Some of Nvidia’s largest customers have been told that prices of servers containing its AI chips will rise by more than 15% in many cases, according to a Bloomberg News report on Saturday. The price hikes, driven by soaring memory chip costs, will go into effect on systems shipped early next year and will impact flagship platforms including Vera Rubin and Grace Blackwell.
The increases will depend on Nvidia’s chip generation and memory configurations, the report said, citing people familiar with the process. Reuters confirmed the report but could not immediately verify it independently. The news sent ripples through a semiconductor market already grappling with supply constraints and escalating demand for AI infrastructure.
Memory Costs Drive the Increases
At the heart of the price escalation is the surging cost of high-bandwidth memory, or HBM, which is essential for Nvidia’s most advanced AI accelerators. HBM prices have climbed sharply throughout 2026 as demand from hyperscale cloud providers outpaces production capacity at Samsung, SK Hynix, and Micron. The three memory manufacturers have collectively invested tens of billions of dollars in expansion, but new fabs take years to come online.
Vera Rubin, Nvidia’s next-generation platform set to ship in the second half of 2026, relies on HBM4 memory that promises substantially higher bandwidth and capacity than the HBM3e used in current Blackwell systems. The transition to newer memory standards adds to per-unit costs at a time when memory makers are already capitalizing on tight supply.
Grace Blackwell, the current-generation data center platform combining Nvidia’s custom Grace CPU with Blackwell GPUs, will also see price increases. The system has become the backbone of AI training deployments at companies including Microsoft, Google, Meta, and Amazon.
Hyperscalers Push Back, But Dependence Remains
Major customers like Amazon, Microsoft, Google, and Meta are all pursuing their own in-house chip programs, but they remain heavily dependent on Nvidia purchases for their data center build-outs. Amazon’s Trainium and Google’s Tensor Processing Units have made inroads, but neither has displaced Nvidia’s dominance in frontier AI training.
“The price increases reflect a broader structural shift in the AI hardware market,” semiconductor analyst at Futurum Group noted, pointing to the combination of memory inflation, advanced packaging bottlenecks, and insatiable demand for inference compute.
The timing of the hikes could complicate capital expenditure plans for hyperscalers, which are already spending record amounts on AI infrastructure. Dell’Oro Group recently forecast that worldwide data center capital spending will surpass $3 trillion by 2030, with AI accelerators accounting for roughly a third of that total.
For cloud providers, the cost increases will likely be passed through to enterprise customers renting AI compute, potentially raising the price of inference and training workloads across the industry. Smaller AI startups that depend on cloud-based GPU access could face the steepest relative impact.
discussion