Smart Glasses Are Learning to See in the Dark

Our last series looked at AI glasses as a product category. This one follows a single capability further. It asks what happens when glasses can see in the dark and understand what they see. A mass base of camera glasses now meets cheaper infrared sensors, on-glass AI, and defense money for digital night vision. The software that must interpret night scenes is the slowest part.

Key points

• More than seven million Meta camera glasses sold in 2025 form a ready base for any night feature.¹
• Display-free models make up 96% of AI glasses and cannot show the wearer an enhanced night image.²
• Wafer-scale quantum dot sensors and on-glass AI chips point to a plausible path toward night vision at eyewear cost and power.³⁴
• The Army is funding digital night vision on eyewear through prototype deals worth more than $350 million.⁵
• Multimodal models still lose accuracy at night, and scene understanding is the weakest link in the chain.⁶

What's changing

A camera base big enough to matter

EssilorLuxottica sold more than seven million Meta glasses in 2025. That is up from about two million across 2023 and 2024 combined.¹ The partners aim for ten million units of annual capacity by the end of 2026.¹ Counterpoint counted 263% growth in AI glasses shipments in the first half of 2026.² Google and Samsung are taking Android XR glasses into optical retail through Warby Parker and Gentle Monster.⁷ Meta already markets its glasses for night use. Its guide cites dim-light optimization and sign reading.⁸ The evidence is strong on volume. It is weak on visual night vision. We read the installed base as ready. The display layer likely lags by two years or more.²⁷

Sensors and chips shrinking toward the frame

Imec has put quantum dot infrared photodiodes on a 300 mm CMOS pilot line.³ Wafer-scale processing points toward infrared sensors priced for consumer goods. Lead content and slow pilot timelines are the main risks.³ Qualcomm's AR1+ chip ran a one-billion-parameter language model entirely on prototype glasses.⁴ On-glass inference avoids the phone round trip that a live night overlay cannot afford.⁴ Apple reportedly plans a camera dedicated to computer vision in its first glasses, now expected in 2027.⁹ That camera is the slot a low-light sensor would fill.⁹ This evidence is early. Battery drain from continuous enhancement remains unsolved.⁴

Defense money for digital night vision

Anduril holds a $159 million prototype deal and Rivet a $195 million deal under the Army's Soldier Borne Mission Command program.⁵ Meta is building with Anduril on that work.⁵ Anduril's EagleEye pairs a daytime display with a digital night vision display.⁵ Anduril credits machine learning for making digital night vision fast enough to use.⁵ The Army does not expect production before 2028.⁵ Transfer to consumers is plausible. Export controls could slow it.⁵ Assistive research offers a civilian proof point. Small AR studies report better low-light mobility for people with night blindness.¹⁰ Those trials were small and uncontrolled.¹⁰

Night perception lags the hardware

The EgoNight benchmark found a consistent day to night accuracy gap in nearly all ten multimodal models tested.⁶ DarkEQA found similar limits in dark rooms.⁶ Meta is testing always-on capture that would feed a model a continuous night data stream.¹¹ The same capture raises privacy risk after dark. Face recognition could draw bans that stall the whole feature set.¹¹ We see this as the binding constraint. Brighter images are close. Reliable scene understanding at night is likely further off.

What to watch

• Whether Meta, Google, or Apple ships a display model with a dedicated low-light mode before the end of 2027.⁷⁹
• Published night performance data from any vendor that markets after-dark use.⁸⁶
• A consumer partner taking Imec's quantum dot sensors past low-volume production.³
• Regulatory action on always-on capture or face recognition in glasses.¹¹