Overall demand for AI infrastructure remains healthy, but subtle changes are occurring within its structure:
1. Token spending for Large Language Models (LLMs) has slowed after strong growth in June, with open-source models eroding market share. For example, the Kimi K3 has entered the top five in spending, breaking the previous monopoly held by closed-source models.
2. Rental prices for the new generation GPU B200 accelerated, increasing by 7.3% quarter-over-quarter to $5.72 per GPU-hour. On the other hand, rental prices for the older H100 model decreased by 1.1% quarter-over-quarter, the first decline in seven months, suggesting a shift in demand towards the latest generation of chips (such as the Blackwell architecture). This generational shift benefits Nvidia (as customers are willing to pay higher rental rates for the new architecture, helping to ramp up B200 sales and extend the product cycle).
3. Price increases in the storage market have narrowed significantly. Contract prices for DRAM and NAND are projected to grow by only +16% and +13% respectively in Q3 2026 (compared to the high increases of +56% and +58% in Q2). This is primarily attributed to the high base effect, weak consumer demand, long-term agreements with cloud service providers (CSPs), and high inventory levels resulting from advance stockpiling in the first half of the year. (Source: JPMorgan Chase)
