Nvidia's largest hyperscaler customers face sharply higher costs next year, with Vera Rubin and Grace Blackwell systems expected to jump in price as surging memory costs ripple through the chip stacks.

Bloomberg cited sources familiar with Nvidia's new pricing regime, indicating that hyperscalers will see a 15% increase in prices for Rubin and Grace Blackwell processors next year. They said the size of the increase will depend on the chip generation and memory configuration.

Contract manufacturers serving major data center projects, including Microsoft, Google and Oracle, have started notifying customers about the incoming increases, those people said.

Nvidia's AI accelerators remain expensive and in short supply because production at Taiwan Semiconductor Manufacturing has struggled to keep pace with soaring demand. Nvidia generates gross margins of about 75%, but the latest increases suggest that even the world's most valuable semiconductor company is unwilling to absorb rising memory costs.

The coming price shock may intensify pressure on hyperscalers already committing hundreds of billions of dollars to AI infrastructure. Goldman has forecast that AI CapEx among hyperscalers will top $1 trillion next year: