Executive Technical Summary

AI infrastructure depends on large-scale data movement between compute, memory, storage, and networking systems. As AI clusters scale from thousands to potentially millions of accelerators, electrical interconnects face growing challenges in power consumption, signal integrity, thermal density, reach, and bandwidth density.

Photonics helps address these problems by replacing or supplementing high-speed electrical links with optical links. Optical communication uses photons instead of electrons to carry information through fiber, waveguides, and photonic integrated circuits.

Photonics is already central to long-distance telecommunications and data-center networking. The next step is moving photonics closer to the compute layer through silicon photonics, optical I/O, co-packaged optics, photonic chiplets, and optical interconnect fabrics.

Recent industry and technical research increasingly identifies photonics as a major scaling technology for AI data centers because of its bandwidth, energy efficiency, and scalability across multiple layers of data-center architecture. A 2026 Nature industry analysis described photonics as a transformative solution for AI data centers, specifically citing bandwidth, energy efficiency, and scalability across the AI data-center stack.

Photonics provides AI infrastructure with several engineering advantages:

The important caveat is that photonics is not a universal replacement for electronics. Electronics remains superior for dense logic, memory, control, switching, and arithmetic. Photonics is strongest where the problem is moving information, not necessarily storing or digitally processing it.

The likely future is not “photonic computers replace electronic computers.” The likely future is:

Electronic compute + photonic data movement + electronic-photonic packaging + optical networking.

Why AI Infrastructure Needs a New Data Movement Layer

AI systems are built on parallelism. Training a large model requires many processors working together, sharing parameters, gradients, activations, embeddings, weights, and intermediate results. Inference at scale also requires high-speed movement between memory, accelerators, servers, and networking systems.

The larger the AI system becomes, the more communication dominates performance.

In many AI workloads, compute units spend part of their time waiting for data. When accelerators are underfed, expensive compute hardware is not fully utilized. That means the infrastructure problem is not only “more chips.” It is also:

The core issue is that AI scaling increases the burden on interconnects.

A simplified AI cluster contains:

GPU / AI Accelerator
      ↓
Package-level interconnect
      ↓
Board-level communication
      ↓
Server network interface
      ↓
Top-of-rack switch
      ↓
Cluster switch fabric
      ↓
Storage / memory / other accelerators
      ↓
Another GPU / AI accelerator

Every layer consumes power, adds latency, creates heat, and introduces potential bottlenecks.

At small scale, electrical interconnects are sufficient. At very large scale, the physical limits of electrical data movement become more obvious.

The Electrical Interconnect Problem

Electrical interconnects move data using electrons and electromagnetic signals through conductive materials such as copper traces, cables, package substrates, printed circuit boards, and interposers.

Electronics is excellent for computation and control. But high-speed electrical data movement faces several hard engineering constraints.

1. Resistive Loss

Copper is conductive, but it is not perfect. Electrical signals lose energy as heat through resistance. As frequency rises and interconnects become denser, resistive losses become more significant.

This heat must be removed by the cooling system.

2. Skin Effect

At high frequencies, current tends to flow near the surface of a conductor. This reduces the effective conducting area and increases resistance.

For high-speed AI interconnects, skin effect contributes to signal attenuation and energy loss.

3. Dielectric Loss

Signals traveling through printed circuit boards, packages, and cables interact with dielectric materials. At high data rates, dielectric losses can become significant.

This limits reach and increases the need for equalization, retiming, and signal conditioning.

4. Crosstalk

Dense electrical traces can interfere with one another. As interconnect density increases, crosstalk becomes harder to manage.

Crosstalk degrades signal integrity and can increase error rates.

5. Electromagnetic Interference

Electrical signals are susceptible to electromagnetic interference. Shielding and careful layout can reduce the problem, but at high speeds and high densities, electromagnetic interactions become increasingly difficult.

6. Equalization and Retiming Power

As electrical links become faster and longer, the receiver often requires equalization, clock recovery, retiming, and digital signal processing to reconstruct the transmitted data.

These circuits consume power.

In other words, electrical links do not only consume power in the wire. They also consume power in the circuitry required to make the signal usable after transmission.

7. Bandwidth Density Limits

Modern AI systems need huge bandwidth through limited physical space. Package edges, board routing channels, switch faceplates, and rack-level connections all create physical bottlenecks.

