[m-hodges]

> Even at $63 per share, we give SpaceX a lot of benefit of the doubt in two of the three scenarios, in which we assume the company can achieve a rapidly reusable Starship rocket enabling multiple launches per week and successfully commercialize data centers in space.

> In our downside scenario, orbital data centers won’t work or offer any advantage over terrestrial ones. We surmise that the company, having invested tens of billions to find this out, would cut bait on the project sometime around 2028, the way management walked away from plans to build multiple small-car factories at Tesla.

[kldavis4]

The orbital data center thing is really strange. Without some kind of massive technological breakthrough the physics of an orbital data center make no sense. Napkin math for a single data center including solar panels, radiators, compute and structure is somewhere on the order of ~6500 tonnes. Let's say Starship can put 100-150 tonnes into orbit per launch, that is somewhere between 40-65 launches for just one single orbital data center.

[overgard]

I'm pretty sure he's full of shit and has zero real intention of building orbital data centers. It makes for a good headline, which is what these AI companies are in the business of. The math doesn't math and the logistics of maintaining and building the datacenter seem insane.

[ericd]

They claim they want to get to a gigawatt per year by the end of next year, so we won't have to wait long to see.

[bandrami]

He also claimed in 2015 that nobody would be driving their own cars by 2018. Meanwhile I've seen a driverless car once in my life.

[ericd]

Sure, I think everyone knows he's not batting 1000. I've used Waymos and FSD, and they were both great, so I don't think we're too far off now.

[Sammi]

My man, it's 2026.

How long will you have to wait for the spacex claims to become "not too far off now"?

[TiredOfLife]

I have seen 3 people use iPhones in 18 years.

[hparadiz]

The video they posted today outlines their specs. They are gonna be smaller. About 100T. Based on Starlink architecture but with way larger radiators and solar panels. Apparently the design is simpler than existing Starlinks because it has fewer different parts.

Specs: https://x.com/SawyerMerritt/status/2064108916611420273?s=20

[amluto]

That cooling system is … aggressive. 110m^2 of radiator to dissipate 120kW average heat. That’s about 1.1kW/m^2, which works out to very close to 100C at the radiator under absolutely perfect conditions (absolute zero ambient temperature, emissivity 1, no adverse geometric effects).

Radiating into 300K ambient, it’s 134C. (300K ambient is about what you get if the sun is visible or if a large fraction of what the radiator can see is the Earth.)

You can slightly fudge the 300K case with a spatially or specially selective system, which would add weight and complexity. (Well, you can’t get rid of the problem of the Earth being warm by spectral selectivity — it’s the same spectrum as the radiator.) You cannot fudge the absolute zero case — the Stefan-Boltzmann law is extremely unforgiving. [0] And you need some headroom for the system that gets the cooling fluid to the radiator.

The thermal qualification temperature of an H100 is 87C and the associated HBM is 95C.

So I don’t see how this can work short of using higher-temp chips. I have no idea how SpaceX expects to source any such thing in any meaningful volume.

[0] This includes heat pumps. A heat pump makes it worse.

[hparadiz]

I assume the radiators will be in the shadow behind the solar panels and never facing the sun. That's seems like the easiest quick win.

[amluto]

This isn’t that big a win. If you put the radiator in the shadow of the solar panels then a good fraction (which depends on the geometry) of the radiated heat from the side facing the panels hits the panels and warms the panels, and the radiated heat from the back of the panels warms the radiator. And the panels themselves are less efficient when they’re warmer.

On the flip side of all this, the Starlink satellites work, and I would expect SpaceX to have some idea what they’re doing. Starlink satellites have largish power systems (smaller than these proposed AI satellites, but not outrageously so). They presumably turn a smaller fraction of their power into low-temperature heat: the ion thrusters and the various transmitters emit a good deal of non-thermal power. A good fraction of Starlink’s electronics way well function at rather higher junction temperatures than 100C. And I imagine that Starlink satellites are economical to operate at an average of much less than 100% power. So I don’t really know what’s going on. It’s plausible that Starlink gets away with a cooling system that’s not so great for a compute satellite.

[fc417fc802]

Is the equilibrium temperature in orbit not somewhere between freezing and room temperature? If so spreading the heat evenly across the solar panels must necessarily be sufficient. Issues only arise if you try to get away with a smaller radiator or expose the radiator to direct sunlight or commit some other own goal.

