This is hiding so, so much complexity behind a simple hand wavy “modular”. I have trained large models on thousands of GPUs, hardware failure happen all the time. Last example in date: an infiniband interface flapping which ultimately had to be physically replaced. What do you do if your DC is in space? Do you just jettison the entire multi million $ DGX pod that contains the faulty 300$ interface before sending a new one? Do you have an army of astronauts + Dragons to do this manually? Do we hope we have achieve super intelligence by then and have robots that can do this for us ?
Waving the “Modular” magic key word doesn’t really cut it for me.
Something tells me that the price of batteries is already cheap enough for terrestrial data centers to make more economic sense than launching a datacenter - which will also need batteries - into space.
Assuming the $165M of panels and batteries last for a decade, and there are no maintenance costs, they'll provide 3,504,000MWh over that time for an energy cost of 4.7 cents per kWh. This is competitive with grid power in some places. It also has the advantage of not needing backup generators. But maintenance costs do exist, and it makes more financial sense to buy power as you use it rather than pay upfront.
> Optimistically assuming 12 hours of sunlight per day, a 40MW datacenter would need 480MWh of batteries to cover the dark period, costing $50 million.
A 40MW data center doesn't run constantly at 40MW. That's its load rating. Like any industrial facility, actual peak loads are probably around 80% and average loads are lower.
Also, why do you assume that the data center has to be off-grid? That's a constraint of a space-based datacenter, not a ground based datacenter.
Datacenters with storage can complement grid power.
> The cheapest batteries today are around $100/kWh.
If we are comparing ground based data centers to hypothetical space based ones, then consider that grid scale iron air batteries are coming soon at $20/kWh.
https://www.wesa.fm/environment-energy/2024-02-19/weirton-fo...
I assumed the battery + solar setup would need to provide 40MW because while datacenters usually do run below capacity, you'd also want some extra capacity to account for cooling systems, battery/panel degradation, and the fact that for some tasks (such as AI training), you actually do get close to 100% capacity for long periods of time. Feel free to cut my numbers by 20%, but I don't think that would change the bottom line: off-grid datacenters could be cost competitive in some regions, but the upfront costs don't make them worthwhile right now. If battery costs go down (as I hope they will), that will likely change.
An orbital datacenter would not need significant batteries because it would be placed in a dawn-dusk sun-synchronous orbit. The panels would only be occluded during solar eclipses, which in low earth orbit last a few seconds. Starcloud is betting that launch costs will plummet but battery costs will not, and that they'll be able to cheaply solve space-specific issues related to cooling, maintenance, and reliability.
If you look at my other comments in this thread, you'll see I predict they will fail. A lot of people are coming to the same conclusion, but for mistaken reasons (eg: thinking that space-based datacenters would need as many batteries as ones on the ground). I'm just trying to correct that.
Yep, we agree on that.
> off-grid datacenters could be cost competitive in some regions, but the upfront costs don't make them worthwhile right now.
I still don't understand why the alternative to space based datacenters being proposed is off-grid datacenters.
Why not compare it to grid-connected datacenters with enough behind the meter generation and storage to avoid peak grid prices? After all the ultimate comparison metric is cost (and ideally C02 emissions)