SpaceX Just Filed for a $55 Billion Chip Plant. Your Inference Bill Doesn’t Care

Filed May 7, 2026. SpaceX putting $55 billion into a Texas fab called Terafab. Grimes County. Could hit $119 billion across phases. That’s not a typo. Intel manufacturing partner, 14A process node, custom GPUs for Grok and Tesla and SpaceX satellites, 1 terawatt per year output target. Weeks before the xAI-SpaceX IPO at $1.25 trillion.

Days after renting 220,000 Nvidia GPUs to Anthropic.

One of the biggest private capital commitments in history.

And it landed while your Copilot subscription just got metered.

Here’s the thing nobody’s leading with: if compute is about to get cheap at the source, why are you paying more per token than you were six months ago?

The IPO Is the Story Right Now

June. Combined xAI-SpaceX entity hits public markets at $1.25 trillion. The Terafab filing is May. Not coincidence.

A $119 billion capex line in a prospectus does something specific: it tells institutional investors you’re building infrastructure that outlasts any single model generation. Compute company. Not a software startup with an API. That narrative supports a premium valuation on the roadshow.

It also means public shareholders are funding the fab build.

If you buy xAI-SpaceX equity in June, you’re partly buying a semiconductor thesis. The subsidized GPU rental to Anthropic. Which looked like Musk playing nice with AI rivals. Starts looking like a loss leader designed to show utilization rates before the IPO.

Once you’re public, you show margins. Loss leaders become priced. The 220,000 GPUs rented to Anthropic at presumably favorable rates?

Either that’s a long-term contract locked in before today, or it ends when shareholder pressure shows up.

Intel Got a Lifeline

Intel’s foundry business has been a mess.

No big-name customers, execution problems, TSMC eating their lunch on advanced nodes. Then SpaceX walks in with $55 billion committed to 14A process manufacturing.

That’s the headline for Intel. A verified customer with real money, betting on their process node for custom AI silicon. Changes the narrative.

Foundry business has a marquee name.

For Nvidia, it’s a warning.

Terafab is purpose-built to get SpaceX off Nvidia GPUs long-term. Custom silicon, purpose-designed for Grok and Tesla workloads, manufactured on a node Intel is betting will be competitive. If it works. And Intel 14A genuinely closes the gap with TSMC’s best processes. Nvidia loses a volume customer who’s currently spending real money on H100s and H200s.

The Ohio fab comparison is the thing SpaceX doesn’t want mentioned. Intel’s Ohio project was announced big, promised big, and has been delayed and over budget repeatedly. Greenfield advanced fabs fail on schedule. Always. SpaceX’s own filing admits it: still dependent on TSMC and Samsung, timeline targets may not be achievable. They published risk factors. SpaceX doesn’t publish risk factors.

What You Do With This

Three-to-five year horizon before Terafab affects anything.

The fab hasn’t broken ground. IPO: this year.

Your inference bill is happening now. US vendors are moving to meters. Copilot’s June 1 change. Claude’s rate limits. The direction is obvious: pay-per-use, not flat subscription. Because they can.

Open-weight models have closed the quality gap for most non-frontier tasks.

DeepSeek V4, Llama variants, Mistral. Running a 7B model on a 4090 is fine for a lot of production work. A 70B model on a cloud GPU instance handles more. The cost difference against OpenAI or Anthropic is 5-10x. The quality gap on routine tasks is negligible.

I’m not saying replace everything. I’m saying test it. Pick one task you’re paying $15 a month for through a closed API. Run it through a self-hosted model. Compare outputs. If the quality is close enough, you’re paying a premium for convenience you might not need.

The compute layer is getting built by people spending more money than most countries. Your inference bill is being set by companies racing to meter usage before that compute layer becomes commoditized. The window where flat-rate subscriptions made sense is closing.

Pay attention to which vendors are still offering them and why.