SpaceX IPO

[Cyclotron]

SpaceX filed an application with the Federal Communications Commission in late January for a constellation of up to 1 million satellites to serve as an orbital data center for AI applications. The company provided a few technical details about the constellation, including the size of the satellites, in that application.

The AI Sat Mini (uh) is suggested to be 170 meters long, provide 100 kW of power, and have a radiator covering about 100 square meters.
SpaceX and Tesla are working to build the chips at a substantial scale for this constellation.  They say they know how to do heat rejection in space as they have 10k satellites doing it right now.

Meanwhile, Aetherflux, Axiom Space, Kepler Communications, Planet, Sophia Space, and Starcloud are using NVIDIA accelerated computing platforms to power next-generation space missions across orbital and ground environments. NVIDIA has announced its NVIDIA Space-1 Vera Rubin Module. NVIDIA IGX Thor and  Jetson Orin platforms.

What is less apparent is the purpose of all of this processing in space. AI processing needs to be done as close to the data as possible. The articles are light on details, but what appears to be one data set for processing is imagery of the Earth, and I assume the universe, from satellites and observatories. That can't be the whole of it, though that data set is enormous.

[Marco]

LLM's have comparatively tiny datastreams and latency really doesn't matter much.

[tom66]

It's likely space-based datacenters in SSO will be used for training, not inference (execution of the weights).  Training new models is extremely expensive, but has essentially no latency penalty.  SSO puts them at about 250ms latency, which is too long for agentic AI to really perform well.  Though perhaps you could have light weight models in terrestrial datacenters and heavier ones in space, and then logic to ferry data between the two?

The future is really strange for sure.

[wraper]

Why would it have lower weight than regular unfolding mechanism?

Orchestrating the unfolding of a bunch of carefully packed rods to a km structure is a lot harder than inflating a thin sheet folded like an accordion.

If only it was that simple

[Marco]

If only it was that simple

NASA was designing inflatable structures yonks ago, but they wanted to cure resin after inflation. Which doesn't make a lot of sense for a radiator filled with refrigerant any way.

There simply has never been much reason to go into extremes of surface to weight ratios, polar orbit data centres with ridiculously high power density electronics are a completely different design space than classical satellites. Closest to it were old concepts for power beaming, exactly what they were working on inflatable structures for.

[brucehoult]

SSO puts them at about 250ms latency, which is too long for agentic AI to really perform well.

I'm on Starlink (which is a satellite-based service), in New Zealand, and my ping time to grok.com or chatgpt.com is on average 25ms (normal variation 21-30 ms)

SSO can be as low as 183km (96º inclination, 88 minute period). That is 1.2ms ping time from the point it is passing over at that moment. At 7.8 km/s it's within 45º of overhead (so 1.7ms ping) for 47 seconds. plenty of time for many queries. Or within 60º of overhead (2.4ms ping) for 54 seconds.

If your individual query takes longer than a minute then ping time is hardly a concern!

[Cyclotron]

It's likely space-based datacenters in SSO will be used for training, not inference (execution of the weights).  Training new models is extremely expensive, but has essentially no latency penalty.  SSO puts them at about 250ms latency, which is too long for agentic AI to really perform well.  Though perhaps you could have light weight models in terrestrial datacenters and heavier ones in space, and then logic to ferry data between the two?

The future is really strange for sure.

Why do you believe it will be used for training? I'm curious because I haven't seen much mentioned about training models in space much.

Everything I have read or discussed is related to Space AI data centers for inference. The NVIDIA hardware announced consists entirely of inference accelerators, with only one GPU type that might be ok for training purposes.

I agree with the ideas of what those things are and do, but all of the articles and even the NVIDIA hardware being developed are for inference.

[ejeffrey]

Space data-centres are such an interesting puzzle. I think too many people knee jerk against it ... anyone who brings up the ISS is not thinking it through for instance. The ISS batteries make the weight of PV and radiators almost irrelevant for the ISS, how can you expect it to be state of the art when they operate in an entirely different design space?

The only reason to compare to the ISS is because it's currently the largest solar array in orbit.

But data centers in space is just a dumb idea promulgated by a complete misrepresentation of the situation.

Electricity is the highest *operating* cost of a data center, but it's just not that expensive.  It's still small compared to the hundreds of billions of dollars of capital expenditures hyperscalers and AI companies are planning to pour into data centers over the next few years.

And stuff in space is less useful than stuff on the ground.  It works worse.  The reason space data centers are a dumb idea is because it's just not possible for the power savings (if there are any: highly debatable) would ever make up for the loss in capability.

