Support NeoGAF

[sono]

Summary
Elon Musk's attempt to do for semiconductors what the original Gigafactory did for batteries.

The Core Elements of Terafab

• Vertical Integration "Under One Roof": Historically, chipmaking is fragmented: one company designs the chip, another makes the lithography machines (ASML), another fabs the silicon (TSMC), and another packages it. Terafab aims to house design, fabrication, memory production, and packaging in one single facility in Austin which he claims is not done elsewhere at the moment.
• The "Terawatt" Scale: The name comes from Musk's goal to produce one terawatt of computing power annually. To put that in perspective, he estimated the current global AI compute at only ~20 gigawatts per year
• Two Classes of Chips:
  
  1. Terrestrial (Inference): 2nm chips optimized for "edge" tasks—specifically for the Optimus humanoid robots and Tesla's FSD (Full Self-Driving) hardware.
  2. Space-Hardened (D3 Chips): Processors built to withstand radiation and extreme heat for SpaceX's orbital AI satellites.
• Joint Venture: It is a rare formal partnership between Tesla, SpaceX, and xAI. SpaceX provides the launch capability for the orbital compute, Tesla provides the manufacturing/robotics expertise, and xAI provides the software architecture.

De-risking the future for Musk's companies.

• Bypassing the TSMC/Samsung Bottleneck: Musk stated that current global suppliers cannot expand fast enough to meet his needs. By building his own fab, he avoids being held hostage by global chip shortages or geopolitical tensions (e.g., Taiwan/China).
• The "Space Data Center" Shift: One of the most radical points was moving compute to space.[2] Musk argued that solar power is 5x more efficient in orbit and vacuum cooling is "free." Terafab is the factory that builds the brains for these orbital data centers.
• Enabling the Robot Economy: Musk believes Tesla will eventually need 100–200 billion chips per year for Optimus robots.[4] No existing factory on Earth could supply that volume, making Terafab a requirement for the "Master Plan Part 3" to succeed.

Vision of a closed-loop ecosystem rather than a single product:

• Energy: Powered by Tesla Solar.
• Intelligence: Trained by xAI (Grok).
• Chips: Fabricated at Terafab.
• Deployment: Launched by Starship or driving as a Tesla.

In Musk's view, these are no longer separate businesses. He is building a single "machine that builds the machine" for a galactic civilization. The confusion usually stems from him jumping between the granular details of 2nm lithography and the "epic" vision of putting data centers on the Moon.

 

[winjer]

The more competition, the better.
TSMC having a monopoly of the top process nodes meant they were able to make several price increases without any pushback, even for older nodes.

 

[Trunx81]

He knows that Taiwan won't be safe for long ...
But the name, bro .. I can imagine Elon sitting in the board meeting, giggling and saying "Can we insert the wort 'fab' somewhere?"
After naming the Tesla models S, E and X ...

 

[AJUMP23]

Competition is good.

 

[StueyDuck]

Elon sure knows how to name things...

(I dont have any comments of the actual fabrication facility, will need to know more)

 

[Trogdor1123]

Isn't cooling in space incredibly inefficient though?

Regardless, that's only part of this. This sounds very interesting.

Has he ever stated all of his steps in his master plan? Or even how many there are? It's like something out of a movie. It seems to actually be happening.

 

[IAmRei]

He knows that Taiwan won't be safe for long ...
But the name, bro .. I can imagine Elon sitting in the board meeting, giggling and saying "Can we insert the wort 'fab' somewhere?"
After naming the Tesla models S, E and X ...

I dont know, he always worst in naming his product

 

[Zathalus]

Being able to compete with both ASML and TSMC is one hell of a task. Competition is always good, but there is a reason everybody failed when compared to those two.

 

[Wheeljack539]

From the title I thought this was going to be about storage drives.

 

[Dacvak]

Competition is always good, but let's see how this actually pans out. Elon is great at promising things and never delivering.

