Nebius Raises Prices by 20% for the Second Time This Year, Signaling a Shift in Pricing Power to GPU Cloud Supply Side
nashnova research
GPU cloud provider Nebius will raise prices across all GPU tiers by roughly 20% from October 1 — the second hike this year — sending shares up nearly 7% after hours and signaling that AI compute demand still far outstrips supply.
How big is this price hike?
All four flagship chips are going up: H100 rises from $3.85 to $4.50 per GPU-hour, up ~16.9%; H200 from $4.50 to $5.40, up 20%.
Newer Blackwell chips see steeper increases: B200 from $7.15 to $8.50, up ~18.9%; B300 from $7.85 to $9.50, up ~21%.
This means → both legacy and cutting-edge architectures are repricing together — the bottleneck is not one chip model but the entire GPU compute pool.
Two rounds in six months — what is the cumulative damage?
In May, Nebius already raised on-demand capacity prices by an average of 29% and preemptible capacity by 51%.
Take the B300: before May it sat around $6.10/hour; now it is $9.50 — a cumulative jump of roughly 56%.
In plain terms = the same card's rental cost has risen more than half in six months, and customers are still lining up to pay.
Customers are willing to bid above list price?
Nebius ran an auction test for scarce Blackwell capacity. Customers paid 15% to 20% above the highest posted price just to secure access.
Management disclosed that clients have booked compute slots into Q1 and Q2 of 2028, with some orders spanning tens of thousands of GPUs. Demand visibility has stretched from roughly eighteen months to over two years.
This reflects a market where the question is no longer "can we negotiate a discount" but "can we get a seat at all."
What does $60 million per megawatt mean?
Analyst Jonah Lupton estimates that Nebius's short-term contract pricing may already exceed $60 million per megawatt.
Future contracts for Vera Rubin architecture — Nvidia's next-generation GPU platform — could top $80 million per megawatt.
This means → compute is no longer priced by "how many cards you rent" but by power capacity — whoever can secure electricity, cooling, and rack space secures the compute.
How did pricing power flip from buyer to seller?
Previously, hyperscale cloud operators dictated terms: near-100% uptime requirements, ultra-strict temperature and humidity standards. One rack going down to a power cut, overheating, or switch failure could cancel six months of rent; accumulated SLA breaches could terminate the entire lease.
Now the tables are turning. One reported contract requires a tenant leasing only a fraction of a large facility to cover the full rent for the entire building if a single payment is late. The landlord's response: "We know it's outrageous… but we can."
In plain terms = it used to be "the customer writes the rules." Now data-center owners set the terms — whoever controls the scarce resource is the one with leverage.
Which stocks are riding the ripple effect?
Nebius jumped nearly 7% after hours; its year-to-date gain is now about 150%. CoreWeave added nearly 4%.
Optical-connectivity names rallied broadly: AAOI and Credo Technology up ~3%, Marvell and Astera Labs up over 2%, Coherent up ~2%.
This means → the market reads Nebius's repricing as a tailwind for the entire AI infrastructure chain — from GPU cards to optical modules to data-center operators, the supply side benefits across the board.
市场有风险，内容仅供研究参考，不构成投资建议。