In my opinion, the biggest problem with data centers in space is not latency, it's not power, it's not cooling, it's not making the computers survive launch, and it's not even the prospect or radiation damage or reduced lifetime.  It's networking.  Got take a look at a modern data center, and count the amount of network bandwidth they have.  Thousands of fibers, many petabits or even exabits per second of aggregate internal bandwidth.  Data center operators don't install all that expensive network hardware because they like giving Cisco and Arista money, they do it because it lets them get more use out of their servers. 

A fleet of satellites each with a rack or a few racks worth of equipment, communicating with optical or microwave links will have orders of magnitude worse internal bandwidth and latency than a modern data center.   Starlink's satellite to satellite links are amazing... for what they are.  But compared to a data center core switch they are a joke.

Again, that's my guess on the more insurmountable problem, but all the other problems I listed above are also serious problems.

It's likely space-based datacenters in SSO will be used for training, not inference (execution of the weights).  Training new models is extremely expensive, but has essentially no latency penalty.

Most everyone I see talking about it says they want it for inference, not training.  Training makes even less sense.

First off, training requires massive, massive clusters, and while it's not sensitive to user latency, it's highly sensitive to internal latency.  Poor latency means slower training which means longer time to market.  So those satellite interconnects are a major problem.

But the more serious problem with this proposal is that the lifecycle costs are already heavily trending towards inference, and that is projected to continue.  So using orbital data centers just for training would be "solving" a small percentage of the problem that is expected to get smaller.

[EEVblog]

But data centers in space is just a dumb idea promulgated by a complete misrepresentation of the situation.

Electricity is the highest *operating* cost of a data center, but it's just not that expensive.  It's still small compared to the hundreds of billions of dollars of capital expenditures hyperscalers and AI companies are planning to pour into data centers over the next few years.

And stuff in space is less useful than stuff on the ground.  It works worse.  The reason space data centers are a dumb idea is because it's just not possible for the power savings (if there are any: highly debatable) would ever make up for the loss in capability.

In my opinion, the biggest problem with data centers in space is not latency, it's not power, it's not cooling, it's not making the computers survive launch, and it's not even the prospect or radiation damage or reduced lifetime.  It's networking.  Got take a look at a modern data center, and count the amount of network bandwidth they have.  Thousands of fibers, many petabits or even exabits per second of aggregate internal bandwidth.  Data center operators don't install all that expensive network hardware because they like giving Cisco and Arista money, they do it because it lets them get more use out of their servers. 

A fleet of satellites each with a rack or a few racks worth of equipment, communicating with optical or microwave links will have orders of magnitude worse internal bandwidth and latency than a modern data center.   Starlink's satellite to satellite links are amazing... for what they are.  But compared to a data center core switch they are a joke.

Again, that's my guess on the more insurmountable problem, but all the other problems I listed above are also serious problems.

The interesting thing is that it's going to happen. So it will be interesting to see if it's useful/profitable or not. They can probably push the hype for the next 5 years until investors realise it's either a winning or a losing idea.
They are already talking about designing the algorithms/compute around the (poor) interconnect limitations of satellites. I'm sure anyone knows if there is a market space there or not yet.
Within the same time frame there will be also be a glut of ground based AI data centres at current construction rates.

[brucehoult]

But data centers in space is just a dumb idea promulgated by a complete misrepresentation of the situation.

Electricity is the highest *operating* cost of a data center, but it's just not that expensive.  It's still small compared to the hundreds of billions of dollars of capital expenditures hyperscalers and AI companies are planning to pour into data centers over the next few years.

On the contrary, I think it is you who does not understand the situation, as understood by the people wanting to do this.

Getting electricity "for free" is NOT the reason they want to do this. The ability to generate electricity in space is simply a necessary prerequisite, not the reason to do it.

The reason to do it is the amount of real estate available to build on, the planning and consenting process to build a data centre on Earth, and simply the delays involved in construction on the ground.

It is not the cost of buying electricity, it is the cost and planning and construction delays in getting new power stations built to provide that electricity.

And water rights.

With a space-based setup you have one factory somewhere making satellites, and a launch pad. Once you have those in place there is no practical limit to the speed you can scale at. No delays. One Superheavy/Starship making one flight a day can launch a LOT of satellites very quickly. All the more so if you build more than one rocket or have them flying more than once a day.

I'm not offering an opinion on whether this is all a good idea or not. I'm rather sceptical about it.

I'm explaining the actual reasons why people are proposing to do this.

[EEVblog]

The reason to do it is the amount of real estate available to build on, the planning and consenting process to build a data centre on Earth, and simply the delays involved in construction on the ground.
It is not the cost of buying electricity, it is the cost and planning and construction delays in getting new power stations built to provide that electricity.
And water rights.

The winners might be those who come into the game in a few years time and scoop up the failed data centres that have already expended all the time, energy and money building them and getting them approved

[Cyclotron]

I believe there are two drivers. Data gravity wells in space that require processing, and the issues of land-based building. The latter are almost all "not in my backyard" issues.