 

[calico]

F.A.B. Virgil

 

[EverydayBeast]

Not surprised Elon is comfortable with machines

 

[poppabk]

This is pretty noble goal, but putting it in to practice? Especially with the current immigration situation.

 

[Alandring]

Good luck, but it's almost certain that it won't work.

 

[Vlodril]

The guy sure dreams big.

 

[Magic Carpet]

So another decade of waiting for a made in USA chip? Or is the Terafab overseas like everything else?

 

[IntentionalPun]

This is pretty noble goal, but putting it in to practice? Especially with the current immigration situation.

What's noble about it?

Building robots and AI datacenters in space + the chips that are needed for them sounds pretty useless to humanity TBH.

 

[John Bilbo]

Twitter seems to be down.

 

[Griffon]

Good.

 

[Insane Metal]

Huh. Doubt this will work but wish him good luck.

 

[Trilobit]

Terafab sounds like some fab clothing brand for gay men.

 

[skneogaf]

The semi conductor company I work for are currently finalising a specific chip for him and tesla, nobody else wanted to make it so we said yes. It's different than our usual so we're installing specific machines for it which I'll have to do maintenance on.

Hopefully he leaves that with us and not wants it anywhere near his new place although I imagine it's far too simple for this new terafab.

 

[Dr Bass]

Everyone making fun of the usage of the word "fab" exposes their complete ignorance of the industry. It's what they've been called for decades.

 

[LyricalShrike]

What's noble about it?

Building robots and AI datacenters in space + the chips that are needed for them sounds pretty useless to humanity TBH.

Developing orbital infrastructure is a key step towards ensuring humanity's survival. I've been to the factories where some of these things are being built and the atmosphere is genuinely inspiring.

 

[Wildebeest]

Should be easy for him considering how this is simple technology to reproduce and every promise he made before he has smashed fully delivering ahead of schedule and without any flaws.

 

[jason10mm]

I really hope me makes it look like one of these.....would be EPIC trolling

 

[Donald mac Ronald]

He needs to get off this stupid idea of space data centres Jesus Christ.

The laws of thermodynamics and physics disagree with him. He says cooling in space is free. This guy is supposed to be smart. Cooling in space is the opposite of free, it's insanely hard. The tech isn't even there for it and to cool shit in space would cost so much more than just build cheaper data centres on earth.

 

[LiquidMetal14]

I'm good with him in this.

 

[jason10mm]

He needs to get off this stupid idea of space data centres Jesus Christ.

The laws of thermodynamics and physics disagree with him. He says cooling in space is free. This guy is supposed to be smart. Cooling in space is the opposite of free, it's insanely hard. The tech isn't even there for it and to cool shit in space would cost so much more than just build cheaper data centres on earth.

Elon owns a rocket company that sends shit into space on the reg. Imma gonna side with him that "it's free" is appropriate for how he means it.

 

[Donald mac Ronald]

Elon owns a rocket company that sends shit into space on the reg. Imma gonna side with him that "it's free" is appropriate for how he means it.

"Every kilowatt of heat generated by an AI chip must be released via radiation.
The Problem: Radiators are much less efficient than liquid cooling. To cool a modern AI cluster (which can pull 100kW+ per rack), you would need massive, wing-like radiator arrays—sometimes larger than the satellite itself—pointed permanently away from the sun."

Just another one of his dumb hype ideas.

 

[EverydayBeast]

My initial gaming response would be could a console manufacture possibly have a Tesla Chip in it powering everything?

 

[IntentionalPun]

Developing orbital infrastructure is a key step towards ensuring humanity's survival. I've been to the factories where some of these things are being built and the atmosphere is genuinely inspiring.

Yeah right, "orbital infrastructure"? You mean GPU farms in space? How is that key to our survival?

 

[Sgt.Asher]

 

[LyricalShrike]

Yeah right, "orbital infrastructure"? You mean GPU farms in space? How is that key to our survival?

