The quoted "1 TW of photovoltaic cells per year, globally" is the peak output, not the average output. They're only about 20% higher peak output in space… well, if you can keep them cool at least.
That wasn't the original question. The head of this thread was quoting Musk's claim, which I repeat here:
> it is possible to put 500 to 1000 TW/year of AI satellites into deep space
This is 500-1000 times as much as current global production.
Musk is talking about building on the Moon 500-1000 times as much factory capacity as currently exists in aggregate across all of Earth, and launching the products electromagnetically.
Given how long PV modules last, that much per year is enough to keep all of Earth's land area paved with contiguous PV. PV doesn't last as long in space, but likewise the Moon would be totally tiled in PV (and much darker as a consequence) at this production rate.
In fact, given it does tile the moon, I suspect Musk may have started from "tile moon with PV" and estimated the maximum productive output of that power supply being used to make more PV.
I mean, don't get me wrong, in the *long term* I buy that. It's just that by "long term" I mean Musk's likely to have buried (given him, in a cryogenic tube) for decades by the time that happens.
Even being optimistic, given the lack of literally any experience building a factory up there and how our lunar mining experience is little more than a dozen people and a handful of rovers picking up interesting looking rocks, versus given how much experience we need down here to get things right, even Musk's organisation skills and ability to enthuse people and raise capital has limits. But these are timescales where those skills don't last (even if he resolves his political toxicity that currently means the next Democrat administration will hate his guts and do what they can to remove most of his power), because he will have died of old age.
> Clearly this person was referencing a financial efficiency predominantly through uptime.
I read the person you are quoting differently, as them misunderstanding and thinking that the current 1 TW-peak/year manufacturing was 1 TW-after-capacity-factor-losses/year.
The 1TW is the rated peak power output. It's essentially the same in space. The thing that changes is the average fraction of this sustained over time (due to day/night/seasons/atmosphere, or the lack of all of the above).
It's still the same 1TW theoretical peak in space, it's just that you can actually use close to that full capacity all the time, whereas on earth you'd need to over-provision substantially and add storage, so 1TW of panels can only drive perhaps a few hundred GW of average load.
The dominant factor is "balance of system" aka soft costs, which are well over 50%.[0]
Orbit gets you the advantage of 1/5th the PV and no large daily smoothing battery, but also no on-site installation cost, no grid interconnect fees, no custom engineering drawings, no environmental permitting fees, no grid of concrete footers, no heavy steel frames to resist wind and snow loads. The "on-site installation" is just the panels unfolding, and during launch they're compact so the support structure can be relatively lightweight.
When you cost building the datacenter alone, it's cheaper on earth. When you cost building the solar + batteries + datacenter, it (can be) cheaper in space, if you build it right and have cheap orbital launch.
I do say it's predicated on cheap orbital launch. Clearly they expect Starship to deliver, and they're "skating to where the puck will be" on overall system cost per unit of compute.
But yeah, I didn't include that delivering all that stuff by truck (including all the personnel) to a terrestrial PV site isn't free either.
Yeah, soft costs like permitting and inspections are supposedly the main reason US residential solar costs $3/watt while Australian residential solar costs $1/watt. It was definitely the worst and least efficient part of our solar install, everything else was pretty straightforward. Also, running a pretty sizable array at our house, the seasonal variation is huge, and seasonal battery storage isn’t really a thing.
Besides making PV much more consistent, the main thing this seems to avoid is just the red tape around developing at huge scale, and basically being totally sovereign, which seems like it might be more important as tensions around this stuff ramp up. There’s clearly a backlash brewing against terrestrial data centers driving up utility bills, at least on the East Coast of the US.
The more I think about it, the more this seems like maybe not a terrible idea.
So far most of the datacenters are built in very convenient places and people will start to build them in inconvenient places like Sahara or Mongolia way before they will building them in space
Maybe. But for SpaceX, it’s more aligned with what they’re trying to do to just learn to manufacture them at scale and lob them into space. And one of the benefits there is the uniformity of it - they can treat them all the same, rather than dealing with a bunch in different geographies with different power issues, governmental issues, etc. That’s been one of the major issues with rolling out solar. In the US, there are >20,000 AHJs, each with different rules and processes. A huge constellation of satellites seems easier to reason about and build systems to maintain en masse, because it’s more uniform.
