In early January, AWS quietly increased the price of EC2 capacity blocks for machine learning. This corresponds to approximately a 15% increase on GPU-based instances used for heavy-duty ML training.
It's worth noting that AWS implemented the change over the weekend without separately announcing it to customers.
Specifically, p5e.48xlarge and p5en.48xlarge instances are now more expensive. According to The Register, these instances with multiple NVIDIA H200 GPUs will be used for large-scale machine learning and AI workloads. In most regions, the hourly price for p5e.48xlarge increased from $34.61 to $39.80. For p5en.48xlarge, the price increased from $36.18 to $41.61 per hour. Prices are even higher in the Western North California region of the United States, reaching nearly $50 per hour for the p5e variant.
Capacity blocks are intended for customers who require guaranteed GPU capacity at a predetermined point in time. Unlike Spot Instances, which can disappear suddenly, organizations use Spot Instances to reserve capacity for fixed time windows that can range from one day to several weeks. This service is primarily used by companies that perform business-critical ML training, where interruptions can have significant financial impact. Therefore, the target group primarily consists of large organizations with large cloud budgets.
According to Amazon, the price adjustment is due to a change in the ratio of supply and demand. Capacity block pricing is adjusted quarterly based on expected market conditions. AWS emphasizes that this dynamic has long been part of the pricing model for this particular service.
No clear pricing policy
This price increase comes several months after AWS announced significant price reductions for GPU instances, specifically on-demand and savings plans. At that time, capacity blocks were not included in these reductions. This highlights that pricing and coordination of the various purchasing models within AWS are done independently of each other.
This change may impact organizations using the Enterprise discount program. These contracts typically offer discounts from published prices, but as the base price increases, the absolute amount customers pay also increases. This could lead to new discussions about pricing agreements between large customers and AWS.
The price increase comes amid a global shortage of advanced GPUs. Demand for computing power for AI and machine learning continues to grow, but supply is limited. Competitors such as Microsoft Azure and Google Cloud are aggressively entering this market, but they also face availability limitations.
