Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Tesla’s $1.3tn market value already assumes a future driven more by artificial intelligence than by electric cars. Delivering on that expectation would require far greater control over the hardware that powers autonomy, robotics and AI training.
Elon Musk has said the company therefore needs to build and operate a so-called TeraFab to manufacture its own chips. Can Tesla actually do this?
That idea should be considered separately from Musk’s more extravagant ambitions. Compared with projects such as Mars colonisation or neural implants, building a chip fabrication plant is grounded in existing industrial practice.
Much of the scepticism around a new entrant to the chip industry rests on a misconception: when many think of chipmaking, they tend to picture TSMC’s most advanced output of 3- and 5-nanometre chips that are used for Nvidia’s flagship AI chips today. Those high-end chips require cutting-edge machines and decades of accumulated trial and error. Against that standard, a new entrant looks hopeless.
Yet that is also the wrong comparison. Chipmaking does not have to match TSMC at the most advanced level to be viable. There exists a narrow middle ground where manufacturing remains technically achievable for a new entrant and still powerful enough to support AI workloads. Any Tesla fab would have to target older technology, around 7 nanometres, which is widely regarded as the last advanced generation of chips before complexity and capital requirements rise sharply.
TSMC first started volume production of these chips in 2018 and the technology remains widely used in AI and data centres today. Even if it is several years behind, Tesla could, in principle, produce usable AI chips for running models for vehicle autonomy and humanoid robots.
That makes 7nm the correct benchmark, but not an easy one. It still requires extreme ultraviolet lithography machines from ASML. It also needs ultra-clean facilities with large-scale power, cooling and water infrastructure, hundreds of precision tools and advanced chip packaging capabilities.
Critically, it would need hundreds of engineers with extensive experience in bringing chip defect rates down, talent that is scarce and already concentrated at TSMC. Initial production would typically take three or more years, with high material losses and prolonged trial and error before reaching usable output.
Even if Tesla clears the technical bar, feasibility alone does not guarantee value. When making chips, the benchmark today is TSMC, which spent more than $40bn in capital expenditure last year. That spending is justified because its risk is spread across a long customer list and thousands of designs, helping to absorb production losses and accelerate learning.
Tesla cannot replicate this model. Musk has said Tesla does not intend to sell chips outside the company. Without global orders, a chip business becomes a fixed-cost operation that would struggle to avoid being structurally lossmaking. Previous investments in US fabs by TSMC suggest per-fab costs of at least $20bn, even accounting for subsidies. Unlike most industrial assets, chip fabs require sustained reinvestment just to remain competitive. Even under generous assumptions, recovering investment would take many decades.
Recommended
Economics aside, execution risk remains the greater challenge. Starting in the mid-2010s, Intel’s troubled transition to its 10-nanometre chips, despite decades of industry experience, was driven by aggressive timelines and internal pressure, leading to years of delays and a permanent loss of leadership.
Tesla’s history of reported inconsistencies in EV build quality, including panel gaps and post-delivery rework, does not automatically imply technical incompetence. But it does reflect a corporate tolerance for shipping products before manufacturing processes are fully stabilised. In cars, such defects are largely cosmetic and can be corrected. In chipmaking, that is not an option.
US-based GlobalFoundries is another instructive case. After acquiring IBM’s lossmaking chip business in 2015, it concluded just three years later that advanced chip manufacturing was not economically viable.
Tesla would face both of these risks at the same time: the cultural pressures Intel faced and the unfavourable economics that forced GlobalFoundries to retreat. History suggests this combination is particularly prone to value destruction, apparent only after the capital has already been spent.
