From the article:
We assumed SpaceX’s Starship was used at a highly optimistic launch cost per kilogram of US $44
Or float it in the ocean.
Well, floating data centers have been proposed as well. Given that the surface solar collectors on land or ocean would have to be around 5X larger, and have bigger batteries to compensate for really bad weather and night, I’d think it’d be cheaper and probably more cost effective to build it on land.
On top of all that, the most optimistic outline for launches in the Ars Technica article is 10 launches per day just for this satellite cluster. The pessimistic one is 40+ launches per day. I imagine it will be long time before Starship can support either of those cadences.
Also, salt water and technology do not get along. Maintenance costs would be much, much larger on a floating data center. Especially since unlike a ship instead of just floating in the salt water, it would be constantly sucking in large quantities of that salt water for cooling.
I saw that, but I’m not sure where he got $44/kg in the first place. Searching around I see aspirational targets of $100-200/kg for SpaceX Starship. I don’t follow SpaceX so I have no idea what the real number is.
But more importantly he doesn’t show how he is comparing payload launch costs ($44/kg) against an terrestrial power rates ($0.20/hr). They’re completely different concepts. His terrestrial number is right in the middle of Google’s range of $810–7,500/kW/y, but what did he estimate for the space datacenter power rates?
And as I said, he didn’t seem to account for amortization, which is a factor of 1/N years.
I’m ignoring that it doesn’t make sense to launch a rack of H100s because that’s more or less what NVIDIA did.
Underwater data centers make more sense than space data centers.
MIcrosoft’s Project Natick had a successful two year run. And Hailanyun / HiCloud Technology seems to have built out a significant data center in China. These address cooling more than leveraging solar power, but cooling and power are related.
This is a decent, recent article on the current state. It’s not a very technical article though.
No, this is a transparent snow job. The actual feasibility doesn’t deserve any more intellectual seriousness than that.
People are tempted to try and backfill the arguments pro and con in a vacuum because it’s intellectually interesting but the clean test is simply whether the proponents are attempting to advance arguments that clarify or muddle the debate.
Let’s take the Google Paper mentioned by @CaveMike as an example because on the surface, it looks like a serious attempt to clarify the debate. Read the actual paper though and hone in on the total time it spends talking about thermal issues:
First, all the language that double dog dare super promises we are definitely going to cover thermal management:
This research initiative is focused on addressing several of the major ingredients required: power generation, high-bandwidth, low latency communication between chips, radiation tolerant compute, a thermal management system, and a data link to ground stations.
Cooling would be achieved through a thermal system of heat pipes and radiators while operating at nominal temperatures.
Other significant challenges such as on-orbit reliability and repair, high bandwidth ground communications, and thermal management are also discussed in this paper
Future space-based experimental milestones should also involve solutions for thermal management
And then, the total sum conversation about thermal management:
Effective thermal management is a critical optimization challenge for power-dense TPUs operating in a vacuum. Advanced thermal interface materials and heat transport mechanisms, preferably passive to maximize reliability, are essential to efficiently move large heat loads from the chips to dedicated radiator surfaces.
Check yourself if you don’t believe me but this isn’t some accidental oopsy, gee, let’s try harder next time. This is a deliberately engineered artefact designed to serve a specific purpose. It’s entirely reliant on convincing a bunch of people that there exists what looks like a super serious paper giving the problem the intellectual weight it deserves and relying on people being too lazy to look into the actual document to assess, lest they realize how clownish of a farce it is. Everything else involved in the paper is pure padding to make it harder to detect that it handwaves what every serious proponent has always established as the greatest implausibility.
Thus, we should be analyzing Space Data Centers from the perspective of other transparent snow jobs: NFTs, Hyperloop, uBeam, Gamestop short squeezes, the tulip mania, etc.
That it is a transparent snow job is not super interesting. The interesting thing, based on other transparent snow jobs is why people are trying to push this transparent snow job at this current point in time.
But one tactic of all transparent snow jobs is that the volume of ancillary conversation becomes a key KPI. Take NFTs for example, as long as people had actual debates about whether NFTs were a transparent scam or the undiscovered cheat code to future generational wealth, it allowed the debate to become muddy enough that people were misled into thinking “Well, there might be some truth to both sides”. It didn’t require everyone to believe in NFTs, just enough so those with a vested interested benefited.
I’ll leave who is pushing this and why they might be pushing this as an exercise for the interested reader but stop treating it as having any more intellectual weight than any of the other aforementioned transparent snow jobs had in retrospect.
My two hypotheses so far:
- Get away from any attempt at content moderation, pollution limitation or laws in general.
- For SpaceX : Make the company “too big to fail” in multiple segments: space access, defense, ISP, AI, data centres, news/propaganda (Twitter), and (since the IPO) a significant portion of everybody’s pension fund. That way, if it fails financially, lawmakers will trip over themselves to bail the whole thing out with public money. If it gets fined or somehow restricted by the laws or the courts, in the U.S. or elsewhere, it will be defended by U.S. politicians of both major parties.
Wow, yeah. Advanced thermal interface materials and heat transport mechanisms already exist, and are already used in all computer systems. That’s the easy part. There’s no point in even mentioning those except to distract away from the real problem of what are you transporting the heat to.