I’m not saying this is a good idea. I’ve got a lot of SpaceX stock, and I wasn’t really happy to hear the news, this is mostly me trying to understand why they might think this is a good idea, and brainstorming out loud, with a dash of coping. Seems most here think that it’s just stupid, but then, most commenters thought Starlink was stupid, iirc, and that turned out to be wildly wrong. But it might also just be stupid this time.
Do you imagine there'd be less red tape involved in launching multiple rockets per day carrying heavy payloads?
Like this argument just gets absurd: you're claiming building a data center on earth will be harder from a permitting perspective than FAA flight approval for multiple heavy lift rocket launch and landing cycles.
Mining companies routinely open and close enormous surface area mines all over the world and manage permitting for that just fine.
There's plenty of land no one will care if your build anything on, and being remote with maybe poor access roads is still going to be enormously cheaper then launching a state of the art heavy lift rocket which doesn't actually exist yet.
> Ok, why are so many being built in Northern Virginia, rather than in the middle of nowhere where there will be no backlash?
Right? So if that's the case why would putting them in Space, far less accessible in every conceivable way, with numerous additional expenses and engineering constraints, be cheaper?
Yeah, I don't know if it wins on cheaper, even with $2M fully reusable starship launches. Maybe rollout speed vs piecing together BD deals with a bunch of different infra providers? The expansion of the grid is going to be hamstrung until congress finally passes energy permitting reform, which they've tried and failed at repeatedly. But they could do non-interconnected microgrids in the desert like Redwood Materials has been trying.
Maybe there's a concentration in VA because there's a set of deals/procedures in place with infra providers there that make it easy to scale up, similar to how DE has well developed corporate infrastructure, so everyone incorporates there. But that stops when the area hits its limit in power provision (which seems to be happening right now). In which case, being able to do this yourself end to end by putting this stuff in space with your own power generation makes it the ultimate scale-up opportunity - no real limits on space or power availability, so once you get that method down, you can mass-scale and get great economies of scale. Maintenance isn't a thing, these will be disposable.
I think that's it, money's not the limiting factor if they can pitch this successfully, which I think they will. They want massive scale without the constraints you hit when doing it on earth. I think he's aiming for scale that we haven't seen in DCs on earth.
>why would putting them in Space, far less accessible in every conceivable way, with numerous additional expenses and engineering constraints, be cheaper?
Because mostly AI is power hungry, and in space you need 1/5th the solar panels, almost no solar panel support structure, and almost no daily load shifting batteries.
As it turns out, this matters more than launch costs or cooling radiators or radiation bit flips. This is why you need to do the math instead of assuming you know what the solution looks like.
Just based on weight, looks like a Block 4 starship should be able to bring up ~150 30 panel pallets of 550W panels, about 2 MW. They're trying to get a starship launch down to $2M with full reuse. GPU DCs are frequently in the neighborhood of 500GW, so maybe 250 launches for just the power generation, or $500M? And then there's radiators, so let's say $1B for launch of power and heat dissipation. For comparison, 500MW of H100 machines retails for >$10B, and the launch cost for those shouldn't be too bad compared to the power, since they're more value dense. And then there's land and ongoing power and cooling spending for the terrestrial version, which you don't have for the space version. So actually, doesn't seem terrible economically? This is obviously very back-of-the-envelope, and predicated on the optimistic scenario for starship launch cost.
You cannot just block out the mass of stuff and declare it'll cost exactly the launch cost, nor can you take the cost of current datacenter servers and go "they'll definitely cost the same to put in space".
A regular set of servers will straight up be destroyed if put on a rocket and launched into space: the motherboards and PCBs aren't mounted or rated to survive the vibration. The connectors and wiring isn't rated for that vibration. Sure, some probably make it, but you will lose machines from just launching them alone. Any electrolytic capacitors in there? If your system exposes them to vacuum or even just low pressure, then those likely die too. Solar panels? We can launch them obviously, there's a reason people send up expensive solar panels: because you're doing a lot of work making sure they'll physically survive the launch.
