Who Bears the Memory Costs?

NVIDIA is not absorbing the rising memory costs of HBM4 and SOCAMM internally, but instead passing them on to customers by reflecting them in AI server prices. As a result, prices for systems based on Vera Rubin and Grace Blackwell, scheduled for release in early next year, are expected to increase by more than 15% depending on the configuration.

NVIDIA has maintained a high gross margin in the mid-70% range despite rising memory prices. This is because, although HBM is the largest cost factor in accelerators such as the B200, its share relative to the final product price is relatively small. However, as memory costs shift to higher tiers with the transition to the Rubin generation, market concerns are growing regarding NVIDIA's pricing power and margin volatility.