(Mar 22, 2026 05:17 AM)Yazata Wrote: [...] https://x.com/SpaceX/status/2035519125284380672 [...] Elon said that he talked to all of Tesla's current suppliers - Taiwan Semiconductor, Samsung, Micron Technology - and told them they needed to ramp up their production many times over. He reassured them that he would buy all the chips. But they were hesitant to take such a big step. So Elon decided that he needed to do it himself.

The new fab will be able to produce any kind of chip, performing all the steps from design, through masking, fabrication, lithography, packaging etc, and testing, all under one roof. Nobody else does this. He anticipates a rapid iteration cycle where concepts are tried and tested, then changes are built into the next iteration. (Elon says he has some ideas for new physics in chip design and manufacturing, that he believes should work, that his new vision should be able to make, test and improve very rapidly. Thus improving the chip state-of-the-art.)

[...] https://x.com/GregAbbott_TX/status/2035584741097767201 [...] https://terafab.ai/

Quote:"Ok. So essentially Terafab is not a plan for a 'standard' semiconductor fab, but includes mask production, test house and packaging under one roof. This alone is a very alien concept. If they only achieve this alone, not even considering reaching the crazy output numbers, they're already in unicorn territory.

All the plans on top what they actually are willing to do with the silicon are essentially Dyson Sphere Program in real life. There is no other way to put this.

In this very case, as this directly affects an industry I happen to work in: I assign a major probability to this either failing or delaying much beyond the intended timeline. Sure you can buy equipment and hire experienced people, but semiconductor manufacturing in advanced nodes is essentially the farthest humanity has ventured down the tech tree. It is so incredibly difficult and complicated that it's essentially impossible to explain to the average joe how it even works without simplifying it into oblivion. 95% I talk to can't even grasp correctly what I do for a living, and digital design is very far up the stack.

I'm saying this not because I'm rooting against Elon or xAI, Tesla or SpaceX, but just to manage expectations.

Starlink has proven with their satellite cadence and PCB in-house pipeline that these companies do understand how to build stuff reliably at scale and down to a sustainable price point.

If there is one guy and his entourage that can actually succeed at this, it's this. Nobody else.

But expect bumps along the way. This might be the single thing in tech more complicated than the literal rocket science."

Quote:"Given that several companies make advanced chips, but no companies have ever made fully reusable rockets or achieved SpaceX scale, I think Starship is harder, but we shall see.

Terafab will technically be two fabs, each making only one chip design. This greatly simplifies process flow and allows more linear, adjacent movement of the FOUP.

A super high production rate allows us to test very quickly what steps can be deleted, simplified or sped up, even after the design is fixed. Current fabs are extremely conservative, operating on rigid historical heuristics, which are mostly, but not all, correct.

Anything that is a rate limiter at the machine level means that machine will be redesigned, unless already at limit of physics.

Having new iterations of a chip design be produced every day in the research fab (with <7 day lag) means being able to try out many high risk, high return ideas.

Etc

In any event, there is no other way to reach extreme scale, so either we make Terafab or we will be stuck at the ~20% chip/memory output growth per year of the current industry."

Quote:"So is the plan to actually design and build the actual fabrication machines like ASML?"

Quote:"As a guideline: this is one of the more famous examples where bottlenecks currently exists. But the supply chain in semiconductors has other nasty ones. Japan for example basically has a monopoly on photoresist. And they probably are not setup to just supply such an endeavor without major waiting times. And ASML has upstream suppliers like Zeiss and Trumpf which probably also do not expect a sudden explosion in demand either.

Essentially it's a 20 to 30 layers deep cobweb of dependencies and just like with cars, you move as fast as the slowest component.

So this will only work if they can manage to simplify stuff. Which would certainly be advantageous for everyone. I'm pretty exited to see how this plays out."

Quote:"It's really outstanding how many commenters clearly have limited to no reading comprehension.

Never did I say they WILL fail, nor did I wish it upon them. (the opposite is the case)

I was just outlining why I think this is an exceptionally difficult endeavor. I'm even saying: if someone will succeed at this, it's gonna be them.

But delays and failures along the way are normal, and many major programs that were hard and started by Elon had them.

Dry electrode and 4680? -> succeeded now, but with major delays
FSD -> turned out far more difficult than initially suspected, delay on original timeline is closing in on one decade
Starship -> coming along nicely, but also had major hiccups on Ship 2 iteration

Manufacturing innovative high-tech things at scale, at a sustainable cost, with sufficient yield is extremely difficult. That is why highly specialized supply chains exist. For most of these things, there is only one supplier, two if you're lucky.
In some cases, it's the companies themselves, and that is especially true for anything founded by Elon.

