Photonic & Neuromorphic Chips
Light instead of copper between chips (co-packaged optics) starts shipping in 2026; brain-like chips remain a research niche.
Open in the interactive tree →Because moving data between chips costs more energy than computing, optical links increasingly replace copper right at the chip package. Neuromorphic chips (for example Intel’s Hala Point, 2024, 1.15 billion artificial neurons) process signals event-driven and frugally but remain research systems. Purely optical arithmetic units are still research.
As of now
Broadcom announced in October 2025 that it is shipping Tomahawk 6 “Davisson”, the first 102.4 Tb/s switch with co-packaged optics (samples to early-access customers), and Nvidia, which partnered with TSMC on its COUPE platform, has begun shipping Spectrum-X switches with co-packaged optics. Ayar Labs raised $500 million in March 2026 to scale production of optical chiplets, and Lightmatter says its Passage L200 (32 Tb/s) will be available in 2026. TSMC says COUPE-based products enter production in 2026. Neuromorphic chips remain a research niche; no commercial mass-market product could be confirmed.
Open steps
- Lasers that survive in switches Low AI leverageCo-packaged optics put lasers next to costly switch chips; external, replaceable laser sources need far lower failure rates than pluggable optics.
- Inverse-designed optical parts High AI leverageAlgorithm-designed photonic components can be tiny, but they must still work when real foundries print them, tolerant to process variation.
- Software and benchmarks for neuromorphic Medium AI leverageEvent-driven chips lack a common programming model and agreed benchmarks, so it is unclear which real tasks they win on energy.
- Precision of optical matrix math Medium AI leverageOptical matrix multipliers are fast but analog; converters, noise and calibration eat much of the energy saving at useful precision.
Where AI could help
Medium AI leverage. Inverse design and automatic tuning shorten photonic component design; lasers, packaging yield and customer qualification set the ramp.
- Inverse-design compact photonic components that survive foundry variation
- Automate thermal tuning and calibration of photonic circuits
- Map neural networks onto optical and neuromorphic hardware
- Co-design links and switches for AI cluster traffic
Shown so far
- In July 2026 Harvard and Max Planck researchers reported inverse-designed silicon nitride photonic components about 500 times smaller than conventional designs and compatible with a commercial foundry process (Nature Communications; mirrors reflect up to 98.5%). source
Prerequisites
- Laser1960
- Optical Fiber1970
- 2 nm Chips & High-NA EUV2025