M&A: xAI, A SpaceX Company

[Vultur]

What would be the bottleneck? There’s plenty of silicon.

Building the physical factories. Building large scale new infrastructure, especially today, tends to be slow and expensive.

[RedLineTrain]

What would be the bottleneck? There’s plenty of silicon.

An uncooperative supply chain for machinery and materials, according to Musk in the interview.  Not something that Tesla hasn't dealt with in the past, mind you.  The way they pushed through it was to create the Gigafactory with Panasonic and then the 4680 cell in-house.

[RedLineTrain]

100-200 billion chips is for Tesla - not clear if Tesla AI specifically or also other Tesla chips?

You expect them to be separate efforts?

[RedLineTrain]

100-200 billion chips per year seems outrageous by at least a couple of orders of magnitude.

Consider that even with 1-2 billion chips (assuming Musk meant that amount), you would need 580k-1.16m million wafer starts per month, for which you would need a cleanroom footprint of about 19m-39m square feet using current methods.  Then you would need to upgrade your equipment and layout every couple of years to stay on the leading process node.

Compare to the Gigafactory Austin footprint of about 4 million square feet (12 million square feet of factory floor).

[Robotbeat]

What would be the bottleneck? There’s plenty of silicon.

Building the physical factories. Building large scale new infrastructure, especially today, tends to be slow and expensive.

I know a guy

[DigitalMan]

100-200 billion chips per year seems outrageous by at least a couple of orders of magnitude.

Consider that even with 1-2 billion chips (assuming Musk meant that amount), you would need 580k-1.16m million wafer starts per month, for which you would need a cleanroom footprint of about 19m-39m square feet using current methods.  Then you would need to upgrade your equipment and layout every couple of years to stay on the leading process node.

Compare to the Gigafactory Austin footprint of about 4 million square feet (12 million square feet of factory floor).

Well, he did say besides buying chips from all the existing suppliers, they were considering building their own fab with 10 lines at 100,000 wafer starts each.

[RedLineTrain]

100-200 billion chips per year seems outrageous by at least a couple of orders of magnitude.

Consider that even with 1-2 billion chips (assuming Musk meant that amount), you would need 580k-1.16m million wafer starts per month, for which you would need a cleanroom footprint of about 19m-39m square feet using current methods.  Then you would need to upgrade your equipment and layout every couple of years to stay on the leading process node.

Compare to the Gigafactory Austin footprint of about 4 million square feet (12 million square feet of factory floor).

Well, he did say besides buying chips from all the existing suppliers, they were considering building their own fab with 10 lines at 100,000 wafer starts each.

Good point.  He did say that at the Tesla Shareholder's meeting.  I guess that solves it.  That's 3 to 4 million square foot footprint apiece for 10 fabs.  1-2 billion chips per year.  He misspoke in the Baron interview.

From a construction standpoint, the first fab seems doable on a two year timeline.  Then you would need all of the packaging facilities and memory.  And all of the equipment.  Supplies.

Incredibly challenging with little margin for error in coordinating it all.  Hope that there would be gigantic returns to scale.  I guess there could also be returns to uniformity of product.  These fabs would be built specifically for the chip as designed.

So that’s why I think, as far as I can see, the only option is to go build some like very big chip fab. And then you go to solve memory and packaging too. But otherwise, you just tap out at whatever the chip production rate is. And so I guess, Tera would be – you’d want to say it’s got to be at least 100,000 wafer starts per month size fab. And maybe that would be one of ten in a complex. So it would ultimately be one million wafer starts per month.

[Vultur]

1-2 billion (not 100-200 billion) might be workable. Assuming rapid demand growth continues, which I don't think is necessarily a safe assumption even over the next few years. There's a lot of moving parts.

[RedLineTrain]

Optimus is going to be ramping rather quickly.  1 million plant to go online next year in Fremont, 10 million plant to go online in two years in Austin.  One or two chips per robot.

Cars will be ramping to 4 million a year in a couple of years.  Two or three chips per car.

xAI may have huge inference needs.  Currently, inference is being outsourced to the hyperscalers, but xAI could bring a portion in house.  Currently, training is at ~1 GW and quickly rising higher.  Using assumed inference:training ratios in the future, we could readily see very high demand for inference and therefore chips there.

SpaceX is a bit of a wildcard.  It could serve inference to Tesla, xAI, and perhaps others.  Rate limited by Starship launch capacity.

[bulkmail]

100-200 billion chips is for Tesla - not clear if Tesla AI specifically or also other Tesla chips?

You expect them to be separate efforts?

Besides big AI chips Tesla uses plenty of smaller microcontrollers and other semiconductor devices. If it's practical to build some of those on the manufacturing process of the new factory (for technical and contractual reasons), then maybe that increases the number. Or it's just a misspoken one as others noted above.

[RedLineTrain]

I think microcontrollers and some of the other semiconductors are well off the latest process node.  For instance, yesterday STMicroelectronics signed a deal with SpaceX for 18nm microcontrollers.  Those will be fabbed in France.

The challenge on AI5/6/7 is that it needs to always be at extreme production scale on the latest process nodes.  AI4 was never held to that standard (7nm process node) and is already pretty long in the tooth, more than 2.5 years old.  Much more forgiving.  Actually, there are not a lot of chips fabbed on the latest process nodes.

