The last few days I’ve been grappling with exactly what to say about Elon Musk’s Terafab announcement.
Elon Musk on Saturday announced that his rocket making company SpaceX and electric vehicle major Tesla will jointly run a new chip manufacturing facility in Austin, Texas, AFP reported.
The chips will be for use in artificial intelligence projects, robotics and data centres for space, with aim to produce 1 terawatt of computing power per year, it added, citing Musk. (1 terawatt = ¬1 trillion watts).
Why is Elon Musk starting his own chip manufacturing facility?
The world’s richest man said the so-called “Terafab” was necessary because Tesla and SpaceX’s demand for computing power is expected to far exceed that of global chip suppliers. While he did not disclose how much initial investment is being made, reports pegged the initial infusion between $20-25 billion, it added.
“We’re very grateful to our existing supply chain, to Samsung, TSMC, Micron, and others… but there’s a maximum rate at which they’re comfortable expanding. That rate is much less than we would like… and we need the chips, so we’re going to build the Terafab. The advanced technology fab in Austin will have the facilities to design, manufacture, test and improve each chip,” Musk said.
Other articles have made clear the the initial pilot line will be in Austin on Tesla’s Gigafab land, but that the eventual full-scale production fab may be located somewhere else.
The project aims to make chips to support 100 to 200 GW of computing power on Earth, and 1 TW in space.
Musk did not share a timeline for the production or output. The AFP report noted that the billionaire has many times in the past announced timelines that his companies failed to meet, i.e. robotaxis, SDF, etc.
According to Musk, the Terafab would ultimately help humanity become a “galactic civilization” capable of harnessing the resources of other planets and stars.
A Bloomberg report added that Musk has said previously that the facility would produce 2 nanometer chips.
It added that the Austin facility is expected to make two types of chips, one of which will be optimized for edge and inference, primarily for his vehicle, robotaxi and Optimus humanoid robots.
The other will be a high-power chip, designed for space that could be used by SpaceX and xAI. Musk said he expects xAI to use the vast majority of the chips, it added.
Musk also unveiled a speculative rendering of a future “mini” AI data center satellite with 100 kW power. This is part of a larger installation he wants SpaceX to build to do complex computing in space.
In January, SpaceX requested a license from the Federal Communications Commission (FCC) to launch one million data center satellites into orbit around Earth. “We expect future satellites to probably go to the megawatt range,” Musk added.
Let’s skip over addressing the feasibility (or desirability) of the space half of this grand vision. Never mind that Musk has evidently been freebasing back issues of The L5 News and G. Harry Stine’s Analog columns circa 1979 to conceive his solar-powered orbital cathedrals of vast, cool machine intellects.
No, I want to focus on the task of getting a purely ground-based, cutting edge, 2nm semiconductor fab up and running. There’s nothing impossible about that part, and $25 billion is right around the current price to build a state-of-the-art fab. Other reports say:
Production targets are specific. The facility is designed to produce between 100 and 200 billion custom AI and memory chips per year, targeting an initial output of 100,000 wafer starts per month with a stated ambition to scale toward one million — roughly 70% of TSMC’s current total output, in a single US facility.
Tesla is targeting 2 nanometre process technology, the most advanced node currently in commercial production. Tesla’s fifth-generation AI chip, AI5, is among the first products Terafab is designed to produce, with small-batch production expected in 2026 and volume production projected for 2027.
And here’s where my deep skepticism kicks in, over both the output and timeline.
Having more cutting-edge fabs here in the continental United States is a great idea, and there’s nothing wrong with vertical integration for companies to combine design, fabrication and test under one roof. Indeed, Musk has merely reinvented Integrated Device Manufacturing, which used to be the norm until the foundry model took hold, and that’s still the norm with companies like Intel (though not necessarily all under one roof). And Intel, Samsung and TSMC already have their own mask manufacturing facilities in-house. So based on the little information available now, Musk’s radical new approach isn’t.
But that timeline of initial production in 2026 and volume production in 2027? Unless Musk has already placed orders for the critical semiconductor manufacturing equipment (especially however many ASML TWINSCAN EXE:5000 or TWINSCAN EXE:5200B EUV lithography systems the fab will need) long before now, that’s simply not happening. The lead time for the most critical machines are a year or more. And that’s just one machine.
There are five essential semiconductor equipment manufacturers:
Applied Materials
ASML
KLA
LAM Research
Tokyo Electron
If you’re building a modern, 2nm fab, chances are pretty good you need all five. You need the ASML litho machines mentioned above. You need KLA inspection tools to raise and maintain yields, and you need, at the very least, one of AMAT, LAM or TEL to provide the rest. Take away all three and you can’t equip a fab, period.
And that’s after you’ve poured the agonizingly precise, ridiculously level concrete slab for the equipment to rest on, installed huge networks of high voltage power systems (ASML machine alone now run at 1,000 watts), ultra-high purity air handling equipment and ultra-high purity de-ionized water handling systems. All that needs to be installed, up and running before you even start installing the tools. And then all the tools need to be brought up and qualified for production. And that all needs to be done before tape-out, mask production and the first wafer is run.
But the timeline isn’t the only thing that’s delusional about Musk’s announcement.
Terafab is designed for an initial output of 100,000 wafer starts per month, with ambitions to scale to 1 million wafer starts per month at full capacity. For context, that full-scale target would represent roughly 70% of TSMC’s entire current global output — from a single facility operated by companies that have never fabricated a chip.
Musk said the facility would produce between 100 and 200 billion custom AI and memory chips per year, powering Tesla’s “Full Self-Driving” software, the Cybercab robotaxi program, and the Optimus humanoid robot line. He also said millions of Optimus robots would help build and operate the facility.
Put aside the swarms of robots for now. A lot of what can be automated for a fab already has been with automated FOUP-handling equipment. What hasn’t been automated is routine maintenance, troubleshooting, etc. I’m guessing it will be quite a while before Frankie Fabbot is ready to take on any of those tasks.
100,000 wafer starts a month is around what TSMC’s cutting edge fabs produce a month. Samsung has a huge fab supposedly capable of 450,000 wafer starts a month. A million starts a month in one extremely big fab is possible, but it’s going to take multiple lines, lots of money, and throughput will still be gated by the lowest capacity machines on the line (again, I’m guessing the ASML litho machine). Modern fabs have already maxed out production speeds for modern high-yield plants. You can’t simple pull a lever to make the line go faster.
Then there are his plans to build memory chips and space-hardened versions of the xAI chips in the same megafab. Neither of those goals is impossible, but it’s going to require more equipment than his $25 billion estimate to essentially run three production lines in the same building. And preparing chips to operate in space generally involves completely different “rad-hard” processes, involving different types of wafer chemistry, that are nowhere near the 2nm process node.
Musk is a smart guy. I bet he can eek a few performance gains out of how his fabs are organized, AI line optimization, etc. But unless he’s smarter than Einstein, Thomas Edison, Tony Stark and Reed Richards combined, I’m hard-pressed to see how he can optimize more than, say, 15% over existing cutting edge fabs.
Nothing Musk has proposed for the earth-bound side of Terafab is impossible, but it is all-but-impossible to implement in the time-frames he’s alleging, for the money he’s claiming. One million wafer starts a month in a single fab? Doable, but not by 2027, and not for $25 billion. Having such a fab up and running full tilt in 2029-2030, for $100 billion, seems at least at the edge of feasibility. Whether the AI boom will still be a “thing” by then is unknown, but a fab actually capable of one million wafer starts a month (robots or no robots) is still a license to print money…
