Nvidia can't eat the memory bill: AI servers jump 15%

Nvidia — the company with more pricing power than anyone in tech — just told its biggest customers that server prices are going up more than 15%, and the reason is memory. Systems built on Vera Rubin and Grace Blackwell, shipping in early 2027, get repriced because Samsung, SK Hynix and Micron now hold the leverage in the AI supply chain. An accelerator is only as useful as the DRAM bolted to it, and DRAM is scarce. Nvidia passing the cost through instead of absorbing it is the tell.

Everyone is paying anyway. Nvidia is reportedly in talks to back Perplexity at a $30B+ valuation — a 50%+ markup in a year, on annualized revenue that climbed from under $250M to over $750M. And Alibaba just closed a $10.21B share placement, Hong Kong's largest-ever follow-on, priced at an 8.4% discount with every dollar earmarked for chips, infrastructure and models.

One day later it shipped Wan3.0, a video model generating 30-second clips — double its predecessor — from text, images, audio, spreadsheets, slides and PDFs. Context: Alibaba's last quarter saw earnings fall 75% year over year on AI capex. The spending isn't slowing; the bill is just getting itemized.

Why it matters: the AI buildout's cost floor is now being set by memory vendors, not GPU vendors — and it's rising faster than anyone's margins.