If the system cannot physically route enough high-speed electrical signals, scaling becomes difficult even if compute chips keep improving.

Why Photonics Fits the AI Infrastructure Problem

Photonics fits because AI infrastructure increasingly needs to move enormous quantities of data through constrained physical space with lower energy and better reach.

Optical communication uses light instead of electrical current as the carrier.

A simplified optical interconnect works like this:

Electrical data
      ↓
Driver circuit
      ↓
Laser or optical carrier
      ↓
Modulator encodes data onto light
      ↓
Fiber or waveguide transports optical signal
      ↓
Photodetector converts light back to electrical signal
      ↓
Receiver electronics recover data

The advantage is not magic. It is physics.

Photons do not experience electrical resistance in the same way electrons do in metal conductors. Optical fibers can carry high-bandwidth signals over long distances with low loss. Optical channels can use multiple wavelengths simultaneously. Optical links are not affected by electromagnetic interference in the same way copper links are.

This is why optical communication already dominates long-distance data transmission.

The next question is: How close can optics move toward the processor?

That is the frontier of AI photonics.

Photonics vs Copper: Where Light Has the Advantage

Photonics is not always better than copper. Copper is cheap, mature, and excellent over short distances. Electronics is still dominant inside chips and packages.

But as distance, bandwidth, and density increase, optical links become more attractive.

Copper Is Strongest When:

Photonics Is Strongest When:

That is why photonics is becoming more important as AI infrastructure scales.

It is not because light is fashionable. It is because the data movement problem increasingly favors optical physics.

The Key Metric: Energy Per Bit

One of the most important metrics in AI infrastructure is energy per bit.

Energy per bit measures how much energy is required to move one bit of information.

Energy per bit = total link energy / number of bits transmitted

This is critical because AI systems move extraordinary amounts of data. Even small improvements in energy per bit can matter at data-center scale.

If a data center moves exabits of information, energy per bit becomes a major part of total power consumption.

Electrical interconnects consume energy through:

Optical interconnects consume energy through:

Photonics is not energy-free. But in the right regimes, optical links can reduce the energy required for high-bandwidth data movement, especially as distance and bandwidth scale.

This is why photonics is important for AI cooling: not because it eliminates heat, but because it can reduce energy wasted in data movement.

Cooling: Why Photonics Matters Thermally

AI data centers are increasingly power-limited and cooling-limited. More compute generates more heat. More memory bandwidth generates more heat. More networking generates more heat. More electrical signal conditioning generates more heat.

Cooling is not a side issue. It is now part of AI system architecture.

Photonics can help reduce cooling pressure in the interconnect layer.

The key point is this:

Photonics does not make AI infrastructure heat-free. It makes high-bandwidth data movement more thermally scalable.

Optical interconnects can help reduce:

However, photonic systems still generate heat from lasers, modulators, drivers, detectors, thermal tuning circuits, and control electronics.

The correct engineering view is balanced:

Photonics reduces certain thermal bottlenecks while introducing its own packaging, laser, and thermal-control challenges.

That is why electronic-photonic co-design is so important.

Bandwidth Density: Why AI Needs More Than Faster Lanes

AI infrastructure needs more than higher per-lane data rates. It needs more total bandwidth through limited physical space.

Bandwidth density measures how much data throughput can be achieved per unit area, per package edge, per board footprint, per rack unit, or per watt.

Photonics helps bandwidth density in three ways.

1. High-Speed Optical Channels

Optical links can support very high data rates per channel.

2. Wavelength-Division Multiplexing

Multiple wavelengths of light can travel through the same fiber or waveguide at the same time. This is called wavelength-division multiplexing, or WDM.

Instead of requiring one physical conductor per channel, optical systems can multiplex many channels onto one optical path.

3. Compact Optical Routing

Integrated photonics can route optical signals through waveguides on chips or packages, allowing compact optical engines, transceivers, and interconnect systems.

The result is a path toward moving more data through less physical space.

That matters because AI systems are increasingly constrained by package escape, switch radix, rack density, faceplate bandwidth, and board routing.

Wavelength-Division Multiplexing: Photonics’ Parallelism Advantage

One of the strongest scientific advantages of photonics is wavelength parallelism.

Light can be separated by wavelength. Different wavelengths can carry different streams of data simultaneously through the same optical fiber or waveguide.

This is wavelength-division multiplexing.

