I didn’t quote Elon at all. Why you so upset?

You’re arguing against a version of the idea that nobody actually proposed. And you’re still doing it.

ELI5:

A computer uses electricity. Almost all of that electricity eventually becomes heat. In space, you don’t cool the computer by blowing air over it. You conduct the heat to a radiator, and the radiator emits infrared radiation into space. That’s literally what spacecraft have been doing for decades.

A heat pump doesn’t magically destroy heat, either. It moves heat from one place to another while consuming additional power. If you have 100 kW of computers and a heat pump consumes 10 kW, you ultimately have roughly 110 kW that has to leave the spacecraft. Nobody is disputing that.

The important question is whether 100-150 kW can be radiated into space. It can. At roughly 400 K, an idealized high-emissivity radiator emits about 1.3 kW/m². So 150 kW requires on the order of 115 m² of radiating surface(they use double-sided radiators so it’s actually half that), before accounting for engineering margins, Earth IR, view factors, etc. At higher radiator temperatures the required area gets dramatically smaller. This isn’t speculative physics.

“Semiconductors need to stay cool” is also not equivalent to “semiconductors must be operated at terrestrial datacenter temperatures.” Semiconductor reliability generally improves at lower temperature, but the allowable junction temperature is much higher than room temperature. Different electronics can also be designed for different temperature ranges. The spacecraft thermal system’s job is to keep the junctions within their specified limits, not to maintain the entire satellite at 21°C because that’s what a conventional server room does.

Radiation is a much more legitimate engineering problem. But “the chips must therefore be old, slow rad-hard CPUs” doesn’t follow. There are several ways to address radiation: shielding, redundancy, ECC, watchdogs, scrubbing, fault-tolerant architectures, selective hardening, shorter mission life, and using commercial silicon where the risk is acceptable. Starlink already demonstrates that SpaceX is willing to use this general philosophy rather than designing every component like a 1980s NASA deep-space probe.

And ASICs aren’t supposed to “fix everything.” They’re useful because AI/datacenter workloads are already extremely amenable to specialized accelerators. An ASIC doesn’t have to be a single-purpose chip that becomes worthless if you change one training algorithm. Modern accelerators contain programmable processors, memory systems, matrix engines, interconnects, etc. Even Nvidia’s own products are effectively highly specialized compute architectures rather than magical general-purpose computers.

“Nvidia is the only company capable of making fast AI chips” is an assertion, not an engineering law. Google has TPUs. Amazon has Trainium/Inferentia. Microsoft has Maia. Meta has MTIA. Jim Keller is behind tenstorrent. Various other companies have developed AI accelerators. The existence of Nvidia does not establish that nobody else can design an accelerator.

Likewise, “100× more expensive and 40× slower” is a number you invented unless you have an actual orbital hardware architecture and benchmark to support it. If the claim is that orbital compute is uneconomical, then compare actual quantities: $/FLOP, FLOP/W, useful compute per kilogram, launch cost/kg, radiator mass per kW, solar-array mass per kW, radiation-induced error rate, expected lifetime, and maintenance/replacement cost. That’s an engineering argument. “It can’t possibly work because terrestrial datacenters don’t do it” isn’t. And the Microsoft underwater-datacenter experiment doesn’t demonstrate that orbital datacenters are impossible. It demonstrated that maintenance accessibility has value. That’s obvious. An orbital data center would have to be designed around that constraint, just as satellites, submarines, aircraft, offshore platforms, and nuclear reactors are.

Finally, the “gigawatt” point is fair in one narrow sense: watts alone don’t tell you useful computational performance. You need performance/W, performance/kg, cost/FLOP, etc. But that doesn’t make gigawatts meaningless. Power is one of the fundamental constraints on a compute system. If somebody puts 1 GW of useful compute power in orbit, the relevant question is exactly what performance they get from it. So yes, there are serious engineering questions about orbital AI datacenters. Radiation, thermal design, mass, launch cadence, degradation, optical communications, power generation, and maintenance are all real problems.

But “there are engineering problems” and “therefore the concept violates physics” are two very different conclusions.

The former is engineering.

The latter is just internet physics.

You just got fucking schooled by Chad GPT. Have a blissful day.