SpaceX Switches Large‑Model Training Stack from JAX to C, Claiming Ten‑fold Speedup

New C‑language training stack

SpaceX has replaced JAX with a training stack written entirely in C. Elon Musk stated that the stack can accelerate large‑model training by an order of magnitude. The Grok‑5 model is already being trained on this stack, and the inference portion will be rewritten in C as well.

Colossus II supercomputer

SpaceX is building a next‑generation supercomputer codenamed Colossus II with a total power budget of 1 GW and an estimated capital cost of about $170 billion. By contrast, the earlier Colossus I built for xAI cost $30‑40 billion, so Colossus II scales the hardware capacity by roughly one order of magnitude.

AI infrastructure is listed in SpaceX’s prospectus as a core business line alongside rocket launches, and the company is actively negotiating compute‑rental deals and planning an orbital data centre.

Compute‑rental agreement with Anthropic

SpaceX disclosed a short‑term compute‑rental contract with Anthropic. The agreement is a 180‑day lease that charges $12.5 billion per month, runs until May 2029, and therefore has a potential total value exceeding $400 billion.

“Any party may terminate the agreement with 90 days’ notice.” – SpaceX prospectus

Musk emphasized that the short‑term term was chosen by SpaceX, not by Anthropic, and that external customers are secondary to SpaceX’s own compute needs. If internal demand rises, external rentals can be reclaimed with the 90‑day notice.

Because Tesla’s autonomous‑driving workloads also compete for the same compute resources, internal demand can displace external tenants at any time.

References: Business Insider article, Elon Musk X post, John Tinsman X post.

Code example

[1]https://www.businessinsider.com/elon-musk-spacex-anthropic-escape-hatch-2026-5
[2]https://x.com/JohnTinsman/status/2059654405964657093
[3]https://x.com/elonmusk/status/2059884150187053488