The techniques used to solve the problems of data centers in space(launch cost, cooling, assembly, etc.) can be leveraged to build other infrastructure such as larger habitats, orbital factories, and orbital solar arrays. Space compute is just the first step.

 

[CGNoire]

Isn't cooling in space incredibly inefficient though?

Regardless, that's only part of this. This sounds very interesting.

Has he ever stated all of his steps in his master plan? Or even how many there are? It's like something out of a movie. It seems to actually be happening.

How so?
Isnt space extremely cold?

 

[CGNoire]

What's noble about it?

Building robots and AI datacenters in space + the chips that are needed for them sounds pretty useless to humanity TBH.

Putting data centers in space should save a huge amount of energy and cooling cost keeping prices lower for civilians down here on earth.

 

[poppabk]

How so?
Isnt space extremely cold?

It's in a vacuum. You have nothing to transfer the heat to. Think about your standard air CPU cooler - now takeaway the air.

 

[Panajev2001a]

Vertical Integration "Under One Roof": Historically, chipmaking is fragmented: one company designs the chip, another makes the lithography machines (ASML), another fabs the silicon (TSMC), and another packages it. Terafab aims to house design, fabrication, memory production, and packaging in one single facility in Austin which he claims is not done elsewhere at the moment.

Would be nice if they tried to make their own machines, but this paragraphs is sorry pure drivel. Having design and fabrication under one roof can be done purchasing equipment from ASML and companies like Intel and if we talk historically AMD too where vertically integrated…

The bit about doing memory, well at some point if this AI chips crisis continues I think Apple will make a big semiconductor purchase and get capability to design and manufacture its own RAM and NAND Flash too. Too much of a volatility risk to their cash cows on the HW side.

 

[sono]

His presentation bears watching more than once, there is so much in there, much broader than chip fabrication.

I think an aspect of this is pre-ipo where he presents his vision for his group for an interplanetary journey. Such a journey needs vast amounts of compute.

 

[Donald mac Ronald]

How so?
Isnt space extremely cold?

Putting data centers in space should save a huge amount of energy and cooling cost keeping prices lower for civilians down here on earth.

Maybe you're Elon?

I can guarantee you he said to a room full of scientists and engineers. Why don't we just put data centres in space, free solar and it's cold. The whole room signed.

People also seem to think it would be some box in space with a few solar panels on it. For a data centre with enough compute you would need miles long solar panels just for one centre...

His fabrication idea sounds good though.

 

[deriks]

So Tesla RAM lr whatever is a possibility in the near future?

 

[CGNoire]

Maybe you're Elon?

I can guarantee you he said to a room full of scientists and engineers. Why don't we just put data centres in space, free solar and it's cold. The whole room signed.

People also seem to think it would be some box in space with a few solar panels on it. For a data centre with enough compute you would need miles long solar panels just for one centre...

His fabrication idea sounds good though.

Ok now that your done loling. You feel like actual enlightening me on why what Im saying is incorrect?

I would think cooling pipes on the outside of the sattlelite would allow the heat to escape into outer space. No?

 

[CGNoire]

It's in a vacuum. You have nothing to transfer the heat to. Think about your standard air CPU cooler - now takeaway the air.

Ok makes sense. So how do electronics on space stations expell heat?

 

[pepeno]

So another decade of waiting for a made in USA chip? Or is the Terafab overseas like everything else?

Knowing his type of engineers, their fab will be operational in 2-3 years. In one year they build up the biggest AI cluster on planet and Elon speciality is mass scale fab production.

I still remember time when his Gigafactory was announced for the first time years ago and all "experts" were saying this is fake shit, Elon has no clue about manufacturing business, it's too large, too many issues etc. etc etc. Then they build it and since then they build like 20 of them with some exceeding those.

He needs to get off this stupid idea of space data centres Jesus Christ.

