All Posts
April 1, 2026
x min read

The Laser Problem in AI Hardware and How POET Can Solve It

Print / Save as PDF

‍

Why AI clusters are divorcing lasers from optical engines—and how POET’s Optical Interposer is built for exactly this moment.

The past two years of artificial intelligence infrastructure build-out have clarified something uncomfortable for systems architects: the hard part is not just more compute. It’s moving more data, more often, across more chips—without turning the data center into a thermal and electrical dead end.

That pain concentrates in the interconnect. Today’s AI clusters depend on enormous numbers of optical modules—800G now, accelerating to 1.6T and beyond—because copper simply cannot deliver the reach and bandwidth that scale-up and scale-out fabrics demand. But pluggable optics at the rack faceplate carry their own scaling wall: rising per-module power, growing cooling burden, and a hard faceplate density ceiling.

This is why co-packaged optics (CPO), near-packaged optics (NPO), and optical I/O chiplets keep appearing in roadmaps. The principle is straightforward: shorten the electrical path between the SerDes and the optics by placing optical engines—modulators and detectors—directly on or beside the switch or accelerator package. Shorter traces mean lower loss, lower latency, and the removal of power-hungry retimers and DSP stages that exist largely to fight the physics of long copper channels.

But shortening the electrical path exposes a new bottleneck: the laser.

Why the Laser Has to Move

Silicon-photonics optical engines need a stable optical carrier—continuous-wave (CW) light that modulators can imprint data onto. In a conventional pluggable transceiver, the laser lives inside the module. That’s convenient when the module sits at the cool faceplate. When you move the optical engine next to a 500W+ AI accelerator, that convenience becomes a liability on two fronts.

“The pluggable external laser has the lowest temperature environment and is generally advantaged in reliability, maintainability/serviceability, and stability.”

International Photonics & Electronics Committee (IPEC), Pluggable External Laser Source Implementation Agreement

Thermal stress. High-power lasers operating continuously near extreme hotspots age faster, drift in wavelength, and fail sooner. The thermal decoupling problem is not merely engineering inconvenience—it’s a fundamental reliability math problem at million-port AI cluster scale.

Serviceability. A failed integrated laser on a co-packaged switch is not something an operator wants to remediate by pulling the entire switch ASIC out of a live pod. The replacement unit must be swappable at the faceplate, independently of the optical engines riding next to the compute die.

External Light Sources (ELS) are the industry’s answer. An ELS is a multi-wavelength laser module that lives outside the hot package—typically at the system faceplate—delivering CW light over fiber to one or more optical engines that handle modulation and detection. The logic is clean: keep optical conversion close to the ASIC, but keep the laser itself in a cooler, independently replaceable location.

Standards Are Making ELS Real

Without interoperability standards, ELS is a bespoke science project for every vendor. The Optical Internetworking Forum (OIF) has addressed this with the External Laser Small Form-Factor Pluggable (ELSFP) implementation agreement—a faceplate pluggable form factor explicitly designed for CPO systems whose optical engines carry no integrated lasers.

What ELSFP Defines

  • Blind-mate optical connectivity enabling safe, hot-swap field replacement of lasers
  • Applicability spanning massive-scale Ethernet switches to optically-enabled ASICs
  • Host-controller-centric management (CMIS-style), keeping laser intelligence at the system level rather than locked inside modules
  • Multi-source design intent—reducing vendor lock-in and enabling competitive supply.

That phrase in the OIF specification—“special-purpose optically enabled ASICs”—is doing significant work. It points directly at the AI accelerator landscape, where optical connectivity is migrating from the network edge into the accelerator’s immediate neighborhood. Optical I/O chiplets bonded next to GPU or XPU packages can keep data in the optical domain for hundreds of meters without the copper penalties, but they require a reliable, standardized, field-serviceable light supply. ELSFP is that supply standard.