[orwin]

The radiators then need to be protected from earth's heat/light if you want to reach 0K. I agree with GP, it doesn't seem great.

[yread]

Maybe they count them twice? That would work out to 41C if both sides are in shade (or 69C if one is shaded and one on the sunny side)

[pryce]

I feel like we are going to relive the hyperloop-decline with the specs of proposed orbital datacentres. But the hyperloop was in one regard a success, in that it built temporary hype. Maybe it was even named after that.

[sylos]

The hyperloop was a huge success...in delaying actual transit options like trains.

[lelandfe]

What, you don't have a Tesla elevator on your street yet? https://www.youtube.com/watch?v=j70GvgXt-lE

[SwellJoe]

It also achieved Musk's actual goal: derailing and delaying mass transit initiatives, the California High-speed Rail project, in particular.

He knew it was bullshit. He's dumber than most people realize, but he's not dumb enough to actually believe all the bullshit he spouts.

[hparadiz]

He's sending the weirdest mixed signals. If we want to hit anywhere near 1 on the Kardashev scale we're gonna need a planet wide high speed rail system. They should just give the major tech companies naming rights to the stations for 35 years for like 2-3 billion a pop and scale up construction 5x.

I don't get the slow roll on this thing. The stations in LA and SF are gonna end up sort of like Tokyo station for the Shinkansen. A giant mall area with a huge amount of commerce. Why isn't this thing done yesterday?

[SwellJoe]

He doesn't actually want progress. He wants power and wealth for himself. Public transit is an equalizer, it serves normal people.

[nine_k]

> Why isn't this thing done yesterday?

Land rights. Environmental reviews. Various suits attempting to exert extra compensations for tangential (at best) issues. Basically the California around.

[ianburrell]

Planet wide high speed rail doesn't make much sense, it is better of medium sized trips. Flights are will always be better for intercontinental trips. Vacuum trains could be faster but would be so expensive to build. Issues would be big deal in evacuated tunnels.

CA HSR is slow partly because the state government has limited funds and giving out slowly. The federal government should have been contributing more. The other problem is it kept getting delayed from legal issues, mainly from acquiring land. Going forward is bad because they decided to do the easy part first and leave the hard tunnels for later.

[NortySpock]

Orbital ring would kickstart the space economy like nothing else, and is technically achievable with existing materials.

[dilyevsky]

Pinning CAHSR jobs program as hyperloop's fault is pretty funny. Maybe he did intend to upset it but since both turned out to be massive boondoggles doesn't seem like it worked - they managed to delay it themselves just fine.

[SwellJoe]

Sure, California is a failed state on a lot of fronts, but Musk literally said he was pushing hyperloop for that reason. https://www.fresnobee.com/opinion/editorials/article26445107...

Whether he was the only reason for high speed rail being derailed and delayed isn't really the point. He intentionally worked against mass transit, and lied repeatedly about hyperloop as part of that.

[NordStreamYacht]

I don't think he's dumb at all, if what you say is correct. He wanted to block something for his own benefit and he achieved it.

[bulbar]

I'm not convinced that's the case. I tend to think he just saw something he found cool and triggered his narcissism so he had to come up with something he thought was cooler. That something turned out to be BS and he just dropped the idea and moved on.

[SwellJoe]

He is ruthlessly cunning. Not unusual in sociopaths and narcissists.

[themafia]

> the California High-speed Rail project

The state government of California was doing a bang up job of destroying this project on their own. He didn't really need to help them. Moreover given the rise of remote work I'm not sure it has as much value as it would have had it been constructed when it was designed.

> but he's not dumb enough to actually believe all the bullshit he spouts.

Yet he's not afraid of the consequences either. This seems more telling to me.

[SwellJoe]

The consequence of becoming the world's richest man? Oh, the horror. The very wealthy do not face consequences for their actions.

He lies every day about his companies, and no one ever does a thing about it.

DOGE killed a few hundred thousand children for no reason at all; it didn't even save money. He did it purely for the joy of killing black and brown children in the global south. No one with any power to deliver has ever suggested he should face consequences for it.

Why should he be afraid of consequences? There will never be any for him, as long as he is protected by unimaginable wealth and a power structure designed around serving those with wealth.

[sumeno]

He has never faced an actual meaningful consequence, of course he isn't afraid of them.