A simplified WDM system looks like this:

λ1 → data channel 1
λ2 → data channel 2
λ3 → data channel 3
λ4 → data channel 4
λ5 → data channel 5
All wavelengths travel through the same fiber.

This is powerful because bandwidth can scale by adding wavelengths.

Electronics generally scales bandwidth by increasing signaling rate, adding lanes, improving encoding, reducing noise, and using more complex equalization. Photonics can also increase lane speed, but it has the additional advantage of wavelength parallelism.

For AI infrastructure, WDM can support:

This does not make WDM trivial. It requires lasers, wavelength control, multiplexers, demultiplexers, thermal stability, filters, and precise optical design.

But it gives photonics a unique scaling dimension.

Optical Interconnects for AI

An optical interconnect is a communication link that uses light to move information between system components.

In AI infrastructure, optical interconnects can exist at multiple levels:

1. Data Center Interconnect

This connects data centers or large facilities over long distances.

Photonics already dominates here.

2. Cluster Interconnect

This connects groups of servers, racks, and switches inside an AI data center.

Optics is already widely used here through transceivers and fiber links.

3. Rack-Level Interconnect

This connects servers and switches within racks or between nearby racks.

Optical links become more attractive as bandwidth increases.

4. Board-Level Interconnect

This connects components inside servers or accelerator systems.

This is an emerging frontier.

5. Chip-to-Chip Interconnect

This connects processors, accelerators, memory, and chiplets.

This is one of the hardest and most important future frontiers.

6. On-Package Optical I/O

This places optical engines very close to ASICs, GPUs, or accelerators.

This is where integrated photonics and advanced packaging become critical.

Silicon Photonics for AI Infrastructure

Silicon photonics uses silicon-based semiconductor manufacturing methods to build photonic integrated circuits.

It matters for AI infrastructure because it can potentially combine:

Silicon photonics is used in optical transceivers, data-center interconnects, coherent optics, sensing, and emerging optical I/O architectures.

A major industry issue is that silicon is not an efficient native light emitter. Silicon has an indirect bandgap, which makes laser integration difficult. As a result, silicon photonic systems often use external lasers, hybrid III-V integration, heterogeneous integration, or co-packaged light sources.

That is why silicon photonics is not only a photonic design challenge. It is a packaging, materials, and manufacturing challenge.

A 2026 Semiconductor Engineering analysis notes that photonic interconnects can increase bandwidth density while reducing power consumption, but also highlights material compatibility, thermal, and mechanical stress challenges.

This is the realistic view: silicon photonics is powerful, but commercialization depends on solving manufacturing and packaging problems.

Co-Packaged Optics: Moving Light Closer to Compute

Co-packaged optics, or CPO, is one of the most important ideas in AI photonics.

Traditional optical networking often uses pluggable optical modules at the front panel of a switch. Electrical signals travel from the switch ASIC across the board to the pluggable module, where they are converted into optical signals.

As bandwidth rises, that electrical path becomes more power-hungry and difficult.

Co-packaged optics moves optical engines closer to the switch ASIC or compute package.

The goal is to reduce the distance high-speed electrical signals must travel before optical conversion.

Traditional Pluggable Optics

ASIC → long electrical board trace → pluggable optical module → fiber

Co-Packaged Optics

ASIC → short electrical connection → optical engine near package → fiber

The advantage is that shorter electrical paths can reduce signal loss, lower power consumption, and improve bandwidth density.

NVIDIA has publicly described co-packaged optics as a way to improve power efficiency and resiliency for large-scale AI factories. Yole Group has also described co-packaged optics as an emerging critical technology for addressing AI-driven bandwidth and energy challenges.

This does not mean CPO is easy. It introduces challenges in:

But the direction is clear: as AI systems scale, optics is moving closer to compute and switching silicon.

Pluggable Optics vs Co-Packaged Optics

Both architectures matter.

Pluggable Optics

Advantages:

Limitations:

Co-Packaged Optics

Advantages:

Limitations:

The likely future is not instant replacement. It is a transition where pluggable optics, linear-drive optics, near-packaged optics, and co-packaged optics coexist depending on the system requirement.

Optical I/O: The Next Step Toward Chip-Level Photonics

Optical I/O means using optical links for input/output near chips, packages, or accelerator systems.