The window for design-win and qualification work is open now. Large-scale ELS deployments are expected to accelerate from approximately 2027 onward as CPO moves from pilot showcases to broad production deployment. The companies that get qualified in 2026-27 own the supplier relationships when the volume ramp hits.

The Energy Math That Forces Change

Energy constraints are not an abstract concern. Consider a representative 800G DR8 optical module: typical power draw around 14.2 watts. At 800 Gbps, that’s approximately 17.75 picojoules per bit—before facility overhead. At AI cluster scale, where back-end fabrics contain hundreds of thousands of optical interfaces, this accumulates rapidly into meaningful megawatts of constraint.

Next-generation near-package silicon-photonics light engines publicly claim sub-5 pJ/bit including laser power. Some CPO demonstrations cite ~4 pJ/bit. Even accounting for optimistic engineering assumptions in vendor claims, the directional gap is compelling—and it maps directly onto the IEA’s projection that data centers could consume ~945 TWh annually by 2030, with networking equipment representing up to ~5% of that load.

Efficiency gains in optical interconnect translate into real infrastructure headroom at any serious AI operator’s scale.

A Market Built Around the Laser Gap

The opportunity isn’t theoretical. LightCounting estimates the combined AI Ethernet optics and CPO market reached $16.5B in 2025, heading to $26B in 2026. The specialized co-packaged optical components segment—closer to POET’s direct addressable space—is forecast to exceed $1.3B in 2025 and grow to $2.7B by 2028, with external lasers explicitly called out as an “immediate demand” component category.

Large-scale ELS deployments are expected to accelerate from approximately 2027 onward, as the CPO transition moves from piloted showcase to broad production deployment. The window for qualification and design-win work is now.

The Competitive Landscape

ELS is not a single-vendor category. Multiple approaches are converging on the same architectural space, with different technology bets, integration levels, and go-to-market strategies.

Where POET Fits: The Manufacturing Platform Bet

POET’s thesis is that external light sources will not scale commercially if they continue to be manufactured like boutique photonics. The entire economic model of photonic packaging—manual active alignment, per-unit test, discrete micro-optic assemblies—is a per-unit cost structure that conflicts with the volume economics demanded by hyperscale AI deployments.

The POET Optical Interposer™ addresses this at the process level. By embedding waveguides and passive devices at wafer scale, POET enables active components—lasers, monitor photodiodes—to be flip-chipped using passive alignment rather than costly active alignment procedures. The result is a manufacturing flow that increasingly resembles semiconductor assembly, rather than artisanal photonics.

POET Starlight™

Multi-Wavelength Light Engine

Integrates lasers and passive components onto the POET Optical Interposer to deliver a complete light source for host-board integration. Designed to supply CW light to silicon-photonics-based optical engines—demonstrating applicability to the industry-standard ELSFP module path.

POET Blazar™

High-Power Multi-Channel Hybrid Laser

POET Blazar was demonstrated successfully at the 2026 OFC Conference in Los Angeles.

Purpose-built for co-packaged optics and high-bandwidth chip-to-chip optical links. Leverages wafer-level chip-scale technology to lower cost, improve reliability, and increase effective indium-phosphide supply efficiency versus traditional DFB-array solutions.

Both products were demonstrated publicly at OFC 2026, positioning POET at the moment when the industry is transitioning from CPO research prototypes to qualification-ready components. POET’s claim of up to 75% lower packaging cost versus competing approaches is rooted in the elimination of active alignment steps—a cost structure advantage that compounds as unit volumes scale.

“POET is powering the age of photonics with innovative products that seamlessly integrate into existing networking infrastructure and give developers the ability to build high-speed and low-latency networks using an optical solution that is highly scalable and at the desired cost structure.”