[mayama]

After articles on how starship is never getting off launchpad, I'm sceptical of news regarding elon, pro and anti. Sure, things like FSD are hopium or scam, but not everything could be. Haven't seen hard science estimates on their new ai satellites.

[zarzavat]

The issue with AI datacenters in orbit is not science or engineering but economics. There is no practical reason why it can't be done, but why would anyone want to buy compute from such a datacenter when they could pay substantially less money for better capacity on Earth?

Orbit is a bad place for a datacenter. Your equipment will be hot and bombarded with radiation. You can't do any repairs. Even with reusable rockets you still have to pay for fuel to launch everything into orbit. You get a small benefit of more efficient power generation but it's not worth all the downsides.

If you want to have solar-powered AI datacenter that you can't service, you'd be way better off building the power generation in a coastal desert and building your datacenter underwater.

[tyre]

With the rockets, it was more, “wow that seems impossible, but incredible if you can pull it off versus existing options.”

With the space data centers, it’s more, “Okay yeah so let’s say you do…why?”

It just seems like a really hard thing to do when the available options are, like, the Chinese building data centers on the Tibetan plateau where it is cold, ample renewable energy, tons of land, and, like, oxygen.

[ianburrell]

150 kW is tiny, that is a single rack. Would have to launch a huge number of them to match the GW data centers being built. There is also a lot of overhead for each one.

If Starlink didn't exist, then it could make sense to put edge data centers in orbit. But Starlink means can put edge computing everywhere on the ground.

[asadotzler]

In 2025, SpaceX put 3,200 Starlink satellites into orbit with 122 Falcon 9 launches. That's about 2,000 tonnes. In about 2 more years, as the Starlink constellation reaches its licensed size limit, it should transition from growth mode to maintenance mode, freeing up over 1,000 tonnes of extremely reliable launch capacity.

What could that 1,000 tonnes per year get your in terms of sell-able data center capacity, and how does that compare to terrestrial build-out time frames?

[ac29]

According the the X post, each 150kW satellite is a bit over 2 tons, so a full year of launches would get you 75MW. Thats not much when AI datacenters are GW-class.

[skybrian]

Would it make any sense for it to orbit next to a Starlink satellite?

[Zigurd]

That projection depends on packaging and cooling breakthroughs that we would already be seeing in terrestrial data centers if they were practical.

[vagab0nd]

I think it comes down to scaling and removing the bottlenecks.

If you build one data center on earth, you did just that.

If you build one in space and make it work and cost effective, you can scale infinitely (which is why SpaceX is uniquely positioned), until you hit the next bottleneck (which is why Tesla is building a fab).

Tangential: If you play Factorio or Satisfactory, this is _all_ you do. Removing bottlenecks.

[lenerdenator]

> If you build one in space and make it work and cost effective, you can scale infinitely

"If we can just keep consuming massive amounts of capital, at some point, we'll turn a profit. Maybe even a big one."

It's not 2019 anymore, but this mindset acts like it still is.

SpaceX, along with its wanna-IPO classmates OpenAI and Anthropic, are going public because there's basically no more private investor money left and there's no political will (outside of whatever the hell is going on in Donald Trump's head) to lower interest rates to what they were in the golden age of blitzscaling.

They have to ask the public for more money. They just don't have enough humility to do it on decent terms, as evidenced by the IPO deals they're floating.

If this fails to produce what you're talking about - which is likely given things like "speed of radio transmissions", "thermal conductivity in a vacuum" and other things governed by those pesky laws of physics - the game's over.

[fc417fc802]

I think that's looking at it the wrong way. Just view it as a fleet of satellites with computers in them that exist for the purpose of selling capacity. I don't see any issues with the physical fundamentals but the economics in terms of dollars either per flop or per bit transferred is where I don't understand how it's supposed to work.

Satellites are really expensive and electricity on the surface is relatively cheap. What hypothetical prospective customers are willing to pay the necessary rates? I understand that bandwidth back to the surface is expensive but are we really expecting so much raw data to be trapped in space that it justifies sending computers into orbit in order to crunch it at the source?

[queenkjuul]

The physical fundamentals don't work. Radiating exhaust heat is exponentially harder in space than in the atmosphere.

