An analysis says rental prices for Nvidia’s H100 graphics processing units (GPUs) have jumped 40 percent in six months, pushing the market into another supply cliff. A surge in demand is driving up prices for artificial intelligence (AI) server components. The knock-on effect is deepening tight supplies of high-bandwidth memory (HBM), raising the likelihood of structural benefits for Samsung Electronics and SK Hynix.

SemiAnalysis said the one-year contract rental rate for an H100 rose from $1.70 per GPU per hour in October last year to $2.35 in March this year. It said allocations scheduled for new deployment through August are already sold out, and on-demand immediate rental capacity has been exhausted across all GPU types. SemiAnalysis also said that even for H100 and H200 eight-node systems, or 64 GPUs, more than half of suppliers reported having no inventory.

The backdrop to the demand surge includes the spread of multi-agent workloads and native media-generation services. When multiple AI agents run several tasks at the same time, token consumption rises exponentially. SemiAnalysis said its employees recently consumed billions of tokens over the past seven days.

An analysis says demand is unlikely to weaken easily even if rental prices rise further, as the return on investment for AI tools reaches 5 to 10 times the cost. It forecast a second wave of token-consumption growth if major companies such as those in the Fortune 500 move to full-scale adoption of agent AI.

On the supply side, a surge in AI server prices has delayed new cluster deployments, worsening supply-demand conditions. Contract prices for LPDDR5 are expected to rise to about four times the level a year earlier, and DDR5 to about six times, prompting server makers to raise AI server prices by more than the increase in component costs.

In addition, some operators facing weaker profitability in cluster investments have delayed or halted new deployments, leaving planned supply unable to reach the market. Shrinking supply combined with rising demand is deepening rental tightness.

This trend also helps explain why demand for the H100 has not weakened even after a shift to next-generation Blackwell GPUs. Contract renewals for H100 and H200 clusters are increasing, and some have signed four-year renewals running to 2028. With lead times for new Blackwell deployments extended to June or July due to surging demand, H100 inventory equipped with HBM2e and HBM3 is also being absorbed quickly. If a full ramp-up of GB300 clusters coincides, demand for HBM4 could come sooner, it said.

◆Memory contract prices seen staying in an uptrend through the fourth quarter

Rising memory prices are also notable as GPUs hit a supply cliff. DS Investment & Securities said the premium of current DDR5 spot prices over contract prices is 60 to 70 percent, while DDR4 stands at 70 to 90 percent. Based on average selling prices, it said the contract-price upcycle is likely to continue through the fourth quarter this year, and demand downside will remain firm as long as cloud service providers (CSPs) maintain AI infrastructure investment at current levels.

It also said the recent expansion of long-term supply agreements (LTAs) is raising the cycle’s downside. The analysis said margins become more stable as price floors and volumes are fixed in advance. It said this differs from past structures where spot-price declines quickly spread to contract prices.

Deepening tight supply of DRAM and HBM is further lifting profitability for the two major memory makers, including Samsung Electronics and SK Hynix. DS Investment & Securities said the two companies’ capital expenditure expansion is focused on HBM and leading-edge processes, keeping constraints on commodity DRAM supply in place.

It said overall memory supply is likely to remain tight through the end of 2027, and a rapid shift to oversupply is likely to be limited. It said the key will be HBM4 yields and supply timing. The pace of stabilisation in HBM4 supply in the second half of the year is expected to be a key variable determining the strength of earnings leverage for Samsung Electronics and SK Hynix.