GPU compute is getting pricier
Hourly rental prices for the latest Nvidia Blackwell GPUs have spiked sharply—reported at $4.08 per hour, up about 48% in two months—while the industry is seeing outages and rationing for high‑end compute. Those price signals reflect surging demand for agentic workloads and tighter capacity across cloud and rental markets. (the-decoder.com) (intellectia.ai)
Renting Nvidia’s newest Blackwell graphics processors is getting more expensive fast, with hourly prices reported at $4.08 after a two-month jump of 48%. (techmeme.com) Those chips are usually rented by the hour through cloud providers, the way companies rent server time instead of buying hardware outright. A pricing tracker cited by The Wall Street Journal put Blackwell rentals at $2.75 two months ago and $4.08 this week. (the-decoder.com) Blackwell is Nvidia’s latest family of data-center graphics processors, built for training and running artificial intelligence models. Listings tracked across 23 providers show wide spreads for Nvidia B200 rentals, from $2.25 an hour on some reserved offers to $8.60 per graphics processor hour on CoreWeave’s on-demand eight-chip cluster. (getdeploying.com) The price spike is showing up alongside service strain at the companies buying the most compute. Anthropic’s Claude status page logged elevated errors on April 10, and The Decoder, citing The Wall Street Journal, said Claude application programming interface uptime was 98.95% over the 90 days ending April 8. (status.claude.com) (the-decoder.com) OpenAI is also reallocating capacity. Its help center says the Sora web and app products will be discontinued on April 26, 2026, and the Sora application programming interface on September 24, 2026. (help.openai.com) The demand surge is tied to “agentic” tools, which means software that does multi-step work on a user’s behalf instead of answering one prompt at a time. The Decoder reported token usage rising from 6 billion to 15 billion per minute between October and March as those workloads spread. (the-decoder.com) Cloud pricing pages show why buyers are chasing the newest chips anyway. Crusoe lists Nvidia H100 at $3.90 per graphics processor hour and Nvidia H200 at $4.29, while its Blackwell systems are not posted with self-serve prices and require sales contact. (crusoe.ai) The industry response has been to build more capacity, but that takes time and power. OpenAI said in January 2025 that its Stargate project planned up to $500 billion of artificial intelligence infrastructure investment over four years, starting with $100 billion. (openai.com) For companies training models or running high-volume artificial intelligence products, the signal is simple: the most sought-after compute is still scarce, and the meter is running faster. (the-decoder.com)
