Nebius raised the rent on AI compute again

01

WHAT ACTUALLY CHANGED

Reuters reported on September 17 that Nebius will increase selected pay-as-you-go prices from October 1. Selected Nvidia GPU rates will rise by 17 to 21 percent, some CPU-only instance rates by 25 percent, and some memory offerings by about 41 percent. These are category ranges reported from the provider's notice, not a claim that every Nebius product, region, customer, or contract receives the same increase.

The move follows another pricing change that took effect June 1. Nebius's public changelog says prices were updated then for Compute virtual machines and Standalone Applications using Nvidia B300, B200, H200, and H100 GPUs. The documentation does not turn those two events into a universal price index. It does establish that list pricing for important accelerator products has changed twice within a short planning window.

Nebius's current pricing page shows how the meter works. Running virtual machines are billed by the second and priced by the hour. Listed on-demand GPU rates include $7.85 for B300, $7.15 for B200, $4.50 for H200, $3.85 for H100, and $1.80 for RTX Pro 6000. The detailed documentation says applicable taxes are excluded, stopped virtual machines do not incur compute charges, and attached storage continues to be billed. Public prices can change, so these figures are a dated snapshot rather than a permanent tariff.

The provider offers another lane for customers willing to commit. Nebius says large-scale clusters reserved for multiple months can receive discounts of up to 35 percent from on-demand rates. That can make steady production work cheaper, but it also transfers forecasting risk to the customer. A team that reserves too little may still buy expensive flexible capacity. A team that reserves too much can pay for idle machines or remain attached to a platform after its workload changes.

Reuters connects the increases to strong demand for AI computing and reports that Nebius signed four customer contracts averaging more than $1 billion each during the quarter. Those contract and demand figures come from the company. Nebius does not publish customer-by-customer prices, utilization, margins, queue times, idle capacity, or the complete resource mix behind those agreements, so the figures cannot show what a typical startup or university actually pays.

02

WHY THIS MATTERS

Smaller buyers often need the flexibility being repriced. A frontier laboratory can negotiate a long cluster reservation and keep it busy. A startup testing product demand, a university waiting on a grant, or a research team running irregular experiments may not know whether it needs eight GPUs next month or none. Pay-as-you-go access protects that uncertainty, which is exactly why the provider can charge more for it.

The GPU-hour is a useful meter and a poor verdict. A faster accelerator may cost more per hour and still finish a fixed job for less money. It may also sit idle while data loads, software fails, a checkpoint writes, or the network waits. The honest comparison is cost to a defined result at a defined quality, with wall-clock time, failed attempts, storage, data movement, support, and human debugging included.

Commitment discounts can quietly become switching costs. A multi-month reservation reduces the visible rate, but a customer also chooses a region, accelerator family, orchestration stack, storage layout, network, monitoring system, and operating routine. Leaving means moving data, retesting numerical behavior, adapting automation, retraining staff, and surviving a period when two platforms may run at once. The discount should be compared with that exit bill.

Sovereign compute does not automatically mean affordable compute. Nebius is headquartered in Amsterdam and operates European infrastructure, including services in Finland and the United Kingdom. That can strengthen regional capacity, jurisdictional choice, and supply diversity. It does not guarantee stable prices, public-interest allocation, transparent margins, or access for organizations too small to sign a long reservation.

Public AI policy needs an affordability layer. Grants and national compute programs often count installed accelerators, investment, or peak capacity. Researchers and smaller companies experience a different reality: application windows, quotas, queue times, changing list prices, minimum commitments, and migration work. A credible European compute strategy should measure who receives usable hours, at what complete cost, with what notice when the terms change.

03

WHERE IT COULD HELP

• Keep a versioned price ledger for every accelerator, CPU, memory tier, storage class, region, reservation term, tax treatment, support plan, and effective date, then attach the exact sheet used to every budget rather than relying on a live webpage that can change
• Benchmark one representative workload on each viable platform and report time to target quality, successful and failed runs, utilization, storage, data transfer, engineering time, and total cost per completed job instead of comparing only the hourly GPU line
• Model three demand cases before accepting a commitment discount: the expected schedule, a demand collapse, and an urgent expansion, including unused reservation cost, on-demand overflow, cancellation rules, renewal price, and the time required to leave
• Build portability before the price changes by using reproducible environments, open data formats, portable checkpoints, infrastructure definitions, export tests, documented dependencies, and a small recurring recovery run on a second provider or owned cluster
• Set budgets and anomaly alerts around projects and completed jobs, tag every resource owner, stop abandoned machines automatically, include persistent storage after compute stops, and require a human review before long reservations or region-specific dependencies