Moving The Cloud to orbit

[Lee Jay]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

What do you mean by space rated solar panels? What do you know about that tech?

Apparently nothing because SpaceX has been making their own Starlink solar panels using regular terrestrial solar cells for this entire time.

And you think they could ramp that production by 3 orders of magnitude in zero time with zero effort, no doubt.

And why are they building their own?  Could it be because commercially-available panels are not suitable for Starlinks?

[Coastal Ron]

Apparently nothing because SpaceX has been making their own Starlink solar panels using regular terrestrial solar cells for this entire time.

Does SpaceX make their own solar panels, or do they source them from a third party?

Doing a quick search, I see references to Taiwan Solar Energy Corp (and a 2021 article about it), and references to future plans of bringing PV panels in-house.

From that article:

Driven by cost reduction as well as the shorter life expectancy of Starlinks satellites, SpaceX opted for the silicon-based solar cells that are less durable in space environment.

So I guess a separate question is what is the expected usable life of the SpaceX "data centers in space"? Because it sounds like they will have to over build in order to compensate for reduced output towards the end of their expected lifespan.

[RedLineTrain]

SpaceX imports cheap, available solar cells and makes them into space panels.  Tesla imports cheap, available solar cells and makes them into terrestrial panels of two types.

At scale, you also want to manufacture cells in-house domestically.  You can probably buy the manufacturing equipment from the Chinese to get started quickly.  We hear from the press that a SpaceX/Tesla team has been traveling in China, we can assume for that purpose.

You probably would want to site the cell manufacturing near your terrestrial data centers because that's where the electricity is cheapest, if you are willing to curtail use at certain times.  But if you can find an empty building in another low electricty-cost locale like Buffalo, you may wish to expedite manufacturing there.  We hear from the press that a SpaceX/Tesla team has been inquiring with local/state governments, we can assume for that purpose.

[Robotbeat]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

What do you mean by space rated solar panels? What do you know about that tech?

Apparently nothing because SpaceX has been making their own Starlink solar panels using regular terrestrial solar cells for this entire time.

And you think they could ramp that production by 3 orders of magnitude in zero time with zero effort, no doubt.

And why are they building their own?  Could it be because commercially-available panels are not suitable for Starlinks?

I find discussing anything with you exhausting because of the contempt dripping from every word. Anyway.

Theyre already launching way more than 7 Megawatts per year, so not 3 orders of magnitude. They bought a terrestrial German solar cell stringer. These just aren’t that high volume compared to terrestrial solar (China makes north of a Terawatt of solar panels per year), so it really isn’t that big of a deal. In-space arrays can be simpler and lighter than terrestrial ones, in fact.

[Robotbeat]

Apparently nothing because SpaceX has been making their own Starlink solar panels using regular terrestrial solar cells for this entire time.

Does SpaceX make their own solar panels, or do they source them from a third party?

Doing a quick search, I see references to Taiwan Solar Energy Corp (and a 2021 article about it), and references to future plans of bringing PV panels in-house.

From that article:

Driven by cost reduction as well as the shorter life expectancy of Starlinks satellites, SpaceX opted for the silicon-based solar cells that are less durable in space environment.

So I guess a separate question is what is the expected usable life of the SpaceX "data centers in space"? Because it sounds like they will have to over build in order to compensate for reduced output towards the end of their expected lifespan.

Tesla makes their own panels with terrestrial solar panel stringer lines that they bought from Germany. They buy the cells from Taiwan and assemble them into panels. You misinterpreted the article, mixing up panels and cells.

But spacex has been making trips overseas to actually bring in-house manufacturing capabilities for the entire cell domestically and they are hiring for perovskite solar cell experts

[InterestedEngineer]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

wait until you learn about how SpaceX already makes their solar panels using regular old monosilicon cells

dangit, you spoiled my trick question by answering it for him.

I'm betting a beer he didn't know that.

It's not custom GaAs specialized stuff - SpaceX don't play that game.

And the volume you'd use for 7-10GW/year in space is 2-3% of yearly solar cell production.  With a ramp rate still over 10% a year, effectively rounding error.

don't spoil my trick questions!

[Lee Jay]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

wait until you learn about how SpaceX already makes their solar panels using regular old monosilicon cells

dangit, you spoiled my trick question by answering it for him.

