If you confidentially filed this past week to do an IPO of your flagship SpaceX in late June, in the week of your 55th birthday, aiming to raise US$75 billion at a valuation of around US$2 trillion, you clearly have the wherewithal and the track record to undertake any project on earth, or indeed in the sky. But to get there, you need a grand plan to build the most powerful processor chips on earth first.
Addressing the chip supply problem
Not surprisingly, that’s just what Musk did last month when he announced that three of his companies were teaming up to spend up to US$25 billion building a Terafab, a giant chip foundry, near Tesla’s Giga Texas campus in Austin. Terafab, as Musk has conceived it, won’t just address the huge global chip shortage the world is witnessing right now but also expand AI processor capabilities. “We either build the Terafab, or we don’t have the chips,” Musk said. His plan is to build a very big fab that includes logic, memory and packaging in the US. “That’s going to be very important to ensure that we are protected against any geopolitical risks.”
Terafab’s facility, which is likely to exceed 100 million sq ft, or roughly the size of 1,350 soccer fields, will integrate chip manufacturing for cars, robots and orbital satellites. It won’t just build Tesla-designed D1 Dojo chips used in Tesla’s Fuel Self Driving (FSD) cars or AI4, AI5 and AI6 inference chips it has designed for real-time processing in Optimus humanoid robots and autonomous driving but also an array of other chips and sensors, including high bandwidth memory or HBM chips designed to deliver extremely high-speed data transfer at low latency for high-performance AI computing.
Musk believes bringing fabrication, memory and packaging under one roof in Texas, Terafab could help create a “recursive design loop”, allowing chips to be designed, printed and tested in days rather than weeks or months. Chips made by TSMC are designed in the US by Nvidia, Apple, AMD, Broadcom and others; manufactured in Taiwan; and packaged and tested in Taiwan, Penang, Singapore or Vietnam. They are then sent out to end manufacturers like Hon Hai Precision, which makes Apple’s iPhones at its plants in Shenzhen. Terafab won’t just be a massive vertically integrated chip facility. It will be located right next door to Tesla’s plants, where Cybercab Robotaxis and Optimus Humanoid robots are being manufactured. Moreover, if the Terafab experiment succeeds, it would give Musk’s firms much greater control over the chip-development cycle.
See also: OpenAI pauses Stargate UK data centre citing energy costs
Here’s why Musk believes the world needs the biggest chip-making plant ever built: current chip supply projections of the big two chip fab operators TSMC and Samsung Electronics — the only companies with the capability to build top-of-the-line 2-nanometre chips — will satisfy just 2% of what the three tech giants that he controls — SpaceX, Tesla and xAI — need. Whether you agree with Musk’s right-wing policies or his grand vision of “colonising” Mars, one thing is certain. If he can build the Terafab, he would fundamentally reshape how the US produces the processors that power AI, electric vehicles (EVs), as well as space exploration. Indeed, everything that a chip processor powers.
The way Musk sees it, the Terafab chip manufacturing project is the logical “next step” towards harnessing the power of the sun. The giant chip fab will run on solar power. Musk said nearly 80% of Terafab’s total output would eventually be dedicated to space-based AI satellites. He said solar irradiance is five times stronger in orbit, potentially making space-based inference cheaper than terrestrial data centres within three years. Even if that takes five years instead of three, space-based data centres will be a better option than their earth-based counterparts.
Musk believes Terafab could also help create a “galactic civilisation”, his long-held vision for the future — one where humans expand beyond Earth to colonise other planets like Mars, creating a massive, orbiting solar-powered network of satellites and data centres in space, powered by constant solar energy. By the way, next year Musk is targeting a Starship flight to Mars that will carry Tesla’s humanoid robots. A successful flight, he has vowed, would pave the way for human landings on Mars probably as soon as 2029, although “2031 is more likely”, he recently tweeted on X. The goal is simple, he said in Austin last month: “Turn science fiction into science fact.”
See also: Thailand seizes US$260m assets linked to alleged scammers
All of this might sound highfalutin sci-fi gobbledygook to you, but companies owned by Musk, as well as those owned by Amazon.com’s founder, Jeff Bezos, are already spending tens of billions of dollars a year to make those space dreams come true. While there is war in the Middle East and families and small businesses fret over supply chain bottlenecks, inflation, and high interest rates, the innovation economy continues to spend hundreds of billions of dollars annually on AI and space-related infrastructure. What are Elon Musk and his peers at Google, Amazon, Meta Platform, OpenAI and Anthropic trying to do? The goal is simple, Musk said in Austin last month: “Turn science fiction into science fact.”
