Every AI model trained in 2026 runs on silicon chips that are approaching their physical limits. The heat, the power draw, the fundamental speed of electrons crawling through copper interconnects — these are walls the industry is hitting simultaneously. The solution being built by startups, national labs, and the world's biggest chip companies uses something radically different: light.
Photonic computing replaces electrons with photons — particles of light — to move and in some cases process information. The numbers are striking: 10 to 50 times more bandwidth than copper, up to 30 times better energy efficiency for AI workloads, and near-zero heat from the transmission medium itself. In 2026, photonic components are already inside data centres on every continent. The question is no longer whether photonics will reshape computing, but how fast and how completely.
The Problem With Silicon
Silicon transistors have powered the computing revolution for sixty years by following Moore's Law — roughly doubling in density every two years. Physical reality is now catching up. When transistors are only a few atoms wide, quantum tunnelling causes electrons to leak through barriers that were supposed to stop them. Heat dissipation becomes critical: the more transistors packed into a chip, the more heat generated, requiring increasingly elaborate cooling that consumes enormous power.
For AI workloads specifically, the bottleneck has shifted from compute to communication. Training a large language model or running inference at scale requires trillions of data transfers between chips every second. The copper wires carrying those signals have limited bandwidth, generate heat, and introduce latency. This is the wall that photonics is purpose-built to break through.
What Photonic Computing Actually Does
To understand photonic computing, it helps to start with what a photon actually is: a massless particle of light that travels at approximately 300,000 km/s in a vacuum — around 200,000 km/s through optical fibre. Photons carry energy and information, generate no electrical resistance, and crucially do not interact with each other. Two beams of light passing through the same waveguide don't collide; they pass straight through.
In a conventional chip, electrons carry data through copper interconnects. In a photonic system, photons carry data through optical waveguides — tiny channels etched into silicon or lithium niobate (a crystal material that interacts especially efficiently with light). The waveguides guide photons just as copper guides electrons, but with far less energy loss and zero resistive heat.
A key technique is wavelength division multiplexing (WDM): sending multiple independent data streams simultaneously on different wavelengths of light through the same waveguide. A single optical fibre can carry dozens of wavelengths at once — each effectively its own data channel — multiplying bandwidth without adding any additional physical connections.
Where Photonics Is Already Deployed
The photonics revolution is not a future forecast. It is already underway, starting with networking rather than computation.
Data centres connect racks of servers with optical fibre transceivers that convert electrical signals from chips into light pulses for transmission. These links are faster, longer-reaching, and more energy-efficient than copper. The optical transceiver market alone is projected to exceed $12 billion by 2026, driven almost entirely by AI infrastructure demand.
NVIDIA has begun commercialising co-packaged optics (CPO) in its Spectrum-X and Quantum-X networking switches — integrating optical components directly onto the same package as the switching silicon, rather than using pluggable modules connected by copper traces. This delivers 3.5 times lower power consumption compared to traditional pluggable transceivers, with large-scale deployment rolling out across 2025 and 2026.
The Companies Building It
Lightmatter (Boston) is the company pushing photonics furthest toward compute. Its Passage M1000, unveiled at Hot Chips 2025, is a 3D photonic interposer — a layer of optical silicon that sits between AI chips, connecting them with light instead of copper. The M1000 delivers 114 terabits per second of total bandwidth across 256 optical fibres, integrating 34 chiplets into a single 4,000 mm² die complex. Lightmatter has raised $850 million at a $4.4 billion valuation, with its L200 CPO product expected to ship in 2026.
Ayar Labs, working with Alchip, demonstrated a prototype at TSMC's Open Innovation Platform in 2025 featuring eight optical engines on a single substrate delivering over 100 Tbps of scale-up bandwidth per AI accelerator. The company's CEO has stated that the on-chip optical I/O market will begin maturing between 2026 and 2028.
Q.ANT, a German startup, is taking a different approach with its NPU 2 photonic processor built on thin-film lithium niobate — a material that modulates light more efficiently than standard silicon waveguides. Q.ANT is targeting shipments in the first half of 2026, positioning the NPU 2 directly for AI inference in data centres.
Intel has maintained a silicon photonics programme since the early 2000s, with commercial transceivers already deployed at significant scale across hyperscale data centres worldwide.