So of course, now you have to build a space-rated server frame, PCBs and GPUs. You ain't going to buying bulk H100's from Nvidia. And you have to package and mount it to get it both survive the launch and physically fit into the payload bay. Then you have to add a deployment system for it, sensors etc. And then you have to add an assembly system, because if it doesn't fit in one launch (you're proposing 250+ launches for power alone) then all of these systems need to be assembled in orbit. How are they going to be assembled? How are they going to be maneuvered? Even if you could rendezvous accurately with the construction orbit, we're talking months of drift from every little thing knocking stuff around, putting it into a spin, etc.
So either each of these is now a fully contained satellite, complete with manoeuvering system and power, or you're also needing to develop a robotic assembly system - with power and manoeuevering in order to manage and assemble all this.
And let's not forget mission control: every single one of these steps is incurring a bunch of labor costs to have people manage it. And not cheap labor costs: you're going from "guys who roll racks in and plug stuff in and can be trained up easily" to "space mission control operators".
Is this doable? Probably. Is this going to be in anyway cheaper then Earth? Not in the slightest, and it's not going to be close.
Heh relax, it’s back of the envelope to try and see if it’s anywhere in the ballpark, and if they get launch to be cheap enough, it seems like it might be. You don’t assemble one massive thing in space, it’s a bunch of disposable individual sats with laser interconnects, like Starlink.
This is really underselling it tbh. Any land that's growing corn in a developed country is likely top 1% of land on earth. Half of the earth is desert and tundra. Which is still incredibly easier to work with than space because you can ship there with a pickup very cheaply. Maybe when nevada and central australia are wall-to-wall solar panels we can check back on space.
The Technology Connections Youtube channel recently did a great video arguing pretty convincingly that the land used to grow corn for cars would be vastly more efficiently used from an energy perspective if we covered it with solar panels.
You just have to remember, most of these people live in high density regions and have little comprehension about how much surface area humanity truly occupies... And that isn't even accounting for offshore constructs.
Realizing the impracticality of it (and that such approaches often collapse under the infeasibility of it) ... wouldn't it be better to... say... cover the Sahara in solar panels instead? That's gotta be cheaper than shipping them into space.
From an engineering perspective, with today’s costs, yes. But don’t forget the political complications of dealing with all those countries that own the Sahara, that’s going to come at it’s own cost.
I think this is all ridiculous, to be clear, but re: this problem couldn't the radiators in theory be oriented so that they vent in opposite directions and cancel out any thrust that would be generated?
Solar modules you can buy for your house usually have quoted power ratings at "max STC" or Standard Testing Conditions, which are based on insolation on Earth's surface.
STC uses an irradiance of irradiance 1000W/m2, in space it seems like you get closer to 1400W/m2. That's definitely better, but also not enormously better.
Seems also like they are rated at 25C, I am certainly not a space engineer but that seems kind of temperate for space where cooling is more of a challenge.
Seems like it might balance out to more like 1.1x to 1.3x more power in space?
Satellites can adjust attitude so that the panels are always normal to the incident rays for maximum energy capture. And no weather/dust.
You also don't usually use the same exact kind of panels as terrestrial solar farms. Since you are going to space, you spend the extra money to get the highest possible efficiency in terms of W/kg. Terrestrial usually optimizes for W/$ nameplate capacity LCOE, which also includes installation and other costs.
For one or a few-off expensive satellites that are intended to last 10-20 years, then yes. But in this case the satellites will be more disposable and the game plan is to launch tons of them at the lowest cost per satellite and let the sheer numbers take care of reliability concerns.
It is similar to the biological tradeoff of having a few offspring and investing heavily in their safety and growth vs having thousands off offspring and investing nothing in their safety and growth.
I think it's because at this scale a significant limit becomes the global production capacity for solar cells, and SpaceX is in the business of cheaper satellites and launch.
You don't even need a particularly large scale, it's efficient resource utilization.
