On March 21, 2026, Elon Musk walked onto a stage inside Austin's defunct Seaholm Power Plant and announced something the semiconductor industry is still trying to process: a $25 billion chip fabrication venture called Terafab.
The name comes from its target: one terawatt of computing power per year. That's one trillion watts. For context, the entire global semiconductor industry currently produces roughly 20 gigawatts annually. Musk is proposing to build fifty times that — in a single facility.
"The most epic chip building exercise in history by far," Musk called it. He's probably right, regardless of whether it succeeds or fails.
The Numbers
Why Build It?
The simple answer: Musk's companies are going to need a lot of chips. An almost incomprehensible amount.
- Tesla needs AI chips for Full Self-Driving, for the Dojo supercomputer training its neural networks, and for its planned fleet of Optimus humanoid robots
- SpaceX needs computing power for Starlink satellites, Starship guidance systems, and — according to Musk — orbiting data centers that will process AI workloads in space
- xAI (which SpaceX acquired in February 2026) needs massive compute to train and run Grok and future AI models
"That rate is much less than we'd like," Musk said of global chip production. "We either build the Terafab, or we don't have the chips."
The math is stark: Musk claims to need 100–200 gigawatts of terrestrial chips annually, plus potentially a full terawatt of space-based compute. Current suppliers — TSMC, Samsung, Micron — simply can't scale fast enough.
What Will It Make?
Unlike traditional fabs that specialize in logic chips, memory, or packaging, Terafab aims to be vertically integrated:
AI Chips
Custom silicon for AI training and inference, including Tesla's fifth-generation AI5 chip
Memory
High-bandwidth memory integrated alongside processors for AI workloads
Advanced Packaging
Chiplets and 3D stacking to combine multiple dies into single packages
Target Node
Musk has previously referenced 2nm — cutting edge, matching TSMC's most advanced processes
The Orbital Vision
Perhaps the most audacious part: Musk wants to put most of this computing power in space.
Why? The math actually makes some sense:
- 5x more solar power — no atmosphere to diffuse sunlight
- Free cooling — vacuum dissipates heat efficiently via radiation
- No grid constraints — U.S. electricity generation is only 0.5 terawatts total
SpaceX has already filed with the FCC for plans involving up to one million orbital data center units. Initial "Mini" satellites would provide 100 kilowatts each, scaling to megawatt-class systems.
"Quantity has a quality all its own." — Elon Musk, quoting Stalin
Why This Is Nearly Impossible
Here's the thing about building a semiconductor fab: it's the most complex manufacturing endeavor humans have ever attempted. And Musk is proposing to do it at a scale nobody has ever tried.
The EUV Problem
Modern chips require Extreme Ultraviolet lithography machines — $200 million devices made by exactly one company on Earth: ASML, in the Netherlands. Each machine:
- Fires a laser at tin droplets 50,000 times per second
- Creates plasma that emits 13.5nm wavelength light
- Bounces that light off mirrors with atomic-level precision
- Takes 12+ months to build
- Weighs 180 tons and ships in 40 freight containers
ASML produces maybe 50 EUV machines per year. A cutting-edge fab needs dozens. There's a global queue.
The Talent Problem
Running a fab requires thousands of specialized engineers — process engineers, equipment engineers, yield engineers, integration experts. This talent pool is finite and concentrated in Taiwan, South Korea, Japan, and specific regions of the U.S. and Europe.
Intel spent years and billions trying to catch up to TSMC. Samsung has world-class resources and still trails. Talent isn't something you can simply buy.
The Yield Problem
A chip fab doesn't just make chips — it makes enough good chips to be economical. At cutting-edge nodes, defect density determines whether you're printing money or burning it.
TSMC has spent decades perfecting yield. Their institutional knowledge is arguably their most valuable asset. It can't be copied or acquired.
The Time Problem
Fabs take 3–4 years to build and another 1–2 years to ramp to high-volume production. The technology moves fast. A process that's cutting-edge when you break ground may be outdated by the time you're shipping.
The Jensen Warning
NVIDIA CEO Jensen Huang, when asked about Musk's plans, was diplomatic but clear: building a fab is "extremely hard." He noted that NVIDIA itself uses TSMC rather than building its own fabs for good reason — the complexity is almost unimaginable.
So Can He Do It?
Skepticism is warranted. But dismissing Musk's ambitions has historically been a losing bet:
- SpaceX was told reusable rockets were impossible. They now land them routinely.
- Tesla was told electric cars couldn't compete. They're now the most valuable automaker.
- Starlink was dismissed as impractical. It now has millions of subscribers.
The difference: those industries didn't have TSMC. Semiconductors may be the one domain where catching up to the leaders requires a generation of accumulated expertise that can't be replicated with sheer capital and willpower.
Then again, that's exactly what people said about rockets and cars.
Watch the Announcement
What Happens Next
Musk gave no timeline for when Terafab will produce its first chips. Construction details remain sparse. The most concrete statement: production ramp targeting 2027.
We're tracking every development. Follow the news or check the timeline for updates.