4 ms·

It's more interesting to ask, "Does AI need to follow the current model of evil megacorps building massive data centres that, collectively, guzzle more energy t

by beloch 4mo ago

It's more interesting to ask, "Does AI need to follow the current model of evil megacorps building massive data centres that, collectively, guzzle more energy than most nations on Earth?" Perhaps LLM's (or something better) will develop to be more efficient and quickly become something most people run on local hardware. Perhaps fad-obsessed management types will move onto the next big thing and AI will start being used more judiciously. Perhaps society will set sane regulatory limits that shape the direction AI is going in, from models that take jobs people want to models that, given the right hardware, can do the jobs few want. Anthropic and OpenAI don't have to succeed for AI to succeed. If they turn out to be a bubble that bursts and torches a lot of investors, it might actually be a fundamentally good thing for everybody else.

vessenes 4mo ago
To be clear, nobody WANTS to have to go build all these datacenters. Well, maybe some pure-play hyperscalers do. But there's an immense amount of economic incentive to be able to do this more efficiently, capital and energy. And, what those hyperscalers want will not matter for a second if there isn't demand for the tokens output by those datacenters - they'll go instantly dark and have to seek new forms of valuable compute to offer. If this current building spree ends in massive solar and other power generation being overbuilt and cutting energy costs, we've had a really good outcome.
s08148692 4mo ago
This is why there is so much interest in space based AI compute. It's not just SpaceX - Google, Anthropic and Nvidia have openly expressed interest If you look at SpaceX plans and ambitions, they hope to deploy massive compute to orbit (multiple Terrawatts, hundreds of thousands of sats). If their ambitions even slightly materialise it would make ground based compute pale in comparison. Whether or not they succeed in their plans is beside the point - the point is they know that terrestrial electric infra can't sustain the growth they need
- joebe89 4mo ago
  How would cooling work in space based computing? To a layman like me it seems like a significant hurdle to overcome.
  - Pingk 4mo ago
    It really doesn't. You're purely relying on radiation fins to carry heat away, which are incredibly inefficient. > The radiator surface area problem also scales uncomfortably. At 838 watts per square meter, rejecting 1 megawatt of waste heat requires roughly 1,200 square meters of radiator. Deploying that much surface area on a satellite is a structural engineering challenge that gets harder with every order of magnitude. The ISS solar arrays span about 2,500 square meters total. So even a 2MW data centre in space requires a cooling array rivalling the international space station. Starcloud launched a single H100 in November and they were unable to run it 24/7 due to heat buildup. Even with novel solutions to make heat transfer to the fins more efficient, like phase-change liquids, the limiting factor is that the vacuum of space is a tremendous insulator. https://thecoolingreport.com/intel/starcloud-orbital-data-center-radiative-cooling.html https://thecoolingreport.com/intel/starcloud-orbital-data-ce... https://satnews.com/2026/03/17/the-physics-wall-orbiting-data-centers-face-a-massive-cooling-challenge/ https://satnews.com/2026/03/17/the-physics-wall-orbiting-dat...
- pjc50 4mo ago
  What is the multiplier in cost for a teraflop of compute in space vs on the ground? 100x? 1,000x? > Whether or not they succeed in their plans is beside the point No, I think that does matter eventually? Maybe for the IPO value?
  - jason_s 4mo ago
    https://spectrum.ieee.org/orbital-data-centers https://spectrum.ieee.org/orbital-data-centers > In an interview, McCalip says his initial rough calculations a few years ago suggested that data centers in space would cost in the range of 7 to 10 times more, per gigawatt of capacity, than their terrestrial counterparts. “It just wasn’t practical,” he says. “Not even close.” But when Elon Musk began publicly backing the idea, McCalip revisited the numbers using publicly available information about Starlink’s and Tesla’s technologies and capabilities.

by beloch 4mo ago

vessenes 4mo ago

To be clear, nobody WANTS to have to go build all these datacenters. Well, maybe some pure-play hyperscalers do. But there's an immense amount of economic incentive to be able to do this more efficiently, capital and energy. And, what those hyperscalers want will not matter for a second if there isn't demand for the tokens output by those datacenters - they'll go instantly dark and have to seek new forms of valuable compute to offer. If this current building spree ends in massive solar and other power generation being overbuilt and cutting energy costs, we've had a really good outcome.

s08148692 4mo ago

This is why there is so much interest in space based AI compute. It's not just SpaceX - Google, Anthropic and Nvidia have openly expressed interest If you look at SpaceX plans and ambitions, they hope to deploy massive compute to orbit (multiple Terrawatts, hundreds of thousands of sats). If their ambitions even slightly materialise it would make ground based compute pale in comparison. Whether or not they succeed in their plans is beside the point - the point is they know that terrestrial electric infra can't sustain the growth they need

joebe89 4mo ago
How would cooling work in space based computing? To a layman like me it seems like a significant hurdle to overcome.
- Pingk 4mo ago
  It really doesn't. You're purely relying on radiation fins to carry heat away, which are incredibly inefficient. > The radiator surface area problem also scales uncomfortably. At 838 watts per square meter, rejecting 1 megawatt of waste heat requires roughly 1,200 square meters of radiator. Deploying that much surface area on a satellite is a structural engineering challenge that gets harder with every order of magnitude. The ISS solar arrays span about 2,500 square meters total. So even a 2MW data centre in space requires a cooling array rivalling the international space station. Starcloud launched a single H100 in November and they were unable to run it 24/7 due to heat buildup. Even with novel solutions to make heat transfer to the fins more efficient, like phase-change liquids, the limiting factor is that the vacuum of space is a tremendous insulator. https://thecoolingreport.com/intel/starcloud-orbital-data-center-radiative-cooling.html https://thecoolingreport.com/intel/starcloud-orbital-data-ce... https://satnews.com/2026/03/17/the-physics-wall-orbiting-data-centers-face-a-massive-cooling-challenge/ https://satnews.com/2026/03/17/the-physics-wall-orbiting-dat...
pjc50 4mo ago
What is the multiplier in cost for a teraflop of compute in space vs on the ground? 100x? 1,000x? > Whether or not they succeed in their plans is beside the point No, I think that does matter eventually? Maybe for the IPO value?
- jason_s 4mo ago
  https://spectrum.ieee.org/orbital-data-centers https://spectrum.ieee.org/orbital-data-centers > In an interview, McCalip says his initial rough calculations a few years ago suggested that data centers in space would cost in the range of 7 to 10 times more, per gigawatt of capacity, than their terrestrial counterparts. “It just wasn’t practical,” he says. “Not even close.” But when Elon Musk began publicly backing the idea, McCalip revisited the numbers using publicly available information about Starlink’s and Tesla’s technologies and capabilities.