Unlocking DWDM at Scale for AI Datacenters
The rapid growth of AI workloads is driving an unprecedented demand for datacenter interconnect bandwidth. As AI clusters scale beyond single racks, the industry is converging on multi-wavelength DWDM optical interconnects as the solution, but the laser source required to power them at scale remains an unsolved challenge.
Higher wavelength counts are needed to meet aggregate bandwidth targets, yet the two primary laser source approaches each face fundamental constraints that prevent them from meeting this demand.
Conventional discrete DFB-based (Distributed Feedback) external laser sources do not scale to higher wavelength counts. Each additional color requires a higher-power laser die to maintain equivalent output at the CPO input, a constraint that compounds yield penalties, increases cost non-linearly, and narrows the supplier pool as channel counts grow. The architecture scales by demanding more from the hardest part of the system.
Comb lasers have been proposed as an alternative path to high wavelength counts, but remain at an early stage of development. Current comb laser implementations face significant challenges in wall-plug efficiency, output power per line, and manufacturability at datacenter scale. Delivering equivalent optical power to the CPO modulators requires substantially higher electrical power than a direct DFB-based ELS architecture, which is an efficiency penalty that is particularly significant at the power densities demanded by AI datacenter deployments.
Critically, the lower per-line power of current comb sources also creates a fundamental tension with the CW input requirements of micro-ring modulator based CPO engines, a system-level constraint that requires substantial advances in MRM sensitivity before comb-based architectures become viable at scale.
These are not insurmountable challenges, but they are not yet solved.
Photon Bridge takes a different approach:
making DFBs scale
Making DFBs scale
Photon Bridge takes a different approach: making DFBs scale.
Our cantilever waveguide integration platform integrates the wavelength multiplexer directly on the silicon interposer alongside the laser array, addressing channel count by addition rather than amplification.
Each DFB operates at modest, thermally manageable power levels - well within the reliable operating envelope of multiple qualified foundries - while the integrated architecture delivers the aggregate multi-wavelength output that next-generation CPO engines require.
The result is a laser source that combines the proven performance, efficiency, and supply chain maturity of DFB technology with a scalable path to 8, 16, and 32 wavelengths that conventional discrete approaches cannot match.
Photon Bridge’s platform delivers three compounding advantages that underpin this approach.
The first is optical power
Photon Bridge preserves full InP laser gain through facet coupling: the laser operates entirely within the InP die, as designed by the foundry, with no integration penalty on gain, threshold, or efficiency. Thick-film waveguide geometry then handles the resulting optical power without nonlinear penalty, supporting power levels exceeding 1W on-chip. Together these properties deliver higher optical power to the fiber output than alternative heterogeneous integration architectures can achieve, a structural advantage that widens as channel counts and modulation speeds increase.
The second is energy efficiency
Photon Bridge’s thick-film waveguide geometry enables engineered thermal co-tracking between the on-chip multiplexer and the InP laser array across the full operating temperature range. This passive alignment eliminates the need for active thermal stabilization at the module level, opening a path to TEC-free ELS operation and meaningful improvement in system-level wall-plug efficiency, a benefit that compounds across the large number of ELS modules deployed in a datacenter-scale AI cluster.
The third is supply chain resilience
By operating at lower per-laser power levels than conventional designs, Photon Bridge accesses a more favorable point on the InP yield curve and enables qualification across multiple foundries, providing procurement flexibility and supply continuity that single-source, high-power incumbent approaches cannot match.
The system-level impact is significant. Photon Bridge targets greater than 30mW per color per fiber at the CPO input, the highest output of any heterogeneous laser integration platform available today. Assuming a 3dB total optical link budget from ELS fiber output to CPO engine input, a single 16-color module can serve up to 112 co-packaged optical engines at 50Gbps per wavelength, scaling to 9.6 Tbps total bandwidth per module at 200Gbps per wavelength, from a single compact, field-deployable unit.
The Laser Source Challenge
Nvidia has publicly committed to CPO-based scale-up switching in its 2028 Feynman platform, with large-scale deployment volumes projected between 2028 and 2030. Co-packaged optics architectures of this class rely on external laser sources to supply the multi-wavelength optical power that CPO engines require, making high-performance, scalable ELS a critical enabler of the transition, not merely a supporting component.
The industry has aligned around this architecture: the OCI-MSA specification, adopted by Nvidia, AMD, Broadcom, Microsoft, and Meta, defines the external laser source as the standard optical power delivery mechanism for co-packaged optics deployments.
