Servers built around NVIDIA AI accelerators may become more expensive for deliveries starting in early 2027. On August 22, Bloomberg reported that some large customers had been told that prices for the affected systems would rise, in many cases by more than 15%. The systems include the Grace Blackwell and Vera Rubin families, and the notices were reportedly relayed by contract server manufacturers that also supply Microsoft, Google, and Oracle.
NVIDIA has not made this customer communication public. This is therefore not an official NVIDIA list-price revision, but a report about unofficial quotes and procurement prices. Even so, in AI servers memory is spread across two tiers, next to the GPU and next to the CPU, and capacity grows with each generation. That makes the report a useful prompt for asking when component supply constraints show up in the purchase price of a rack.
What the "15%+" notice covers
Bloomberg said the size of the increase varies by NVIDIA chip generation and memory configuration. It did not say that one rate applies to every customer and every SKU. Because the notices were reportedly passed along through manufacturers, it cannot be read as NVIDIA uniformly changing the price it charges every hyperscaler directly.
Separately, The Information, citing two people at companies that received the notices, reported that some Grace Blackwell 300 systems shipping next year, and a flagship system the outlet calls Vera Rubin 200, would rise by about 17%. Bloomberg's "more than 15% in many cases" and The Information's "about 17% for some flagship systems" differ in both scope and wording. They should not be merged into a single claim of a 15–17% increase across all products.
The reports concern the customer purchase price of server systems that incorporate AI chips, not the published list price of standalone GPUs. Memory is cited as a background factor, but The Information also mentioned components other than memory. There is no basis for attributing the entire increase to HBM alone.
Two memory tiers behind the price differences
NVIDIA's GB300 NVL72 combines 72 Blackwell Ultra GPUs with 36 Grace CPUs. In the official configuration, it has 20TB of HBM3E on the GPU side and 17TB of LPDDR5X on the CPU side, for a total of 37TB of fast memory. Memory pricing and supply are not determined by HBM, which sits close to the GPU, alone.
The Vera Rubin NVL72, which NVIDIA publishes with preliminary specifications, pairs 72 Rubin GPUs and 36 Vera CPUs with 20.7TB of HBM4 and 54TB of LPDDR5X. The total is 74.7TB, roughly 2.0 times the GB300 NVL72's 37TB. This compares official reference/maximum configurations and does not represent the standard configuration every customer will receive. NVIDIA also says the Vera Rubin figures are subject to change.
| Official reference configuration | GPU-side memory | CPU-side memory | Total fast memory |
|---|---|---|---|
| GB300 NVL72 | HBM3E 20TB | LPDDR5X 17TB | 37TB |
| Vera Rubin NVL72 (preliminary spec) | HBM4 20.7TB | LPDDR5X 54TB | 74.7TB |
The gap means next-generation racks are exposed to supply conditions for both HBM4 and LPDDR5X. It also helps explain, from a specifications standpoint, why reports say the size of the increase varies by configuration. However, roughly double the capacity does not mean double the rate of price increase. Contract terms and unit component prices matter, and so do non-memory server components in the quote.
Component costs have moved into rack quotes
NVIDIA's 10-Q for the quarter ended April 26, 2026 states that cost of revenue includes the cost of memory and other components. The company says it does not normally adjust product and solution prices for short-term cost fluctuations, but has raised prices on some products in response to supplier increases and may do so again. It also explains that reduced availability or higher prices for components such as memory can raise data center build costs, and that a supply constraint on a single component can prevent sales of complete systems.
There is also evidence of tightness on the semiconductor memory side. In its April–June 2026 results, Samsung Electronics cited limited production capacity and industry-wide price increases, and said supply shortages would continue in the second half amid rising demand for server DRAM, enterprise SSDs, and HBM. This does not corroborate the notices attributed to NVIDIA. Still, Samsung does not expect the memory market to ease in the second half of 2026.
The path by which prices reach customers also has several links: component makers, NVIDIA's platform, the manufacturers who assemble the bill of materials, and finally cloud and data center buyers. What Bloomberg and The Information described are notices to customers through server manufacturers. Memory-maker market conditions alone do not determine how much the final quote rises.
In the same 10-Q, NVIDIA disclosed $119 billion in manufacturing, supply, and capacity commitments as of April 26, 2026, of which $95 billion falls due during the rest of fiscal 2027. This is neither the amount of the reported price increase nor the cause of it. It indicates the scale of long-term component procurement the company must secure.
In that quarter, NVIDIA recorded revenue of $81.615 billion, data center revenue of $75.246 billion, and a GAAP gross margin of 74.9%. These are past quarterly results, however, and do not directly indicate pricing power or gross margin on systems delivered in early 2027. Current profitability cannot be used to estimate configuration-specific prices for next-generation systems.
A 15% higher price is a separate question from cost per token
If purchase prices rise as reported, capital spending and margins at cloud providers could be affected. But rental rates, training costs, and per-token inference costs will not necessarily rise by the same 15% or 17%. Utilization, power, networking, and contract terms all matter, as does the performance actually delivered.
Based on current system prices, The Information estimated that chip system costs for a 1GW data center could increase by at least $5 billion. It also calculated a Vera Rubin rack with 72 GPUs at roughly $8 million. Neither figure is an official NVIDIA price, and neither is a universal estimate. Even at the same power capacity, required spending will differ depending on the racks, configurations, and procurement contracts adopted.
NVIDIA claims the Rubin platform can cut inference token costs to as little as one-tenth of Blackwell's. That is a future performance and cost claim by NVIDIA, and it has not been confirmed to offset the reported 15%-plus price increase. Even so, it shows that a rack's purchase price and the operating cost per useful token should be judged on separate yardsticks.
NVIDIA is scheduled to announce FY2027 Q2 results and hold a conference call on August 26. Any remarks on component costs and supply conditions, or on the impact on gross margin and pricing, could help gauge the scope of these reports. However, that quarter precedes early-2027 deliveries, and actual results alone cannot confirm the contents of the notices or configuration-specific quoted prices. Even if the 15%-plus figure is confirmed, without knowing which rack configurations and delivery terms it applies to, it is not possible to judge the unit cost of AI compute.