Free-space photonics is the use of light beams through air or vacuum, rather than fibre, as the communications channel, and the bulk-optical hardware that shapes those beams; 2026 work shows it moving in two directions at once, becoming the assumed interconnect for orbital compute clusters while being designed out of instruments wherever a chip can replace an optical bench.
Summary
What it is. Free-space photonics covers two things that are easy to conflate. The first is the communications channel: modulated light propagating through atmosphere or vacuum between two apertures, competing with fibre (no right-of-way, no trenching) and with millimetre-wave radio (far more bandwidth, far tighter beams). The second is the physical apparatus of free-space optics: lenses, gratings, spatial light modulators and dispersive elements arranged on a bench, used to shape, split and multiplex beams. The link is a growth story. The bench is, in several of the 2026 papers here, explicitly the problem to be removed.
How the link works. A laser is modulated, collimated through a telescope, and captured by a receiver aperture that must stay pointed within a small angular tolerance. Everything hard follows from that: pointing and tracking, beam wander, and bandwidth limits in the transmit/receive electro-optics. One 2026 demonstration built a hybrid transmitter that drives both a free-space optical channel and a D-band (110-170 GHz) millimetre-wave channel from a single optically harmonically locked laser pair, giving low RF phase noise and narrow optical linewidth from the same source, and characterised the combined capacity against beam angle misalignment at over 100 Gb/s signalling. A separate strand attacks the receiver side with machine learning: a neural network learns a residual deformation of the analytically derived matched filter for carrier-less amplitude and phase modulation, using 16 time-, frequency- and memory-domain features, and beats fixed matched filtering under severe bandwidth constraints in a hardware-in-the-loop experiment.
Why it matters now. The demand pull is data movement. Copper reach shrinks as bandwidth rises, pushing optics from the data centre floor into the rack and onto the board ref. Above that, a space-based AI infrastructure design study proposes fleets of satellites carrying solar arrays and TPU accelerators, connected by inter-satellite free-space optical links, flown in close formation specifically to make those links high-bandwidth and low-latency; it illustrates the approach with an 81-satellite cluster of 1 km radius and reports that Trillium TPUs survived a total ionising dose equivalent to a five-year mission without permanent failures. Trapped-ion roadmaps similarly depend on networking many ion-trap chips over photonic links to get past ten thousand qubits ref.
The parameters that decide it. Link budget and pointing tolerance (the misalignment sensitivity measured in the hybrid D-band work), available modulation bandwidth and how much of it can be recovered by digital or learned equalisation, and the size, weight, power and alignment stability of the optics. The last of these is where integration bites: high-speed flow imaging is being rebuilt entirely on-chip precisely because the free-space version relies on bulky optics, long dispersive elements and is prone to alignment instabilities, and orbital angular momentum switching is being pushed into a 46 nm van der Waals film for the same reason.
Evidence base
Open questions
- What link availability and bit error rate do hybrid FSO/D-band transmitters achieve over real outdoor paths and across weather, as opposed to the laboratory misalignment sweeps reported at over 100 Gb/s?
- Can inter-satellite free-space optical links hold the required pointing under machine-learning-based formation control at the 1 km, 81-satellite scale, and what is the demonstrated link budget rather than the assumed one?
- How much of the bandwidth limitation in free-space links can be recovered digitally before the equalisation power budget eats the advantage over intensity-based fibre links, given that coherent DSPs already cost three to five times more and consume three to four times more power ref?
- Do integrated mode-shaping elements such as glass-encapsulated MPLCs reach telecom wavelengths, insertion loss and mode-count figures good enough to replace bulk free-space multiplexers, or do they remain visible-wavelength proof-of-concepts?
Assessment drafted 2026-08-31 from up to 10 KB sources using the technology-scorecard framework; scores are a draft read pending review.