What happened
The Korea Herald reports that server manufacturers have warned major data-center customers of price increases for Nvidia-based systems built on the Grace Blackwell and Vera Rubin platforms. Bloomberg, as cited by the Herald, reported increases of more than 15% in many cases, while The Information reportedly cited increases of about 17%. The Herald attributes the pressure largely to tight supplies of high-bandwidth memory and conventional server DRAM, which are essential to AI infrastructure. The reported customer warnings and price levels have not been independently confirmed in the source by Nvidia, the server manufacturers, or the named cloud companies.
The Korea Herald reported on Aug. 24 that Samsung Electronics and SK hynix were gaining bargaining power as higher memory costs fed into prices for Nvidia-based AI server systems. Citing Bloomberg, the Herald said server manufacturers had begun notifying major data-center customers, including Microsoft, Google, and Oracle, that prices would rise by more than 15% in many cases for shipments starting in early 2027. The reported increases would affect systems using Nvidia’s next-generation Vera Rubin platform and its current Grace Blackwell platform. The article said the size of the increase varies by chip generation and memory configuration.
The Herald identified memory as a central source of the pressure. Nvidia AI accelerators use high-bandwidth memory, or HBM, while AI data centers also consume large quantities of conventional server DRAM. According to Counterpoint Research data cited in the report, SK hynix held 58% of global HBM revenue in the first quarter. Samsung led the broader DRAM market in the second quarter with a 39% share, and the two Korean companies together accounted for 65% of global DRAM revenue. The report said most HBM revenue in the first quarter still came from HBM3E, with HBM4 expected to become more significant in the second half of the year.
The report also cited The Information, which said some server suppliers had warned customers of increases of about 17% for major Nvidia-based systems. The Information estimated that such an increase could add at least $5 billion to the chip-system cost of a 1-gigawatt data center. The Herald emphasized that this estimate concerns chip-system costs rather than the total cost of constructing a data center. The customer warnings, the exact price increases, and the $5 billion estimate were not independently confirmed in the source by Nvidia, the server manufacturers, Microsoft, Google, Oracle, or other customers.
Read the primary source: koreaherald.com ↗
Why it matters
The report points to a direct cost and supply constraint on the infrastructure behind large-scale AI services. It also suggests that Samsung and SK hynix may have more negotiating leverage because they control significant shares of HBM and broader DRAM revenue. Higher system prices could affect the economics and pace of new AI data-center projects, although the source does not establish that any planned project has been canceled or delayed.
The reported price increases matter because AI computing depends on tightly integrated combinations of accelerators, memory, networking, and power infrastructure. If the price of complete Nvidia-based systems rises, companies building new capacity may need to spend more before generating revenue from AI services. The Herald quoted Shin Joong-ho, head of research at LS Securities, who said that major technology companies could absorb higher costs temporarily but might reassess the expected returns of new AI data-center projects if memory inflation continued to push system prices higher.
The story also illustrates how the economics of AI infrastructure are distributed across the supply chain. Nvidia is the most visible supplier of the accelerator platforms, but memory makers can exert influence when available HBM and server DRAM capacity is constrained. The Herald cited TrendForce’s expectation that server DRAM contract prices would rise another 13% to 18% in the third quarter from the previous quarter as AI demand absorbed available capacity. These figures describe market expectations and reported contract pricing, not a universal price paid by every buyer.
The pressure extends beyond Nvidia’s own systems. The Herald reported that Amazon, Microsoft, Google, and Meta are developing more of their own AI processors to reduce reliance on Nvidia, but said those systems also require high-performance memory. That means custom chips may diversify accelerator supply without eliminating the memory bottleneck. The article does not show how much of the reported pricing power translates into profit for Samsung or SK hynix, nor does it establish that higher server prices will materially reduce consumer access to AI services.
What to watch next
Watch Nvidia’s upcoming financial results, the evolution of HBM4 shipments, and whether server suppliers publicly confirm the reported increases. The key unknowns are the precise prices paid by individual customers, the share of each system’s cost attributable to memory, whether higher costs are passed through uniformly, and how cloud companies adjust AI infrastructure spending. Custom AI accelerators from Amazon, Microsoft, Google, and Meta may reduce dependence on Nvidia, but the Herald reports that they still require high-performance memory.
Nvidia’s fiscal second-quarter results, due Wednesday according to the Herald, may provide the next public evidence about demand, supply, and the cost of deploying its newer systems. Important signals would include commentary on Vera Rubin and Grace Blackwell production, memory availability, customer commitments, and whether Nvidia or its manufacturing partners acknowledge broad price changes. The source does not provide Nvidia’s response to the reported customer warnings.
HBM4 adoption is another key indicator. The Herald said HBM4 shipments were expected to become more meaningful in the second half of the year, while SK hynix had an early lead in HBM and Samsung was expanding in advanced HBM products. Future market-share and pricing data could show whether the two Korean suppliers retain leverage as newer memory standards scale, or whether additional capacity and competing suppliers ease the constraint.
Finally, observers should distinguish reported system-price increases from the broader cost of building and operating AI data centers. The article does not identify the affected customers’ order volumes, contract terms, delivery schedules, or final purchase prices. It also does not establish whether the reported increases are temporary, whether all Nvidia configurations are affected, or whether cloud companies will delay, resize, or redirect projects. Those facts will determine whether the development is mainly a windfall for memory suppliers or a broader test of AI infrastructure economics.