I'm betting a beer he didn't know that.

It's not custom GaAs specialized stuff - SpaceX don't play that game.

And the volume you'd use for 7-10GW/year in space is 2-3% of yearly solar cell production.  With a ramp rate still over 10% a year, effectively rounding error.

don't spoil my trick questions!

"7GW a year of space rated solar panels".  Not cells, "panels", is what I said.

And where do we have specs on the capacity of these panels?  How much capacity are they making per year?  If it's more than 7MW, it's not a lot more.  Certainly not 70, or 700.

[RedLineTrain]

It's not custom GaAs specialized stuff - SpaceX don't play that game.

Musk Inc. never plays the game of exotic materials for solar cells.  This should be "no more complicated than drywall."  However, at 100-GW apiece for SpaceX and Tesla, it could make good sense to iterate a bit on the fundamental technology.

[Twark_Main]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

wait until you learn about how SpaceX already makes their solar panels using regular old monosilicon cells

dangit, you spoiled my trick question by answering it for him.

I'm betting a beer he didn't know that.

It's not custom GaAs specialized stuff - SpaceX don't play that game.

And the volume you'd use for 7-10GW/year in space is 2-3% of yearly solar cell production.  With a ramp rate still over 10% a year, effectively rounding error.

don't spoil my trick questions!

"7GW a year of space rated solar panels".  Not cells, "panels", is what I said.

And where do we have specs on the capacity of these panels?  How much capacity are they making per year?  If it's more than 7MW, it's not a lot more.  Certainly not 70, or 700.

Currently it's around twice that, by my rough calculations.

Oh no, Elon will have to ramp a production line!  No way Musk has ever successfully done that before, right guys?

[InterestedEngineer]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

wait until you learn about how SpaceX already makes their solar panels using regular old monosilicon cells

dangit, you spoiled my trick question by answering it for him.

I'm betting a beer he didn't know that.

It's not custom GaAs specialized stuff - SpaceX don't play that game.

And the volume you'd use for 7-10GW/year in space is 2-3% of yearly solar cell production.  With a ramp rate still over 10% a year, effectively rounding error.

don't spoil my trick questions!

"7GW a year of space rated solar panels".  Not cells, "panels", is what I said.

And where do we have specs on the capacity of these panels?  How much capacity are they making per year?  If it's more than 7MW, it's not a lot more.  Certainly not 70, or 700.

I wish there was a "let me prompt that for you" equivalent to "let me google that for you", but I don't think the economics would work.

https://www.teslaoracle.com/2025/03/06/spacex-bastrop-factory-ramps-up-starlink-production-to-70000-per-week-in-just-20-months/

"Less than 2 years ago this was an anti-building with no employees and we built a factory from scratch. We've gone from producing 0 kits to 70,000 a week and we've gone from 0 employees to over a thousand."

I think you should give up and figure out another angle of attack, this one is hilariously wrong, in case all the sarcasm keeps zooming past your head.

[TheRadicalModerate]

I wish there was a "let me prompt that for you" equivalent to "let me google that for you", but I don't think the economics would work.

https://www.teslaoracle.com/2025/03/06/spacex-bastrop-factory-ramps-up-starlink-production-to-70000-per-week-in-just-20-months/

"Less than 2 years ago this was an anti-building with no employees and we built a factory from scratch. We've gone from producing 0 kits to 70,000 a week and we've gone from 0 employees to over a thousand."

What does this have to do with solar cell capacity?  It's a story that's mostly about ground station kits--mostly.  It's also pretty clearly AI slop, except for the quote, which is kind of a non sequitur.

Not really disagreeing with your main thesis, but you've chosen a terrible citation.

[InterestedEngineer]

I wish there was a "let me prompt that for you" equivalent to "let me google that for you", but I don't think the economics would work.

https://www.teslaoracle.com/2025/03/06/spacex-bastrop-factory-ramps-up-starlink-production-to-70000-per-week-in-just-20-months/

"Less than 2 years ago this was an anti-building with no employees and we built a factory from scratch. We've gone from producing 0 kits to 70,000 a week and we've gone from 0 employees to over a thousand."

What does this have to do with solar cell capacity?  It's a story that's mostly about ground station kits--mostly.  It's also pretty clearly AI slop, except for the quote, which is kind of a non sequitur.

