Nvidia has told some of its largest customers that AI server prices will rise more than 15% for units delivered in early 2027, Bloomberg reported, as surging memory costs finally pass through to the most expensive systems in the industry. On a 1GW AI data center, this single round of increases could add at least $5 billion to the build bill — before a single GPU draws power. For the supply-side view — how pricing power shifted from GPUs to memory — see Memory, Not GPUs, Now Drives AI Server Price Hikes.
What Nvidia told its biggest customers
Bloomberg said several large Nvidia customers were notified that AI server prices could climb more than 15% starting with deliveries next year. The increases hit both the current Grace Blackwell line and the next flagship, Vera Rubin, with the exact figure varying by GPU generation and memory configuration.
The Information went further: some GB300 and Vera Rubin 200 systems are expected to rise around 17%. On current pricing, that moves a 72-GPU Vera Rubin rack from roughly $7 million to about $8 million.
2026 already had two price waves before this one
This is not the first time Nvidia hardware has gotten more expensive in 2026. The consumer market went first: median retail prices rose about 39% for the RTX 5060 Ti 16GB, 36% for the RTX 5070, 27% for the RTX 5060 and 9% for the RTX 5090 (retail numbers mix supply, demand and channel inventory, so not all of it is official pricing).
Then came pro GPUs. The 96GB RTX PRO 6000 Blackwell went from about $8,565 in early 2025 to $13,250 in June 2026 and $16,000 by August. Now servers join the wave — and server internals were already volatile: some components, from TSMC wafers and advanced packaging to networking, cooling and memory, have moved up to 40% within a single week, with racks rising 2-3% weekly at points.
The real cause is memory, not markup
The mechanism is a feedback loop, not simple margin-grabbing. AI servers drove HBM demand to new highs, and HBM is far more wafer-hungry than standard DRAM: bigger dies, more stacking layers, more complex packaging, so the same wafer input yields proportionally fewer bits. As HBM expands, it squeezes the wafer capacity left for the regular server, PC and phone DRAM those same fabs used to run.
TrendForce numbers make the squeeze concrete: Q2 2026 traditional DRAM contract prices were up 58-63% quarter over quarter and NAND up 70-75%; Q3 server DRAM is still climbing 13-18%, and prices may keep rising quarter over quarter through 2027. Server RDIMM bit supply in 2027 could grow only 15-20% year over year, well behind server CPU shipment growth.
Nvidia is adapting to its own shortage. Because LPDDR5X is tight, it is halving the SOCAMM memory in the next Vera Rubin Superchip, and suppliers can currently cover only about 60% of Nvidia's estimated LPDRAM demand. It has also reopened Rubin Ultra's HBM evaluation beyond the original 12-Hi HBM4E to include 8-Hi HBM4E, 12-Hi HBM4 and 8-Hi HBM4, with no final spec set.
What it means: a squeeze from both sides
Here is the uncomfortable contrast: while frontier inference prices keep falling (OpenAI just cut GPT-5.6 Sol by 20%+), the hardware underneath is getting more expensive. Cloud providers are getting squeezed on both ends of the margin sandwich, and the $5B-per-GW shock lands just as power and land were already becoming the next bottleneck.
The response is already visible. Amazon, Microsoft, Google and Meta are all ramping in-house AI chips to cut dependence on Nvidia. Nvidia, for its part, is moving upstream: it took a minority stake in data-center developer Cloverleaf Infrastructure on August 21, invested $1.5 billion in SoftBank's SB Energy, and is working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to pull more than $500 billion of third-party capital into AI infrastructure. The fight has shifted from "can we build the chip" to "can the customer get power, land and grid access."
What builders and operators should do now
- Lock in capacity and pricing early: units delivered in early 2027 will cost 15-17% more, and TrendForce sees memory prices rising into late 2027.
- Model capex with memory cost bands, not a single point estimate — this cycle is moving quarter to quarter.
- Watch Nvidia's HBM decision on Rubin Ultra: a spec change (12-Hi HBM4E versus wider options) changes per-rack performance and total cost of ownership.
- Use configuration as a short-term lever: cloud providers are already swapping 96GB/128GB RDIMMs for 32GB/64GB modules to cut procurement cost.
- Read Nvidia's infrastructure investments as a signal: when the chipmaker starts buying power and land, the bottleneck has moved beyond chips.
FAQ
Q: Why are Nvidia AI server prices going up?
A: Memory. HBM expansion is consuming the wafer capacity that regular DRAM needs, pushing contract prices up (server DRAM +13-18% in Q3 2026 alone). Nvidia is passing that through: servers delivered in early 2027 will cost more than 15% extra, and some GB300 and Vera Rubin 200 systems around 17%.
Q: How much more will an AI data center cost?
A: Roughly $5 billion more per 1GW from this round alone, since a 72-GPU Vera Rubin rack rises from about $7 million to $8 million.
Q: Should we lock in Nvidia capacity now or wait?
A: Lock in early if you can. TrendForce expects server memory prices to keep climbing quarter over quarter into 2027, and Nvidia has said increases start with early-2027 deliveries; waiting likely means paying more.
