Dashboard
Silicon Photonics & Photonic AI investment tracker
Photonic AI — Current Development Stages
3 parallel layers · TRL 1–9Layer 1 — Interconnect
1.6T transceivers entering volume deployment at hyperscalers. CPO at 51.2 Tb/s switches now standard in new datacenter builds. NVIDIA, Google, Amazon, and Microsoft committed to optical I/O-first architectures through 2028 capex cycles.
NVIDIA Spectrum-X Co-Packaged Optics launched (TSMC, Coherent, Corning, Lumentum consortium). Direct photonic connection to chipset — 3.5× electrical efficiency, 4× fewer lasers, 63× better signal integrity.
Lightmatter Passage L200/L200X entering customer chip integration in 2026 via GlobalFoundries, ASE, and Amkor; production systems targeted for 2027. Celestial AI photonic fabric in customer evaluation. Photonic interposer connecting GPU clusters at rack scale — targeting 1 Pb/s aggregate bandwidth.
Board-to-board optical inside servers — CPU/GPU boards connected directly via fiber, eliminating electrical signaling at board level. Si₃N₄ optical backplanes targeted.
Layer 2 — Photonic Switch
ASIC + CPO switches now at third-generation CPO: Broadcom shipped Tomahawk 6 – Davisson (102.4 Tb/s CPO) in October 2025, doubling the bandwidth of Tomahawk 5 (51.2 Tb/s). Marvell Teralynx 10 remains in production deployment. CPO is now economically preferred over pluggables at hyperscale density. AMD acquired Enosemi (May 2025) to accelerate photonic interconnect manufacturing.
IPRONICS ONE and Lumentum fully photonic switch PICs sampling to Tier 1 OEMs. Photon routing instead of electron routing — sub-nanosecond switching latency demonstrated in lab.
Photonic AI switch becomes standard in hyperscale spine/leaf architecture. AI-optimised optical flow steering — major reduction in inter-server latency and network power consumption.
Layer 3 — Photonic Processor
MIT executes a full DNN entirely in the optical domain (Nature Photonics). First single chip running a deep neural network purely in light with no mid-layer electronic conversion — the most important proof-of-concept of the decade.
Taichi chip — 160 TOPS/W (Science 2024). Large-scale photonic chiplet achieves 160 TOPS/W, 50–160× more efficient than current GPUs. Runs image recognition. Lab chip with peer-reviewed performance.
LightIn — programmable photonic processor (arXiv). 40 MZI programmable unit cells, 3.8×3 mm². Runs matrix multiply, image recognition, channel switching, and PUF security on one chip.
64×64 photonic accelerator — 1 GHz MAC, 7.61-bit precision (Nature). 65nm SiPh + 28nm CMOS co-packaged, 16,000+ photonic components. Runs max-cut optimization.
Celestial AI in A/B customer tape-outs. No production silicon yet — still pre-revenue. EO polymer modulators (LWLG) and TFLN platforms at pilot-scale validation for 200Gbaud+ applications.
First production photonic AI accelerators for inference-specific niche workloads. Lightmatter Passage and Celestial AI targeting early revenue at specialized HPC/AI deployments.
Photonic TPU in hyperscale AI datacenters — the ultimate goal. Silicon-photonic AI accelerators standard for bandwidth-bound training and inference at scale.
Technical challenges
Precision limited to 7–8 bit
Versus GPU 16–32 bit. Sufficient for inference workloads, not yet sufficient for large LLM training.
Noise accumulation
Builds across multiple waveguide layers — requires continuous calibration that is difficult to automate at scale.
Analog-digital conversion overhead
Each optical-to-electrical conversion consumes power, eroding the efficiency advantage. Many benchmark figures measure only the optical portion, excluding ADC, laser source, and electronic control.
Upcoming Catalysts
COHR Q4 FY2026 Earnings
800G/1.6T transceiver ramp and InP vertical integration margin update
MRVL Q2 FY2027 Earnings
Teralynx 10 CPO volume shipments and Celestial AI custom silicon progress
AVGO Q3 FY2026 Earnings
AI ASIC revenue cadence, CPO deployment volumes, and 1.6T switch pipeline update
ECOC 2026 Conference
52nd European Conference on Optical Communications, Málaga, Spain — 1.6T/3.2T standards, TFLN commercialization updates
1.6T Transceiver Volume Ramp
Hyperscaler volume deployments of 1.6T transceivers; key revenue inflection for optical manufacturers
LITE Q1 FY2027 Earnings
ROADMs and pump laser shipment growth; 3.2T coherent design-win update. Exact date not yet announced — estimated from prior-year cadence.
POET Volume Production Milestone
POET optical interposer volume shipment ramp via DenseLight/POET Singapore partnership
LWLG Strategic Partnership
EO polymer commercialization agreement with Tier 1 component supplier expected
OFC 2027 Conference
Optical Fiber Conference, Los Angeles — next-gen CPO architectures, EOP modulators, and fully photonic processor previews
Disclaimer: This is research commentary, not financial advice. All investments carry risk.
Last reviewed: August 2026