This matters because AI accelerators need to move enormous amounts of data off-chip. Package-level and board-level I/O are becoming bottlenecks.

Electrical I/O faces limits in:

Optical I/O offers a way to move data with high bandwidth and lower distance-related loss.

The long-term direction is:

GPU / AI accelerator
      ↓
short electrical path
      ↓
photonic I/O engine
      ↓
fiber or optical waveguide
      ↓
remote accelerator / switch / memory system

This could help enable:

Research and industry development are moving toward tighter integration of electronics and photonics. Columbia Engineering reported a 3D photonic-electronic platform in 2025 aimed at high bandwidth density and energy efficiency for next-generation AI hardware.

Photonic Chiplets and Advanced Packaging

A chiplet is a small functional die integrated with other dies in an advanced package. The semiconductor industry is increasingly moving toward chiplet architectures because monolithic scaling is becoming more difficult and expensive.

Photonic chiplets could provide optical communication functions inside advanced packages.

A future AI package might include:

The photonic chiplet would handle high-bandwidth optical communication while electronic chiplets handle computation and control.

This is one of the most important long-term possibilities for AI infrastructure.

Optical Switching and AI Network Fabrics

Today’s AI networks often rely heavily on electronic switching. Optical signals are converted into electrical signals, switched electronically, and sometimes converted back into optical signals.

That optical-electrical-optical conversion can add power, latency, and complexity.

Optical switching aims to route light more directly.

Potential advantages include:

However, optical switching has challenges:

Optical switching may become important in AI-scale networks, but it is not a simple replacement for electronic switching. It will likely appear in specific layers where optical circuit switching or hybrid packet/circuit architectures make sense.

Photonics for AI Infrastructure vs Photonic Computing

This distinction is extremely important.

Photonics for AI Infrastructure

This means using light to move data in and around AI systems.

Examples:

This is already commercially important and likely to grow rapidly.

Photonic Computing

This means using light to perform computation itself.

Examples:

Photonic computing is promising, especially for certain linear algebra and signal-processing operations, but it has different challenges:

The near-term AI infrastructure opportunity is stronger and more immediate than fully replacing digital electronic computation.

In simple terms:

Photonics will likely enter AI first as a data movement technology, then selectively as a computing acceleration technology.

That is the serious, non-hype version.

Why Photonics Can Beat Electronics in AI Data Movement

Photonics has strong advantages in specific regimes. The key is to compare by function, not by ideology.

1. Speed and Bandwidth

Optical carriers operate at extremely high frequencies. More importantly, photonics supports wavelength parallelism. Multiple wavelengths can carry independent data streams through the same physical channel.

This makes photonics extremely strong for high-bandwidth communication.

2. Distance Scaling

Electrical links degrade rapidly at high speeds over distance. Optical links can maintain signal integrity over longer distances with lower loss.

That is why fiber is used for long-distance communication and why optics becomes increasingly attractive as AI clusters scale across racks and facilities.

3. Lower Transmission Loss

Optical channels can have lower distance-related transmission loss than copper links in high-speed regimes. This helps reduce the need for heavy electrical equalization and signal conditioning.

4. Better Electromagnetic Isolation

Photons are electrically neutral. Optical fibers do not radiate or pick up electromagnetic interference like copper conductors.

This improves signal integrity and reduces crosstalk problems.

5. Wavelength Multiplexing

WDM gives photonics a unique parallel scaling mechanism.

A single fiber can carry many wavelengths. Each wavelength can carry high-speed data. This is a powerful bandwidth multiplier.

6. Cooling Advantage in Interconnects

Photonics reduces some heat associated with electrical data movement. This matters when interconnect power becomes a large part of system power.

7. Physical Infrastructure Compatibility

Optical fiber is already widely deployed in telecommunications and data centers. Photonics builds on a mature communication foundation.

Where Electronics Still Wins

A credible technical page should say this clearly:

Photonics is not superior everywhere.

Electronics still wins in:

Photons are excellent for movement and measurement. Electrons are excellent for logic and storage.

That is why the future is hybrid.

The AI Data Center Stack: Where Photonics Adds Value

Facility Layer

At the facility level, photonics supports high-capacity interconnection between data centers, campuses, and cloud regions.

Cluster Layer

At the cluster level, optics connects racks, switches, and groups of servers.

Rack Layer

At the rack layer, optical links can support high-bandwidth communication between AI servers and top-of-rack switches.