Steve Johansson · Managing Director, AI Breakthrough Awards

The Opportunity for POET

As the market for ELS products accelerates, POET expects to capture a meaningful share of revenue. For a platform company in the qualification and ramp phase of a new product cycle, the significant near-term opportunity is in positioning for scale. A company that demonstrates ELSFP-ecosystem integration, achieves independent qualification data, and scales laser-die supply relationships during the 2026-2027 window will be well-placed as CPO transitions from pilot to production across the hyperscale AI infrastructure stack.

Takeaways

For AI Hardware Manufacturers

  • Treat ELS as an architectural decision—it changes packaging, management software, and field-service workflows, not just a component swap
  • Prioritize ELSFP-aligned multi-sourcing and explicit eye-safety requirements to reduce deployment friction at scale
  • Benchmark interconnect options at system-level pJ/bit, including laser power and any DSP/retimer requirements

For Data Center Operators

  • Evaluate ELS and CPO first where optics power and faceplate density are already limiting cluster scaling
  • Model MW-level savings across your fabric, not just per-module watt savings
  • Demand clear operational semantics: redundancy groups, host-controller management, telemetry, and replace-in-place field procedures

POET’s Critical Proof Points

External light sources are not the whole CPO story—but they are the enabling layer without which optical engines cannot scale. Once lasers are standardized, serviceable, and multi-sourced, optical engines can evolve faster and integrate more tightly with the AI compute stack. POET’s optical interposer was designed for exactly this transition.

Key Numbers

$26B  AI optics market 2026E

4×  efficiency gain CPO vs. pluggables

75%  POET claimed packaging cost reduction

$2.7B  CPO components market 2028E

Disclosure: This article contains forward-looking statements and market estimates derived from third-party industry research. Market sizing figures (LightCounting, Research & Markets) are sourced from publicly available summaries and are provided for informational context only. TAM/SAM/SOM scenarios are illustrative analyses and do not constitute company guidance, revenue forecasts, or investment advice. Competitive comparisons reflect publicly available information at the time of writing and are subject to change. POET Technologies is a publicly traded company (NASDAQ: POET); readers should refer to official SEC/SEDAR filings for material financial information.

© 2026 POET Technologies Inc.

‍

Share This Article

The Laser Problem in AI Hardware and How POET Can Solve It

Published on
April 1, 2026

‍

Why AI clusters are divorcing lasers from optical engines—and how POET’s Optical Interposer is built for exactly this moment.

The past two years of artificial intelligence infrastructure build-out have clarified something uncomfortable for systems architects: the hard part is not just more compute. It’s moving more data, more often, across more chips—without turning the data center into a thermal and electrical dead end.

That pain concentrates in the interconnect. Today’s AI clusters depend on enormous numbers of optical modules—800G now, accelerating to 1.6T and beyond—because copper simply cannot deliver the reach and bandwidth that scale-up and scale-out fabrics demand. But pluggable optics at the rack faceplate carry their own scaling wall: rising per-module power, growing cooling burden, and a hard faceplate density ceiling.

This is why co-packaged optics (CPO), near-packaged optics (NPO), and optical I/O chiplets keep appearing in roadmaps. The principle is straightforward: shorten the electrical path between the SerDes and the optics by placing optical engines—modulators and detectors—directly on or beside the switch or accelerator package. Shorter traces mean lower loss, lower latency, and the removal of power-hungry retimers and DSP stages that exist largely to fight the physics of long copper channels.

But shortening the electrical path exposes a new bottleneck: the laser.

Why the Laser Has to Move

Silicon-photonics optical engines need a stable optical carrier—continuous-wave (CW) light that modulators can imprint data onto. In a conventional pluggable transceiver, the laser lives inside the module. That’s convenient when the module sits at the cool faceplate. When you move the optical engine next to a 500W+ AI accelerator, that convenience becomes a liability on two fronts.

“The pluggable external laser has the lowest temperature environment and is generally advantaged in reliability, maintainability/serviceability, and stability.”

