A $20 billion opening bill sounds enormous, yet it would barely fund one advanced logic fab. Bernstein says Elon Musk’s Terafab ambition belongs on a far larger ledger entirely.
Its math starts with wafer demand, power needs, and fab throughput rather than showy bravado alone. From that base, a semiconductor capacity model links future AI hardware demand to wafer volumes and fab counts, pushing projected chip manufacturing costs beyond $5 trillion, while Tesla and SpaceX hiring posts hint that the idea has moved past chatter.
Why Bernstein’s estimate reaches the $5 trillion mark
Bernstein frames TeraFab as a funding mismatch on a spectacular scale. Its analysts argue that the roughly $20 billion discussed by Elon Musk would scarcely pay for one 7nm-class logic plant, whereas a 1 TW annual AI target implies far larger chip output once the power-to-silicon math is applied.
The note builds from rack power, die size, yields, and system mix rather than a single headline figure. Using wafer capacity assumptions based on 120 kW Rubin and 600 kW Rubin Ultra racks, Bernstein says the capital expenditure scale climbs toward $5 trillion, with fab construction costs alone rising past 70% of the US federal budget.
From 22 million GPU wafers to hundreds of fabs
Bernstein’s model turns Musk’s chip ambition into a volume problem with little room for poetry. To deliver 1 TW of accelerator output each year, the analysts estimate about 22.4 million Rubin Ultra wafers, a number that dwarfs anything now seen in AI infrastructure.
The rest of the stack is no footnote. Bernstein pegs Vera CPU volumes at roughly 2.716 million wafers and HBM4E production needs at 15.824 million wafers, then translates the load into an annual fab count of about 142 to 358 sites.
Tesla and SpaceX begin hiring for a chipmaking push
Recruiting gives the project a more concrete outline. Tesla has posted jobs in Palo Alto and Austin, including a California opening for a lithography process engineer at $88,000 to $240,000 and process integration roles tied to advanced logic, with base pay up to $338,280.
The descriptions point to factory work, not a lab exercise. Applicants are asked for at least 10 years in semiconductors, plus on-call manufacturing support for 24/7 operations, while SpaceX expands silicon division hiring through roughly 60 openings across Texas, Washington, and California.
One roof for logic, memory, and packaging is a rare bet
Musk’s pitch departs from the way chip supply chains are usually split across multiple specialists. Under his concept, logic memory integration would sit beside an advanced packaging line at one site, reducing the back-and-forth that comes with moving wafers and stacked memory between companies.
During Saturday’s announcement, Musk said faster iteration was part of the appeal. He linked in-house mask making to shorter design loops and cast TeraFab as a vertical manufacturing model, a structure few semiconductor groups attempt under one roof.
