Google has signed a deal deal with SpaceX for access to graphics processing unit compute amounting to about $31.5 billion for three years, suggesting just how massive Google believes its own artificial intelligence computing needs may be, especially to support enterprise customers.
The deal gives Google access to some 110,000 Nvidia GPUs, beginning in October 2026.
Some will see this as a bridge to the time when Google can bring its owned compute facilities online at the scale Google believes is necessary.
Gigawatt-scale data centers take years to build, so the deal is a hedge that allows Google to supply
capacity other “compute as a service” suppliers might otherwise take.
Enterprise customers prefer NVIDIA’s CUDA ecosystem. So leasing Nvidia capacity from SpaceX allows Google Cloud to sell what enterprise clients want without diverting its custom accelerator (tensor processing units) supply away from its own model development and inferencing.
By some estimates, Nvidia GPUs, such as the H100 SXM5 nodes, will remain in short supply until 2027.
Google is paying about $8,363 per GPU per month, or roughly $11.45 per hour assuming the machines run around the clock.
Top-tier NVIDIA chips on major cloud platforms currently go for perhaps $5 to $15 per hour, depending upon instance types, with newer architectures like the Blackwell or Reubin LPX for Agentic AI applications perhaps costing from $15 to $50 per hour.
By some estimates estimates, Google will layer its own software, support, and service guarantees on top of the SpaceX infrastructure, then charge enterprise customers somewhere between $18 and $70 per GPU per hour.
For SpaceX, set to go public on June 12, 2026, the additional revenue will undoubtedly provide an underpinning for the new public firm’s valuation. The deal adds about $950 million per month of committed revenue.
Many expect the SpaceX IPO will be the largest in history.
No comments:
Post a Comment