STMicroelectronics introduces the industry’s first 18nm microcontroller for high-performance applications
Nov 18, 2025  Geneva, Switzerland
STMicroelectronics (NYSE: STM), a global semiconductor leader serving customers across the spectrum of electronics applications, has unveiled the STM32V8, a new generation of high-performance microcontrollers (MCUs) for demanding industrial applications. The STM32V8 is designed with ST’s most advanced 18nm process technology with best-in-class embedded phase-change memory (PCM). It is manufactured in ST’s 300mm fab in Crolles, France, and also in collaboration with Samsung Foundry. MCUs are foundational chips, with the broad portfolio of STM32 devices powering billions of devices on the planet across consumer devices, home appliances, industrial applications, medical devices and communication nodes.

https://newsroom.st.com/media-center/press-item.html/p4733.html

[RedLineTrain]

At the Saudi Investment Conference, Musk talks about the relative attractiveness of terrestrial versus deep space data centers.  See video @ 2:00.

My estimate is that. . . the cost effectiveness of AI in space will be overwhelmingly better than AI on the ground. . . even perhaps in the 4 to 5 year timeframe.

https://twitter.com/SawyerMerritt/status/1991191209507770850

[RedLineTrain]

Follow up detail from Musk...

Elon Musk @elonmusk
Starship should be able to deliver around 300 GW per year of solar-powered AI satellites to orbit, maybe 500 GW. The “per year” part is what makes this such a big deal.

Average US electricity consumption is around 500 GW, so at 300 GW/year, AI in space would exceed the entire US economy just in intelligence processing every 2 years.

Solar cell production for terrestrial applications is already far above this number at >1500 GW of annual production capacity.

Tonnage to orbit will be solved by Starship.

Chip production is therefore the major piece of the puzzle to be solved.

The Tesla Terafab is needed for this, as there is otherwise no solution at sufficient scale.

These numbers are immense by earthly standards, but don’t register materially on the Kardashev II scale. For that, lunar production of solar-powered AI satellites is needed to take things to the 100+ TW/year level.
1:43 PM · Nov 19, 2025

https://x.com/elonmusk/status/1991215836959228322

[meekGee]

Follow up detail from Musk...

Elon Musk @elonmusk
Starship should be able to deliver around 300 GW per year of solar-powered AI satellites to orbit, maybe 500 GW. The “per year” part is what makes this such a big deal.

Average US electricity consumption is around 500 GW, so at 300 GW/year, AI in space would exceed the entire US economy just in intelligence processing every 2 years.

Solar cell production for terrestrial applications is already far above this number at >1500 GW of annual production capacity.

Tonnage to orbit will be solved by Starship.

Chip production is therefore the major piece of the puzzle to be solved.

The Tesla Terafab is needed for this, as there is otherwise no solution at sufficient scale.

These numbers are immense by earthly standards, but don’t register materially on the Kardashev II scale. For that, lunar production of solar-powered AI satellites is needed to take things to the 100+ TW/year level.
1:43 PM · Nov 19, 2025

https://x.com/elonmusk/status/1991215836959228322

Went up from 100 GWatt/yr to 300-500

I got 100 using 10 ships/day/tower and 10 towers, but that was just from extrapolating from Starlink.

 I wonder what they're thinking long term for the design.

[RedLineTrain]

Add'l timeline information from Elon.

Elon Musk @elonmusk
I can see a path to doing 300GW per year within 8 years
3:30 PM · Nov 23, 2025 164.1K Views

Adam Lowisz @AdamLowisz
Nov 23
Our GDP growth will be limited by our ability to produce electricity. We are going to have to make a lot more electricity.

Elon Musk @elonmusk
Solar in space
3:39 PM · Nov 23, 2025 44.6K Views

https://x.com/elonmusk/status/1992692184253092259?s=20

[Ludus]

https://eu.36kr.com/en/p/3567865887390593

This is a brief summary of the idea. Elon Musk has been offering comments and posts for a few weeks on the topic.

I post the topic in with Starship because it’s clearly based on the projected capacity of Starship and the Starship based expansion of Starlink. Maybe it should be elsewhere.

This seems like a new landscape of possibilities he’s considering, a bit like the first exploration of Starlink. It also involves a convergence between X.ai, Tesla and SpaceX. Before this Musk rejected any important role for Space based solar power. He explicitly did not see that as a driver for profitably using the launch capacity of Starship. The enormous investments in scaling AI and the power requirements changed this.

He seems to have decided on a first principles basis that Space based solar powered AI clusters make sense and that this ought to be a priority.

This is an existing topic about the idea generally: https://forum.nasaspaceflight.com/index.php?topic=63177.0

This is about Google Moonshot with the same focus. https://forum.nasaspaceflight.com/index.php?topic=63819.0

[Twark_Main]

The idea, clearly, is that AI training tasks let you beam the value back to Earth (ie the finished trained model) without beaming the BTUs back to Earth (and vaporizing the planet, in the limiting growth case).

I admit this something I contemplated, but I never thought someone would be crazy enough to actually attempt it!

It's so funny seeing people attribute this to some sort of marketing hype. As usual, when you see too far people think you're mad... 

[DigitalMan]

Elon has spoken about inference tasks, with the reason being that a short delay is acceptable

[leovinus]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

[DanClemmensen]

Sounds like the 1930s predictions of “gazillions of airships and zeppelins” . As someone who trained and optimized AI/ML models which are in the hands of millions, my money is on better algorithms and hardware which will make a trillion in market cap go “poof”. Just my 2ct

Yep. Remember the huge render farms that were needed in the year 2000 to create CGI that we would now consider barely adequate?