Once you get into that area of tech it is basically vertical integration or don't do it. But vertical integration also means you might need to re-learn some of the things the supply chain has already learned implicitly before you. On the other hand: you can also avoid entrenched mistakes or inefficiencies.

A lot of words for: This is not going to be a walk in the park."

Quote:It’s been 65 years since I heard John F. Kennedy’s speech in Spring of 1961, when he committed America to go to the moon before the end of the decade. When I heard that, I decided to skip graduate school at CCNY and apply straight to NASA.

No speech since has moved me the way it did — until this past weekend.

What we saw this weekend — which, of course, we didn’t see much of in the traditional media — was, I believe, one of the most consequential industrial announcements in modern American history. And it wasn’t because of lunar mass drivers or space datacenters. Allow me to explain.

What Musk Laid Out

Tesla, SpaceX, and xAI – recently tucked in to SpaceX – intend to build the largest semiconductor fabrication facility ever attempted. A 100M sqft, $20-25B plant in North America. Specifically, on the North Campus of Giga Texas in Austin, targeting 2-nanometer (2nm) process technology, the most advanced node on Earth. Characteristically, they’re calling it “Terafab.” The goal is to produce 1 TW of compute output per year. For proportion: every advanced chip fab on Earth combined currently produces ~20 GW. Musk is after 50 times that...

The unofficial conglomerate will pull two families of chips under one roof: the AI5 and 6, for Tesla’s vehicles and Optimus humanoid robots, and the D3: a radiation-hardened chip built for the hostility of space (high-energy ions, photons, electron buildup, and extreme thermal cycling).

Run It Hotter

I know a few things about building systems for space. It is a punishing environment. Every direction you turn, a new constraint wants to kill your mission.

Rather than adapt existing commercial silicon for orbit, SpaceX is instead working towards chips that run at much higher temperatures, which means more efficient heat rejection, which means dramatically lighter radiators. Every kilogram you save changes the entire mission architecture. Anyone who has designed a spacecraft knows what that trade can buy you. As Elon said: “You want to optimize it for space and run it a little hotter than you would normally run a chip on Earth to minimize the radiator mass.” Heat dissipation scales with the fourth power of temperature...

...The physics argument is roughly this: the sun hits a solar panel with 5x the energy it gets on the ground, heat rejection is roughly 5x more energy per square meter in orbit than on the ground, and it is far more efficient than on Earth. At the launch costs and timelines the Starship program aspires to reach, Elon optimistically projects that orbital AI compute could undercut terrestrial alternatives on cost within two to three years.

But, again, none of this is what really stopped me in my tracks.

What Stopped Me Cold

What I most admired from Elon’s speech this weekend is discovering that an American group at the cutting-edge of science and technology – an elite team at SpaceX and Tesla – is going back to physics, designing from first principles, and taking on the status quo in semiconductors.

“We’re not just going to do conventional compute,” Elon said. “There’s some very interesting new physics… we’re going to push the limit of physics in compute and try a bunch of wild and crazy things.”

The AI industry is locked into the GPU/CUDA paradigm — a general-purpose architecture originally designed for graphics that accidentally became the workhorse for neural networks. It works, but it’s brute force, power-hungry, and reaching its limits. Musk is gesturing at a very important question: is the GPU actually the right architecture for inference in robots, autonomy and vehicles, and AI in orbit? Or do you go back to the physics and find something better?

And…what possibilities open up when you can make what you design? It has been roughly 30 years since just about anyone in a position of authority seriously proposed building a leading-edge logic foundry in this country. Thirty years. Of course the person who finally did isn’t proposing to copy/paste what already exists, or work off of the same process nodes and same set of instructions as everyone else….

...Like Elon, I’ve been on the other side of “this will never work.” I know what it’s like.

The naysayers said Faster Better Cheaper was doomed. That reusable rocketry was impossible. That Commercial Crew would never work. That end-to-end autonomy was a pipe dream. Wrong, every time. Musk has a habit of doing things the experts say cannot be done. The timelines are often off. But the things get done.

Here is how I reconcile the idealist and the realist: even if the orbital datacenter vision doesn’t fully pan out — and that is a high bar, given how much hype has pile-driven into the concept at this point — if Tesla and SpaceX are to take a serious crack at it, the byproducts alone would be worth the effort. A leading-edge American fab! A new compute architecture designed from first principles! Sovereign chip manufacturing capacity for the first time in a generation! Radiation-hardened silicon for national security! If all we “get” out of Terafab is those things, it will have been a tremendously great deal for this country...