Humanity has a finite (and too small) capacity for building solar panels. AI requires lots of power already. So the question is, do you want AI to consume X (where X is a pretty big chunk of the pie), or five times X, from that total supply?
Using less PV is great, but only if the total cost ends up cheaper than installing 5X the capacity as terrestrial PV farms, along with daily smoothing batteries.
SpaceX is only skating to where they predict the cost puck will be.
That's still a smaller ratio than the ~4X gain in irradiance over LEO. But if you're doing it at scale you could use orbital tugs with ion drives or something, and use much less fuel per transfer.
It's probably not competitive at all without having fully reusable launch rockets, so the cost to LEO is a lot lower.
8 tons over 22 is a little over 1/3rd the original payload to LEO. If 4x the solar generation potential (not irradiance - the sun is not 4x brighter in space at Earth's orbital radius) is the reward, that's putting an incredible premium on a 3x multiplier on launch costs per kg (at minimum - likely higher, you're also inheriting a worse radiation environment).
But those two parameters are not equals: 3x the cost per kg is a much higher number then 4x the solar power.
Here in Maine in the depths of winter (late December), 1 m^2 of ground can collect 4 kwh per day (weird units).
That's why people are trying to build solar here. Our power is expensive due partially to failing to build basically any new generation, and some land is very cheap, and the operational cost of a solar farm is minuscule.
Solar farming is basically an idle game in real life and my addiction is making me itchy.
You can overprovision, and you should with how stupidly cheap solar is.
That we aren't spending billions of Federal dollars building solar anywhere we can, as much as we can, is pathetic and stupid and a national tragedy.
We got so excited about dam building that there's no where to build useful dams anymore, and there is significant value to be gained by removing those dams, yet somehow we aren't deploying as much solar as we possibly can?
It's a national security issue. China knows this, and is building appropriately.
The southwest should be generating so much solar power that we sequester carbon from the atmosphere simply because there is nothing else left to do with the power.
> And then there’s that pesky night time and those annoying seasons.
The two options there are cluttering up the dawn dusk polar orbit more or going to high earth orbit so that you stay out of the shadow of the earth... and geostationary orbits are also in rather high demand.
I grew up on a rural farm in California with a dial-up connection that significantly hampered my ability to participate in the internet as a teenager. I got Starlink installed at my parents' house about five years ago, and it's resulted in me being able to spend considerably more time at home.
Even with their cheapest home plan, we're getting like 100 Mbps down and maybe 20 to 50 up. So it's just not true at all that you would have connections that are a megabit or two per second.
That's not what I'm suggesting. The post says "deep space". If you're going to try to harvest even a tiny percentage of the sun's energy, you're not doing that in Earth's orbit. The comparison is a webcam feed from Mars.
The intractable problem is heat dissipation. There is to little matter in space to absorb excess heat. You'd need thermal fins bigger than the solar cells. The satellite's mass would be dominated by the solar panels and heat fins such that maybe 1% of the mass would be usable compute. It would be 1000x easier to leave them on the moon and dissipate into the ground and 100000x easier to just keep making them on earth.
> There is to little matter in space to absorb excess heat.
If that were true the Earth would have overheated, molten and turned to plasma long ago. Earth cools by.... radiative cooling. Dark space is 4 K, thats -267.15 deg C or -452.47 deg Fahrenheit. Stefan-Boltzmann law can cool your satellite just fine.
> You'd need thermal fins bigger than the solar cells.
Correct, my pessimistic calculation results in a factor of 3,...
but also Incorrect, there wouldn't be "fins" thats only useful for heat conduction and convection.
That's pretty much a solved problem. We've had geostationary constellations for TV broadcast at hundreds of megabytes for decades now, and lasers for sat-to-sat comms seems to be making decent progress as well.
> it is possible to put 500 to 1000 TW/year of AI satellites into deep space, meaningfully ascend the Kardashev scale and harness a non-trivial percentage of the Sun’s power
3. vLLMa exist and take video and images as input.
4. When a new model checkpoint needs to go up, are we supposed to wait months for it to transfer?
5. A one million token context window is ~4MB. That's a few milliseconds terrestrially. Assuming zero packet loss, that's many seconds
6. You're not using TCP for this because the round trip time is so high. So you can't cancel any jobs if a user disconnects.
