Rising memory prices are making AI infrastructure more expensive — and Nvidia’s biggest customers could be in for serious sticker shock.
Last month, Meta CEO Mark Zuckerberg and Microsoft CFO Amy Hood cited rising memory costs for chips and data centers as a key factor in their increased AI spending. The surge has helped fuel a boom in memory stocks.
Research firm Bernstein said in a note to investors this week that the impact could become more pronounced with Nvidia’s next-generation AI systems. For Nvidia, these rising costs could be a substantial growth driver.
The firm estimates Nvidia’s coming Vera Rubin NVL72 system — used to train and run advanced AI models — will cost about $9.1 million per rack.
Memory makes up a big part of that pricetag. Bernstein says high-bandwidth memory prices will more than triple by the time Rubin ships at scale, pushing memory and storage costs to $3.2 million per rack — or more than a third of the total cost.
That could mean Nvidia’s biggest customers will rack up higher bills. Bernstein estimates a 1-gigawatt Vera Rubin data center would cost about $47 billion to build, compared to $40.5 billion for building data centers with Vera Rubin’s predecessor chips.
It’s not just memory. Bernstein also flagged that networking, cooling, and power-delivery costs are rising and contributing to higher prices.
Given demand for its AI systems and limited competition, Bernstein says Nvidia has enough pricing power to pass along higher memory costs to customers.
“We believe that Nvidia likely has some form of dynamic pricing mechanism, and will pass on this increase to customers instead of absorbing it as a hit to margin,” the firm wrote.
Nvidia CEO Jensen Huang has previously warned that memory shortages could last “quite a few years.”
The report comes as Nvidia and Korean semiconductor supplier SK Hynix announced a multiyear partnership earlier this week to develop next-generation memory for future systems, including Vera Rubin. SK Hynix recently hit a market capitalization of $1 trillion in late May.
Nvidia did not immediately respond to a request for comment from Business Insider.
Have a tip? Contact this reporter via email at gweiss@businessinsider.com or Signal at @geoffweiss.25. Use a personal email address, a nonwork WiFi network, and a nonwork device; here’s our guide to sharing information securely.
