Nvidia raises AI server prices by more than 15 percent — memory is the reason
It is not the processors driving the increase but DRAM and HBM. Microsoft, Google and Oracle are already being passed the surcharge on systems shipping in early 2027.
Illustration · AI-generated (AI IN LIFE)
At a glance
- Price increase: more than 15 percent on part of the AI server line-up
- Platforms affected: Vera Rubin and Grace Blackwell, shipping from early 2027
- Customers already affected: Microsoft, Google, Oracle
- Cause: shortage of DRAM and HBM; suppliers are Samsung, SK Hynix and Micron
- Forecasts: Gartner expects the shortage to last into 2027, Deloitte a quadrupling of AI server DRAM prices over the year
Nvidia has told its server manufacturers that prices for a portion of its AI server systems will rise by more than 15 percent. The increase affects systems built on the Vera Rubin and Grace Blackwell platforms, shipping from early next year. Manufacturers are already passing the surcharge on to large data centre operators, among them Microsoft, Google and Oracle.
The cause is the notable part. The increase does not stem from the graphics processors themselves but from the memory surrounding them: conventional working memory (DRAM) and the high bandwidth memory (HBM) attached directly to the chip. Both have been in short supply for months and prices have climbed accordingly. The exact size of the surcharge depends on the chip generation and the memory configuration of a given system.
The shortage has a structural cause. Modern AI accelerators need considerably more memory per server than earlier generations — models and their intermediate states have to sit as close to the compute units as possible, otherwise the processor starves for data. Demand for memory is therefore growing faster than demand for compute itself, while manufacturing capacity at the three large suppliers Samsung, SK Hynix and Micron expands only slowly.
Micron has already stated that industry supply is likely to remain substantially below demand through and beyond 2026. Analysts at Gartner expect the shortage to last into 2027. Deloitte projects that AI server DRAM prices will quadruple from their starting point over the course of the year.
For operators this is not an emergency stop but a shift in the arithmetic. A 15 percent surcharge does not end the investment boom — it simply raises the capital required to deploy the same amount of computing capacity. Anyone budgeting in dollars rather than in compute will get measurably less for their money from 2027 onwards.
That pushes a question to the fore that has so far been asked mainly on the stock markets: when does the compute pay for itself? Economists at the European Central Bank recently warned of a valuation bubble in AI stocks. Rising hardware prices hit that debate at its most sensitive point — they increase the cost without changing the return.
FAQ
Why do prices rise if the chips themselves are not more expensive?
An AI server is more than its processor. A large amount of working memory and high bandwidth memory sits around the accelerator, and those are the components currently in short supply. Their price increase feeds through to the whole system even though the compute die is unchanged.
What is the difference between DRAM and HBM?
DRAM is conventional working memory, the kind found in ordinary computers. HBM stands for High Bandwidth Memory and is stacked directly beside the processor to deliver very high data rates. AI accelerators need HBM because otherwise they could compute faster than data can be fed to them.
Does this slow down data centre construction?
Not immediately. Announced investment programmes continue. The effect is different: the same budget buys less compute from 2027 onwards. How strongly that bites depends on whether the memory shortage really does persist into 2027.
