For decades the dream of "optical computing" stayed in the journal. A chip that does math with light instead of electricity promises bandwidth and energy advantages that silicon hits a wall on — but turning a physics paper into a shipped product has defeated dozens of startups. One Chinese company just crossed that line in public.
Lightelligence (曦智科技) began as an MIT spinout and now bills itself as a "photonic-electronic hybrid compute" provider. Its April 2026 listing on the Hong Kong Stock Exchange (ticker 1879) makes it one of the few optical-computing firms whose financials are now auditable.
From a Nature paper to a product
The origin story is unusually clean. In 2017, MIT PhD Shen Yichen (沈亦晨) was lead author on a Nature Photonics paper, "Deep learning with coherent nanophotonic circuits," which argued that matrix operations — the heart of neural networks — could run on light. That same year he co-founded Lightelligence in Boston with MIT professor Marin Soljačić and others, later moving the headquarters to Shanghai. Early backers included Tencent and Sequoia.
The product arc:
- 2019 — a first optical AI accelerator demo (COMET) that recognized handwritten digits.
- 2021 — PACE (Photonic Arithmetic Computing Engine), an integrated photonic-electronic platform with more than 10,000 photonic devices and a 64×64 matrix.
- 2023 — Hummingbird, an on-chip optical network (oNOC) processor for domain-specific AI workloads.
- 2025 — PACE 2 (天枢), described by the company as the first optically accelerated card to run complex commercial AI models, with a 128×128 photonic matrix.
- 2026 — PACE 3 in development, targeting a 256×256 matrix aimed at low-latency large model (大模型) inference.
What the company actually sells
Lightelligence organizes its tech around three pillars: optical matrix compute (oMAC), on-chip optical network (oNOC), and chip-to-chip optical network (oNET). The first attacks the dense matrix multiplication that dominates inference; the second and third move data between and inside chips using light, addressing the "bandwidth wall" that power-hungry electrical links hit.
The IPO prospectus gives a rare honest financial picture for a deep-tech optical company: 2025 revenue of about CN¥106.37 million (≈ US$15M / HK$115M) against a net loss of CN¥1.34 billion — typical for a pre-scale hardware firm investing in fabs, packaging, and software. The point for observers is not profit but proof that someone is paying for optical compute cards and systems at all.
The company also pushes an optical-interconnect supernode, LightSphereX, which uses optical switching to link GPUs — a separate but complementary bet that AI clusters, not just single cards, need light to keep scaling. A 128-card commercial version was shown in early 2026 with partners including Shanghai Electric and ZTE.
Why photons, why now
Electronic chips are brilliant at logic and memory but inefficient at shuttling huge matrices of multiply-accumulate operations. Light can, in principle, do those in parallel with far lower transport loss. The catch is that a usable AI system cannot be all-optical: electronics still handle storage, control, and conversion. Lightelligence's hybrid design accepts that — photons for the heavy linear algebra, electrons for everything else.
That hybrid framing is where the commercial case lives. Rather than asking customers to rewrite their entire stack, the company packages optical acceleration as a card or an appliance (the "Tianshu Light Cube" edge system) that drops into existing pipelines. Its first long-running real-world deployment is almost comically mundane — an office access-control camera doing object detection — but that is exactly the kind of "boring, always-on" workload where low latency and low power matter.
The same low-power profile is attractive for humanoid (人形机器人) edge inference and even for always-on brain-computer interface (脑机接口) front-ends, where heat and battery life, not peak teraflops, decide what ships.
The hype vs the hardware
Optical computing has a long history of impressive bench results that never reached volume. Lightelligence's edge is that it is now a public company with disclosed revenue and a product cadence (64×64 → 128×128 → 256×256). Skeptics will note that "matrix size" is a vanity metric if the surrounding system — drivers, calibration, yield, software — does not hold up under real workloads. NVIDIA's Tensor Cores and Google's TPUs stabilized at 64×64 and 256×256 for years precisely because system maturity, not raw matrix size, wins deployments.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, photonics will earn its place by cutting inference energy, not by beating GPUs on a benchmark.
Honest limitations
- PACE 2's "128×128" and PACE 3's "256×256" are company-stated specifications, not independently benchmarked against electronic accelerators.
- Revenue is small and the firm is deeply loss-making; commercial traction is early, not proven at scale.
- Optical compute still depends on electronic control and conversion — it complements, not replaces, conventional AI silicon.
- The access-control demo proves stability, not superiority; real large-model inference performance for PACE 3 remains to be verified in deployment.
What readers can do now
- Judge by shipped systems, not matrix size — watch for third-party benchmarks of PACE 2/3 under real inference loads.
- Separate compute from interconnect — LightSphereX (optical GPU links) may commercialize faster than optical compute itself.
- Follow the energy argument — the durable thesis is watts-per-inference, the metric where photons genuinely differ from electrons.