Not really disagreeing with your main thesis, but you've chosen a terrible citation.

Solar cell capacity has nothing to do with it, because Musk's space proposal would consume 2-3% of 2025 capacity, which is rounding error.

it's *making* the satellites including the panels in large quantities that was being questioned.  And zero to 1500/week (7GW/year rate) in 3 years is not a stretch at all if you can go from 0 to 70,000 terminals a week in 20 months.

They are currently at 50 Starlinks/week, and are blocked by launch rate not build capacity.   1500/week of any large electronics package is easy by modern standards, and only 30x what they are doing now.  (comparative examples being prebuilt data center racks, Tesla Model Y (at 20x that rate). and Tesla Megapacks.

And no I'm not going  to quote the source for 50/week.  If I can't quote a quote out of some article for fear the article is AI generated (why did you read it?  I didn't, I searched to make sure the quote was real), I'm just not going to bother, I'm just going assert what you can google or AI yourself.

[Robotbeat]

Yeah?  You're assuming Starship gets working soon, and that manufacturing all those satellites is possible.  Where are you going to get >7GW a year of space rated solar panels, for example?  That's got to be several orders of magnitude more than can currently be manufactured.

wait until you learn about how SpaceX already makes their solar panels using regular old monosilicon cells

dangit, you spoiled my trick question by answering it for him.

I'm betting a beer he didn't know that.

It's not custom GaAs specialized stuff - SpaceX don't play that game.

And the volume you'd use for 7-10GW/year in space is 2-3% of yearly solar cell production.  With a ramp rate still over 10% a year, effectively rounding error.

don't spoil my trick questions!

"7GW a year of space rated solar panels".  Not cells, "panels", is what I said.

And where do we have specs on the capacity of these panels?  How much capacity are they making per year?  If it's more than 7MW, it's not a lot more.  Certainly not 70, or 700.

It is about 70MW/year. They’re doing 50MWh/yr for batteries from a recent presentation (& you need about 1 hour of battery to handle the ~40 minute eclipse time in LEO with margin, and your solar panels for Starlink’s orbit need to be about 1.75x your average power usage—of course AI would be in sun synchronous, 100% sunshine), and they’re spoiling up for more capacity for V3.
https://x.com/xdnibor/status/2023485719189344528?s=46


You can look up the specs for V2 mini and do a rough calculation based on the size of the arrays.

[edzieba]

Now I've proven and industry stats have shown that Claude Code makes a good developer more than 2x as productive, and the cost is rounding error in the above numbers.   So that's $280B worth of value.

Turns out that "2x" is closer to 1.005x to 1.037x (0.5% to an optimistic 3.7%). That puts the 'value' closer to the $1.1B to $5.2B range, orders of magnitude from the investments thus far.
::EDIT:: NBER source paper attached.

[Robotbeat]

That is a paper based on CEO surveys across a broad range of industries. It’s not actually looking at the objective effects on software development with the latest models.

InterestedEngineer’s claim of 2x and the paper’s claim could both be accurate without there being any contradiction.

[InterestedEngineer]

That is a paper based on CEO surveys across a broad range of industries. It’s not actually looking at the objective effects on software development with the latest models.

InterestedEngineer’s claim of 2x and the paper’s claim could both be accurate without there being any contradiction.

I'm now at an AI related startup.  7 people are doing the work of 30 old-school style.   We don't even bother to hire anyone (including me) that hasn't spent at least a month using claude code on serious work - their productivity would be so low they'd be a rounding error or impediment.

Concept to shipping MVP is going to be something like 4-5 months.

The problem the CEOs are facing is that AI assisted programming is a competence multiplier.

And competence is a scale that includes negative signs.

So it's more of an indictment of their hiring practices or knowledge of actual productivity that is' they are claiming 1.03 or whatever.

On a team well trained on the tools it's at least 4x, and given the velocity, probably higher.  For smaller teams the multiplier is bigger, because the coding agents overcome the mythical man month problem of diminishing returns with team size.  Which is another problem the CEOs with very large development teams are going to face, they aren't getting the same multiplier that small teams are getting.  They are going to have to reorganize their organizations.  Fun time to be a VP of engineering.

TL;DR - the competence distribution was always bimodal, and AI increases the distance between the modes.  The average may or may not improve.

This is worth watching - the levels of AI coding competency: 

[Vultur]

That's launch sites.

