Insight and analysis on the data center space from industry thought leaders.
With AI workloads spreading across terrestrial and orbital compute, the atmosphere is emerging as the key bottleneck for moving data between ground and space.
July 27, 2026
AI workloads are no longer confined to the data center. Training and inference workloads are expanding outward to distributed terrestrial campuses, edge sites, and soon, sensing and compute platforms in low Earth orbit (LEO).
Earth-observation constellations, broadband networks, and proposed orbital data centers all generate data that must reach the ground quickly and in volumes that conventional radio-frequency links were never designed to carry.
Optical, or laser, communications offer a way out. A single free-space optical feeder link can move far more data than an RF equivalent, with system architects targeting sustained capacities above 100 Gb/s and a path toward 200–300 Gb/s per link.
This bandwidth enables orbital compute and dense LEO constellations to serve as credible partners to terrestrial AI infrastructure rather than isolated islands.
But there is a catch, and it is not the laser. It is the sky.
The Atmosphere Is the Bottleneck #
Although a ground-to-LEO path runs hundreds of kilometers, most of the signal damage happens in the lowest slice of the troposphere. Turbulence, cloud cover, aerosols, water vapor, and a shifting refractive index all distort an optical beam – scattering it, bending it, and causing the rapid intensity swings engineers call scintillation.
The result is attenuation and dropouts, which can throttle an otherwise high-capacity link down to a fraction of its potential.
Today’s optical ground stations fight back with adaptive optics, forward error correction, coherent detection, and site diversity, spreading stations across geography so that if clouds close in over one, another can pick up the beam. These techniques are effective but share a key limitation: they primarily react, correcting conditions only after link degradation has begun.
However, as demand for space-to-ground AI capacity grows, a familiar question emerges: can a reactive model scale?
Predicting the Channel Instead of Chasing It #
One answer treats prediction itself as a network resource. The idea is to build an atmospheric digital twin, a continuously updated model of the sky above each ground station, and use it to forecast not the weather, but the communications performance: expected throughput, signal-to-noise ratio, attenuation, and the probability of holding a reliable link over the next seconds, minutes, or hours.
The twin is fed from two primary sources. From above and around, it ingests weather-satellite data, radar, lidar, cloud cameras, and numerical weather prediction. From the network itself, every live link is treated as a sensor: received optical power, adaptive-optics corrections, pointing errors, and bit-error rates all describe the atmosphere in real time. Physics-informed machine learning fuses these streams into a forecast, each one carrying a confidence estimate.
Those forecasts then drive decisions. Rather than waiting for a cloud to degrade a link, the network can shift traffic to a clearer station before the degradation occurs – a predictive handover. It can weight-distribute apertures toward the paths expected to perform best, assign wavelengths accordingly, and schedule bandwidth-heavy transfers into windows the model expects to remain clear. Every completed session is then compared against what was predicted, so the twin sharpens over time.
A predictive optical network forecasts atmospheric link quality and re-routes traffic before a cloud degrades the beam, shifting load to a clearer ground station ahead of the disruption.
Building the Infrastructure for Predictive Optical Networks #
Much of this borrows from tools the data center world already knows. Digital twins, software-defined orchestration, telemetry-driven control, and physics-informed machine learning are all familiar. What is new is pointing them at the atmosphere and treating a laser feeder link as a schedulable, predictable resource rather than a fragile best-effort pipe.
The trade-offs are real. Forecasts are probabilistic, not certain, which is why confidence has to be built into scheduling rather than bolted on afterward. A distributed model depends on fast, reliable terrestrial fiber connecting ground stations, cloud infrastructure, and the twins’ computing. And as with any laser system operating in shared airspace, deployment must remain within applicable laser-safety limits.
None of this has been proven at scale yet; the architecture suggests high-fidelity propagation simulation, hardware-in-the-loop testing, and regional demonstrations as necessary steps before operational trials.
Why It Matters for AI Infrastructure #
The strategic point is straightforward. AI is expanding faster than any single site can comfortably absorb, and the value of distributed geothermal-based or orbital compute depends on the ability to move data reliably between surface installations and satellites, and vice versa.
If the atmosphere remains an unpredictable gatekeeper, high-capacity optical links will stay impressive in the laboratory and unreliable in production. Making predictable changes to the channel alters that calculus. It turns a weather-dependent gamble into something closer to a managed resource; operators can plan around the way they already plan around power availability, water use, and cooling.
As islanded geothermal-powered compute expands and compute reaches for orbit, the advantage may fall less to those with the most powerful lasers and more to those who can best anticipate the sky.