Well sorta.
We are presently using the affordable edge, not the bleeding edge, of thermal transport tech in our ground-based high tech equipment. That’s substantially a tautology given the way progress works in the IT biz.
And it’s only by moving first the bleeding edge, and later the affordable edge, upwards and onwards that high power density orbiting anything becomes possible.
What the article is in effect saying is that “We’ll need the best, and that best will have to be better than current best.” It’s an anodyne call to nebulous action. A variation on “Tech will find a way!”
But it’s also not wrong.
Except that current thermal bonding and heat transfer tech already isn’t the bottleneck. I mean, sure, it’d be nice if it were even better, but when people say that “cooling is a problem that needs to be solved”, that’s not what they’re referring to.
There’s cooling and then there’s cooling. You need to get 100% of the heat generated inside your equipment sent to 100% outside of your equipment.
Whether that “equipment” is layer 27 of a chip, the exterior surface of multi-chip packaged IC, or the outside of your building or space station. The heat’s not thrown “away” until it’s all the way “away”.
Right now we do the first few centimeters really really good. The rest is harder. Much harder. Access to terrestrial or oceanic volumes of water make that a lot easier. Sadly those things are absent in orbit.
My employer sends old servers and other obsolete electronic hardware to a company that either resells or recycles it. So we are supposed to be getting money back and the stuff stays out of the landfill. That would not be possible with orbiting data centers.
A while ago I did a back of the envelope check on limitations on Starship launches. The cost of fuel I got was about $25 per kg to LEO. From memory that cost was a pretty even split between LOX and liquid methane.
Airlines run with fuel accounting for about a third of running costs. So the mooted costs to orbit aren’t insane. Just highly improbable. Aircraft last for tens of thousands of cycles. Falcon-9 boosters are clocking up near a hundred cycles. Difficult to imagine Starship ever reaching aircraft cycle numbers. They will however be cheaper than commercial passenger aircraft. But very unlikely to be enough cheaper.
Elon talks of thousands or tens of thousands of Starship launches a year. Really. It strikes me that right now, the supply chain for LOX is going to pace launches for a while. Down at Boca Chica they are building an air separator plant that will deliver LOX. So they might get the price down a bit, but they are going to be rate limited even so. Eventually the cost and rate of delivery of LOX is limited by power. Elon’s vision of launch cadence is going to need a lot more capacity than that plant.
Methane is easier. We already ship LNG around the planet in large quantities. The logistics of supply are well sorted. Whether the Raptor engines can run on commercial LNG, or need purified methane is another question. If you believe Elon’s wilder statements on the number of Starship launches, SpaceX would account for up to 5% of the entire USA consumption of LNG.
Hundreds of dollars per kg to LEO for payloads that launch in such large numbers that all the handling and integration costs are fully amortised might be possible. So Starlink, the mooted server sats, or other identical satellites launching by the thousand.
Everything depends upon serious volume to make the numbers work. Even Starlink has a question mark over it. There comes a point where the market is saturated for any service. And it isn’t just SpaceX in that market. Whether Starlink can pay its way isn’t clear, and if that falters, the house of cards will be looking very shaky. Other than taking customer’s money for a self driving system that was never delivered, it isn’t clear any of Elon’s ventures have ever made money from an AI offering. This appears true for all the other players as well. Something significant will need to change, and change soon.
It’s not the same thing as an orbiting data center but I wonder about the power requirements for the Starlink satellites.
I think this means strip off the gold traces, dispose of hazardous chemicals, and landfill the rest.
What happens to electronic waste would make foe a good FQ itself.
My point is that for these orbiting data centers, nothing is recovered.
There’s a Canadian comedy group The Frantics and a quote from their album Boot to the Head would be appropriate here… that solving FSD (or any AI) “…is not a path to a door, but a road leading forever to the horizon”.
There will always be a stranger edge case. To be fair, we expect AI to be 100% when no human comes close to that.
But I was reading something on another site that due to the small volume of SpaceX shares actually listed, the percentage of shares in any index fund would be substantially less than 1%. That index funds go on the total value of shares in play, not those also locked against being sold, and the total shares available to the public from the IPO was far less than $100B.
In a few months more shars will come on the market (as do-not-sell rules expire and then as employees vest) but perhaps by then, based on current trends, the share values will reach their proper price.
Presumably orbital data centers will also stay out of the landfill. We just need a way to recover their elements from the atmosphere.
I saw an article the other day on China having built an extremely powerful microwave beam weapon built in their desert that could be aimed at satellites, presumably to fry their electronics; although some were skeptical about the power output claimed.
ISTM the business’s real concern about non-recyclability is the loss of revenue when the company can’t sell their obsolete gear.
Back in the day there was a robust used market for corporate desktops, laptops, and for server infrastructure. When I built my biz in the early 2000s we equipped the whole place with last-gen servers, switches, etc. for a fraction of what new gear then cost.
I don’t know how much that kind of used market still exists, and especially not for specialized datacenter and specifically AI datacenter gear. But for darn sure anything reentering the atmosphere won’t be resold for reuse.