Photonic chips how computing with light could change data centres

[Megan63]

TL;DR

• Photonic chips use light rather than electrical signals to move or process information, promising higher bandwidth and lower energy per bit
• The most immediate impact is in data centre networking: co-packaged optics place optical connections right next to switch and processor chips, cutting power and boosting bandwidth
• NVIDIA announced silicon photonics switches with co-packaged optics in March 2025, and companies such as Lightmatter and Ayar Labs are building optical interconnects for AI systems
• Photonic computing, using light to perform calculations, has advanced in research, including a photonic processor running neural networks described in Nature in 2025, but general purpose optical computers remain distant
• Photonics also underpins quantum networking and some quantum computers, linking this field to Q-Day topics

Why light

Modern computing runs on electrons moving through copper wires and transistors. That approach has served us extraordinarily well, but it faces growing limits. As data rates rise, electrical signals over copper consume more power, generate more heat and travel shorter distances before degrading. In large AI clusters, moving data between chips can consume as much energy as computing with it

Light offers a way around some of these limits. Optical signals can carry enormous amounts of data over long distances with little loss, and many wavelengths of light can share a single fibre without interfering. Telecommunications networks switched to fibre optics decades ago for exactly these reasons. The question now is how close to the chip that light can go

Physics explains part of the appeal. Photons do not interact with each other easily, so many signals can share the same waveguide at different wavelengths without interference, a technique called wavelength division multiplexing. Electrical signals in copper suffer from resistance and capacitance effects that worsen at higher frequencies. Those properties give light a natural advantage for moving lots of data

Photonic chips, built using techniques related to conventional semiconductor manufacturing, integrate optical components such as waveguides, modulators and detectors onto silicon or other materials. They can convert electrical signals to light and back, route light across a chip and, in some designs, perform calculations with light. This thread explains the main approaches, recent progress and what they could mean for data centres

The data centre bandwidth problem

AI training and inference rely on thousands of GPUs or accelerators working together, constantly exchanging data. The speed of these connections often limits how efficiently a cluster can run. Within a server, copper connections work well over short distances, but between racks and across large clusters, optical links are already standard. Those links typically use pluggable optical transceivers, modules inserted into the front of switches

Pluggable transceivers have a drawback. The electrical signal must travel from the switch chip across the circuit board to the module before conversion to light, and that journey consumes significant power at today's speeds. In large AI data centres with hundreds of thousands of links, transceiver power adds up to a substantial share of networking energy. As speeds double again, the problem worsens

Copper still has a role. Within racks, short copper cables remain cheaper and simpler, and some AI systems use extensive copper connections between closely packed accelerators. The boundary between copper and optics keeps moving inwards as speeds rise. Engineers choose based on distance, power, cost and reliability

Co-packaged optics

Co-packaged optics address this by placing the optical engines in the same package as the switch or processor chip, right next to it. Electrical signals travel only millimetres before converting to light, saving power and enabling far more bandwidth from the chip. The concept has been researched for years, and 2025 and 2026 brought major commercial steps. Many in the industry see it as essential for the next generations of AI infrastructure

In March 2025, at its GTC conference, NVIDIA announced Spectrum-X and Quantum-X silicon photonics networking switches using co-packaged optics, saying they would substantially reduce power and improve resilience compared with pluggable transceivers. NVIDIA said the Quantum-X InfiniBand switches would become available first, followed by Spectrum-X Ethernet versions. Broadcom and other networking companies have also developed co-packaged optics switches. The technology is moving from labs into production data centres

Startups are pushing optical interconnects further. Ayar Labs develops optical input and output chiplets that let processors communicate directly using light. Lightmatter builds Passage, a photonic interconnect platform designed to connect large numbers of chips with very high bandwidth, and has announced collaborations with major chip companies including joining NVIDIA's NVLink Fusion ecosystem. These approaches aim to make light the default way chips talk to each other

Memory is another target. AI accelerators are often limited by how quickly they can fetch data from memory, and optical links could connect processors to larger pools of memory across a rack. This approach, sometimes called memory disaggregation, could make clusters more flexible. It remains at an earlier stage than optical networking

Photonic computing

Beyond moving data, light can also perform certain calculations. Many AI workloads boil down to matrix multiplications, and optical systems can perform these by passing light through arrangements of modulators and interferometers, where interference effectively computes weighted sums. In principle, such operations can be very fast and energy efficient because light propagates without resistance. Research on optical neural networks has grown rapidly

