Welcome back! This is Canaan's weekly update on Bitcoin mining, energy, and compute infrastructure.

As SpaceX’s IPO sets new records, Elon Musk is talking about a constellation of orbital racks of compute meant to move the AI build-out off the planet.

The reference craft, AI1, is one NVIDIA rack in orbit: 150 kW of peak power, a wingspan wider than a Boeing 747, a 110-square-meter liquid radiator, and laser links back to the Starlink network. SpaceX now wants to launch up to a million of them and is building a Texas mega-factory to do it.

And it’s not just Musk. Google's Project Suncatcher is testing TPUs built for orbit, with prototype satellites also slated for early 2027; Nvidia-backed Starcloud has already flown an H100 in low-Earth orbit and is planning a Blackwell-class follow-on; Nvidia has introduced space-grade compute modules; and Blue Origin, Axiom Space and others are circling the same idea.

The case for data centers in space is mostly a case about Earth's limits. Terrestrial data centers are increasingly gated by power, water, cooling, and the patience of the communities next door. This week alone, a Washington state utility paused a 500 MW data-center request after local pushback.

In space, the sun never sets, so the solar energy is almost continuous and at near-zero marginal cost. There is no land to permit, no water to draw, and no interconnection queue to join. Nvidia's Jensen Huang called the economics poor today but granted that space offers "an abundance of energy" and room to scale.

The case against celestial data centers is a case about physics and money. Take heat, for example. In a vacuum, there is no air or water to carry warmth away; a satellite can only radiate it as infrared. Starcloud's own white-paper math states that a two-sided radiator near room temperature sheds only about 633 watts per square meter, which means that a one-megawatt facility would require a radiator the size of a hockey rink. Radiation is another significant challenge since it degrades the advanced, non-hardened silicon that AI workloads demand. Without a viable repair-or-upgrade path, these satellites could quickly turn into space junk rather than a serviceable asset. OpenAI's Sam Altman has been blunter than most, calling the idea "ridiculous."

While the engineers mostly agree that data centers in space are feasible, the hardest objection seems to be economic rather than technical or physical.

The orbital debate resembles the terrestrial one. The premise of putting servers in space is that compute is an energy-and-heat problem. The solution to such a problem could be to escape Earth's energy and thermal constraints by leaving, but we could also simply get better at solving the energy-and-heat problem. The latter could imply stranded loads, flexible power, better grid integration, heat reuse, or better policies. Canaan's hash-to-heat deployment, an 8 MW project piping roughly 80°C water into a Nordic district-heating network, is the exact inverse of an orbital radiator: on the ground, waste heat is not a liability to dump into the void but a product with a buyer.

Both bets start from the same physics. One concludes the answer is to leave the planet; the other, to use it more intelligently. The 2027 prototype launches will begin to turn that disagreement from argument into data.

Released in April 2017, the AvalonMiner 741 (sometimes called the Avalon A7) was the refinement of the 16nm generation that the Avalon721 had opened a year earlier. It kept the same A3212 chip but packed 88 of them, pushing reliable output to 7.3 TH/s (operators routinely saw 7.5–7.9 in the field) at roughly 1,150 watts, or about 0.16 J/GH at the wall.

The thermals here were the most important part. Canaan introduced an "Airforming" cooling design to keep all 88 chips stable under sustained load: a decisive fix that earned the 741 a reputation for running quietly and simply not falling over. It arrived as a direct answer to its competitors, and it carried forward the open-hardware DNA of the ROQ ("Reliable Open Quality") line. Units were managed through the AUC3 converter and an AvalonMiner Controller that could be a stock Raspberry Pi 3, while large fleets ran on Canaan's openly developed AvalonMiner Management System.

Twenty 741s chained together cleared 146 TH/s - a number that felt enormous in 2017, and a useful yardstick for how far per-machine efficiency has traveled since. The 741 spent its generation fighting for tenths of a joule. Today's hydro-cooled Avalon fleet operates near 17.9 J/TH.