IMPLICATOR
Contract server builders have told some of Nvidia’s largest customers that prices for AI servers containing its chips shipping in early 2027 will rise by more than 15% in many cases. Rising prices and shortages for DRAM, LPDDR and high-bandwidth memory are increasing the cost of systems built around Nvidia’s newest accelerators. The notices add another cost risk to data-center projects already facing project delays, labor shortages, tighter capital markets and community resistance.
Disclosed on Aug. 22, the communications cover systems using Nvidia’s Vera Rubin and Grace Blackwell platforms, and the size of each increase depends on the chip generation and memory configuration. Nvidia did not respond to requests for comment on the communications, and the contract server builders and the people familiar with them were not named in the report. More than 15% is a range across generations and configurations rather than a published price sheet.
What Changed
- Contract server builders told some of Nvidia's largest customers that AI server prices will rise by more than 15% in many cases for systems shipping in early 2027.
- The communications cover Vera Rubin and Grace Blackwell systems, with the increase varying by chip generation and memory configuration.
- GF Securities analysis cited in a July 28 TrendForce item put memory near 29% of a roughly $2.1 million Vera Rubin VR200 system bill of materials without configuration cuts.
- TrendForce projected that HBM bit shipments would grow 50% to 60% year over year in 2027 and still fail to match demand growth.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Memory in the bill
GF Securities analysis, via Wccftech, appeared in a July 28 TrendForce item and put memory at about 29% of a Vera Rubin VR200 system’s roughly $2.1 million bill of materials if Nvidia made no configuration cuts. Nvidia’s preferred memory share of the system bill of materials was 20%.
The same cited analysis said Nvidia may reduce the SOCAMM memory attached to Vera CPUs from about 55 terabytes to 28 terabytes per rack, while leaving GPU HBM4 capacity at 20.7 terabytes. That would lower the amount of costly LPDDR5X memory in each system without cutting the high-bandwidth memory feeding the GPUs.
The shortage reaches conventional DRAM, LPDDR and HBM. Samsung, SK Hynix and Micron produce most of the world’s DRAM, and the two Korean suppliers warned in April 2026 that shortages could persist through at least 2027. On Aug. 4, TrendForce projected that HBM bit shipments would grow 50% to 60% year over year in 2027 and still fail to match demand growth.
Other price signals
A June-to-August 2026 survey of median Newegg listings found that Nvidia’s RTX 5060 Ti 16GB graphics card rose 39%, while the RTX 5070 rose 36%. The median RTX 5070 Ti price was unchanged over the same period, showing that the increases were uneven across Nvidia’s Blackwell gaming range.
FREE WEEKDAY MORNING BRIEFING
Don’t miss the next AI story that matters.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
Enterprise memory points in the same direction. Counterpoint projected that 64GB DDR5 RDIMM modules used in data centers could cost twice as much by the end of 2026 as they did in early 2025, even though the same forecast expected DRAM output to grow more than 20% in 2026.
Buyers have unequal options
Nvidia’s largest cloud customers are developing in-house accelerators, but those chips still depend on memory from the same suppliers. Hyperscalers can reserve production through long-term commitments and direct investments. Smaller enterprises usually buy complete systems from Dell, Lenovo, HPE or Supermicro and have less control over the memory supplier or contract price.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
Procurement teams at smaller companies may have to accept the configuration and supplier their vendor can secure, then spread a deployment over time to reduce exposure to one price spike.
In late July, Nvidia said it had secured AI memory supply from SK Hynix as part of a broader agreement that could be worth $500 billion over multiple years. That scale is unavailable to most corporate buyers.
“In 2026, there’s going to be issues around semiconductor supplies, and it’s going to affect everyone, not just Samsung,” Samsung global marketing president Wonjin Lee said.
Frequently Asked Questions
How much are Nvidia AI server prices expected to rise?
More than 15% in many cases for systems shipping in early 2027. The figure is a range across chip generations and memory configurations, not a published price sheet.
Which Nvidia systems are covered by the communications?
The communications cover systems built around Nvidia's Vera Rubin and Grace Blackwell platforms.
Why are the server prices rising?
DRAM, LPDDR and high-bandwidth memory are becoming more expensive and harder to secure. Samsung, SK Hynix and Micron produce most global DRAM, and suppliers have warned that shortages could persist through at least 2027.
How might Nvidia reduce the memory cost?
The cited analysis said Nvidia may cut SOCAMM memory attached to Vera CPUs from about 55 terabytes to 28 terabytes per rack while leaving GPU HBM4 capacity at 20.7 terabytes.
What does the price increase mean for buyers?
Large cloud operators can reserve memory capacity through long-term commitments and direct investments. Smaller enterprises often rely on system vendors for the available memory supplier and configuration, leaving them with less control over contract prices.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