Server Layer

At the server layer, optics may increasingly support high-speed connections between accelerators and network interfaces.

Package Layer

At the package layer, optical I/O and photonic chiplets may eventually connect compute dies, switch dies, and accelerator packages.

Chip Layer

Inside the chip, electronics still dominates. But photonic structures may eventually play a role in specialized clocking, optical I/O, or certain analog acceleration functions.

The key architectural direction is that photonics moves closer and closer to the compute engine as bandwidth and power pressures increase.

Practical Example: Why a GPU Cluster Needs Photonics

Imagine an AI training cluster with thousands of accelerators.

Each accelerator performs matrix operations, but it must constantly exchange data with other accelerators. If communication is slow, the accelerators wait. If communication consumes too much power, the system becomes power-limited. If communication creates too much heat, the data center becomes cooling-limited. If the network cannot scale, the cluster cannot efficiently train larger models.

Photonics helps by increasing the capacity and efficiency of the communication layer.

It does not make the GPU itself magically faster. Instead, it helps the whole AI system operate more efficiently by reducing communication bottlenecks.

This is why photonics is infrastructure-level technology.

It improves the system around the compute.

Practical Example: Why Co-Packaged Optics Matters for AI Switches

AI clusters require large switch fabrics. As switch ASIC bandwidth increases, the electrical path from the ASIC to front-panel pluggable optics becomes more difficult.

At very high bandwidths, that electrical path consumes more power and becomes harder to equalize.

Co-packaged optics moves optical engines closer to the ASIC, reducing the electrical distance.

This can improve:

NVIDIA’s 2025 Spectrum-X Photonics announcement claimed major improvements in power efficiency, signal integrity, resiliency, and deployment compared with traditional methods for large-scale AI networking. Claims from vendors should always be interpreted in context, but the direction supports the broader industry trend: AI networking is pushing optics closer to the switching silicon.

The Role of Lasers in AI Photonics

Lasers are essential because photonic systems need optical carriers.

In optical communication, the laser provides the light. Modulators encode data onto that light. Detectors recover the signal.

Laser architecture is one of the most important design choices.

Options include:

Laser reliability, efficiency, wavelength stability, thermal behavior, and coupling efficiency all affect system performance.

One reason silicon photonics is challenging is that silicon does not naturally make efficient lasers. This forces system designers to solve laser integration through other materials and packaging strategies.

Modulators: Encoding AI Data Onto Light

A modulator converts electrical data into optical data.

It changes the optical carrier in response to an electrical signal.

Common modulation approaches include:

For AI infrastructure, modulator performance matters because it affects:

A high-performance optical link requires efficient modulators that can operate at high speeds with low energy and acceptable loss.

Photodetectors: Converting Light Back Into Electrical Data

A photodetector receives optical signals and converts them into electrical current.

Detector performance affects:

Common detector materials include germanium, indium phosphide, and other III-V semiconductors. In silicon photonics, germanium-on-silicon detectors are widely used because germanium can absorb telecom-band light more effectively than silicon.

The detector is where the optical signal re-enters the electronic system.

This is why electronic-photonic integration is essential. Photonics does not operate in isolation. It must interface cleanly with electronics.

Link Budget: The System-Level Reality

Every optical link has a link budget.

The link budget accounts for gains and losses from source to receiver.

Losses may come from:

A simplified link budget must ensure:

Received optical power > receiver sensitivity + system margin

If the link budget is poor, the system may need more laser power, better detectors, lower-loss components, or stronger signal processing.

This matters for AI infrastructure because power is precious. A photonic system with poor coupling and high loss may lose its energy advantage.

That is why packaging and coupling are not minor details. They are central to whether AI photonics works commercially.

Packaging: The Hard Part Nobody Should Ignore

In photonics, packaging is often harder than the chip.

A photonic chip can perform well in a lab, but commercial deployment requires stable, manufacturable, reliable packaging.

Photonics packaging must manage:

For AI infrastructure, packaging is especially difficult because systems require high density, high reliability, low cost, and field serviceability.

This is one reason adoption takes time. The physics is strong, but the engineering must be solved at scale.

Thermal Drift in Photonic Systems

Photonics can reduce some thermal problems, but photonic devices can also be temperature-sensitive.

