Nebius GPU prices rise across the board, up to 21%; H100 cumulative gain of 53% over five months
A notable shift is emerging in the AI computing power market: while cloud providers continue to expand their data centers, rental prices for certain high-end GPUs are not declining with increased supply—they continue to rise.
According to a price adjustment notice sent by Nebius to its customers, effective October 1, 2026, the company plans to increase prices for multiple on-demand GPU instances by approximately 17% to 21%. The proposed hourly rates per GPU are: H100 rising from $3.85 to $4.50 (a 16.9% increase), H200 from $4.50 to $5.40 (20% increase), B200 from $7.15 to $8.50 (18.9% increase), and B300 from $7.85 to $9.50 (21% increase).
This price adjustment sends a significant signal: the pace of growth in AI computing demand may still outstrip the rate at which high-end GPUs, power, and data center capacity are coming online. Unlike traditional cloud computing, the core scarce resources for AI cloud services extend beyond servers to include GPU availability, power access, liquid cooling systems, networking equipment, and data centers capable of supporting high-density racks. If any single component encounters a bottleneck, new computing power cannot be rapidly deployed.
At the Goldman Sachs Technology Conference, Nebius management stated that the company's visibility into customer demand has extended from approximately 18 months to over 24 months, with some customers already reserving compute capacity for the first half of 2028. The company previously conducted price testing for scarce Blackwell capacity, achieving transaction prices approximately 15% higher than previous peak prices and roughly 20% higher than standard sales pipeline quotes.
The significance of Nebius's price increase lies not merely in the ability to charge approximately 20% more per GPU hour, but in confirming that supply and demand in the AI computing market still favor suppliers. A primary past market concern was that simultaneous launches of numerous data centers could eventually lead to GPU leasing market oversupply, potentially resulting in lower utilization rates, extended payback periods, and asset impairment for companies built under assumptions of higher prices. Nebius signals the opposite: the company is raising prices not only for new B200 and B300 models but also for H100 and H200, which have been on the market for years. This indicates that demand extends beyond the latest architecture, with previous-generation GPUs still offering sufficient performance and cost efficiency for model fine-tuning, inference, scientific computing, and certain enterprise workloads.
If implemented successfully without significant declines in GPU utilization, Nebius will achieve two improvements: higher revenue per unit of computing power and potentially higher economic lifespan and residual value for existing GPU assets than previously anticipated.
However, the on-demand price increase will not uniformly raise all of Nebius's revenue by 20%. Large customers typically sign multi-year contracts with fixed prices, minimum purchase volumes, and service terms that cannot be immediately repriced according to new public rates. This adjustment most directly impacts on-demand and short-term compute contracts rather than long-term capacity agreements already signed by clients like Meta.
The price increase may also involve cost pass-through, as NVIDIA systems, networking equipment, power, and data center construction costs may also rise. Therefore, higher capital returns will materialize only if Nebius's unit revenue growth exceeds its unit investment and operating cost growth.
Nebius's primary advantage is that orders, demand, and financing capabilities have formed a positive feedback loop. The company has signed contracts with Meta worth up to $27 billion, received a $2 billion investment from NVIDIA, and secured four large deals in the second quarter, each averaging over $1 billion. Customer prepayments can cover approximately 50% to 60% of capital expenditures for related projects, reducing reliance on balance sheet expansion alone.
Yet this remains an extremely capital-intensive business. Nebius expects capital expenditures of $20 billion to $25 billion in 2026, far exceeding its revenue guidance of $3 billion to $3.4 billion. The company completed a $5.75 billion convertible bond financing in August. Even with continued price increases, risks including data center construction timelines, debt costs, potential equity dilution, and customer concentration remain significant investor concerns.
Three indicators warrant close observation. First, whether GPU utilization remains high after the price increase—sustained acceptance signals that computing power supply bottlenecks genuinely persist. Second, whether short-term high-price contracts convert into sustainable revenue; on-demand computing carries lower visibility than multi-year contracts and may see rapid price declines if demand cools. Third, how much revenue and cash flow can be generated per unit of capital invested, as the core competition among AI cloud companies centers not on GPU ownership but on achieving high utilization rates and recovering investments before equipment upgrades.
Nebius's price increase indicates that the AI infrastructure industry has not yet entered a price war. What remains truly scarce is high-end computing power that delivers on schedule, operates stably, and provides comprehensive software services. How long pricing power persists will depend on whether demand growth continues to outpace GPU, electricity, and data center expansion. For Nebius, the increase demonstrates sufficient current demand; the next phase is proving this strong demand converts into sustainable free cash flow and shareholder returns.