Free-space optics sends data as modulated light through air or vacuum instead of fibre, and the current evidence base is laboratory-scale: >100 Gb/s hybrid optical/millimetre-wave transmitters, machine-learning receivers to fight bandwidth limits, and design studies that make inter-satellite laser links the backbone of proposed orbital AI compute clusters.
Summary
Free-space optics (FSO) is optical communication without a waveguide. A laser is modulated with data, the beam crosses air or vacuum, and a receiver aperture collects it. The physics is the same as fibre transmission, so the same modulation formats, coherent detection tricks and DSP apply; what changes is the channel. Instead of a controlled glass core you have a path with atmospheric loss, turbulence, and, above all, a pointing problem: the transmitter and receiver must stay aligned to within a fraction of the beam divergence. The recent literature treats misalignment as the first-order design variable, studying combined link capacity as a function of beam angle misalignment in a hybrid FSO plus D-band (110-170 GHz) millimetre-wave transmitter driven by a single phase-locked laser pair, at >100 Gb/s signalling.
Two distinct engineering programmes sit under the same phrase, and they pull in opposite directions. The first is FSO as a link technology: laser beams as an alternative to fibre or radio. The second is “free-space optics” as the traditional bulk-optical bench, and here the whole thrust of current work is to eliminate it. On-chip high-speed imaging is presented explicitly as a replacement for approaches that “rely on bulky free-space optics, long dispersive elements, and are prone to alignment instabilities”, and dynamic control of orbital angular momentum is described as still dependent on “bulky free-space optics or cascaded architectures”, motivating an ultrathin 46 nm van der Waals alternative. Related work miniaturises multi-plane light converters into a fused silica chip by 3D direct laser writing. The lesson transfers: alignment stability is the recurring failure mode of anything that propagates light through free space, and it is what determines whether FSO links are operationally practical rather than merely demonstrable.
The parameters that decide FSO are therefore: available link margin (transmit power, aperture, range, atmospheric or vacuum loss), pointing and tracking accuracy relative to beam divergence, the electrical bandwidth of the transceiver front end, and cost per bit against the incumbent. On bandwidth, a neural-network-assisted deformable matched filter has been shown experimentally, in a hardware-in-the-loop carrier-less amplitude and phase system, to outperform conventional fixed matched filtering under severe bandwidth constraints without decision feedback. On cost, the wider optical interconnect economics are unfavourable to the most capable formats: coherent systems need DSPs consuming 3-4x more power and costing 3-5x more than intensity-modulated ones ref. The largest single demand pull in these sources is orbital: a proposed space-based AI compute system uses fleets of satellites with solar arrays, Google TPUs and inter-satellite links using free-space optics, flown in close proximity to obtain high bandwidth and low latency, illustrated with an 81-satellite cluster of 1 km radius.
Evidence base
- 3 June 2026: a hybrid FSO and D-band (110-170 GHz) millimetre-wave transmitter driven by a single phase-locked laser pair was demonstrated at >100 Gb/s signalling, with combined capacity studied against beam angle misalignment.
- 16 June 2026: a neural-network deformable matched filter, using 16 time, frequency and memory features and trained on an EVM loss, outperformed fixed matched filtering under severe bandwidth limits in a hardware-in-the-loop CAP system without decision feedback.
- 18 June 2026: a space-based AI compute architecture specifies inter-satellite links using free-space optics with satellites flown in close proximity, illustrated by an 81-satellite cluster of 1 km radius; Trillium TPUs survived a five-year-mission-equivalent total ionising dose without permanent failures.
- 29 May 2026: an on-chip high-speed imaging platform was introduced specifically to replace approaches that rely on bulky free-space optics and are prone to alignment instabilities.
- 14 August 2026: dynamic orbital-angular-momentum control was described as still dependent on bulky free-space optics, motivating an all-optical switch in a 46 nm 3R-MoS2 film.
- 2 February 2026: coherent optical systems are reported to require DSPs consuming 3-4x more power and costing 3-5x more than intensity-modulated systems, constraining high-capacity optical terminal economics ref.
Open questions
- What link distance, availability and bit error rate do the demonstrated >100 Gb/s hybrid FSO/D-band transmitters achieve over a real outdoor path, as opposed to a bench with controlled misalignment?
- In the hybrid architecture, how much capacity does the millimetre-wave path retain when the optical path drops out, and does that hand-off meet carrier availability requirements?
- What pointing accuracy and station-keeping precision does a close-formation compute constellation actually require, and does the ML-based constellation control approach hold at 81 satellites and beyond?
- Can FSO terminals reach the required capacity without coherent detection, given its 3-4x power and 3-5x cost penalty, or does high-capacity FSO inherit that penalty permanently?
Assessment drafted 2026-08-31 from up to 10 KB sources using the technology-scorecard framework; scores are a draft read pending review.