7. How do you scale this? How many megabits has anyone actually ever successfully sent per second over the distances in question? We literally don't know how to get a data center worth of throughput to something not in our orbit, let alone more than double digit megabits per second.
and, of course and inter-satellite comms and earth base station links to get the data up and down. Starlink is one thing at just above LEO a few hundred km and 20km apart, but spreading these around 10s of thousands of km and thosands of km apart is another thing
Looks like they are distributing a dividend to shareholders (14$, well below $25), and then diluting everyone.
Has anyone heard a bull case on being a minority shareholder in Scale.ai after this deal? Seems a bit like late stake investors are getting thrown under a bus.
I don't think that's the case here. Windsurf wasn't leading the agentic coding market. They were doing a decent job but others are bigger. Cursor has the brand recognition and Claude is getting a lot of recognition too. MS has github copilot which is still a good brand and Google has been catching up with Gemini.
OpenAI has a new thing called codex but it isn't very good yet. I tried it and it's super flaky. Lot's of errors and it gets stuck when that happens. OpenAI needs something good urgently because agentic coding is the key AI feature right now and the blue print for non coding agentic solutions later. Cursor is probably too expensive currently and windsurf looks like their models are a bit better.
So, OpenAI gains something they don't have: a credible developer option with an active user base and some core IP in the form of training data and know how as well as custom models that they can fold into openai.
3 billion is a lot but not if you consider that world + dog in the enterprise world will be spending big time on AI subscriptions for their developers. This stops being optional in 2025. Millions of developers will be on paid subscriptions permanently very soon. If you start a new job you can expect to get a laptop and a paid subscription to whatever is the agentic coding tool of choice in your new company.
OpenAI wants double digit percentages of that revenue. 1M users paying something like 50$/month would amount to 600M revenue per year. I think the prices will go up and the amount of active users as well. Reason: as these tools are getting better they start saving non trivial amounts of engineering time. At that point you have to value the tool in terms of developer cost. Not 1 to 1. But it's worth a sizable portion of that.
I work in a small startup as the CTO. This is an no-brainer for us. We're cash strapped so we only spend on important things. This would be one of those things. We're doing things I previously would have needed to expand the team for because I would have had no capacity to do those things in the current team. So, in terms of value for money spending on these tools is easy to justify.
I get lots of people are skeptical about AI stuff here. But I would say that a lot of those people suffer from a short term focus and bias. Three years ago none of this stuff existed. Now it's a multi billion$ market that is set to grow rapidly. Stuff is getting better at a very rapid pace. Just stating facts here. 3 billion is a bargain if Openai can make this acquisition work for them. They are buying time to market here. They don't have a year to figure it out. In a year or so this market will be carved up and locked into hard to change year long SAAS contracts. At that point getting people to switch tools will get harder and harder.
> OpenAI has a new thing called codex but it isn't very good yet. I tried it and it's super flaky. Lot's of errors and it gets stuck when that happens.
I agree with this, not sure the experience of everyone else but I felt like Claude Code is more useful.
Meanwhile, I'm keeping tabs on Aider and open-codex, what other options are there?
Thanks for mentioning open-codex. Did not notice that there is a codex fork which is open to other models (update: totally missed that original codex allows that too now). How do you like it? Especially in comparison to Claude Code?
Thanks for answering! I also skipeed that the original codex now allows for other models (perhaps they pulled the open-codex part???). To be fair, I prefer Claude Code to both codexes.
Not replace but it allows me to scale what we do for things we previously would have dropped because it would require growing the team, which we can't really afford. It's a case of getting a bit more out of developers in terms of quantity and scope (mostly this) of what they do. Not a full developer but enough for it to be meaningful. But it's not nothing either. Worth paying for. AI is a lot cheaper than a developer is so I don't need to replace my developers. I prefer people that are multi disciplinary and able to pick up new skills as they are needed. Agentic AIs are good for that because they give you enough to work with that you can get productive with whatever you need to wrap your head around in little to no time.