PsiQuantum sits at the intersection of photonics and quantum computing: it uses photons not just for communication but as the qubits themselves. Where Google's Willow chip — covered in the Google Quantum AI 2026 roadmap — uses superconducting circuits, PsiQuantum bets on photonic qubits fabricated using standard semiconductor manufacturing.
The Hard Limits: What Photonics Cannot Do Yet
Photonics has a genuine weakness: logic. Silicon transistors switch between on and off states billions of times per second to perform computation. Photons do not naturally do this — there is no photonic transistor that is simple, cheap, and nanoscale. Building optical logic gates requires complex structures that are sensitive to temperature and mechanical misalignment.
This means that today's photonic systems are hybrid: light handles communication between chips at enormous speed and bandwidth, while silicon still performs the actual computation. Every time data crosses the boundary between the photonic and electronic domains, it costs energy — this is the conversion penalty, and eliminating it is one of the field's central engineering challenges.
There is also a precision problem. Optical waveguides and interference structures must be aligned to within nanometres. Temperature fluctuations expand materials and shift that alignment. Managing this in production data centre environments — which are warm, vibrating, and running continuously — is an engineering challenge companies are actively solving but have not yet fully cracked.
Photonic systems also excel at specific workloads — matrix multiplication, AI inference, signal processing — and are less suited to the sequential, branching logic that general-purpose CPUs handle. A photonic system that replaces your laptop processor is not on any credible roadmap for this decade.
The Realistic Timeline
The honest picture in 2026:
- Now (2025–2026): Photonics is mainstream in networking. Co-packaged optics are entering data centre switches at scale. Optical transceivers are everywhere.
- 2026–2028: Co-packaged optics across AI accelerator clusters. Photonic interposers like Lightmatter's Passage connecting AI chips. On-chip optical I/O market matures.
- 2028–2032: Hybrid photonic-electronic compute at scale. Early photonic AI inference accelerators in commercial deployment.
- 2030s and beyond: Fully photonic compute — where logic as well as data transfer happens in light — remains speculative. It may not arrive this decade.
The near-term winner is clear: AI infrastructure. The bandwidth demands of large-scale AI training and inference are precisely the problem photonics solves best. It will not replace the silicon in your laptop soon. But it is already reshaping the infrastructure that runs the AI models your devices connect to every day.
Frequently Asked Questions
Is photonic computing faster than silicon?
For data transfer, yes — photons travel through optical fibre at roughly 200,000 km/s with near-zero signal loss and no resistive heat. For raw computation (arithmetic, logic operations), silicon transistors remain faster and more practical. Today's photonic systems use light for communication and silicon for computation.
Will photonic chips replace CPUs and GPUs?
Not in the near term. The foreseeable path is hybrid systems where photonic components handle high-speed interconnects between chips while silicon transistors continue to perform computation. Fully photonic processors capable of replacing GPUs remain a research challenge with no firm commercial timeline this decade.
What companies are leading photonic computing in 2026?
Lightmatter (Passage M1000, 114 Tbps photonic interconnect), Ayar Labs (optical I/O for AI accelerators, 100 Tbps+ per chip), Q.ANT (thin-film lithium niobate photonic processor), and Intel (commercial silicon photonics transceivers) are among the key players. PsiQuantum is pursuing photonic quantum computing specifically.
What is co-packaged optics?
Co-packaged optics (CPO) integrates optical components directly onto the same package as a chip, eliminating the copper traces and pluggable transceiver modules used in older designs. This reduces power consumption by up to 3.5 times and increases bandwidth density significantly. NVIDIA is commercialising CPO in its Spectrum-X and Quantum-X networking switches.
How does photonic computing relate to quantum computing?
Some quantum computing architectures — including PsiQuantum's — use photons as qubits, the basic unit of quantum information. Photons are attractive because they maintain quantum coherence more easily than electrons in many environments. The waveguide fabrication techniques being refined for classical photonic computing also overlap significantly with those needed to build photonic quantum processors at scale.
Sources
- Lightmatter unveils Passage M1000 — World's Fastest AI Interconnect (Quantum Zeitgeist)
- Lightmatter Passage M1000 at Hot Chips 2025 (ServeTheHome)
- How photonic computing can move towards commercialization (World Economic Forum)
- Q.ANT debuts energy-efficient photonic tech in US market (Data Centre Magazine)
- Photonic chips are going mainstream — first in data-center networking, then in compute (Intelligent Living)
