Google will pay SpaceX $920M a month for compute — renting GPUs from its own AI rival's landlord
From October 2026 through June 2029, Google rents roughly 110,000 NVIDIA GPUs from SpaceX for $920M monthly. Anthropic already signed a similar deal at $1.25B a month. SpaceX is quietly becoming the compute landlord for the companies it competes with.
The deal
Per TechCrunch, Google will pay SpaceX about $920 million per month from October 2026 through June 2029 for compute — roughly 110,000 NVIDIA GPUs plus CPUs and memory, at a terrestrial data center (not, despite SpaceX's orbital ambitions, in space). Google called it "a short-term, timely agreement to ensure bridge capacity" for "unexpected demand" on its Gemini Enterprise platform. There's a ramp-up at reduced rates through September, and a mutual 90-day exit option after Dec 31, 2026, plus an out if SpaceX misses GPU delivery by Sept 30, 2026.
The context the headline number hides
The HN thread is thin — it was largely flagged as a duplicate of an earlier discussion — so the real story is in the comparison the source draws. Google's $920M/month buys roughly half what Anthropic secured from SpaceX in a separate May 2026 arrangement reported at $1.25 billion per month through 2029.
Read those two deals together and the picture flips. The headline is "Google spends big on compute." The actual development is that SpaceX has become the compute landlord to the AI industry — leasing GPU capacity to Gemini's Google and to Claude's Anthropic at the same time — even as it reportedly reserved its Colossus 2 facility for xAI.
Why now
The timing isn't incidental: the deal lands as SpaceX courts an anticipated IPO targeting a $75B raise at a ~$1.75 trillion valuation. Recurring, multi-year, billion-dollar compute contracts from the most credit-worthy names in tech are exactly the kind of revenue story that underwrites a number like that. The frontier-AI arms race needs GPUs faster than anyone can build data centers — and SpaceX just turned that shortage into a recurring rent check signed by its own rivals.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
fork() was a 1970s trick to swap a program to disk — and Linux is finally trying to move past it
Everyone treats fork()+exec() as timeless Unix elegance. John Nagle showed up to explain it was a hack so a forking program could be swapped to disk — and a Redis benchmark showed fork() taking 5.67 seconds on a 25GB process.