Near-term CPO deployments are expected to operate at 4–8 wavelengths per fiber. The segment that will dominate through the mid-2030s, however, is the 8–32 wavelength range, the point at which bandwidth density and energy efficiency requirements converge, and at which laser source scalability becomes the critical constraint. Realizing this transition requires laser sources that meet strict technical demands: high optical output power across increasing wavelength counts, precise wavelength control, manufacturability at datacenter scale, high wall-plug efficiency, and long-term reliability.
These requirements define the design challenge and the competitive landscape.
Why Architecture Matters:
The Case for Integrated Multiplexing
The transition from copper to optical interconnects in AI datacenters is not simply a component substitution. The central challenge is wavelength count: DWDM systems require more colors to meet aggregate bandwidth targets, yet the dominant conventional approach - discrete high-power DFB lasers feeding a separate star coupler - carries scaling penalties that become increasingly severe as channel counts grow. Comb lasers address channel count differently but introduce their own constraints, as discussed below. Photon Bridge’s integrated multiplexing architecture resolves the scaling problem at its root.
1. The Conventional Architecture and Its Limits
In a conventional N-color external laser source, each wavelength is produced by a separate high-power discrete laser diode. These diodes connect via polarization-maintaining fiber to a discrete star coupler, which combines and redistributes the wavelengths across output fibers to the CPO engines.
The deeper problem is how this architecture scales. Moving from 8 to 16 colors requires doubling the optical power per laser diode to maintain equivalent per-color power at the CPO input. Moving to 32 colors requires quadrupling it. High-power single-mode InP lasers operate at the demanding end of the InP yield curve — thermally stressed, yield-limited, and increasingly difficult to source from multiple suppliers as channel counts grow. The conventional architecture scales by demanding more from the hardest part of the system: the laser die itself.
This creates two compounding problems.
First, cost scales non-linearly: higher power die carry yield penalties that translate directly into module cost as channel counts increase.
Second, supply chain resilience degrades: the pool of qualified foundries capable of supplying high-power, yield-limited die is narrow, creating single-source dependencies that are a structural vulnerability for large-scale datacenter deployments.
2. Integrated Multiplexing: a Better System
As noted, comb laser architectures face complementary constraints at the system level that remain unsolved at datacenter scale. Photon Bridge’s approach eliminates the discrete star coupler entirely by integrating the wavelength multiplexer directly on the silicon interposer alongside the laser array. This integration changes the scaling dynamic fundamentally.
In the conventional architecture, scaling wavelength count requires increasing per-laser power and the system grows by amplification. In Photon Bridge’s architecture, scaling wavelength count is addressed at the platform level without increasing per-laser power demands. The system grows by addition, not amplification, and the yield, cost, and supply chain constraints that punish the conventional approach at 16 and 32 colors do not apply. The full implications of operating at lower per-laser power levels - yield curve position, multi-foundry qualification, and cost structure - are developed in the next section.
The on-chip multiplexer uses Photon Bridge’s thick-film silicon waveguides, which require no active tuning to maintain channel alignment — a result of the engineered thermal co-tracking described in below. Fewer active components means fewer failure points: a meaningful reliability advantage at the system level.
Once integrated multiplexing is established as the correct architectural foundation, the question becomes what additional advantages different implementations deliver. This is where Photon Bridge’s cantilever waveguide platform provides two further decisive advantages: higher optical power delivery and superior energy efficiency.
Photon Bridge Technology Platform
Photon Bridge’s laser integration platform achieves optical precision and manufacturing scalability simultaneously through mechanical design. The cantilever waveguide interface enables high-yield passive alignment of III-V lasers to silicon, unlocking a class of integrated DWDM sources that competing approaches cannot replicate.
1. Cantilever-Enabled Integration: Precision Without Active Alignment
At the core of the Photon Bridge platform is a mechanically compliant silicon cantilever waveguide that resolves the fundamental tension between optical precision and flip-chip assembly tolerances. Standard OSAT flip-chip bonders achieve placement accuracy of approximately 2µm. Photon Bridge’s cantilever waveguides flex and compress during bonding, converting this mechanical variance into less than 200nm optical alignment accuracy at the laser-silicon interface — without active alignment, post-bond tuning, or specialized tooling.
Photon Bridge integrated device
InP DFB laser array bonded to silicon photonics interposer via cantilever waveguide interface.