The laws of thermodynamics and physics disagree with him. He says cooling in space is free. This guy is supposed to be smart. Cooling in space is the opposite of free, it's insanely hard. The tech isn't even there for it and to cool shit in space would cost so much more than just build cheaper data centres on earth.

I think dude who owns biggest satelite network in humanity history knows a thing or two about cooling in space considering all of his satellites do that.

But I was too very skeptical about it.The thing is that he doesn't plan to run chips at normal earth like temperature. He plans to run them very very hot like 100s of C hot close to melting point. The thing with radiation cooling is that it has square law in it. Meaning if you double the temperature you double cooling efficiency. Then there is the second square law, meaning the bigger surface area of radiator the faster cooling rate you get.

The reason why you need huge radiators on ISS is because humans love 22-24C temperature to live in. Those satelites/data centres will be operating at 100s of C increasing by multitudes cooling capacity.

On earth radiators are big problem because you have finite space. Even if you have huge factory usually you don't see big blocks of coolers but instead using liquid cooling sollutions, because actually size matters. In space there is no limit to size. They are only limited by build cost and rocket tonage/volume not by space radiator occupies.

 

[pepeno]

Honestly he once again shown the logical pathway. His vision really is something else. That wasn't just vision to meh, let's put space datacentres but actually reason why and scaling from that, why you need moon based city to scale up and so on. Once optimus will go operational we pretty much unlock infinite resources.

 

[poppabk]

Ok makes sense. So how do electronics on space stations expell heat?

By radiation. The advantage to being in space is that radiated heat is just gone - there is nothing for it to be readsorbed by.

 

[jason10mm]

"Every kilowatt of heat generated by an AI chip must be released via radiation.
The Problem: Radiators are much less efficient than liquid cooling. To cool a modern AI cluster (which can pull 100kW+ per rack), you would need massive, wing-like radiator arrays—sometimes larger than the satellite itself—pointed permanently away from the sun."

Just another one of his dumb hype ideas.

And my guess is that his "it's free" comment refers to passive radiative cooling, after it is built into the architecture of the satellite and is up and running, is "free" in that it is passive in nature and doesn't require active cooling (which is just moving the heat somewhere else and dumping it). So you could build radiators to offload the heat generated by the pc, itself powered by solar, and the entire system at that point is isolated and running with no additional energy requirements.

Thus the cooling is "free" in space in ways it usually isn't on Earth unless you are in a place where you can use geothermal or a very cold exterior environment.

I have ZERO belief that Elon thinks space is "naturally cold" and doesn't understand how heat transfer and radiatiors work on space craft like you suggest, that would be a truly absurd notion.

 

[Sophist]

Joint Venture: It is a rare formal partnership between Tesla, SpaceX, and xAI.

A rare partnership between Elon Musk, Elon Musk, and Elon Musk.

 

[EternalSunshine1976]

I have ZERO belief that Elon thinks space is "naturally cold" and doesn't understand how heat transfer and radiatiors work on space craft like you suggest, that would be a truly absurd notion.

Especially as Elon has already worked on this problem on a smaller scale with Starlink satellites. I have a lot of trust in the engineers at SpaceX, they've been just doing the impossible for so long, that designing a way to efficiently radiate heat away from satellites seems very doable, almost pedestrian in a way.

 

[IDKFA]

How so?
Isnt space extremely cold?

Rather than laughing at you like the other member, I'll just outline the issue.

Your thought process is correct and is what a lot of people would rightly think. Space is cold at around -270C, but the problem is that space is a vacuum. This means the traditional cooling methods don't work. Without air to provide convection or conduction, the heat generated would need to be removed via radiation. As there is no air in space, the coldness is very inefficient and your left with a situation where components could easily overheat. The answer to this is to build radiator panels to release heat as infrared light. This solves the cooling problem and also means you wouldn't need vast amounts of water.

It's not impossible. It just requires some extreme engineering to build and deploy a big enough radiator in space, not to mention the cost would be in the billions. It's not a question of would it work, but if it would be cost more effective than having them just on earth.