International Photonics & Electronics Committee (IPEC), Pluggable External Laser Source Implementation Agreement

Thermal stress. High-power lasers operating continuously near extreme hotspots age faster, drift in wavelength, and fail sooner. The thermal decoupling problem is not merely engineering inconvenience—it’s a fundamental reliability math problem at million-port AI cluster scale.

Serviceability. A failed integrated laser on a co-packaged switch is not something an operator wants to remediate by pulling the entire switch ASIC out of a live pod. The replacement unit must be swappable at the faceplate, independently of the optical engines riding next to the compute die.

External Light Sources (ELS) are the industry’s answer. An ELS is a multi-wavelength laser module that lives outside the hot package—typically at the system faceplate—delivering CW light over fiber to one or more optical engines that handle modulation and detection. The logic is clean: keep optical conversion close to the ASIC, but keep the laser itself in a cooler, independently replaceable location.

Standards Are Making ELS Real

Without interoperability standards, ELS is a bespoke science project for every vendor. The Optical Internetworking Forum (OIF) has addressed this with the External Laser Small Form-Factor Pluggable (ELSFP) implementation agreement—a faceplate pluggable form factor explicitly designed for CPO systems whose optical engines carry no integrated lasers.

What ELSFP Defines

  • Blind-mate optical connectivity enabling safe, hot-swap field replacement of lasers
  • Applicability spanning massive-scale Ethernet switches to optically-enabled ASICs
  • Host-controller-centric management (CMIS-style), keeping laser intelligence at the system level rather than locked inside modules
  • Multi-source design intent—reducing vendor lock-in and enabling competitive supply.

That phrase in the OIF specification—“special-purpose optically enabled ASICs”—is doing significant work. It points directly at the AI accelerator landscape, where optical connectivity is migrating from the network edge into the accelerator’s immediate neighborhood. Optical I/O chiplets bonded next to GPU or XPU packages can keep data in the optical domain for hundreds of meters without the copper penalties, but they require a reliable, standardized, field-serviceable light supply. ELSFP is that supply standard.

The window for design-win and qualification work is open now. Large-scale ELS deployments are expected to accelerate from approximately 2027 onward as CPO moves from pilot showcases to broad production deployment. The companies that get qualified in 2026-27 own the supplier relationships when the volume ramp hits.

The Energy Math That Forces Change

Energy constraints are not an abstract concern. Consider a representative 800G DR8 optical module: typical power draw around 14.2 watts. At 800 Gbps, that’s approximately 17.75 picojoules per bit—before facility overhead. At AI cluster scale, where back-end fabrics contain hundreds of thousands of optical interfaces, this accumulates rapidly into meaningful megawatts of constraint.

Next-generation near-package silicon-photonics light engines publicly claim sub-5 pJ/bit including laser power. Some CPO demonstrations cite ~4 pJ/bit. Even accounting for optimistic engineering assumptions in vendor claims, the directional gap is compelling—and it maps directly onto the IEA’s projection that data centers could consume ~945 TWh annually by 2030, with networking equipment representing up to ~5% of that load.

Efficiency gains in optical interconnect translate into real infrastructure headroom at any serious AI operator’s scale.

A Market Built Around the Laser Gap

The opportunity isn’t theoretical. LightCounting estimates the combined AI Ethernet optics and CPO market reached $16.5B in 2025, heading to $26B in 2026. The specialized co-packaged optical components segment—closer to POET’s direct addressable space—is forecast to exceed $1.3B in 2025 and grow to $2.7B by 2028, with external lasers explicitly called out as an “immediate demand” component category.

Large-scale ELS deployments are expected to accelerate from approximately 2027 onward, as the CPO transition moves from piloted showcase to broad production deployment. The window for qualification and design-win work is now.

The Competitive Landscape

ELS is not a single-vendor category. Multiple approaches are converging on the same architectural space, with different technology bets, integration levels, and go-to-market strategies.

