Demand for NVIDIA Corp. (NASDAQ:NVDA) GPUs has become so intense that even cloud providers with badly functioning infrastructure can still find customers willing to pay healthy margins, according to SemiAnalysis.
“If you’ve got GPUs,” buyers will pay, researcher Sam Harshe said, even if a provider’s Kubernetes, the software used to manage workloads across a cluster, is “completely broken” and its storage “barely works.”
Harshe described “ridiculous backwardation” in the market, with customers paying steep premiums for compute available within weeks rather than waiting several months. The findings come from months of testing 77 GPU-cloud providers and interviews with more than 200 customers for SemiAnalysis’ ClusterMAX 3.0 report.
• NVDA shares are showing limited movement. What’s the outlook for NVDA shares?
Bad Infrastructure Isn’t Stopping Sales
GPU-cloud companies rent GPU clusters to AI developers, handling the networking, storage and software around the chips.
SemiAnalysis ranks them from Platinum to Underperforming based on factors including reliability, performance and support. Just 19 of the 77 providers tested earned a Platinum-through-Bronze “Medallion” rating.
CoreWeave Inc. (NASDAQ:CRWV) and Nebius Group N.V. (NASDAQ:NBIS) ranked Platinum, while Alphabet Inc.’s (NASDAQ:GOOGL) Google Cloud and Oracle Corp. (NYSE:ORCL) earned Gold. Microsoft Corp.’s (NASDAQ:MSFT) Azure ranked Silver, while Amazon.com Inc.’s (NASDAQ:AMZN) AWS and Crusoe were Bronze.
Crusoe, a privately held Nvidia-backed AI infrastructure company valued at $30.9 billion last week, received one of the harsher reviews. SemiAnalysis said one Crusoe H100 cluster produced more genuine GPU errors in five days than it saw across the rest of its testing combined, and called Crusoe’s managed clusters "exceptionally unreliable."
SemiAnalysis said Crusoe subsequently fixed several of the issues.
Paying Up for Compute
The wider market shows why customers may tolerate those shortcomings.
Microsoft CFO Amy Hood said in July that demand continues to exceed available supply, describing the imbalance as a “relatively extreme moment.”
Oracle said this month that demand for AI cloud training and inference continues to grow faster than supply, even after delivering more than 300,000 GPUs since the end of its previous quarter, nearly triple the capacity delivered in the fourth quarter.
Nvidia, meanwhile, reported Data Center revenue of $89 billion last quarter, up 117% year over year. CEO Jensen Huang said AI demand was “accelerating.”
What About GPU Depreciation?
SemiAnalysis said companies are signing four-year contracts for H100 capacity, evidence that older Nvidia GPUs can retain significant earning power while compute remains scarce.
That appears to push against investor Michael Burry’s argument that AI companies are depreciating GPUs too slowly.
Prediction traders see limited near-term bust risk. Polymarket puts the chance of an AI-industry downturn by Dec. 31 at about 10%, with roughly $2.4 million traded.
The market resolves “Yes” if at least three of six stress events occur within 90 days, including Nvidia falling 50% from its high or H100 rental prices staying below $1 an hour for five days.
The four-year H100 contracts SemiAnalysis describes point toward continued demand, rather than the collapse implied by Polymarket’s sub-$1 trigger.
Image: Shutterstock
Login to comment