[fc417fc802]

What is the equilibrium temperature of a panel of black sheet metal hanging out at an orbit of 1 AU? Because if the thing is solar powered then it cannot exceed that temperature on average (ie across the entire mass) due to energy being conserved. You can do goofy things like using solar power coupled with a heating element (such as a GPU) to shift more heat to one part of the panel or another (ie employ an insufficient heat spreader design) but you fundamentally cannot exceed the equilibrium temperature on average.

[queenkjuul]

Well the equilibrium temperature of outer space is roughly 0K.

So how do you remove megawatts of heat in 0K?

[numpad0]

There are people arguing it has to be a signal processing backend for a realtime radar system based on Starlink Direct-to-Cell antennas.

IMO, that lines up with a weird remark from Musk around stealth jet obsolescence few years ago, and makes sense.

[JoshTko]

It's a litmus test of who understand basic physics and thermodynamics

[wmf]

I see it as a hedge against NIMBYs. If terrestrial data centers literally can't be built then orbital is cheaper by definition. If not, just don't build them.

[Atotalnoob]

Worse, it’s “don’t worry, the next data centers will be in space, so it’s okay if the ones we are building right now destroy the environment, there won’t be anymore!”

It’s the exact tactic he used with California HSR

[wmf]

I don't see anyone making that argument and we shouldn't allow it. The "temporary" generators are bad enough.

[jiggawatts]

The actual “plan” is to just put a couple of GPUs in each Starlink satellite for inference.

They’re not launching something the size of a building!

[SwellJoe]

Which is irrelevant to SpaceX bottom line. Launching GPUs into space, whether by twos or by tons is not going to be a profitable venture.

It's just a wildly expensive bad idea, and it is obvious to anyone that understands how much it costs to put things in space. A man who runs a rocket company knows how much it costs. He knows it's dumb as hell. But, he also knows the average investor is even dumber, because TSLA still trades at ~357 PE.

[kldavis4]

That just doesn't seem to justify anywhere near the valuation though does it?

[jiggawatts]

Disclaimer: I suspect "data centers in space" is just blatant market manipulation by Elon, exactly the same as he'd done dozens of times before with self-driving, etc.

Having said that... it might justify the valuation if every other data centre build-out gets blocked by insufficient power supply.

Might.

[fc417fc802]

Bear in mind that you could take the solar panels that you would have launched into orbit, use the launch costs to purchase additional solar panels, batteries, and some additional land, and build a data center plus solar farm plus grid scale battery bank on the ground for probably (almost certainly) less than the combined launch costs.

I think the only way compute in space works out financially is if there's so much bulk data already up there that downlink bandwidth becomes a serious bottleneck.

[jbxntuehineoh]

shhhhhh the line WILL go up, whether you like it or not

[credit_guy]

> Let's say Starship can put 100-150 tonnes into orbit per launch, that is somewhere between 40-65 launches for just one single orbital data center.

Ok. And why does it follow from this that the physics of an orbital data center makes no sense?

[adgjlsfhk1]

Because 50 rocket launches seems hard to make cheaper than a building.

[tekla]

The cheapness of the building is not the problem that is being dealt with.

The construction of the building w/ zoning and the political fights and utilities arguments and years of time and whatever is what is being dealt with by putting it in orbit.

[ink_13]

I bet you could build a fully off-grid DC that didn't guzzle water for less money than launching a rocket today, and there would be thousands of people lining up down the block to sell you more than enough land to do it a thousand times for cheap. This just isn't happening yet because it is still quite possible to build on-grid DCs supplied with fresh water for cooling despite what some people would have you believe.

And then when you build on the ground you could even send people to it to swap hard drives and GPUs when they inevitably fail or upgrade them to keep them current. At lower rates of failure than they would in space because we have a planetary magnetosphere protecting us from cosmic rays.

[kaashif]

It would be extremely funny if putting things in space is cheaper and faster than dealing with zoning and local politics.

It implies that China, which can cut through much red tape and has great (and improving) utilities and infrastructure, lots of energy, can just build normal datacenters and save the cost of dozens of space flights.

[eeixlk]

Building in space adds engineering complexity. Heat output requires radiators. Radiation damages equipment without shielding. Repairs require a spacewalk. All this for what, better solar panels? If you think local zoning challenges are an obstacle, then you have 194 other countries to build in. Going to space is probably the worst option. Even putting a data center on a boat is probably a better choice and that still sounds like a stupid idea.

