiency of an ELSFP is only 10% to 15%. That is, only 10% to 15% of the input electrical power is converted into optical power; the rest becomes heat.
Coherent’s ELS (external laser source), like all current ELSFPs, uses a pluggable design.
Lasers for CWDM and DWDM: Scaling Bandwidth with Wavelengths
CWDM = Coarse Wavelength Division Multiplexing.
DWDM = Dense Wavelength Division Multiplexing.
Initially, lasers for CPO used a single wavelength: 1310nm. This wavelength can carry enormous bandwidth. But more bandwidth is always better. Optical fibers can transmit multiple wavelengths in both directions. The limiting factors are the O-band’s width (1260nm to 1360nm, giving 100nm of bandwidth) and the precision of laser tuning and centering to avoid wavelength overlap (precision improving from 2nm to 1nm over time). Having more wavelengths increases bandwidth but also adds complexity and cost.
The recently released OCI-MSA (Optical Computer Interconnect Multi-Source Agreement) — jointly published by AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI — is an open, interoperable optical interconnect specification designed to accelerate AI scaling. The specification proposes using 8 wavelengths, 4 in each direction.
One dense group is centered around 1311 nm with spacing of approximately 2.3 nm; another dense group is centered around 1331 nm. The specific spacing is given in GHz and varies slightly with wavelength. The deviation between the minimum and maximum values for each wavelength can be as large as ±0.2 nm. Therefore, even with dense distribution, there is a clear separation between each wavelength.
To build an ELSFP, 8 InP lasers are needed, each tuned to a specific wavelength.
Wavelength Drift and Temperature Control
InP has a wavelength drift of 0.1 nm/°C. This seems small, but over a 20°C range, the drift can reach 2 nm — which could exceed the spacing between adjacent wavelengths if the operating temperatures of neighboring lasers vary significantly, leading to crosstalk. To avoid this, all lasers in an ELSFP are connected to the same TEC, which keeps the temperature within 20°C, ensuring all lasers operate at roughly the same temperature. The laser frequency can be further adjusted by fine-tuning the operating power.
At the Optical Fiber Communication Conference (OFC), Lumentum demonstrated 16 DWDM channels with 200 GHz spacing, centered at 1310nm. This is not yet a product, just a capability demonstration.
This is not the first time. At OFC 2025 last year, Ayar Labs demonstrated a 16λ laser with 200 GHz spacing centered at 1300nm.
At the same OFC show, Scintil Photonics presented an interesting, higher-density approach to building DWDM lasers for 8 or 16 channels with 100 GHz spacing. The company assembles InP chips on a silicon photonics chip, using the silicon photonics chip to control the wavelength more precisely. This approach may sacrifice some efficiency and output power. But if the goal is precise wavelength control, it could be an ideal solution. Scintil is currently applying for qualification for small-volume production in 2027.
