$55 billion — Friday, May 8, 2026

When The Verge reported SpaceX's Terafab filing last week, the number that stood out was not $55 billion. It was the qualifier next to it: minimum. SpaceX disclosed $55 billion as the floor, expandable to $119 billion, to build a chip fabrication plant in Texas targeting 200 gigawatts per year of computing power on Earth and one terawatt in space. That range — from floor to ceiling — tells you more about the moment than the headline number does.

For context: $55 billion is roughly what the US government committed to domestic semiconductor manufacturing under the CHIPS Act, spread across the entire country over five years. SpaceX is disclosing that as a starting point for one facility. The comparison is not to embarrass the CHIPS Act. It is to show how fast the private capital flowing into compute infrastructure has lapped the public sector's response to the same problem.

The more interesting thing about Terafab is what it represents structurally. SpaceX already operates Colossus 1 in Memphis — the cluster Anthropic just signed for all 300 megawatts of. That makes SpaceX Anthropic's landlord. Terafab, if it gets built, makes SpaceX the supplier to the landlord. The company would own a meaningful slice of the supply chain from silicon fabrication through to the data centre floor, and it would rent access to labs and hyperscalers who cannot or will not build their own. That is a different business than launching rockets.

The labs have noticed. Anthropic signed four separate compute deals in roughly six months — Amazon, Google, Microsoft, and now SpaceX — because the megawatts are wherever the megawatts are, not wherever the preferred strategic partner is. When capacity is scarce enough that you rent from a rival, the entity controlling the capacity has leverage that compounds over time. Terafab is SpaceX's bet that this leverage is worth $55 to $119 billion to establish permanently.

The 1 TW in space figure gets less coverage and deserves more scrutiny. One terawatt is approximately the entire installed electricity generation capacity of the United States. Whether that number is aspirational marketing or an engineering target with a timeline behind it changes the interpretation of everything else in the filing. For now it sits in the document without a date. Watch whether it acquires one.

The same instinct is playing out one layer up. On the same week SpaceX filed for Terafab, every major lab shipped its own agent runtime. Anthropic put finance agents inside Microsoft Office. Mistral added remote agents to its Vibe coding tool. Perplexity opened its Personal Computer agent to all Mac users. xAI launched Grok Connectors. Each runs inside the vendor's own runtime — different connectors, different memory, different deployment surface. A buyer evaluating an AI agent for their team is choosing a vendor's runtime, not a capability layer. The infrastructure fragmentation and the software fragmentation are the same story at different altitudes.