Remix.run Logo
▲ dr_dshiv an hour ago

It is much worse than I thought. I’m a power user. I have some reflections to make.

Hope for solar datacenters in the Middle East?

▲skissane an hour ago | parent | next [-]

The Australia–Asia Power Link (AAPowerLink), aka SunCable, is a proposed undersea HVDC cable to export solar power from northern Australia to Singapore via Indonesia. Due to doubts about the economic feasibility of constructing the cable, and question marks over the willingness of Indonesian and Singaporean governments to approve it, they've been exploring pivoting to building solar-powered AI data centres in northern Australia instead.

▲Doohickey-d an hour ago | parent [-]

I somehow suspect that datacenters are only economically viable if you run them round the clock, right? So at night they'd still be using gas. I guess the cost of compute would have to come down a lot, in order to power down a DC and spin up one where sun is shining instead.

Although I guess for current AI workloads, it could be viable, since the additional latency of sending your (small, text only) prompt around the world for processing is little compared to the time taken to process (whereas for other stuff being close to users matters more)

▲skissane an hour ago | parent | next [-]

> I somehow suspect that datacenters are only economically viable if you run them round the clock, right? So at night they'd still be using gas.

You could build batteries, charge the batteries during the day, run the DCs off it at night.

SunCable / AAPowerLink is likely going to be connected to the Darwin grid, which is currently not connected to the main East Australian grid; but building a connection to the main Australian grid would be a lot cheaper than undersea HVDC to Singapore. Once that happens, there would also be the potential to use pumped hydro (including Snowy 2.0) for storage.

▲DaedalusII an hour ago | parent | prev [-]

solar during day plus battery park, just like factorio

▲happosai 19 minutes ago | parent | prev | next [-]

They could run on solar in US throughout the daylight hours. Just make the tokens cost 5x after the sun sets That would reduce the overnight demand enough the data centers could run through night on batteries.

▲msdz an hour ago | parent | prev | next [-]

> datacenters in the Middle East

Given this year’s events [0], even Sahara seems more likely.

But you’d still have to build it out from ≈scratch, including cooling, internet connection infrastructure, and maintenance, so I’m not sure what the time horizon on amortization would look like, even with the cheap solar power.

[0] https://www.reuters.com/world/middle-east/amazon-cloud-unit-...

▲throwaway473825 11 minutes ago | parent | next [-]

There have been many proposals to build solar in Africa for exports to Europe, but none of them has gone anywhere. That's because the cost of capital is too high. It's cheaper to just build in Spain.

Globally, places like Australia and India should be more competitive for solar than the Sahara, although both Australia and India have a lot of dirty power generation to replace as well.

▲hdgvhicv an hour ago | parent | prev [-]

Aside from a dirt road from the nearest port and a few fibres to various IXPd what do you need to build in Morocco or Egypt that you don’t need elsewhere?

Obviously security is a concern which limits the number of locations, but there’s still plenty of places with stability, security, land, and the lack of taxation and tarrifs that the US puts on energy.

▲dgoldstein0 an hour ago | parent | prev | next [-]

Do our expert controls allow top of the line chips in those countries?

▲TeMPOraL an hour ago | parent | prev | next [-]

There's always an option of... wait for it... datacenters in space.

*runs*

But then of course that's another front-page submission with its own long thread of commenters feeling smart for calling it stupid, despite recently becoming a worked out plan and addressing exactly the (subset of reasonable) things people are complaining about here. In fact, we had those a week or two ago.

I swear the tech space turned into thought-space NIMBYs. "I don't understand any of it, but something about it pattern-matches to something I don't like, therefore it's stupid and people behind it are evil". In this thread, and in every other thread recently, too.

▲user43928 18 minutes ago | parent | next [-]

I looked into it just now, and it appears SpaceX is targeting deployment of the first satellites in 2027.

To deploy 1 GW of capacity, apparently 100 Starship launches would be needed.

On earth, 23 GW are currently under construction, 100-132 GW is already active, and about 100 GW should be added by 2030.

SpaceX's CFO supposedly cited 2029 as the timeline for multiple GW of compute in operation in space. Musk's vision seems to be launching 100GW/year or 1 million ton of satellites.

This would require 10k launches of Starship per year under ideal conditions, or 5k if Starship Block 4 can double the payload capacity to 200t as planned. This would mean 14 launches per day.

It does sound like a lot. Starbase Louisiana, at a $100B buildout, is planned to have 12 launch pads that could handle 30 launches per day.

Obviously there's much more needed than some launch pads. Anyway, that's the numbers and projections I found.

▲latexr 12 minutes ago | parent [-]

And that’s the optimistic take, assuming Musk’s projections are correct. Yet they are so often wrong there’s a full Wikipedia page just for things he said would be coming (often “next year”) and didn’t.

▲user43928 9 minutes ago | parent [-]

Yes, and I am curious about the financial viability of that compute.

From a capacity perspective, if SpaceX remains successful in scaling up their Starship rocket, it seems putting AI compute into space could at least realistically happen at significant scale.

▲latexr 15 minutes ago | parent | prev [-]

> "I don't understand any of it, but it pattern-matches to something I don't like, therefore it's stupid and people behind it are evil".

This is pretty bad-faith. What is the point of your comment? You’re sowing hate and attacking a nondescript group, for what? What does that achieve apart from being hateful?

Yours is one of the few usernames I recognise on HN. I remember seeing some comment from yours (I think it involved Star Trek) which I found interesting at the time and somehow the username stuck with me. Since then, every time I bump into your comments they have gotten more hateful and generally worse and more tribalistic takes.

It genuinely saddens me. Especially as you have in your bio “I'm constantly looking for opportunities to do something actually useful for humanity”, but your words reveal the opposite.

I’m not going to be specific about past posts because frankly I no longer remember their content, only the disappointment I felt. But I can take this one as an example. In addition to every thing I mentioned, I find it a pretty bad take to claim that adding more shit to orbit improves humanity. We need less crap up there, not more. I also find it nonsensical to defend (if we can call it that; you just attacked detractors with no explanation) something that is purely theoretical (which it objectively is; it doesn’t exist yet).

Anyone who has paid attention in the last decades will see all the unkept promises. You can see them in this very submission. Blockchain and AI were full of proponents making the argument that “all this energy consumption is good, actually, because it’ll increase investment in renewables”, but obviously what we’ve seen is an even bigger spend on fossil fuels which are exacerbating the problem. Believing the next promise from proven liars who make everything worse every time is not a smart strategy.

Anyway, I’m not entirely sure why I felt I should write to you specifically. I guess I once saw the possibility of something better and am hopeful that being respectfully called out might help you to reflect. Those of us capable of introspection sometimes need a little nudge to get back on track to the person we want to be. I know that’s true for me. Sometimes I help others, other times they help me. I hope this helps you.

▲dv_dt an hour ago | parent | prev | next [-]

You can find inference providers that are renewable powered

▲riedel an hour ago | parent | prev | next [-]

Hope for better regulations to set incentives?

▲newsclues an hour ago | parent | prev [-]

Nuclear power is the future.

▲dan_grec 42 minutes ago | parent | next [-]

Sweet! Power in 2046 at the earliest of we start today

That is simply not workable, and the suggestion needs to die because it sucks motivation from solutions that can work in a year or two.

▲dalyons 36 minutes ago | parent | prev [-]

In some alternate universe where cost doesn’t matter, sure.