Advantage 1
Higher Optical Power
Photon Bridge delivers higher optical power at the CPO input than competing membrane silicon platforms through two compounding mechanisms, both of which are structural consequences of the integration architecture rather than design optimizations.
The first is full preservation of laser gain.
Evanescently-coupled heterogeneous integration approaches rely on on-chip gratings that cause the lasing mode to interact with the silicon, making laser gain a function of the integrated structure rather than the InP die alone. Photon Bridge’s facet-coupled architecture avoids this entirely. The laser operates as designed by the InP foundry, with no integration penalty on gain, threshold, or efficiency. The silicon integration begins where the InP die ends: the two are optically sequential, not overlapping.
The second is waveguide power handling.
Photon Bridge uses thick-film silicon waveguides, compared to the sub-micron membrane waveguides used in competing platforms. Based on internal simulations applying published nonlinear loss parameters to both geometries under equivalent loss criteria, thick-film waveguides support optical power levels exceeding 1W, more than 50× the power handling of membrane silicon alternatives (simulation-based; methodology and literature references available on request). Membrane silicon waveguides arrive at their nonlinear power handling limit already carrying a linear propagation loss penalty that thick-film waveguides do not: a compounding disadvantage that widens as power levels increase.
The result is higher optical power delivered to the fiber output, a structural advantage that widens as channel counts and modulation speeds increase. Photon Bridge targets greater than 30mW per color per fiber at the CPO input, enabled by this compounding architectural foundation.
The measured LI characteristics of Photon Bridge O-band DFB laser devices demonstrate the device performance that underpins the platform power targets.
Single-facet output reaches approximately 50mW at 25°C, with output remaining above 29mW at 65°C junction temperature, confirming that Photon Bridge laser devices operate in the power regime required for high-performance DWDM interconnects. Specific system output power depends on operating point, thermal environment, and optical path design.
Photon Bridge O-Band DFB Typical Output Power Curves
DFB laser LI characteristic (single facet): measured optical output power across operating temperature range. Device performance confirms operation in the power regime required for the Photon Bridge platform.
Advantage 2
Energy Efficiency
Photon Bridge’s thick-film waveguide geometry delivers an additional and distinct efficiency advantage through engineered thermal co-tracking.
The thermal drift rate of the on-chip multiplexer is engineered to match that of the InP laser array across the full operating temperature range, a consequence of the reduced optical confinement in thick-film waveguides relative to membrane silicon alternatives, which exhibit significantly higher thermal sensitivity. Measurements on fabricated devices confirm passive laser-to-multiplexer alignment is maintained across the operating temperature range without active thermal control.
This engineered co-tracking opens a path to TEC-free ELS module operation, removing a significant parasitic power consumer from the module power budget. Quantification of the full system efficiency benefit is ongoing; directional analysis indicates a meaningful improvement in wall-plug efficiency that compounds across the large number of ELS modules deployed in a datacenter-scale AI cluster.
Passive thermal alignment also supports compatibility with floating wavelength architectures, an emerging requirement in next-generation CPO deployments.
Advantage 3
Multi-Foundry Supply Resilience
In a supply-constrained market for InP components, the ability to qualify and switch between multiple laser foundries is a strategic advantage that is largely invisible in component-level specifications but critical to datacenter operators planning multi-year deployments.
The urgency of this constraint is already visible in the market. Leading AI accelerator vendors have secured significant portions of available supply for ultra-high-power laser diodes, creating procurement pressure for other Tier-1 players and prompting active evaluation of alternative architectures that do not depend on these scarce components. Supply chain resilience is not a future risk, it is a present commercial reality.
Photon Bridge’s platform preserves this flexibility through its architecture. By operating at lower per-laser power levels than conventional external laser source designs, Photon Bridge accesses a more favorable point on the InP yield curve: reducing cost per channel and broadening the pool of qualified foundry sources.
Because Photon Bridge integrates fully manufactured, tested, and burned-in laser die, qualification is performed at the finished component level. There is no co-processing dependency between the InP and silicon steps, and no requirement to re-qualify the full process when switching foundry sources.
This means Photon Bridge customers are not captive to a single supplier. As AI datacenter buildouts accelerate and InP supply tightens, this architectural flexibility translates directly into procurement leverage and supply continuity that single-source, high-power incumbent approaches cannot match.
The Photon Bridge External Laser Source
Photon Bridge’s platform translates directly into a scalable, field-deployable External Laser Source in a TOSA form factor (Transmit Optical Sub-Assembly), a standard, well-understood packaging format that simplifies integration into downstream optical sub-assemblies and reduces system-level integration risk for OEM customers.
