The Largest Chip Manufacturing Plan in History Launches: Why Is Musk's Appetite So Big?

The world’s richest person and Tesla CEO Elon Musk has announced one of his most ambitious visions to date — Terafab.

On March 22, Musk held a launch event for the Terafab project in Austin, Texas, officially announcing the joint effort by Tesla, SpaceX, and xAI to build a 2-nanometer wafer factory. This project is seen as a key move for Musk to break through the global chip supply bottleneck.

This factory, dubbed by Tesla as “the largest chip manufacturing plant in history,” aims to produce 1 terawatt (TW) of AI computing power annually, primarily for space deployment. Musk stated that the current global AI computing output is about 20 gigawatts, and Terafab’s annual capacity will be 50 times that.

Tesla and SpaceX’s computing demands far exceed current supply

The scale of Terafab is so grand that Musk himself described his plan using words like “crazy” and “physical limits” during his speech.

Behind this is the reality of a global chip capacity gap, as well as Musk’s long-standing goal of deploying space-based computing power and advancing multi-planetary civilization. His blueprint prioritizes solving short-term chip shortages to support mass production of Optimus robots and space AI satellite networks; mid-term, leveraging low-cost space computing to expand applications and boost Earth’s economy; and long-term, using lunar bases to achieve computing leaps, pushing humanity toward becoming a multi-planet species and a “galactic civilization.”

Two wafer factories with unprecedented full-process closed-loop production

Why go into chip manufacturing at all? In Musk’s view, current global chip capacity simply cannot meet his future needs.

Although Musk has clearly stated he will continue purchasing chips from existing supply chain giants like Samsung, TSMC, and Micron, and has expressed gratitude to these suppliers, their expansion speeds are far from enough to meet his project’s demands. He bluntly said, “Either we build Terafab, or we will have no chips to use.”

Musk explained that Terafab will have two wafer factories, each focusing on a different type of chip, and both will achieve full-process closed-loop production.

Notably, Terafab will break the current division of labor in global chip manufacturing by consolidating lithography masks, chip fabrication, packaging, and testing within a single facility, creating a rapid iteration cycle of “mask making — chip manufacturing — testing — mask optimization — re-manufacturing.”

Musk revealed that no existing facility worldwide can integrate logic, memory, packaging, testing, and lithography masks all in one place, enabling this full-process, high-speed iteration that surpasses conventional production lines by an order of magnitude. This will support extreme process testing for computing chips and new physics research.

“We’re not just producing computing chips in the traditional way. I believe some very interesting new physical directions are feasible. Given time, we will succeed. We will truly push computing chips to the physical limits,” Musk added.

Regarding applications, he explained that the two wafer factories will have clear divisions of labor, focusing on two types of chips to meet different scenario needs.

Musk plans to produce multiple types of chips

The first type is edge inference optimization chips, mainly for Optimus humanoid robots and Tesla’s autonomous driving systems, with robots being the core market. Musk predicts that global annual vehicle production will reach about 100 million units, and future humanoid robots could reach 1 to 10 billion units annually—demanding 10 to 100 times that of cars. Tesla aims to capture a significant share of this market, which will also drive increased chip production.

The second type is high-power custom chips for space, designed to withstand extreme space environments and deployed in SpaceX’s orbital AI data center network. Space presents radiation challenges such as high-energy ions, photons, and electrons, so these chips will have higher resistance to interference, aging, and radiation than ground-based products. To reduce the payload weight of space radiators, their operating temperatures will be slightly higher than terrestrial chips, with process parameters and fault tolerance standards specially tailored.

Shifting computing deployment to space, with costs expected to fall below ground within 2-3 years

Terafab’s focus on space deployment stems from Musk’s belief that Earth’s energy and computing capacity are inherently limited.

Data he presented shows that Earth receives only one five-hundred-millionth of the Sun’s total energy; the Sun accounts for 99.8% of the total mass in the solar system. Humanity’s annual electricity output is only about one trillionth of the Sun’s total energy. Even if energy production increases by 1 million times, it would only reach a millionth of the Sun’s energy, capping Earth’s computing capacity expansion.

