Elon Musk is set to produce chips: a trillion-scale high-stakes gamble by an "outsider"
Recently, during the live broadcast of the "X Takeover" event outside Tesla's Gigafactory, Elon Musk spoke to the camera with an extremely candid yet arrogant statement:
"I still have a lot to learn about chip manufacturing. To be honest, I don't know how to make chips at all. I have no idea how chips are made, but we will figure it out. There has to be a starting point. It's just like how I once had no idea how to build cars or rockets before."
Shortly before and after this "layman" speech, a plan that is enough to shake the global semiconductor industry has surfaced: Musk has officially launched Terafab, an ultra-large-scale chip factory project. This project is jointly promoted by Tesla, SpaceX and xAI, with its site selected in Texas. The initial investment is 16.8 billion US dollars, and the total final investment is expected to reach a staggering 1.19 trillion US dollars.
His goal is to build a super fab that integrates design, manufacturing, packaging and testing, to achieve an annual AI computing power output of more than 1 terawatt (TW), which is equivalent to 50 times the total annual output of AI chips of the current global semiconductor industry. To speed up iteration, he even plans to manufacture and test about 10 different chip designs at Terafab every year.
As soon as the news came out, traditional chip giants such as TSMC and Intel all paid close attention. TSMC said it is willing to cooperate, on the premise that the factory must be operated by TSMC; Intel is more straightforward, announcing that it is willing to join the project. Musk's response still carries his signature arrogance: the factory will be built and operated by Tesla and SpaceX themselves, and TSMC "can at most sublease a part of it."
Extremely ravenous demand for computing power and the ambition of closed-loop layout
Why would a person who builds cars and rockets suddenly cross the border to tackle the hardest nut in the semiconductor industry?
The most direct driving force is the explosively growing hunger for computing power. Musk said bluntly at the event: "We either build Terafab, or we won't have enough chips."
According to his plan, 80% of the computing power produced by this super factory will be used for space data centers, and 20% will be used for Tesla's FSD autonomous driving, Optimus humanoid robots and Robotaxi fleets on the ground. These businesses have an astronomical demand for chips. Take Optimus as an example, if the annual output reaches the level of 10 million units in the future, the customized AI inference chips required inside it will be a bottomless pit of demand.
The more core anxiety lies in supply chain security. At present, the lifeline of AI chips is almost entirely in the hands of NVIDIA and TSMC. Musk is well aware that pinning the future of Tesla and xAI on others' capacity expansion is not only costly, but also faces the risk of being "strangled" at any time. Especially at a time when geopolitics is full of uncertainty, realizing "computing power sovereignty" and "chip freedom" has become the only way for him to build a closed loop of his business empire.
In addition, Musk is extremely dissatisfied with the speed of traditional fabs. He criticized that traditional chip manufacturing relies too much on standardized equipment, and it often takes two years for a chip design to go from conception to mass production. He hopes to develop self-developed automated equipment, relying on the technical accumulation of Tesla's automation department in Germany, to shorten the "design-manufacturing-test" feedback cycle from several months to several weeks. In other words, he wants to directly apply the rapid iteration logic of rocket manufacturing to chip manufacturing.
Chip Manufacturing: A Mountain Harder to Climb Than Mars
However, semiconductor manufacturing is the most complex and capital-intensive industrial system in the world, bar none. Will Musk's "First Principles" work here?
The first natural moat is equipment. Advanced process chips (such as 2nm and below) are highly dependent on extreme ultraviolet (EUV) lithography machines, and the only global supplier is ASML from the Netherlands. The delivery cycle of EUV equipment usually exceeds 12 months, and the supply is highly concentrated. Analysts point out bluntly that obtaining these key equipment is the primary bottleneck faced by any new entrant. Even if Musk can build the plant, he cannot etch advanced silicon wafers without ASML's machines.
The second natural moat is yield. A person familiar with TSMC's internal situation pointed out that building a fab itself is not the most difficult part. The real challenge lies in continuously producing chips with stable quality and sufficiently high yield. Yield control for 2nm process requires long-term experience accumulation and process data precipitation, which cannot be solved simply by pouring money or relying on a few genius engineers. Stacy Rasgon, an analyst at Bernstein, even believes that the difficulty of building Terafab is "greater than sending a rocket to Mars".
The third natural moat is talent and supply chain. The United States currently has a shortage of about 30,000 semiconductor engineers. Musk has posted a recruitment advertisement on X, inviting world-class chip design and production engineers to join. In addition, the highly integrated model proposed by Musk that "integrates logic chips, memory chips and advanced packaging in the same factory" may instead increase system complexity in the eyes of the industry. The division of labor evolution of the semiconductor industry over the past few decades has its profound economic and technical logic, and forced integration often brings catastrophic yield problems.
Finally, there is the huge capital black hole. Morgan Stanley estimates that the cost of building a fab with a monthly output of 100,000 advanced logic chips may be as high as 45 billion US dollars. Tesla has invested about 20 billion US dollars in the fields of autonomous driving and humanoid robots in 2026, and this budget does not include the expenditure for Terafab. The superposition of multiple high-investment projects will pose a severe test to the company's cash flow.
Timeline and Echoes of History
More critically, Musk has not yet announced a clear construction timeline and capacity ramp-up path for Terafab.
As a reference, Micron's fab in Idaho started construction at the end of 2022 and is expected to start shipping chips in mid-2027 — which is already the pace of a relatively mature semiconductor company. Musk's SpaceX Starship project has also been delayed for at least two years in R&D since it was selected by NASA as a lunar landing craft in 2021. The long cycle of high-complexity, asset-heavy projects is a physical law that Musk must face.
But he obviously doesn't care about these conventional industry warnings. The sentence "I don't know how to make chips at all, but we will figure it out" perfectly sums up his entrepreneurial philosophy: set a grand goal first, then force yourself to learn, integrate resources, and break industry conventions.
The Ultimate Gamble on "Time Sovereignty"
If you only regard Terafab as a chip factory, you may underestimate Musk's intention. Some analysis points out that what he really wants is "time sovereignty" in the AI era — to compress the feedback cycle of chip iteration to the extreme, so as to obtain an asymmetric speed advantage in the computing power competition.
The plan to deploy 80% of computing power in space reveals ambition in another dimension. The average output of the U.S. national power grid is only 0.5 terawatts, while Terafab's target is 1 terawatt. The huge physical constraint of energy consumption forces him to "send computing power into space". Musk is building a two-layer architecture of "ground verification - space deployment" to lay infrastructure for his Starlink, space data centers and even the more ambitious Mars colonization plan.
The success or failure of Terafab will test a fundamental question: whether Musk's methodology of "cross-border dimensionality reduction strike, learning in actual combat" is still effective in the semiconductor industry, the crown of human industrial civilization.
If it works, the power map of the global chip industry will be rewritten, and NVIDIA and TSMC will face the most unorthodox opponent ever; if it fails, this will be the most expensive "tuition" he has paid so far.
Ten years ago, when he said he would build electric cars and recyclable rockets, countless senior experts also asserted that he did not understand manufacturing and aerospace. Today, when this person who "doesn't understand chips" stands in the sun of Texas again, the entire Silicon Valley and Wall Street have to hold their breath. After all, the last time he said "we will figure it out", the rocket really landed steadily back at the launch site.
This article is from the WeChat official account "Tech Business", author: Tech Business, published with authorization from 36Kr.