Nvidia warns customers that AI server prices will rise more than 15%

Nvidia has told some of its largest customers that prices for servers equipped with its AI chips will rise more than 15% in many cases, according to a Bloomberg report published by the South China Morning Post. The price increase takes effect for systems that will ship in early next year.

What's going up, and why

The increase affects systems that use the company's top-tier chips, including the Vera Rubin and Grace Blackwell lines, according to people familiar with the process who spoke to Bloomberg on condition of anonymity because they were discussing communications not yet made public. The size of the increase varies by chip generation and memory configuration.

The companies that build the servers under contract for major data center operators, such as Microsoft, Google (Alphabet), and Oracle, have already notified their customers about the coming increases. Nvidia did not respond to requests for comment.

The root of the problem isn't the accelerator chip itself, but memory. Nvidia's processors are the heart of the machines that build and run AI software, but their effectiveness depends on how much DRAM (dynamic random-access memory) is paired with them. With demand for AI infrastructure soaring, the cost of memory chips has risen sharply.

Who holds the leverage now

The fact that the industry's most dominant company can't hold the line on prices or absorb rising costs shows how much bargaining power memory makers have gained. The leading DRAM suppliers are Samsung Electronics, SK Hynix, and Micron Technology, all currently benefiting from the race for AI infrastructure.

This isn't an isolated move. According to the same report, major tech companies such as Apple and Qualcomm have recently said they were forced to charge more for their products because of chip shortages. In other words, cost pressure from memory is spreading across the entire hardware chain, not just AI servers.

What changes for developers in Brazil

For Brazilian developers, the effect won't show up as a GPU invoice, but in the hourly price of compute. Most model training and inference here runs in the cloud (AWS, Google Cloud, Azure, Oracle) or through on-demand GPU providers. If the cost of the servers these providers buy rises more than 15%, it's reasonable to expect part of that to be passed on to GPU instance pricing throughout 2027, as the more expensive systems begin shipping.

This reinforces a barrier that already weighs on Brazil: computational cost as an entry obstacle. Startups and small teams that depend on A100, H100, or the new Blackwell accelerators to run AI workloads tend to feel the difference exactly at the most sensitive stage, product validation, when there's no revenue yet to offset infrastructure spending.

Some mitigation approaches are already part of the toolkit for people who build software, and deserve extra attention in this scenario:

• Smaller models and quantization: running quantized versions (INT8, INT4) reduces dependence on more expensive memory configurations, precisely the component driving the price increase.
• Optimized inference: batching techniques, KV caching, and efficient runtimes help extract more throughput from the same rented GPU.
• Advance contracts: reserved instances and long-term plans can lock in prices before the pass-through reaches on-demand pricing tables.
• Hardware alternatives: part of the inference workload can migrate to accelerators that don't depend on the same premium DRAM supply chain.

What remains open

Some points remain unanswered. Nvidia has not officially confirmed the figures or the exact date of the increases, and the report relies on anonymous sources. It's also unclear how much of the increase hyperscalers will absorb and how much will be passed on to the final price of cloud instances, or at what pace. Memory scarcity is the central bottleneck, so the behavior of Samsung, SK Hynix, and Micron over the coming quarters should determine whether this is a one-time spike or the beginning of a new cost baseline for all AI infrastructure.