Bloomberg reported that some of Nvidia's biggest customers, the contract manufacturers who assemble AI servers for hyperscalers like Microsoft, Google, and Oracle, have been told that system prices are set to rise by more than 15% on units shipping in early 2027. The increase traces back to soaring memory prices: HBM and other DRAM used in Vera Rubin and Grace Blackwell configurations has become scarce as every major AI lab and cloud provider competes for the same limited fabrication capacity. Because GPU servers are typically priced as a bundle of compute, memory, networking, and cooling, a memory shortage translates almost directly into a sticker-price jump for the finished machine, even though Nvidia's own GPU die pricing hasn't necessarily moved. For teams planning infrastructure budgets, this matters beyond the headline number. Anyone forecasting the cost of training runs, inference clusters, or on-prem GPU purchases for 2027 now has to build in a wider margin of error, since the increase varies by chip generation and memory configuration and isn't uniform across vendors. It also reinforces a trend that's been building all year: compute cost per token has been falling thanks to model efficiency gains, but the absolute cost of acquiring frontier-class hardware keeps climbing, which pushes more workloads toward rented cloud GPU capacity rather than capital purchases for companies that can't absorb sudden 15%+ swings. Engineering leaders sizing multi-year GPU commitments, or negotiating reserved-instance contracts with cloud providers, should treat this as a signal to lock in pricing sooner rather than later and to build memory-cost volatility into any total-cost-of-ownership model for AI infrastructure.