The shortage of computing power is escalating! The price of the one-year contract for H100 has skyrocketed by 40%. The wave of demand for AI is once again sparking a "GPU battle".

With the explosive growth of the global artificial intelligence (AI) industry demand, the GPU computing power shortage situation has further intensified, leading to a significant increase in rental prices for core computing power chips. According to the latest data released by industry research firm SemiAnalysis, the one-year rental price of NVIDIA Corporation (NVDA.US) H100 GPU has experienced a substantial surge. As of March 2026, the one-year contract rent has increased from $1.70 per GPU per hour in October 2025 to $2.35, a nearly 40% increase. The entire industry's GPU computing resources have been almost completely sold out. The H100 rental price has increased by nearly 40% in half a year, and the spot market is in a state of emergency. SemiAnalysis's index data shows that the increase in H100 rental prices began in the fourth quarter of 2025 and accelerated significantly in 2026. By the end of January 2026, the one-year contract price of H100 exceeded $2 per hour per card for the first time; the month-on-month increase in February reached 15%-20%; by the end of March, the price had climbed to $2.35 per hour per card, nearly 40% higher than the low point six months ago, with a monthly increase expected to remain at a high level of 15%-20%. The tightness of supply and demand in the spot market is even more extreme. The report points out that the on-demand rental capacity of all types of GPUs is completely sold out, and even though prices have continued to rise recently, users who have locked in on-demand instances are unwilling to release capacity back to the market. SemiAnalysis commented that at the beginning of 2026, the difficulty of purchasing GPU computing power on the market is comparable to the rush to buy last-minute airline tickets during peak season. Not only are prices high, but the availability of spot resources is almost completely depleted. SemiAnalysis also mentioned that in the extreme imbalance of supply and demand, many unconventional phenomena have appeared in the market. Some users are willing to pay as much as $14 per hour per card to get AWS's p6-b200 spot instances; leading new cloud companies have stopped selling single-node computing power; a large number of H100 rental contracts signed 2-3 years ago are being renewed at the original price, with some contracts even being directly renewed for 4 years until 2028, locking in long-term capacity. Some computing power tenants are splitting and subletting their leased clusters, similar to the premium subletting of apartments during major events. The shortage of the entire chip supply has exacerbated the tightness of computing power supply. The report points out that January 2026 became an important turning point for storage prices, with the prices of DRAM and NAND flash memory, which had been rising sharply for several consecutive quarters, experiencing parabolic jumps in the first quarter of 2026. According to SemiAnalysis's storage model calculations, in the first quarter of 2026, LPDDR5 and DDR5 contract prices are expected to increase by about 4 times and 5 times, respectively, compared to the same period last year. The skyrocketing prices of core components such as storage have further pushed up the overall cost of AI servers. To hedge against the risk of gross profit margin, server OEMs have significantly increased the prices of AI servers, far exceeding the increase in component costs. This directly compresses the expected returns of computing cluster projects, forcing a large number of operators to slow down or even abandon their plans to deploy new clusters. The additional supply that was supposed to enter the market is now on hold, further tightening the supply situation in the rental market. The full-scale outbreak of demand is the core driver of the current computing power shortage. On one hand, the rapid popularization of media-generated AI tools by companies such as ByteDance and Alphabet Inc. Class C (GOOGL.US), as well as the surging demand for models such as Anthropic's Claude 4.6 Opus and Claude Code, with annualized revenue ARR soaring from $9 billion to over $25 billion in just one quarter. Coupled with the explosive growth in the use of open-source models such as GLM and Kimi K2.5, and the massive purchase of computing power after large financings by companies such as OpenAI and Anthropic, the demand for GPUs continues to rise. On the other hand, the scale application of multi-agent workflows has led to an exponential increase in token consumption, and the extremely high return on investment for AI tools makes computing power demand highly rigid, further exacerbating the supply-demand gap. The market landscape is undergoing a deep restructuring, and there is still room for short-term price increases. The report points out that there is a clear deviation between the secondary market and the industrial fundamentals. Although the GPU market supply continues to tighten and prices have risen significantly, directly benefiting new cloud companies by expanding gross profit margins and extending the lifespan of assets, the market sentiment towards leading new cloud companies such as CoreWeave (CRWV.US), Nebius (NBIS.US), and IREN (IREN.US) remains pessimistic, with their stock prices already at the lower end of the 6-12 month trading range. The market is still anchored in the narrative of "GPU will eventually oversupply and become commoditized," which sharply contrasts with the reality of ongoing shortages and continued bargaining power for manufacturers. SemiAnalysis points out that the future trend of GPU rental prices will depend on three key factors: the progress of the large-scale deployment of the GB300 cluster in 2026, to see if the increase in computing power supply can alleviate the current shortage or if the growth in token demand will continue to outpace the new supply; the possible further deterioration of semiconductor supply constraints in the industry chain, with a focus on the capacity limitations of key components such as Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR (TSM.US) N3 advanced processes, HBM, and DRAM/NAND storage; and the pace of ARR growth in AI labs, as well as the speed of token consumption growth brought about by the proliferation of AI tools. In conclusion, SemiAnalysis gives a clear judgment in the report: under the support of multiple factors, the probability of GPU rental prices continuing to rise is high, and this trend has formed a self-reinforcing cycle - new cloud companies, under the expectation of supply tightening and rising prices, will lock in more hardware capacity in advance, further exacerbating the supply squeeze and pushing prices higher. This round of price increases will directly increase the capital return rate of new cloud companies that are already deployed, while also extending the economic life of existing GPUs. Short-term contracts, large-scale installations with H100, and manufacturers with recent new capacity deployments will be the most direct beneficiaries.