For example, ring resonators and interferometers can shift their behavior as temperature changes. A small temperature shift can move a resonance wavelength, affecting filtering or modulation.

This creates a control problem.

Thermal drift may require:

In dense AI systems, thermal drift is a serious issue because data centers are already thermally stressed.

This is why “photonics solves cooling” is too simplistic.

A better statement is:

Photonics can reduce interconnect-related heat, but dense photonic systems still require careful thermal engineering.

Reliability and Serviceability

AI infrastructure requires high uptime.

A photonic system must be reliable under real operating conditions, not only impressive in a laboratory.

Important reliability questions include:

These questions determine whether a technology becomes infrastructure.

Security Benefits of Photonics in AI Infrastructure

Photonics can contribute to security, but again, precision matters.

Fiber Physical Security

Optical fiber can be monitored for disturbances, bending, tapping, or intrusion. Fiber sensing can detect changes along the fiber path.

Electromagnetic Isolation

Optical fiber does not radiate electromagnetic signals like copper conductors. This can reduce some electromagnetic leakage risks and improve isolation.

Quantum Communication

Photonics is essential to quantum communication and quantum key distribution. However, QKD is not a universal replacement for cryptography. It has practical limitations related to distance, authentication, trusted nodes, hardware security, cost, and integration.

Secure Timing and Sensing

Photonic systems can support secure time transfer, precision sensing, and infrastructure monitoring.

The strongest near-term security benefit for AI infrastructure is probably not “quantum encryption everywhere.” It is physical-layer integrity, electromagnetic isolation, high-quality fiber infrastructure, and eventually specialized quantum-secure links where justified.

The Engineering Benefits of Photonics Over Other Technologies

Photonics competes and cooperates with several technologies: copper interconnects, wireless links, electronic switching, and conventional pluggable optics.

Photonics vs Copper

Photonics wins in high-bandwidth, longer-reach, high-density, signal-integrity-sensitive environments.

Copper wins in short-reach, low-cost, mature, highly integrated environments.

Photonics vs Wireless

Photonics wins in bandwidth, latency, reliability, security, and controlled infrastructure.

Wireless wins in mobility and deployment flexibility.

AI data centers need deterministic, high-capacity, low-latency communication. That strongly favors fiber and photonic links over wireless.

Co-Packaged Optics vs Pluggable Optics

CPO can improve power and bandwidth density by moving optics closer to ASICs.

Pluggable optics remains easier to service and more mature.

The transition will be workload- and architecture-dependent.

Photonics vs Purely Electronic Scaling

Electronic scaling is still essential, but interconnect power and bandwidth are increasingly limiting. Photonics offers a different scaling path: use light for data movement instead of forcing copper and electrical I/O to handle every layer of communication.

What Photonics Does Not Solve

To stay credible, it is important to explain what photonics does not solve.

Photonics does not automatically solve:

Photonics is a powerful infrastructure technology, but it must be integrated into a full system architecture.

The correct claim is not:

Photonics solves AI.

The correct claim is:

Photonics helps solve one of AI infrastructure’s hardest problems: scalable, high-bandwidth, energy-efficient data movement.

That is a much stronger and more credible position.

The Future of AI Infrastructure Is Electronic-Photonic

AI infrastructure is moving toward systems where electronics and photonics are co-designed.

Future systems may include:

The architecture may evolve from:

Compute chip → electrical board → pluggable optics → fiber

Toward:

Compute chip → photonic I/O → optical fabric → compute/memory/network resource

That transition is technically difficult, but it is the direction the physics and infrastructure pressures point toward.

QCLS Perspective: Why Photonics Matters for AI

AI is not only a software revolution. It is an infrastructure revolution.

The next generation of AI depends on the ability to move data faster, farther, denser, and more efficiently. Compute chips will continue to improve, but the systems around them must also evolve.

Photonics matters because it addresses the interconnect problem at the physical layer.

Light can carry enormous bandwidth.
Light can travel through fiber with low loss.
Light can support many wavelengths in parallel.
Light avoids many electrical signal-integrity problems.
Light can reduce data-movement energy in the right regimes.
Light can help AI infrastructure scale beyond what copper alone can support.

This is why photonics is becoming foundational to AI data centers, optical interconnects, co-packaged optics, silicon photonics, and next-generation computing infrastructure.

The serious version is simple:

AI needs compute. But compute needs communication. Photonics is how communication scales.