Companies can be a bit slow to update their hiring processes to their needs. But good developers should be ahead of the curve in any case. For this, just be proficient with the tools.
Be ready for the inevitable interview question "so, AI ... explain me how you are using it and what you are doing with it?". Much easier to answer that question if you have some meaningful time of routinely using this stuff behind you and can articulate what works and doesn't work for you.
And if they don't ask, that's actually a great question to ask back if you get the opportunity "I've been using agentic tooling, how are you guys using that a <company name>? Also I would like a subscription to <my favorite AI tool> if I work for you". Stuff like that makes you stand out as ambitious and interested in the future. There are of course going to be places that maybe don't like that. But then ask yourself whether you'd want to work there. So, either way, you learn something.
I would look forward to the next 20 years and not backward to the last 20.
The whole frontend/backend distinction did not really exist until the web. And infrastructure is definitely something that should be automated far more than it currently is. If it needs babysitting by a team of devops, you just created a lot of work rather than automating/solving it. Tedious and repetitive. It has "AI will make this a lot easier" written all over it.
So, just be ready for the ambition level to be raised for developers. Learn to build the whole system, not just bits and pieces of the system. Lean on AI to get stuff done and figure things out. It's all just code. None of it is really that hard. But it can be a lot of work if you do all of it manually.
And let's be honest, agentic tools are showing promise and great progress but they are nowhere close to independently working on existing code bases. That's not how I use them. But they are great for problem solving, debugging, prototyping, exploring some new languages and APIs, and generally taking care of more tedious coding tasks.
> I wanna pick your brain a lil. Are you saying agentic AI has helped you replace devs that you would otherwise need for your startup?
I need fewer devs to get more work done... but interestingly it has put a premium on experience because a lot of the "human work" is debugging and fixing where the LLM missed the mark. So less headcount, higher skill required.
Isn't retaliation for chargebacks a violation of most merchant contracts with the credit card companies?
The one time I have had to request a chargeback I was basically told to contact the financial institution[1] in case the merchant tried anything funny.
[1] In this case Nets A/S who acted as a middle man between my bank and VISA. But I assume my bank would have said the same thing if they had the agreement directly with the CC company.
Suppose it is true, I would like to know what you would do after your account is banned. You are not a part of that contract so you cannot sue the merchant. And credit card company likely won't bother with a lawsuit because there is one person affected by this.
I think the point is to establish some precedent that what Kraken was offering actually was an unregulated security, something that might not have been completely clear before this. I doubt that the Fed is deeply concerned about Kraken beyond that.
I think that's the point? The SEC doesn't want an issue where Kraken loses all the funds because they didn't actually stake anything at all and instead were investing in tulip futures.
I hope it was sarcastic. At this point, are there any exchanges that aren't doing untoward things with their customer's assets? We basically have exchanges that failed doing this and other exchanges with ambiguous descriptions of what they have done that haven't failed yet.
I doubt it will push a significant chunk of these users to self-custody. Self-custody is significantly more complicated than leaving assets on an exchange—which is why most people don't do self-custody. I'm sure this will drive some to self-custody. But I don't imagine it'll be a very large percentage.
Is it tho? A MetaMask wallet staking directly with the protocols/ yield aggregators is easier than creating a kraken account and doing KYC verification.
self custody vs. centralized exchange data would be relevant here. Instead of just saying “people don’t do it”. Because people do… and I would venture to say more people do it than keep their coins on a centralized exchange.
That may indeed be "easier", but self-custody requires a high level of confidence in your own ability to set things up properly and securely. And equally high confidence that your machine and wallet won't be compromised. I'd also say that getting to that point is difficult, unless someone is willing to play around with a test-net and not learning with their real balance.
There's definitely an appeal when it comes to having an exchange manage this all for you, even when they take a fee for doing so.
This may come off as… impractical… but I don’t understand why.
The best way to obtain coins without going through KYC is by using the protocols. The most cost effective is running a filecoin node but there are thousands of ways of getting coins by participating in the ecosystem.
Right. I self custody most of my stuff, but I do stake a few coins with Kraken because the amount is relatively small and it's easier. Taking away this service is annoying.
reply