Photonic Computing

Photonic computing is the use of light — laser light guided through microscopic waveguides etched into silicon — to carry data between chips and, in its more ambitious form, to perform computation itself. Because photons travel without resistive heating and can be multiplexed at many wavelengths through one waveguide, photonic hardware promises far higher interconnect bandwidth and lower energy per bit than the copper wiring it replaces, targeting the bottleneck that now limits large AI systems.^[1][2]

History and Background

Optical computing was researched heavily from the 1960s, then largely shelved when electronic transistors kept scaling. What revived it was not logic but packaging and interconnect: as AI models grew, the cost of moving data between accelerators and between chips and memory — the so-called memory wall and the interconnect wall — began to dominate total power use, and silicon photonics had matured through the telecom industry into a manufacturable CMOS-compatible process.^[2]

Lightmatter, spun out of MIT research in 2017, became the sector's bellwether: a US$400 million Series D in October 2024 at a US$4.4 billion valuation brought total funding to US$850 million, with T. Rowe Price leading and Fidelity and Alphabet's GV participating.^[3] In March–April 2025 the company released an optical interposer (manufactured by GlobalFoundries) and a chiplet designed to be packaged directly with AI accelerators, ahead of its Passage optical interconnect system; its CEO has claimed the technology can cut data-centre energy used for a given workload to between one-sixth and one-twentieth of conventional levels.^[1][2] A wider investment wave followed — Celestial AI raised US$250 million in early 2025 for optical fabric — while established players pursued co-packaged optics on their own roadmaps.

Key Concepts and Technology

Photonic systems are usually described in two tiers. Optical interconnect replaces electrical SerDes links between chips, packages and racks: electrical signals are converted to modulated laser light by microring resonators, multiplexed across wavelengths, routed through waveguides or fibres, then converted back by photodetectors. This is the near-term, shipping category, and it is what Lightmatter's Passage interposer does — it sits under chips in a package and connects neighbouring dies with light paths instead of copper.^[2]

Optical computation goes further, performing matrix multiplications — the workload that dominates neural-network inference — directly in the optical domain: modulators encode inputs onto light, passive optical components weight them, and photodetectors sum the results, so an entire multiply-accumulate happens as light propagates, in effect at the speed of light. The trade-offs are real. Photonic arithmetic is analog, so it suffers from thermal drift, fabrication variation and limited precision; laser sources must be supplied and stabilised; and today's devices are best at the linear algebra of inference rather than the nonlinear, high-precision operations of training. Most commercial designs therefore target inference acceleration and interconnect, not general-purpose replacement of GPUs.^[1]

The strategic appeal is thermodynamic. Data centres for AI are now limited by power delivery and heat removal; every bit that never has to be converted back into electricity inside a copper trace is energy not turned into heat.

Applications and Impact

Near-term deployments are in AI cluster fabrics — linking thousands of accelerators so that a single model can be sharded across them without the network becoming the rate limiter — and in co-packaged optics for switches and servers. Longer-term, photonic matrix engines are being evaluated for data-centre inference, where throughput per watt rather than peak FLOPS decides operating cost. Analysts broadly treat photonics as an inevitable direction for high-performance interconnect, with timing (not principle) being the open question.

>See Also

• NVIDIA NVLink
• High Bandwidth Memory
• Liquid Cooling
• National Semiconductor Strategy
• Johor Data Centres