That’s interesting, but what’s the point? If it’s like 2 DGX boxes in each satellite, spaced out, the interconnect between them is going to be very slow, and the individual computational power of each satellite will not be that impressive.

And if you connect them all in one constructed mesh and wire them together, well, you’ve made a 200MW datacenter! The economies remain the same.

If hardware gets more power efficient, well… Then why do you need to go to space anymore?

100kW? Nvidia BGX 200 servers are 14kW each, not counting the interconnect, or anything else. According to nuggets I’ve read online, we’re talking 200 megawatts for an Earth-based AI datacenter these days, without something exotic like underclocked Cerebras WSEs (which would be pretty neat, actually…)

Plugging 200 megawatts into this:

https://www.calctool.org/quantum-mechanics/stefan-boltzmann-law

I get about 0.46 square kilometers, depending on the coolant temperature, and ultimate efficiency of the system (with how you orient the thing relative to solar panels, how you circulate coolant…)

I have no clue what the construction of such a huge structure would look like, but if it was a simple 0.5 inch aluminum sheet, it would weigh like 15,000 metric tons. Even much thinner, that’s still on the order of “mass of a cargo ship”

Why is that, though?

Well, something like the ISS doesn’t generate much heat, and hypothetical rockets that need big radiators have very hot coolant to dissipate heat quickly. But space data centers are the sinister combination of “tons of waste heat” and “needs a low coolant temperature.”

Well, that’s just it. Facebook makes so much money their Metaverse failure doesn’t even matter, they can just write it off and Zuckerberg faces zero consequences.

This is the same. These guys could burn many billions on space data centers, get zero returns, yet face zero consequences, essentially leaving the rest of the world with the bill/wasted work.

That’s tragically dysfunctional.

Fair.

I’m pretty sure that’s a plot point in some cyberpunk-ish sci fi.

On radiators, plugging it into this formula:

https://projectrho.com/public_html/rocket/heatrad.php

I get a circular radiator at least a kilometer wide, assuming the radiator is quite efficient, a rather modest datacenter, and very hot coolant (70C).

…Realistically, the coolant temperature would need to be much lower. See how it’s a power of four in the formula? That means the radiator area gets very large real quick.

I cannot emphasize how expensive a functional 1km+ radiator would be in space. It’s mind bogglingly expensive.

If a space datacenter is in LEO like Starlink, then it’s in Earth’s shadow a lot of the time, and would have to be “part” of the starlink network constantly zooming over the ground. If it’s geosynchronous, then laser communication (or any communication) gets real tricky, and latency is limited by the speed of light. I’m not saying it’s impossible, but reliable high data rates would be an expensive engineering challenge.

Heat is radiated into the vacuum for free

When you combine that with a mesh network like Starlink, the need for laying fiber lines disappears entirely

Citation needed.

And on water usage, I will point out that gas generators and evaporative cooling are only used on Earth because other methods (geothermal, big radiators, heatpumps) are somewhat more expensive… Not, like, orders of magnitude more expensive like pure radiative cooling in space.

I mean, making fuel is one thing. A future where we can manufacture electronics and spaceflight tech off Earth is far, far away. Like so far a lot of other technologies are going to scramble the economics anyway.

I’m pretty sure that’s the point TBH.

Kinda scary when the context is these kinds of corporate systems, though.

That’s because science literacy is pretty low.

And to be fair, the average person doesn’t need to understand vacuum thermodynamics. The issues is when a few of those “average people” are now billionaires making unilateral decisions, surrounded by yes men and feeds instead of experts informing them of reality.

It doesn’t matter.

It’s a fantasy in billionaire’s heads, a self-perpetuating meme, and no one is telling them no. So they’re going to fund it, whether it makes any sense or not.

Reality doesn’t really matter anymore.

It’s not speculation. Nvidia themselves have run experiments with GPUs in orbit, and the issue gets worse with smaller lithography (eg newer chips).

So do all the Microsoft subscriptions they already buy, yet they’re extremely popular anyway?

Businesses will adore this. I can guarantee a lot of us will be forced to use these at work, like Teams and CoPilot, as a further mega deal with Microsoft.

…But honestly, I think “home” buyers who don’t really care about PC stuff, aka most people, would pick tablets over this.

Especially for businesses. At work you’ll have no choice but to use something like this.

But the point is that Valve could easily be Amazon some day. All these little companies taking their first anticompetitive steps could.

Of course everyone loves them when they’re small, and nice, and growing, until they get so big it’s way too late to do anything about it. But many will still feel loyalty, like they do to Amazon today.

Eh, the context I was thinking of is that they are constantly playing “safety theatre” where it absolutely doesn’t matter. They’ve tried to kill open models and basically capture regulators by misleading or outright lying, for their benefit.

In other words, this is a case of “a broken clock is right sometimes,” and I think they knew Trump will back down.

Yeah… Microsoft and Google have a list of employees to fire now.

Trump will back off to some extent, to avoid inflaming stock markets (and his Big Tech friends heavily invested in Anthropic tooling).

Anthropic will fire a few people. OpenAI will raise money somehow.

That’s about it.

That’s basically always been the case anyway, except for Gemini for long context, perhaps.

And, of couse, prioritize open models over API. It’s:

Open training models > open weights models > restrictively licensed open models > open weights models over API > Claude/Gemini

Anthropic is self righteous and self serving.

Like, I like local LLMs more than most, and Anthropic models are great at the moment, but don’t mistake Anthropic for having a spine.

Talk in other comment sections suggests this means Nvidia (or AMD, or Cerebras, or Intel or anyone) can’t legally supply Anthropic with hardware.

So it effectively kills Anthropic? And that would sink the stock market.

As at the other extreme, a lot of Big Tech (like Palantir) is heavily reliant on Anthropic, which would unsettle the oligarchs and likely force Trump to walk back.

I don’t know if this is true, but this does smell like one of Trumps many “unenforceable tweets”