Some of Nvidia’s largest customers have been told that AI servers using the company’s chips could become more than 15% more expensive, with the increases expected to affect systems shipped from early 2027. The reported changes include configurations built around Nvidia’s Grace Blackwell and Vera Rubin platforms.
The main pressure is coming from the rising cost of memory. Nvidia’s AI accelerators depend heavily on high-performance memory, while demand from data centers has pushed suppliers such as Samsung, SK Hynix and Micron to operate in an increasingly tight market. TrendForce expects conventional DRAM contract prices to rise another 13–18% in the third quarter of 2026.
Why AI Infrastructure Is Getting More Expensive
The development reveals an important shift in the AI infrastructure race.
For years, the biggest bottleneck was access to GPUs. Now, memory is becoming an equally important constraint. As AI models become larger and data centers deploy more accelerators, the amount of memory required by each system continues to increase.
That gives memory manufacturers greater pricing power—and forces higher costs through the entire AI infrastructure chain.
The Cost Reaches The World’s Largest Data Centers
Server manufacturers supplying companies such as Microsoft, Google and Oracle have reportedly begun informing customers about the coming increases. The exact impact will depend on the Nvidia generation and memory configuration of each system.
For the companies building massive AI data centers, even a 15% increase can represent billions of dollars in additional infrastructure spending.
The Bigger Picture
The AI boom is creating an unusual economic dynamic.
Demand for computing is rising so quickly that the components required to build AI infrastructure are becoming scarce—and more expensive.
AI is becoming more powerful. But for now, building the machines behind it is becoming more expensive too.
