GPU spot prices up 68% since January

Social posts on May 17 said spot prices for rented AI graphics processors had risen 68% since January across cloud providers, citing marketplace and trader data. The posts tied the move to a tightening market for short-term compute, where companies rent Nvidia-heavy capacity by the hour or month rather than signing long contracts. Publicly available industry tracking and recent Bloomberg reporting support the broader picture of rising AI infrastructure demand, though the exact 68% figure appears to come from proprietary market data rather than a company filing or exchange disclosure. The result is a market where GPU rental prices, data-center construction and hyperscaler capital spending are being read together as signals of near-term supply strain. ### Where did the 68% number come from? SemiAnalysis publishes an H100 daily and contract pricing index that tracks on-demand and term rental rates for Nvidia H100 capacity across the market. Bloomberg reported last year that Silicon Data had launched what it described as a daily index for GPU rental costs, tracking the hourly price of renting AI chips used in cloud computing. Those services have become reference points for traders, brokers and startups trying to price short-duration access to scarce AI hardware. (api.semianalysis.com) The May 17 posts did not point to a public exchange settlement or regulator filing. Instead, they cited marketplace and trader data, which means the 68% jump should be understood as a spot-market reading from private data providers rather than an audited industry benchmark. SemiAnalysis and Bloomberg both describe GPU rental pricing as an emerging market signal, not a standardized commodity index. (api.semianalysis.com) ### Which companies are driving demand for rented GPU capacity? Bloomberg reported on February 6 that Amazon, Alphabet, Meta and Microsoft together had forecast capital expenditures of about $650 billion in 2026 for new data centers and related equipment. A later Bloomberg report on April 30 said those plans had risen to as much as $725 billion this year after Alphabet and Meta increased guidance and Microsoft gave its first estimate through December. (api.semianalysis.com) Those companies are the largest buyers of AI infrastructure and are widely treated by investors and suppliers as the core source of demand for advanced accelerators. Bloomberg Intelligence said in a separate market note that rising hyperscaler AI spending was supporting demand for GPUs and application-specific chips. That does not prove a one-to-one link between any single capex forecast and spot rental prices, but it does place the price move inside a market already absorbing very large infrastructure orders. (bloomberg.com) ### Why would spot prices rise if more data centers are being built? CBRE said, in a Bloomberg report published February 25, that U.S. data-center capacity under construction fell to 5.99 gigawatts at the end of 2025 from 6.35 gigawatts a year earlier. The report said permitting, zoning and power procurement delays were slowing new projects even as demand for AI computing kept rising. That matters because spot GPU markets react to usable capacity now, not just to facilities announced for future delivery. (bloomberg.com) Bloomberg also reported on April 1 and March 16 that developers were struggling to source electrical equipment and secure enough grid capacity for large new facilities. Wood Mackenzie, cited by Bloomberg, said U.S. data-center development had slowed because the power grid was reaching its limit in some areas. Those constraints help explain why short-term compute can stay expensive even when spending plans are accelerating. (bloomberg.com) ### What does the GPU rental market measure that chip sales do not? Bloomberg said in its 2025 report on Silicon Data that the company’s index tracked the hourly cost of renting a GPU. That makes the rental market different from semiconductor sales data, which captures chip shipments but not the price users pay to access already-installed hardware in the cloud. Rental prices can move faster because they reflect immediate availability, utilization and contract length. (bloomberg.com) SemiAnalysis says its pricing dashboard follows daily, on-demand and contract rates over time. In practice, that means a rise in spot pricing can show that uncommitted capacity is getting harder to find, even before broader supply statistics catch up. That is an inference from how spot markets function, based on the structure of the pricing services described by SemiAnalysis and Bloomberg. (bloomberg.com) ### What should readers watch next? Amazon, Alphabet, Meta and Microsoft will provide the next major public readout on AI infrastructure demand in their quarterly earnings reports and capital-spending updates. GPU pricing services such as SemiAnalysis’s H100 index and Silicon Data’s rental benchmarks are likely to remain the clearest public market gauges of whether short-term compute is easing or tightening. CBRE’s next updates on data-center construction and power availability will also matter because they show how much physical capacity is moving from plan to build. (api.semianalysis.com)