Where POET Fits: The Manufacturing Platform Bet

POET’s thesis is that external light sources will not scale commercially if they continue to be manufactured like boutique photonics. The entire economic model of photonic packaging—manual active alignment, per-unit test, discrete micro-optic assemblies—is a per-unit cost structure that conflicts with the volume economics demanded by hyperscale AI deployments.

The POET Optical Interposer™ addresses this at the process level. By embedding waveguides and passive devices at wafer scale, POET enables active components—lasers, monitor photodiodes—to be flip-chipped using passive alignment rather than costly active alignment procedures. The result is a manufacturing flow that increasingly resembles semiconductor assembly, rather than artisanal photonics.

POET Starlight™

Multi-Wavelength Light Engine

Integrates lasers and passive components onto the POET Optical Interposer to deliver a complete light source for host-board integration. Designed to supply CW light to silicon-photonics-based optical engines—demonstrating applicability to the industry-standard ELSFP module path.

POET Blazar™

High-Power Multi-Channel Hybrid Laser

POET Blazar was demonstrated successfully at the 2026 OFC Conference in Los Angeles.

Purpose-built for co-packaged optics and high-bandwidth chip-to-chip optical links. Leverages wafer-level chip-scale technology to lower cost, improve reliability, and increase effective indium-phosphide supply efficiency versus traditional DFB-array solutions.

Both products were demonstrated publicly at OFC 2026, positioning POET at the moment when the industry is transitioning from CPO research prototypes to qualification-ready components. POET’s claim of up to 75% lower packaging cost versus competing approaches is rooted in the elimination of active alignment steps—a cost structure advantage that compounds as unit volumes scale.

“POET is powering the age of photonics with innovative products that seamlessly integrate into existing networking infrastructure and give developers the ability to build high-speed and low-latency networks using an optical solution that is highly scalable and at the desired cost structure.”

Steve Johansson · Managing Director, AI Breakthrough Awards

The Opportunity for POET

As the market for ELS products accelerates, POET expects to capture a meaningful share of revenue. For a platform company in the qualification and ramp phase of a new product cycle, the significant near-term opportunity is in positioning for scale. A company that demonstrates ELSFP-ecosystem integration, achieves independent qualification data, and scales laser-die supply relationships during the 2026-2027 window will be well-placed as CPO transitions from pilot to production across the hyperscale AI infrastructure stack.

Takeaways

For AI Hardware Manufacturers

  • Treat ELS as an architectural decision—it changes packaging, management software, and field-service workflows, not just a component swap
  • Prioritize ELSFP-aligned multi-sourcing and explicit eye-safety requirements to reduce deployment friction at scale
  • Benchmark interconnect options at system-level pJ/bit, including laser power and any DSP/retimer requirements

For Data Center Operators

  • Evaluate ELS and CPO first where optics power and faceplate density are already limiting cluster scaling
  • Model MW-level savings across your fabric, not just per-module watt savings
  • Demand clear operational semantics: redundancy groups, host-controller management, telemetry, and replace-in-place field procedures

POET’s Critical Proof Points

External light sources are not the whole CPO story—but they are the enabling layer without which optical engines cannot scale. Once lasers are standardized, serviceable, and multi-sourced, optical engines can evolve faster and integrate more tightly with the AI compute stack. POET’s optical interposer was designed for exactly this transition.

Key Numbers

$26B  AI optics market 2026E

4×  efficiency gain CPO vs. pluggables

75%  POET claimed packaging cost reduction

$2.7B  CPO components market 2028E

Disclosure: This article contains forward-looking statements and market estimates derived from third-party industry research. Market sizing figures (LightCounting, Research & Markets) are sourced from publicly available summaries and are provided for informational context only. TAM/SAM/SOM scenarios are illustrative analyses and do not constitute company guidance, revenue forecasts, or investment advice. Competitive comparisons reflect publicly available information at the time of writing and are subject to change. POET Technologies is a publicly traded company (NASDAQ: POET); readers should refer to official SEC/SEDAR filings for material financial information.

© 2026 POET Technologies Inc.

‍