Costs and economic viability
Essentially, Musk is proposing to build only an in-house chip manufacturing plant owned by three firms he runs as CEO — Tesla, SpaceX and its new AI start-up, xAI. The trio eventually aims to reach 1 terawatt (TW) of compute per year. That’s one trillion watts or 1,000 gigawatts. Let me give you a better sense of just how huge 1TW really is. One terawatt, or 1TW, is 50 times the current global compute capacity.
The chips will initially be used only by other Musk firms, which are the giant chip fab’s main investors. For someone who is almost a trillionaire and whose flagship firm, SpaceX, is readying the largest IPO ever for late June, the US$25 billion initial cost of Terafab is small change, with three separate entities pooling money to get the project going. While SpaceX might soon have access to a boatload of money, Tesla is not in the greatest of financial shape. The EV pioneer generated US$6.2 billion in free cash flow last year, but Goldman Sachs estimates that with EV sales tanking, Tesla’s free cash flow could turn negative this year.
Analysts say just the first phase of construction could cost up to US$35 billion. To reach 1TW, Terafab would need to build up to 360 foundries the size of TSMC’s first fab in Arizona, Stacey Rasgon, Bernstein’s chip analyst, noted in a recent report. Here’s his math: “1TW of annual compute would require somewhere between seven and 18 million 300mm wafer starts per month (WSPM), dominated by High Bandwidth Memory (HBM). This would be equivalent to 140 to 360 new 50kW WSPM factories, or US$5 trillion to US$13 trillion of capex spend at US$35 billion per fab-equivalent,” he noted in his report. Put another way, “the required 1TW capacity would be on the order of the entire current global installed (chip) capacity base,” he noted.
Building chips in a 2nm-capable foundry from scratch is one of the most capital-intensive bets in industrial history. It requires manipulating materials at the atomic scale with near-perfect yield while coordinating thousands of steps using machinery that pushes the boundaries of known physics. A 2nm chip is so small that technology interacts with individual atoms. Foundry engineers are dealing with tolerances measured in fractions of an atom. Only two companies — TSMC and, more recently, Samsung Electronics — have successfully operated a 2nm foundry. The world’s No 3 player, Intel, has splurged tens of billions of dollars over the years and has yet to mass-produce 2 nm chips. Its main problem is low yields or a higher level of defective chips. Chinese foundries are at a much lower 7nm level mainly because they are banned from accessing chip equipment and tools made by Dutch giant ASML Holdings, which is the sole provider of such sophisticated tools.
Tesla’s detractors say Terafab is just another Musk distraction. Six years ago, the Tesla CEO promised to deliver 20 million EVs a year by 2030. Last year, Tesla delivered just 1.6 million EVs. Musk officially abandoned his 20 million electric cars target in May 2024 and made a hard pivot to delivering robotaxis and humanoid robots. Robotaxis are an old obsession of Musk. In 2018, Musk proudly announced that there could be a million Tesla robotaxis by the end of 2019. Seven years on, Tesla operates just ten unsupervised or driverless robotaxis in just one city — Austin, Texas. Other Tesla robotaxis operate only with a safety driver. Its rival, Waymo, majority-owned by Google, currently has over 3,500 driverless robotaxis plying in 10 US cities. Waymo expects to launch driverless robotaxi services in London and Tokyo by the end of September. And WeRide recently began pilot robotaxi services in Singapore and Dubai.
To be sure, Musk has taken on complex challenges before, like sending 10,000 satellites into orbit, going from near bankruptcy a decade ago to a net worth of nearly US$1 trillion. Yet building a profitable chip Terafab might be the most challenging task he has ever undertaken, even harder than colonising Mars. Musk has repeatedly proven naysayers wrong far too many times, so you don’t want to bet against him. “Why is Tesla doing this on their own,’ wonders Barclays chip analyst Dan Levy. ‘How will Tesla handle the risks ahead (like) steep learning curve with no experience manufacturing chips at scale, lack of IP on making leading edge nodes, complexity in packages, as well as long lead-times on key equipment like ASML’s extreme ultraviolet lithography?” For now, Terafab remains a “show-me story,” Levy argues. He expects to see much smaller-scale aspirations at least in the near to mid-term. For his part, Musk, who easily gets bored with mundane things, might ultimately move on to other, more formidable challenges on earth or in space.
Assif Shameen is a technology and business writer based in North America
