Nvidia to raise AI server prices by more than 15%, Bloomberg says

Nvidia has notified some of its largest customers that prices for servers with its AI chips will rise by more than 15% in many cases, according to a Bloomberg report published on Saturday (23rd) and picked up by Reuters. The main trigger, according to the report, is the surge in the cost of memory chips.

The increases take effect for systems shipped starting early next year and affect, among others, models with the flagship Vera Rubin and Grace Blackwell chips. According to people familiar with the matter cited by Bloomberg, the exact percentage will vary depending on the chip generation and each system's memory configuration.

What the source says, and what it doesn't confirm

It's important to separate what's been reported from what's still open. Reuters said it could not independently verify Bloomberg's report, and Nvidia did not immediately respond to a request for comment made outside business hours.

Another relevant point: the price adjustments were communicated to customers by the companies that build servers under contract (the so-called integrators/ODMs) for large data center operators such as Microsoft, Google (Alphabet), and Oracle. In other words, the increase reaches hyperscalers through the assembly chain, not necessarily as a price list published directly by Nvidia.

The report also doesn't detail whether and how each cloud provider will pass the cost on to the final price of GPU instances. That remains undefined.

Why memory became the villain

The angle pointed to as the cause is the cost of memory chips. Modern AI servers rely on large amounts of HBM (High Bandwidth Memory), the stacked memory attached to GPUs, in addition to conventional DRAM. With the data center construction race driving up demand for these components, memory prices have been pushing up the total cost of each rack, and it's this pass-through that is now showing up in server prices.

Since the adjustment varies by memory configuration, beefier systems, precisely the ones used for training large models, tend to feel the biggest percentage impact.

What changes for those building AI in Brazil

Most Brazilian startups and engineering teams don't buy Nvidia servers directly: they consume GPU capacity via the cloud, whether from major global providers or from national and regional providers that resell or operate this hardware. That's precisely the link.

If the acquisition cost of Vera Rubin and Grace Blackwell servers rises by more than 15% for data center operators, there are two possible paths for the market, and both affect Brazilian developers:

• Pass-through in the GPU hourly rate. Instances based on the new generation may arrive more expensive than the downward curve the market usually expects with each hardware leap. For those doing fine-tuning, training, or heavy LLM inference, this goes straight into the infrastructure bill.
• Dollar-denominated prices. AI cloud costs are, in practice, dollarized. An increase at the source adds to the exchange rate effect, amplifying the impact on the budget of those who bill in reais.

In practice, the cost-per-token or cost-per-training-experiment calculation that many teams use to decide on architecture may need to be redone, factoring in a higher hardware floor for the next generation.

Some mitigation approaches that are already part of the playbook for cost-conscious teams, and which become even more relevant in this scenario:

• Quantization and smaller models to reduce dependence on top-of-the-line instances for inference.
• Capacity reservation (reserved instances/committed use) to lock in pricing ahead of possible adjustments, when volume justifies it.
• Taking advantage of previous GPU generations, which tend to become relatively more attractive in terms of cost-benefit as the new line comes in more expensive.
• Evaluating alternative hardware for specific workloads, depending on what the provider offers.

The context: Nvidia as a barometer for the entire sector

The timing of the leak is not trivial. Nvidia releases its second-quarter results on August 26, just days after the report. The company has become a kind of proxy for the AI ecosystem as a whole, spanning everything from chipmakers to the companies financing the accelerated expansion of data center capacity, which is why any signal about pricing reverberates beyond it.

One possible reading, within what the source provides: the memory cost pass-through suggests that the phase of "each generation is more powerful and cheaper per FLOP" may hit friction in the short term, with scarce components pushing prices up right when demand for model training shows no sign of slowing down.

For the Brazilian market, the practical takeaway is to watch how cloud providers will price the new generation's instances when they arrive in 2027, and not to assume the next chip will automatically be cheaper. Until then, what exists is the notification reported by Bloomberg, without official confirmation from Nvidia.