Google's first data center satellite is operational and, according to TechCrunch, it is already writing poetry. On board there are four TPUs and a power generation system capable of up to 1 kW. The detail that matters, however, is different: due to power limits, the company can switch on only one TPU at a time for about 20 minutes. On Reddit the news immediately became a joke: 'I spent a lot on my rig but I can only run Gemma 4'. One user suggests trying Qwen 3.8 27b, as if the problem were merely model choice.

The issue is more structural than a joke. In a terrestrial data center, power is a cost but it is continuously available. In orbit, energy is an intermittent and limited resource: a satellite generating at most 1 kW must also power onboard systems, communications, and thermal control. If an accelerator requires a significant share of that budget, the consequence is duty cycling: compute must be switched on intermittently, not by choice but by physical constraint. This reduces useful throughput far below what the presence of four TPUs would suggest.

For those working on edge and on-premise deployments, the orbital case is a useful extreme. Many local architectures have constrained power and cooling budgets: remote sites, vehicles, cabinets without dedicated cooling. The difference is that on Earth the margin is wider, but the logic is the same: adding hardware does not increase capacity if the system cannot power it and dissipate its heat. Capacity planning in these contexts must start from the energy budget before the number of accelerators.

There is also a signal for the model market. The ironic suggestion on Reddit to move to a smaller model is not trivial. Compact models and quantization techniques reduce peak power and allow useful inference on hardware that cannot remain always on. In scenarios where cooling is passive or energy is limited, model choice becomes part of the system design, not an afterthought.

Google's stated goal, according to the source, is to pave the way for networks of close-flying satellites performing large inference workloads in orbit. If inference happens on board, the need to transmit raw data to the ground is reduced; only results are sent. But the 20-minute constraint complicates space-to-ground communications and narrows the useful compute window. It is therefore not only a hardware problem: it is a problem of scheduling, thermal management, and network architecture.

For those evaluating on-premise or edge deployments, AI-RADAR has a section on /llm-onpremise dedicated to assessing these trade-offs: it analyzes frameworks and scenarios for understanding when local compute makes sense compared to the cloud. Google's satellite does not answer that question, but it shows how much energy constraints can redefine what is technically possible.