You were comparing terrestrial data centres with orbital equivalents.

Show me that the cost of and delay from launch and communication licences for a million satellite constellation, for example (presumably with the 10k+ launches per year that Musk keeps talking about), is lower and less than the one-time permitting/etc for an equivalent terrestrial data centre.

cost is near irrelevant.

compare schedules.

Did you watch Elon's latest interview?  He laid it all out.   Years for permits.  4 years for turbine parts

This makes sense ... But only if the space data centers can come online fast enough. That remains a big open question; even a small delay in space terms can be significant on the timescale of the current boom.

And I still think that post-boom there will be a fairly hard ceiling on data center demand.

[Robotbeat]

In the US, about 7% of electricity use is by datacenters, or about 30-35GW. There’s like 100GW planned or in the construction phase as we speak.

I don’t think SpaceX will be near the limit in demand if they actually do build 10-100GW of datacenters in orbit and do so cheaper than terrestrial datacenters. Jevons Paradox means efficiency improvements are likely to increase demand, if anything, as AI becomes more useful.

I really do think people should try using the most advanced, multi-agent models to get a feel for what AI can do, ie for software development or whatever. Anthropic Claude Opus 4.6 Max. GPT 5.3-Codex-Spark. Gemini 3 Pro/DeepThink, etc. you’ll quickly run out of tokens as they’re timeshared, but you’ll see the value.
These things use as much energy as a car accelerating on a freeway, or a single AI satellite. It’s easy to see a million of them being used.

[Vultur]

In the US, about 7% of electricity use is by datacenters, or about 30-35GW. There’s like 100GW planned or in the construction phase as we speak.

I don’t think SpaceX will be near the limit in demand if they actually do build 10-100GW of datacenters in orbit and do so cheaper than terrestrial datacenters

That's true IF there's not a decrease in demand/glut in supply, the way there was for fiber in the early 2000s.

(I am personally quite skeptical that that 100GW currently planned/in construction capacity will all come to fruition.)

I really do think people should try using the most advanced, multi-agent models to get a feel for what AI can do, ie for software development or whatever.

Capability isn't enough by itself though. You need profitability . And that's a lot harder to establish.

[Coastal Ron]

In the US, about 7% of electricity use is by datacenters, or about 30-35GW. There’s like 100GW planned or in the construction phase as we speak.

I don't have the figures at hand, but recently I saw how much energy the gas liquefaction industry uses. It was somewhere in the double digits of % of total energy consumed in the U.S.

So for launching data centers into space, there would be a significant energy cost that would have to be absorbed for generating propellant for Starship launches before it could be replaced by space based data centers.

I don’t think SpaceX will be near the limit in demand if they actually do build 10-100GW of datacenters in orbit and do so cheaper than terrestrial datacenters. Jevons Paradox means efficiency improvements are likely to increase demand, if anything, as AI becomes more useful.

I've been advocating that A) SpaceX didn't need to buy xAI in order to get into the "data centers in space" business, and that B) that xAI may not provide enough demand (and revenue) to support the minimum amount of "data centers in space" that would be required to pay back the investment. Meaning that I think if this concept is successful, that SpaceX will be in the Data Center as a Service (DCaaS) business, and not just reserving compute capacity for themselves.

I really do think people should try using the most advanced, multi-agent models to get a feel for what AI can do, ie for software development or whatever. Anthropic Claude Opus 4.6 Max. GPT 5.3-Codex-Spark. Gemini 3 Pro/DeepThink, etc. you’ll quickly run out of tokens as they’re timeshared, but you’ll see the value.
These things use as much energy as a car accelerating on a freeway, or a single AI satellite. It’s easy to see a million of them being used.

This was part of the topic of a roundtable discussion I had this morning, and we need to remember that at least as of last year that large companies were not figuring out how to incorporate A.I. into their workflows yet. For the developer world, yes, it has been transformative, and no one will want to go back to the old ways. But for everyone else they are still experimenting with large scale solutions.

One other point, and this kind of related to your mention of the Jevons Paradox, is that the advent of technologies like OpenClaw (a free and open-source autonomous AI agent) could change things in unexpected ways, including reducing the amount of data center use, since OpenClaw can be used on devices a small as a Raspberry Pi. So essentially edge computing could change the demand for data centers for certain applications.