The module targets greater than 30mW per color per fiber at the CPO input. The Photon Bridge platform supports 8-, 16-, and 32-wavelength configurations.
The core integration platform, waveguide architecture, and foundry process are designed to scale across these configurations without fundamental redesign, addressing increasing channel counts while maintaining consistent performance and manufacturability at each wavelength tier.
Photon Bridge External Laser Source: TOSA module with multi-wavelength DWDM fiber outputs.
This output power level is not merely compliant with current interconnect specifications, it is the enabler of next-generation ones. Emerging DWDM interconnect standards are currently specified at 50Gbps per wavelength, a modulation rate constrained in part by the optical power limitations of available laser sources. Higher per-color power unlocks higher modulation speeds on the same channel plan, a transition already indicated by leading AI accelerator vendors as their next development target.
The system-level implications of greater than 30mW per color are significant. Assuming a 3dB total optical link budget from ELS fiber output to CPO engine input - accounting for connectors, fiber-to-chip coupling, and passive splitting -approximately 15mW per color is available at the engine input. This enables a single fiber color to serve multiple CPO engines simultaneously via passive split, without active power management or additional optical amplification. Across a 16-color module, the resulting bandwidth capacity scales as follows:
Modulation Speed | Engines per color | Total Engines | Total Bandwidth |
|---|---|---|---|
50 Gbps NRZ | 7 | 112 | ~5.6 Tbps |
100 Gbps PAM4* | 5 | 80 | ~8 Tbps |
200 Gbps PAM4* | 3 | 48 | ~9.6 Tbps |
* 100 and 200Gbps figures based on modeled estimates pending published specifications. Assumes 2mW, 3mW, and 5mW CW input per engine respectively. 3dB link budget assumes connectors, fiber-to-chip coupling, and passive splitter excess loss.
Future outlook
Optical scale-up in AI datacenters
The optical scale-up transition is underway, but how next-generation DWDM architectures will evolve remains an open question: more wavelengths, higher power per wavelength, higher modulation speeds, or combinations of all three.
The one point of consensus is that optical power delivery will be a critical constraint as the transition progresses. Near-term deployments at 4–8 wavelengths are already underway; the segment expected to dominate through the mid-2030s is the 8–32 wavelength range, where bandwidth density and energy efficiency requirements converge and where laser source scalability becomes the decisive constraint.
Photon Bridge is designed specifically for this window.
Photon Bridge’s platform is designed to address this uncertainty directly. High optical power per wavelength and a scalable path to higher channel counts are valuable regardless of which direction the market takes. More power per color enables higher modulation speeds on existing channel plans. More colors enables higher aggregate bandwidth on existing fiber infrastructure.
Both directions are served by the same architectural foundation.
As deployment volumes grow, supply chain resilience will become an increasingly visible constraint. Architectures dependent on high-power, yield-limited die from a narrow supplier pool will face procurement pressure that compounds with scale.
Photon Bridge’s multi-foundry qualification approach is designed to stay ahead of this, providing supply continuity at the volumes that large-scale deployments require.
What this means
Photon Bridge is built to scale laser source technology to meet the bandwidth demands of next-generation AI datacenters in performance, volume, and cost simultaneously.
By integrating the multiplexer on-chip, Photon Bridge removes the scaling constraint that forces conventional architectures to demand ever-higher power from individual laser diodes, accessing a more favorable point on the InP yield curve and enabling multi-foundry qualification that protects customers from the supply dependencies that increasingly threaten large-scale deployments.
On optical power, Photon Bridge’s facet-coupled integration preserves full laser gain, decoupling laser performance entirely from the silicon process, while thick-film waveguide geometry handles the resulting power without nonlinear penalty. Together, these properties deliver greater than 30mW per color per fiber at the CPO input, enabling a single 16-color module to serve tens to over one hundred CPO engines depending on modulation speed.
On energy efficiency, engineered thermal co-tracking between the on-chip multiplexer and the InP laser array eliminates the need for active thermal stabilization, opening a path to TEC-free operation and meaningful improvement in system-level wall-plug efficiency at datacenter scale.
As AI datacenter architectures evolve, Photon Bridge ensures that laser source technology will be an enabler of growth, not a limiting factor.
As AI datacenter architectures evolve, Photon Bridge ensures laser source technology will be an enabler of growth, not a limiting factor.