I need to find some time to research the former. I can imagine a lot of efficiencies in processing satellite imagery, such as in environmental and cartographic applications. Spy needs aside, there is a huge amount of image data floating out there. What percentages might be targeted at these data sources in space vs. upload for processing? That's a question I'm curious about.

[Analog Kid]

Put all them satellites up there, sure.
Then maybe Iran can shoot them all down one day.
One blessed day.

[ejeffrey]

The interesting thing is that it's going to happen. So it will be interesting to see if it's useful/profitable or not. They can probably push the hype for the next 5 years until investors realise it's either a winning or a losing idea.

I don't think it's going to happen, at least not in any real sense.  A couple of companies (mostly likely SpaceX or Google) might launch a couple of test satellites, run some strategically chosen benchmarks, and get a nice press release out of it.

But I think it's all PR and stock manipulation.  And for that, other than hiring a handful of token engineers to write white papers and build mockups, there is no need to do the actual work, or distract from their core goal of deploying as many GPUs on the ground as they can buy.  It's not like there will be any penalty from investors if in 5 years they say "turns out, we could build all we needed on the ground"  The only important thing is to avoid the perception that they are falling behind today.

[ejeffrey]

On the contrary, I think it is you who does not understand the situation, as understood by the people wanting to do this.

Getting electricity "for free" is NOT the reason they want to do this. The ability to generate electricity in space is simply a necessary prerequisite, not the reason to do it.

The reason to do it is the amount of real estate available to build on, the planning and consenting process to build a data centre on Earth, and simply the delays involved in construction on the ground.

Real estate is an even stupider reason, and one that exactly illustrates the problem.

There is no shortage of land to build data centers on and data centers take an absolutely tiny amount of land. The cost of land and permitting is negligible in a data center project.  Power availability is definitely an issue, but there are certainly places you can build that have adequate power or are developing it rapidly.

What people get misled by is the "anti data center" pushback.  It makes good headlines but it doesn't mean what people think it means.  It doesn't mean there is nowhere to build data centers.  It means that data center operators are not being allowed to build where they wanted to at the price they thought they could get it for.  Hyperscalers don't necessarily want to build where land is the cheapest and permitting the easiest -- because those are negligible costs.  They want to build where it's convenient, and where they can get the most out of it, generally by locating near their users.

But they can just select less convenient locations.  And they are.  Because pretty much anywhere is more convenient than in space.  If you want to build a data center in Virginia to be close to customers in the eastern seaboard, putting it in space doesn't solve the problem.

And water rights.

This is basically a non-issue.  I mean, it's an issue that data centers are using a lot of water in certain places that don't have any to spare, but they don't *have* to do so.  Data centers use evaporative cooling because it's cheap.  The could build the same heat pump radiators they claim they will use in space on the ground and they would work just fine with no water consumption.

[Psi]

They will be designing these things in a modular way.  There's the big solar + cooling bus parts, and then there's a central module with a docking system where the "compute" modules plugs in for power and cooling.
It makes the most sense to do things like this because you will probably want to replace the compute module when newer/faster AI chips come out. The power and cooling part will last way longer than the chips will stay cutting edge.

Have lots of these systems in space has a lot of other potential uses even if the demand for AI disappeared. You could dock a pharmaceutical module for fab of drugs in zero G, or some sort of a 3D printing space manufacturing system that spits out large frames for space stations.

Having a lot of solar and cooling in space that is modular is quite universally useful.

[Cyclotron]

The interesting thing is that it's going to happen. So it will be interesting to see if it's useful/profitable or not. They can probably push the hype for the next 5 years until investors realise it's either a winning or a losing idea.

I don't think it's going to happen, at least not in any real sense.  A couple of companies (mostly likely SpaceX or Google) might launch a couple of test satellites, run some strategically chosen benchmarks, and get a nice press release out of it.

But I think it's all PR and stock manipulation.  And for that, other than hiring a handful of token engineers to write white papers and build mockups, there is no need to do the actual work, or distract from their core goal of deploying as many GPUs on the ground as they can buy.  It's not like there will be any penalty from investors if in 5 years they say "turns out, we could build all we needed on the ground"  The only important thing is to avoid the perception that they are falling behind today.

Of course, I see what you are saying. SpaceX and Google need to manipulate their stocks. Is SpaceX also not going to build their processors, you think?

What then is Starcloud?  Just stealing millions from dumb VCs?  There are several others that are not public companies. Are all of these just stealing cash?
Why do you think NVIDIA is investing? What's their return on this?  Is that also stock manipulation?