In contrast, deploying computing power in space offers significant advantages: space has no atmospheric attenuation or day-night cycle, satellites always face the Sun, and solar energy collection efficiency is over five times that of Earth. Space solar panels do not require heavy glass or frames to withstand extreme weather, reducing hardware costs. On Earth, high-quality deployment sites are increasingly scarce, and expansion costs keep rising. Space, however, allows virtually unlimited scaling, with unit costs decreasing as size increases.

Musk predicts that within 2-3 years, space-based AI computing costs will be lower than terrestrial ones. “Once the cost of getting into orbit drops, placing AI computing in space becomes almost obviously worthwhile. As scale increases, space will become cheaper and easier; on Earth, as more computing is deployed, available space becomes more limited.”

He believes computing resources will be allocated by scenario, with limited power supply on Earth supporting only 100-200 gigawatts annually (about 20% of total capacity), while the remaining terawatt-level main computing power (~80% of total capacity) will be sent into orbit.

“To reach 1 terawatt of annual computing, we need to send about 10 million tons of payload into orbit each year, assuming 100 kW per ton. We are confident we can do this without any new physical laws. It’s not an impossible task. I believe SpaceX can deliver 10 million tons into orbit each year,” Musk said.

Beyond 1 terawatt, lunar bases aim for a thousandfold increase in capacity

Terafab is not Musk’s ultimate goal.

He also revealed long-term plans to build electromagnetic mass drivers on the Moon to achieve a thousandfold increase in computing capacity. The Moon’s lack of atmosphere and gravity only one-sixth that of Earth means no rockets are needed; payloads could be directly accelerated to escape velocity via these drivers, boosting capacity from 1 TW to 1,000 TW (zettawatt level), drastically lowering deep-space deployment costs.

Musk showed a video screenshot of the lunar electromagnetic mass driver.

“I really hope to witness the completion of the lunar mass driver in my lifetime—that would be spectacular,” Musk said.

He sees this plan as creating clear economic and capacity growth: once the lunar driver is operational, humanity could utilize a tiny fraction of the Sun’s energy—enough to expand Earth’s economy by 1 million times. “Then we will continue expanding to other planets and stars, creating the most exciting future I can imagine.”

Musk envisions humanity reaching “astonishing prosperity” in the future.

Threats to TSMC? Analysts say it still faces many challenges like yield issues

The emergence of the Terafab project has drawn intense attention within the semiconductor industry, even raising concerns about potential pressure on industry giants like TSMC.

However, analysts believe the plan may face significant technical, financial, and structural obstacles.

Building a wafer factory from scratch is widely regarded as one of the most challenging engineering feats in modern industry. Morgan Stanley analysts described it as a “monumental task,” estimating costs could exceed $20 billion and take several years to complete.

The semiconductor industry is highly specialized, with clear distinctions between fabless companies like NVIDIA and dedicated foundries. Musk’s proposal to integrate logic, memory, and advanced packaging technologies runs counter to decades of industry specialization.

Chip manufacturing requires not only capital but also deep process expertise accumulated over years of large-scale production.

Some industry insiders note that starting with a 2-nanometer process is extremely difficult; the factory’s construction may not be the biggest challenge—yield control will be the ultimate hurdle. Yield depends on stable demand and continuous iteration, which even mature companies struggle to maintain at very high levels.

Additionally, Terafab faces structural issues, such as equipment supply. Foreign media have pointed out that advanced EUV lithography systems depend on a few suppliers, with long lead times and high procurement costs. Talent is another bottleneck; the US still lags behind Asia in semiconductor engineering talent, wafer fab experience, and supply chain maturity.

Nevertheless, some analysts believe that if Musk approaches from packaging and supply chain integration, and collaborates with Samsung, Intel, and others, there is still a long-term possibility of reshaping the global chip industry landscape.

— Wu Yuli, The Paper News