[credit_guy]

Why? It seems to me to be much cheaper than a building. Musk's aim is for the cost of one Starship launch to be under $10 million in the long run. 50 rocket launches would be less than half a billion dollars. Can you get a datacenter built on the ground for half a billion dollars?

[sumeno]

Half a billion in JUST the rocket launches, you've still got to build the actual data center too.

That's assuming he can get launches down to $10 million which I'm sure will be right behind his flying roadster and cybertruck boat

[adgjlsfhk1]

The comparison isn't half a billion for a datacenter, it's half a billion for a big shed (and some solar panels). The problem of space based data centers is that the only effort they save you is building a big shed to put the GPUs in (and potentially cheaper power at the cost of more expensive cooling)

[Symbiote]

Using easily available numbers, a high quality distribution center building (heavy duty concrete floors, HVAC etc) is $100 per sqft.

A very large 1,000,000 sqft data centre is therefore $100 million.

[adammarples]

You haven't built a data center for half a billion mate, you've just launched rockets. You still have to build the data centre! Only not from bricks, it's going to be complex modular units made of radiators and solar panels and you have to buy the gpus too, and the whole thing will be bombarded with cosmic background radiation but you can't replace or repair any part of it.

[LearnYouALisp]

Are you for real? HN yep

[credit_guy]

Yes, I am for real. The Colossus datacenter that xAI built cost $7 billion.

[1] https://www.globaldatacenterhub.com/p/can-elon-musks-7b-colo...

[Terr_]

No, you've got the burden of proof backwards: We haven't even begun to talk about maintenance issues, space-proofing the equipment, power-generation, cooling, etc.

[datadrivenangel]

They'll do smaller satellites, and if they can get an order of magnitude improvement on cooling and power generation to weight ratios, it'll be close to the same order of magnitude cost-wise as a terrestrial datacenter. 10x the cost is likely not enough to make it all work, so it's a bad idea, but it's possibly less bad than using lead ballast for starship launches... maybe.

[amluto]

I still don’t see it. The equipment is very, very expensive, is subject to shortages, and has residual value after five years. But these very low Earth orbit satellites literally burn up when they run out of fuel for station keeping.

Even if you pay a bit more to position them in a higher, more stable orbit, you still can’t physically get to them to repurpose equipment.

Check this datacenter gadget out:

https://www.marvell.com/blogs/sustainable-computing-with-cxl...

That trick is a complete nonstarter if the modules you’re trying to reuse are in space.

[datadrivenangel]

Even with order of magnitude improvements on the state of the art they're still looking at ~10x the price per installed and powered unit of compute.

[toomuchtodo]

“Supervised Full Self Driving.” Buy the hope, sell reality.

[jmalicki]

Is there some reason significant improvement on cooling and power generation to weight generation ratios is even remotely possible?

[datadrivenangel]

Better integration allows for a lot of potential improvements, so if the cooling and power and compute are tightly integrated and use structural materials it may be plausible.

Still not going to be cheaper, but if you get moderately optimistic it's cheap enough to maayyyybe start looking like a source of demand for starship if you can find suckers (government) to pay 10x the price per compute.

[ink_13]

> In our downside scenario, orbital data centers won’t work or offer any advantage over terrestrial ones

Technologically, will they work? Shovel enough money at the problem and there's no reason they shouldn't. Rack up some GPUs in a shipping container, attach a couple of solar panels to the outside, zip-tie a Starlink to the door. Except for the little problem of heat. Space isn't cold, it's more of a giant Thermos bottle, and they gotta park in full sunlight, which is famously the only thing supplying the heat that keeps the entire planet alive. Good luck to them with their radiator setup.

But economically? There's no chance. Too expensive to get into space, too many auxiliary systems required, no maintenance is possible, and, given the lead time of prepping payloads for launch versus the rate of new developments in AI hardware, likely badly behind the times if not outright obsolete on the day of liftoff.

There is no shortage of land despite what some in the media would have you believe. Those same solar panels work at ground level. DCs don't have to be water guzzlers or grid destabilizers. They'd be better off by every metric solving terrestrial problems for the same money.

Of course they won't work. It's ludicrous that we have to even entertain the thought of otherwise.

[protocolture]

Do they have a projection for what the company will be valued when he gets to mars by 2017?