[wilfred]

This is basically a non-issue.  I mean, it's an issue that data centers are using a lot of water in certain places that don't have any to spare, but they don't *have* to do so.  Data centers use evaporative cooling because it's cheap.  The could build the same heat pump radiators they claim they will use in space on the ground and they would work just fine with no water consumption.

That's what I don't get. If they want to use water for cooling but a more expensive alternative is available on the ground, how is it cheaper to launch it into space?

[wilfred]

Another thing I don't get is why bundle Space-X with other stuff? Taking out the trash springs to mind. There are a few reasons a company lists as a public company. One is to allow insiders to sell some of their stock. Another is to get access to capital. With cracks appearing in private credit or cockroaches as Jamie Dimon CEO of JP Morgan said maybe Musk sees a closing window to make a grab for cash.  I'll be interested to see how much of the stock is set aside for retail and existing Tesla investors.

[Psi]

Another thing I don't get is why bundle Space-X with other stuff? Taking out the trash springs to mind. There are a few reasons a company lists as a public company. One is to allow insiders to sell some of their stock. Another is to get access to capital. With cracks appearing in private credit or cockroaches as Jamie Dimon CEO of JP Morgan said maybe Musk sees a closing window to make a grab for cash.  I'll be interested to see how much of the stock is set aside for retail and existing Tesla investors.

Musk has already said 30% for retail investors.
Or at least that is the plan, it's not official yet.

[brucehoult]

Another thing I don't get is why bundle Space-X with other stuff? Taking out the trash springs to mind. There are a few reasons a company lists as a public company. One is to allow insiders to sell some of their stock. Another is to get access to capital. With cracks appearing in private credit or cockroaches as Jamie Dimon CEO of JP Morgan said maybe Musk sees a closing window to make a grab for cash.  I'll be interested to see how much of the stock is set aside for retail and existing Tesla investors.

Musk has already said 30% for retail investors.
Or at least that is the plan, it's not official yet.

There will be those who will say that's because retail investors are naive and easily fooled.

While if it was less, the same people would say small investors are being cut off from potential big gains.

[Psi]

IPOs are always extremely volatile.
Probably going to shoot up super high and then crash before finding a stable point. That's pretty common. (not investment advice)

Despite actually wanting to invest in SpaceX I'm not going near it for a while.
I'm grabbing my popcorn and watching what happens.   

[Simmed]

i think most of you have seen this 1

https://www.tomshardware.com/tech-industry/artificial-intelligence/spacex-says-it-is-going-to-begin-manufacturing-gpus-usd1-75-trillion-ipo-listing-reportedly-includes-in-house-gpu-production

The S-1 form seen by Reuters mentions 'manufacturing our own GPUs' as one of the reasons of the 'substantial capital expenditures' the company will incur in the future as it appears to lack ' long-term contracts with many of our ​direct chip suppliers.'

it appears the reason is spacex will be making "GPUs"
hmmm buying ASML machines mr elon ?
this will be interesting
 

and so on youtube, there is now spatters of of "spacex ipo" is dangerous / lousy kinds of video
stirring shit so that ipo will fail ?
 

[tom66]

Tesla / xAI already announced they were going to fab their own AI processor for the newer generation of cars, and also AI datacenters.  Previously the processors were fabbed by Samsung in Texas.
https://en.wikipedia.org/wiki/Terafab

So I'm guessing that's what they mean when they say they will make their own GPUs -- although I'll take a punt these are more likely to be similar to TPUs. GPUs tend to have the capability to work with multiple types of integers and floating point numbers, whereas TPUs tend to specialise on something like int8 or fp8.  GPUs also tend to have engines optimised for video operations, e.g. image scalers, texture mapping, video decode/encode that all gets deleted on a TPU to save die area.

[wilfred]

and so on youtube, there is now spatters of of "spacex ipo" is dangerous / lousy kinds of video
stirring shit so that ipo will fail ?
 

One of the ones I have seen is put out by Patrick Boyle. It covers a lot of ground, but I don't remember him mentioning GPU's. I'm not planning on participating so I haven't paid attention to whether institutional investors go first for their allocation to set the final price which will then be used for retain investors. I have a feeling it is the reverse here but I'm not sure.

I'm wondering how much cash will be sucked out of the market to buy Space-X. From what I've heard Starlink is the only part of Space-x that makes money. Which makes me wonder why it must be saddled with rockets, AI and whatever else is being tacked on. If Twitter/X is included then I definitely think the IPO is a bad idea. Not because of anything I have against it but it is like investing in a conglomerate. I think you'd be better off investing in rockets, social media ,AI, internet and GPU's? via separate best in class companies.

But Index funds due to the rushed inclusion on the NASDAQ will have to buy into the IPO sooner than they might have otherwise, which I think will be bad for those investors.