SpaceX Colossus lands $26B in multi-year GPU contracts from Google, Anthropic; TSMC ships first US-made Blackwell
SpaceX has monetized its Colossus data center infrastructure into a commercial GPU-rental business, landing two massive capacity deals: Google will pay $920 million per month ($33 billion total) from October 2026 through June 2029 for access to approximately 110,000 NVIDIA GPUs; Anthropic locked $1.25 billion per month through 2029 for full access to Colossus 1's 220,000+ NVIDIA processors (near Memphis). A third deal with Reflection AI adds $150 million per month (~$6.3 billion through 2029). Combined, SpaceX's compute contracts total over $39 billion in committed revenue, cementing the company's transition from aerospace/launch services into a hyperscale AI infrastructure provider.
Concurrently, NVIDIA and TSMC celebrated the first production Blackwell wafer manufactured at TSMC's Phoenix facility (Fab 21), reaching volume production of what Jensen Huang called 'the single most important chip' now being built on US soil. The wafers still require transport to Taiwan for advanced CoWoS-L packaging and HBM3E integration, but the onshoring milestone signals TSMC's successful ramp of 4NP (custom 4nm) for Blackwell and reduces long-term dependence on Taiwan-based packaging (Amkor and TSMC are building U.S. packaging capacity for 2028+).
For architects and infrastructure teams: SpaceX's $39B compute commitment backlog telegraphs continued hyperscaler confidence in sustained AI capex, despite Samsung's recent earnings miss and market jitters. The requirement that Google receive 110,000 GPUs by September 30 or face contract termination is also a hard constraint on SpaceX's execution—miss it and capacity gets reassigned, squeezing Colossus economics. On the supply side, TSMC's Arizona milestone de-risks geopolitical concentration, though full supply-chain sovereignty (packaging + HBM) remains 18–24 months away. Monitor SpaceX's Nasdaq 100 inclusion (7-Jul) for passive index flow impact.
Sources
- Primary source
- cnbc.com
“Reflection will get immediate access to Nvidia GB300s, top-of-the-line AI chips used to train and run advanced models, and has agreed to pay SpaceX $150 million per month”
- blogs.nvidia.com
“It's the very first time in recent American history that the single most important chip is being manufactured here in the United States by the most advanced fab, by TSMC”
- techtimes.com
“The arrangement, which nears $1 billion a month, is one of the largest cloud-compute commitments ever disclosed”