The link budget, in one paragraph
Radio power spreads. A transmitter radiating into a beam of angular width θ spreads its energy over an area that grows as the square of distance, so received power falls as 1/d². You claw it back with antenna gain, focusing the beam and gain rises as you shrink the wavelength or grow the aperture. That is the entire argument for optical links: at infrared wavelengths, a modest telescope produces a beam divergence measured in microradians rather than degrees, so almost all the transmitted photons arrive at the receiver instead of warming the sky. Optical links also sidestep spectrum licensing entirely, which is not a physics advantage but is a very real commercial one. The price is pointing. Microradian beams require the two spacecraft to know where each other are and to hold aim through vibration and thermal distortion. Optical terminals are, in practice, precision mechanisms with a laser
attached, and that is why the terminal suppliers in Chapter 14 are a real business rather than a commodity.
How much data does an orbital data centre actually move? This is the question that gets skipped, and the answer is far more encouraging than intuition suggests, for some workloads, and far worse for others.
Figure 7.1 — Bandwidth is only a problem for some workloads
Work it through: Selling inference. A rack serving, say, 10,000 output tokens per second at roughly four bytes per token generates about 40 kB/s, 0.32 megabits per second. A thousand racks would need 320 Mbps. Text inference is, in bandwidth terms, nothing. You could downlink a substantial orbital inference business over a link that would embarrass a home broadband connection. Training checkpoints. A 500 GB checkpoint every hour is about 1.1 Gbps sustained. Real, manageable, and mostly avoidable by checkpointing to storage in orbit rather than to the ground. Getting training data up. Here is the wall. Ten petabytes of training corpus over a 100 Gbps link takes about nine days of continuous transmission, assuming you have contact with a ground station the whole time, which you do not. Training on Earth-resident data in orbit is a data-logistics problem long before it is a compute problem. Data that is already up there. A high-resolution imaging satellite generates raw data far faster than it can ever downlink it, which is why so much of it is discarded or heavily compressed on board. This is the strongest nearterm case in the sector: put the compute next to the sensor and downlink conclusions rather than pixels. So the ranking of natural first workloads falls straight out of the arithmetic: process data that is already in orbit; serve inference where the output is small; run batch training on resident datasets; and only last, and reluctantly, ship large terrestrial corpora upward.
Latency, which is not the problem people think
A spacecraft at 550 km is 1.8 milliseconds away at the speed of light straight overhead, and perhaps 3, 7 ms at a realistic slant range. Add ground segment and backhaul and an orbital node sits, in latency terms, roughly where a data centre in a neighbouring city sits. That is fine for essentially all inference, fine for training, and not fine for latency-critical trading or real-time control loops. The real availability problem is not distance but geometry: any given spacecraft is above any given ground station for only a few minutes per pass. Continuous service therefore requires either many ground stations, or inter-satellite links that relay traffic across the constellation to whichever node currently has a ground contact. The second option is what turns a fleet of computers into one computer, and it is why the relay layer is strategically more important than its revenue currently suggests.
Where bandwidth breaks Cloud cover, which blocks optical ground links and forces geographically diverse station networks; pointing and acquisition under thermal distortion; the capital cost of a ground segment nobody counts in the spacecraft budget; and the regulatory question of who is allowed to downlink what, where, which Chapter 17 takes up when we leave the United States.