In April 2025, Lightmatter researchers published a paper in Nature describing a photonic processor that ran neural network models, including image classification and language tasks, with accuracy close to conventional electronic hardware. Other groups have published optical computing results in leading journals as well. These demonstrations show photonic computing moving beyond toy problems. They also reveal the remaining hurdles

Those hurdles are significant. Optical computing is analogue, so noise and precision limit accuracy, and converting between electrical and optical domains costs energy that can erase gains. Nonlinear operations, memory and control still rely largely on electronics. Most experts expect hybrid systems, with photonics handling specific operations, rather than fully optical computers in the near term

Some researchers explore neuromorphic and photonic ideas together, using light to emulate neurons and synapses. Others pursue optical approaches to specific tasks such as Fourier transforms or signal processing, where light's natural behaviour matches the mathematics. These niche applications may arrive sooner than general accelerators. They show the breadth of the field

Manufacturing and materials

Silicon photonics uses manufacturing processes compatible with standard chip fabs, which helps with cost and scale. Silicon is excellent for waveguides and modulators but poor at emitting light, so lasers are often made from other materials such as indium phosphide and attached or integrated separately. Thin film lithium niobate has emerged as a promising material for very fast, efficient modulators. Each material choice brings trade offs in performance, cost and manufacturability

Packaging is a major challenge. Aligning optical fibres and lasers with microscopic precision, managing heat and ensuring reliability over years of operation are demanding engineering problems. Advances in packaging are a large part of why co-packaged optics became commercially viable. Foundries and packaging companies are investing heavily

Reliability and serviceability matter to data centre operators. Pluggable transceivers can be swapped easily if they fail, while co-packaged optics are integrated with expensive chips. Designs often use external laser sources that can be replaced separately to address this. Operators will watch real world failure rates closely

Supply chains are expanding. Major foundries offer silicon photonics processes, and packaging companies are building capacity for co-packaged optics. Investment has increased sharply as AI demand grows. That scale should help reduce costs over time

Energy and sustainability

Energy efficiency is the main motivation. Data centres' electricity use is rising quickly with AI, and the International Energy Agency has projected that data centre electricity demand could double by 2030. Networking is a meaningful share of that consumption, and co-packaged optics can cut networking power substantially. Even modest percentage savings add up across millions of links

Photonic computing could, in principle, offer larger savings for specific workloads if accuracy and conversion overheads can be managed. That remains to be proven at scale. For now, interconnects are where photonics delivers clear, measurable benefits. Claims about dramatic AI energy savings from optical computing deserve careful scrutiny

Photonics and quantum technology

Photonics is central to several quantum technologies discussed on this forum. Quantum networks send entangled photons through fibre, and quantum key distribution uses photons to establish keys. Photonic quantum computers, pursued by companies such as PsiQuantum and Xanadu, encode qubits in light. Many other quantum platforms use lasers and optical components for control and readout

Advances in integrated photonics therefore benefit quantum computing too. Better waveguides, modulators, detectors and packaging help build larger, more stable quantum systems and interconnect quantum modules. Photonic interconnects may eventually link quantum processors into larger machines. The overlap between classical and quantum photonics is substantial

What to watch

• Deployment and real world reliability of co-packaged optics switches in AI data centres
• Optical interconnects connecting processors and memory directly
• Peer reviewed photonic computing results on realistic AI workloads with full energy accounting
• Progress in integrated lasers and new materials such as thin film lithium niobate
• Standards for co-packaged optics to enable multi vendor ecosystems
• Cross over between photonic data centre technology and quantum networking

Closing thoughts

Light is moving steadily closer to the heart of computing. Co-packaged optics and optical interconnects are entering production to tackle the bandwidth and power challenges of AI data centres, while photonic computing research shows intriguing possibilities. The near term story is about moving data more efficiently rather than replacing electronic processors. Over the longer term, hybrid electronic and photonic systems could reshape how we build computers

If you work in optical networking, chip packaging or photonics research, please share your perspective below. Practical insights on deployment, reliability and costs would be especially valuable. Questions from curious newcomers are always welcome


Key references for citation

• NVIDIA, Spectrum-X and Quantum-X silicon photonics switches announcement, GTC, March 2025
• Lightmatter, photonic processor paper, Nature, April 2025
• International Energy Agency, Energy and AI report, April 2025, and 2026 update
• Ayar Labs and Lightmatter product documentation for optical interconnects