NVIDIA H200 rental prices are up about 160% from the start of the year, and A100 rents are also higher than at the start of the year

Old GPUs are getting more expensive to rent! NVIDIA H200 rental prices surge 143%: A100 rents are rising too Source: Bandaozongheng Published: 2026-10-26 13:50 NVIDIA compute-generated poster AI inference demand continues to grow, the GPU rental market is seeing a recovery, and the compute value of both new and old models is once again drawing attention.

Old GPUs are also rising in rent

As global AI token consumption grows, prices in NVIDIA's data center GPU rental market have begun to recover. H200 rents have risen notably, while products such as B200 and H100 have also shown varying degrees of increase; even the A100, which has been on the market for years, now rents for more than it did at the start of the year.

According to relevant industry analysis, since 2026 GPU rental prices have generally trended upward. Among them, H200 rents have risen about 160% from the start of the year, B200 about 120%, RTX PRO 6000 about 65%, and H100 about 35%. Even the A100, released years ago, now rents for more than it did at the start of the year.

Looking at the course of rental price changes, this round of increases did not rise continuously from the start of the year. In the first quarter of 2026, rental prices for several GPUs briefly fell back, with H100 and B200 prices at one point about 23% below their levels at the start of the year. After April and May, prices began to rebound noticeably. H200 spiked in late May, then went through a few weeks of pullback before rising again in October.

From another perspective, using a six-year straight-line depreciation model as a reference, current rents for some GPUs are up to about 143% higher than the model's estimated values. This means the actual rental prices of some GPUs are performing noticeably stronger than the levels corresponding to the theoretical depreciation curve.

Looking at per-card rental prices, the relevant transaction price index shows a similar trend. Taking the data up to now as an example, the rental price of H200 is $6.30 per GPU hour, B200 is $9.70, H100 is $3.03, and A100 is $1.09. Among them, H200 has risen about 24% in the past month, and B200 about 20.7%. These data come from the GPU rental market transaction price index and do not represent uniform quotes from all cloud service providers.

Price trends differ across GPU models. Memory capacity, computing power, interconnect capability, and applicable workloads all affect customers' rental choices. As AI application scenarios expand, market demand for different tiers of compute is also changing.

Why the H200 draws more attention

The rise in H200 rents is related to its product positioning for large-model training and inference tasks.

According to public product information, the H200 uses the Hopper architecture, is equipped with 141GB of HBM3e memory, and has a memory bandwidth of 4.8TB/s. Compared with the H100, the H200's memory capacity is nearly double and its bandwidth is about 40% higher. The product targets workloads such as generative AI, large language models, and high-performance computing, and has performance advantages in some large-model inference tasks.

The H200 has larger memory capacity and higher bandwidth, enabling it to handle some tasks with high memory resource requirements. For companies providing GPU cloud services, if customers need to run large models or handle high inference loads, the H200 has an opportunity to capture more rental demand.

AI token growth drives compute demand

Behind the rebound in GPU rents, changes in the scale of AI applications are an important direction for observation.

The large-model industry is expanding from model training to everyday inference calls. As models are integrated into more software, development tools, and enterprise businesses, GPUs need to continuously process user requests, making inference computing an important component of compute demand.

According to statistics published by a model-call aggregation platform, the number of tokens the platform processes each week grew from about 0.5 trillion in mid-January 2025 to 128.4 trillion in mid-September 2026, an increase of about 250 times in roughly 20 months. The relevant statistics also show that more autonomous AI Agents are driving further growth in token usage.

The growth in model calls directly increases the number of inference tasks executed. Unlike one-off Q&A, an AI Agent often needs to call models continuously, execute tools, process results, and continue reasoning based on feedback. A single task may involve multiple rounds of model calls, and token consumption rises accordingly. As such applications expand, demand for computing resources may also continue to grow.

According to relevant industry analysis, since the spring of 2026, model call volumes have grown noticeably, and GPU rents also began to rise during this period. However, existing data do not yet clearly break down the specific sources of rental demand, so it is impossible to confirm how much incremental demand AI Agents, enterprise application deployments, and other scenarios each contributed.

Supply-side factors also affect rental prices. Data center capacity, power supply, and the actual number of GPU devices available for rent can all affect market supply and demand. Growth in token demand and rising GPU rents are correlated in time, but price changes are also affected by multiple factors together.

Old cards still have their own market

Rising rents for the new generation of GPUs do not mean older products have lost their competitiveness. The case of the A100 shows that demand in the GPU market is not entirely determined by the latest products.

The A100 uses NVIDIA's Ampere architecture, with some models equipped with 80GB of HBM2e memory and support for features such as multi-instance GPU. Although its product generation lags behind the H100, H200, and B200, it still has practical value for some model fine-tuning, inference services, and other software workloads adapted to existing hardware.

In addition to the A100, the earlier Tesla V100 also has a chance to play a role in specific tasks. According to relevant reports, a four-card V100 system with NVLink interconnect remains somewhat competitive in some medium-scale large language model tasks. Related tests also showed a performance advantage of up to about 230% compared with the M5 Ultra and DGX Spark, and some high-memory versions have also seen a premium of about $200 on the second-hand market.

Such comparisons are closely related to specific test tasks, software adaptation, and hardware configurations, and cannot be directly extended to all AI workloads. However, it reflects that older GPUs still have their own application space. As long as they can meet the performance, memory, and cost requirements of specific tasks, old cards have a chance to continue participating in AI computing.

For cloud service providers, GPUs already put into use can still generate revenue by renting out computing resources. Some tasks do not require the latest generation of GPUs, and older devices still have a chance to attract customers with suitable rental prices. The demand brought by performance improvements of new cards may not necessarily completely squeeze out the use space of old cards; products of different generations can serve different workloads and budgets.

Higher GPU rents can increase revenue per unit time of equipment, but whether profits can ultimately improve still depends on equipment utilization and overall operating costs. Expenses such as electricity, cooling, data center space, networking, maintenance, and idle equipment all affect operators' actual returns.

Related industry analysis also shows that for a long period previously, the rental prices of A100 and H100 were lower than the level estimated according to a six-year straight-line depreciation model. With the price recovery since the spring of 2026, the rental revenue performance of some GPUs has improved, but supply and demand conditions still differ among models, and future price trends will also depend on changes in computing power demand and supply.

From the rise in rents for H200 and B200 to the continued application space for A100 and V100, the GPU rental market shows a noteworthy change: AI computing demand is covering more hardware generations. The new generation of products takes on demanding tasks with larger memory and higher bandwidth, while older products continue to serve workloads they can accommodate.

For the AI computing power industry chain, in addition to paying attention to the performance and shipment volume of new GPUs, equipment utilization, changes in inference demand, and the cost-performance matching among different models will also become important indicators for observing the computing power market.

This article is reproduced from media reports or online platforms and represents the author's personal position or views. Our reproduction is solely for sharing and does not mean that we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service in a timely manner, and as a neutral platform service provider, we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

DB HiTek renews IATF 16949 certification, focuses on automotive-grade chips 2026-10-26 13:54

Open-source tool unlocks hidden AI computing power, RTX 30 series old cards "revived" 2026-10-26 13:45

Samsung's self-test shows Exynos 2700 performance is about 20% higher than Qualcomm's flagship chip 2026-10-26 11:48