The Californer - eRacks Opens Its Purchasing Records on the AI Hardware Squeeze: the Same 96GB GPU Is Up 87%

Trending...

The open-source server maker published the numbers behind this year's component inflation - an 87 percent GPU list-price increase in under 18 months, server memory roughly four times higher per gigabyte, flash premiums of 30 to 50 percent - and the

CAMPBELL, Calif. - Californer -- eRacks Open Source Systems today published figures from its own purchasing records showing what the AI buildout has done to hardware prices in 2026 - and what those prices mean for teams deciding between renting AI capacity by the month and owning the server that runs it.

The numbers, taken from the company's component purchasing and price tracking this year: NVIDIA's RTX PRO 6000 Blackwell, the 96GB card that serious AI shops standardize on, launched at $8,565, was repriced to $13,250, and now lists at $16,000 - an increase of 87 percent in under 18 months. Server memory (ECC RDIMMs, the error-correcting kind server boards require) has roughly quadrupled per gigabyte in the 2026 shortage. Mainstream enterprise NVMe drives (fast solid-state storage) are out of stock at major distributors, and in-stock units carry premiums of 30 to 50 percent.

The hyperscalers (the biggest cloud operators) buy GPUs, memory, and flash by the container; everyone downstream pays the new price.

More on The Californer

Rising hardware prices do not favor renting, the company argues. Cloud GPU rates ride the same scarcity, and rent never converts into a machine the team owns. For a team running AI workloads daily, an owned server typically pays for itself inside a year - and waiting has a cost of its own, because the same configuration has cost more every quarter this year and the shortage driving that has not eased.

"We watch these numbers because we pay them - every card, every stick of memory, every drive that goes into a customer build," said Joseph Wolff, founder and CTO of eRacks Systems. "The same GPU we bought at $8,565 lists at $16,000 today, and nothing in the supply picture has eased. If owning is where your team lands eventually, sooner costs less than later. That is not a sales line, it is our purchasing history."

eRacks prices its configurator from live component costs, and the price a customer configures today is the price paid at order. The company publishes the rent-versus-own calculator it uses internally at https://eracks.com/tco/ - browser-based, no signup, comparing a team's current AI spending against owning at live configured prices. The AI line runs from 2-GPU value systems to 96GB-class 8-GPU flagships, built to order with open-source stacks and no Windows tax.