Nebius GPU Prices Up 21% Oct. 1 — AI Compute Shortage
Update, Sept. 18
Update, Sept. 18: The market read the hike as a demand signal: NBIS shares spiked to $232.81 in the first half hour of trading Thursday before closing at $217.82, up 4% on the day, and neocloud peers moved on the read-through — IREN rose roughly 5–6% while CoreWeave slipped about 2%, weighed by its debt load and fixed-price contracts it cannot raise (TradingView/Seeking Alpha; FinanceFeeds, Sept 17). One detail worth naming: independent pricing analysis confirmed the H100/H200/B200 figures against Nebius's published price list — and found the steepest increase of all is not any accelerator but memory (+41%), with AMD EPYC Genoa CPU rates (+25%) also outpacing the GPUs (AI Stock Wire, Sept 17).
File photo: Carl Lender / Wikimedia Commons (CC BY 2.0). Not a Nebius facility.
If you rent Nvidia's most coveted chips by the hour, Thursday brought bad news: your compute bill is about to get bigger. Again.
Nebius, the Amsterdam-based neocloud that rents out Nvidia GPU capacity to AI companies, told customers it will raise pay-as-you-go prices for leasing selected Nvidia chips starting October 1, Reuters reported Thursday. It is the company's second price increase in three months — and the latest sign that the AI compute shortage is not easing. It is getting worse.
The move landed on the same morning rival CoreWeave said it is signing new short-term compute contracts at roughly $40 million per megawatt on an annualized basis, telling investors it has "continued to contract new compute capacity at higher prices." Two of the biggest GPU landlords in the business, within hours of each other, both telling the market the same thing: demand is outrunning supply, and the price of a flop keeps climbing.
Nebius GPU price hike: H100, H200, B200, B300 per-model rates
Here is what Nebius said, via the Reuters report:
Rates for selected Nvidia graphics processing units will rise between 17% and 21%. On-demand rates for some CPUs made by AMD and Intel, plus memory offerings, are going up too. CPU-only instance prices rise 25%; memory offerings jump about 41%. And customers reserving large-scale clusters for multiple months can get commitment discounts — the reserved-capacity escape hatch for anyone willing to lock in.
Screenshots of customer pricing communications circulating on X and Reddit, reported by Seeking Alpha and Stocktwits, put per-model figures on the hike: the H100 moving from $3.85 to $4.50 per GPU-hour (about 16.9%), the H200 from $4.50 to $5.40 (20%), the B200 from $7.15 to $8.50 (about 18.9%), and the B300 from $7.85 to $9.50 (roughly 21%). Nebius has not publicly confirmed those per-model numbers, so treat them as reported rather than official — but they line up neatly with the 17–21% range Reuters attributed to the company. The same reports say AMD EPYC Genoa CPU rates rise from $0.012 to $0.015 per vCPU-hour (25%) and Genoa memory from $0.0032 to $0.0045 per GiB-hour (about 41%), again matching the company's stated figures.
The context that matters: this is a second hike, not a first. In May, Nebius notified customers of an average 29% increase on on-demand capacity and a 51% jump on preemptible capacity, effective June 1, according to customer emails shared at the time and reported by Stocktwits. Take the B300 as the example: its hourly rate has climbed from roughly $6.10 before the May increase to $9.50 after this one, per market reporting — about 56% in four months.
A 56% price increase in four months, on the single most important input to the AI economy. Let that sit for a moment.
It's not just Nebius — the whole cohort is raising
Nebius is not moving alone, which is precisely why this matters. When one landlord raises the rent, it might be margin repair. When the whole street raises it at once, it is scarcity.
CoreWeave's Thursday announcement — made in the same breath as its plan to raise $3 billion in convertible debt plus an at-the-market program for up to 35 million shares — disclosed that its new three-to-six-month contracts are pricing at approximately $40 million per megawatt, measured as annualized revenue divided by the power needed to run the clusters. The company said it has expanded contracted power to about 4.2 gigawatts as of August 11, up from 3.7 GW at the end of June, and added more than $25 billion in net new customer commitments early in the third quarter. Barron's noted CoreWeave had already raised its own prices by roughly 25% back in July. (We covered CoreWeave's financing blitz — the convertible offering and share program that together could raise up to about $6.4 billion — in detail earlier today.)
Nvidia itself keeps confirming the trend from the other side of the trade. CFO Colette Kress told investors on the company's earnings call that H100 rental rates are up 20% year-to-date, with A100 rates up nearly 15%, and that customers are "generating profitable revenue beyond the depreciable life of their GPUs."
And the demand visibility keeps stretching further out. Nebius management has said some customers have already booked compute for the first and second quarters of 2028, with demand visibility exceeding 24 months, and that a capacity auction for scarce Blackwell chips cleared 15% to 20% above previous peak prices, per market reporting. This is not a spot-market blip. It is a forward curve.
File photo: Rsparks3 / Wikimedia Commons (CC0, public domain). Not a Nebius facility.
Why pay-as-you-go prices are the market's loudest signal
Here is the analytical frame worth understanding, because it will keep mattering: on-demand GPU pricing is becoming the spot market for the AI economy — the closest thing this industry has to a barrel-of-oil price.
Most neocloud revenue is locked into long, large contracts with their own negotiated terms. Barron's made the important caution explicit: investors should not assume an on-demand price hike flows straight into reported revenue, because much of it is tied to multiyear deals priced separately. Fair. But on-demand is the marginal price — the price of the next unit of compute, paid by whoever needs capacity right now and cannot wait. That is exactly why it deserves attention: marginal prices move first, and they reveal where supply and demand actually clear.
The second detail that matters is the discount structure. Nebius is raising the rack rate while explicitly offering commitment discounts to customers who reserve large clusters for multiple months. That is a term structure — cheap if you commit, punishing if you stay flexible — and it tells you the company's real objective: convert today's scarcity pricing into tomorrow's contracted revenue. CoreWeave is playing the same game from the other direction, preferring three-to-six-month contracts to capture rising short-term rates rather than locking in long ones. Everybody is trying to surf the same wave; they just disagree about which way it breaks.
What this does — and doesn't — say about the AI bubble debate
Few data points cut as directly into the "AI bubble" argument as this one.
The sharpest version of the skeptical case, associated with "Big Short" investor Michael Burry, holds that cloud companies are overstating the long-term value of Nvidia chips by depreciating them over six years — systematically understating their true annual cost. If that were right, you would expect rental rates on older silicon to decay as newer chips arrive. Instead, Nebius is raising the rental price of the H100 — a chip Nvidia first released in 2022 — by roughly 17%, alongside the newest Blackwell parts. Pricing power that persists across generations is the strongest available rebuttal to the depreciation thesis, and it is exactly what Nvidia's Kress was getting at: customers are earning profitable returns on these chips beyond their accounting lives.
That said, honesty requires the other side of the ledger. Rising rental prices prove demand exceeds supply today; they do not prove the demand is rational, or that every dollar of it converts into durable AI revenue. Neoclouds are simultaneously the biggest beneficiaries of the shortage and the most leveraged bet on it continuing — CoreWeave is raising billions in debt to build the very capacity whose scarcity justifies its pricing, and Nebius's own stock, for all its 150%-plus gain this year, has slumped 27% in the past three months, a reminder that the market is far from settled on the story. Pricing power is evidence, not proof.
Who feels it first
Startups feel it first, because startups live on pay-as-you-go. A 29% hike in June followed by another 17–21% in October compounds into a meaningfully higher burn rate for any team training or serving models on rented GPUs — and the 41% jump in memory pricing lands hardest on data-heavy workloads like retrieval-augmented generation and long-context inference, where memory is the binding constraint.
The incentive design is doing quiet work here, too. Every on-demand increase widens the gap between flexible and committed pricing, nudging customers toward multi-month reservations — which is to say, toward lock-in. That is rational for the neoclouds, and it is worth naming plainly: the shortage is becoming a customer-retention mechanism.
Hyperscalers, meanwhile, get the opposite read. Every neocloud price hike is a data point validating their own enormous capital expenditures — if rented H100s are getting 17% more expensive, owning the silicon looks smarter by the month. Expect the next round of hyperscaler earnings calls to cite exactly this.
Nebius GPU price hike FAQ
When do the new rates take effect? — October 1, 2026.
How much does an H100 cost now? — $4.50/GPU-hour, up from $3.85 (reported per-model figures, not company-confirmed).
What rose most? — Memory (+41%) and AMD EPYC Genoa CPUs (+25%), steeper than any accelerator.
Is there a way to avoid the hike? — Commitment discounts for multi-month large-cluster reservations.
The financing wave behind the shortage
None of this is happening in a vacuum. The price hikes are the revenue-side echo of a capital-expenditure boom that has been building all year: CoreWeave's Thursday double-tap of debt and equity; the reported $22 billion chip loan backing Google and Blackstone's Crux AI; Crusoe's reported $3.9 billion Series F at a $30.9 billion valuation; and Nebius itself, which posted second-quarter revenue of $582 million, up 454% year over year, signed four customer contracts averaging more than $1 billion each according to Reuters, and broke ground in May on its first gigawatt-scale AI factory campus in Independence, Missouri.
Put the two sides together and the picture is almost too tidy: Wall Street is lending tens of billions to build AI data centers, and the companies renting out the chips inside them keep raising prices because there still is not enough to go around. That loop — capital chasing scarcity, scarcity justifying capital — is the defining dynamic of the AI infrastructure market in 2026. It holds until it doesn't, and the on-demand price is the gauge to watch.
The new rates take effect October 1. The thing to watch next is whether the rest of the neocloud cohort matches the move — and what the hyperscalers say about capex on their next earnings calls. If compute is becoming a commodity, its price just printed a new high.
Sources: Reuters' September 17, 2026 coverage of the Nebius announcement; CoreWeave's official statement via Business Wire; Dow Jones/Morningstar and Barron's coverage of both companies' Thursday announcements; market reporting from Seeking Alpha and Stocktwits on per-model pricing; Nebius Q2 2026 earnings reporting.
Comments
Post a Comment