Nvidia rental prices are rising even for last-generation H100s. Nebius has posted an October 1, 2026 on-demand card that lifts selected Nvidia GPU-hour rates by 16.9% to 21.0%. CoreWeave said on September 17 that it is still signing compute at higher prices, including short contracts at about $40 million per megawatt of annualized revenue. A landlord can charge more for a chip it does not make only while that chip stays scarce. That is the signal behind Nvidia, and it is the opposite of a capacity glut.
Nebius’s price page, checked on October 2, still prints two columns. On-demand H100 is $3.85 per GPU-hour. The column marked effective October 1 is $4.50. The same split runs through the H200, the B200, and the B300. Reuters reported on September 17 that this was the second pay-as-you-go increase in about three months, with selected Nvidia GPU rates up 17% to 21%. The card on nebius.com/prices matches that range. It does not show every customer already paying the new rate.
The card Nebius posted
Four Nvidia rows have a public before and after. GB200 NVL72 and GB300 NVL72 racks are still “contact us” in both columns. RTX PRO 6000 stays at $1.80. L40S “from” rates do not move. The hike lands on the accelerators customers are queued for.
| Nvidia GPU | Prior on-demand column | Announced October 1 card | Change |
|---|---|---|---|
| HGX H100 | $3.85 | $4.50 | +$0.65, or 16.9% |
| HGX H200 | $4.50 | $5.40 | +$0.90, or 20.0% |
| HGX B200 | $7.15 | $8.50 | +$1.35, or 18.9% |
| HGX B300 | $7.85 | $9.50 | +$1.65, or 21.0% |
Reuters’ “17% to 21%” is this range, with 16.9% rounded up. “Up to 21%” is the B300. The smaller H100 move is the one that changes the reading. Hopper is not the generation Nvidia is leading with. A prior-generation accelerator getting more expensive to rent is not what surplus hardware looks like.
Put the H100 on a bill. Eight GPUs go from $30.80 an hour to $36.00. Over a 720-hour month the list charge goes from $22,176 to $25,920. An 8-GPU B300 box moves from $45,216 to $54,720. Those are walk-up numbers, before tax, storage, or a commitment discount, and they are not invoices.
CPU floors moved too. AMD EPYC Genoa goes from $0.10 to $0.13 an hour, 30% on that “from” rate. Intel Ice Lake goes from $0.05 to $0.06, 20%. Reuters described some CPU-only instances as up about 25% and some memory offerings about 41%. Neither figure is a line on the public card. There is no memory before-and-after. Storage prices are single figures, starting at $0.08 per GiB-month for shared filesystem.
The announced column is not yet the only column
October 2 is after the effective date, and the page still shows both numbers. The left-hand label is still “On-demand, GPU-hour.” The right-hand label is “GPU-hour (Effective October 1, 2026).” A customer cannot tell, from the page alone, which figure the console will bill.
GPU Finder’s Nebius page, with availability checked at 06:23 UTC on October 2, still listed one-GPU on-demand rates of $3.85 for the H100, $7.15 for the B200, and $7.85 for the B300. Those are the old column. The same check already showed the H200 at $5.40, the new column, with a spot print of $2.45 beside it. The H100 spot print on that row was $2.15. The tracker has not frozen. It has moved one of the four headline GPUs and left the other three on the pre-October number. Treating the whole book as repriced at midnight on October 1 would be tidier, and it would be wrong.
Commitment discounts can run up to 35% below on-demand for large clusters reserved for multiple months. Full discount on the new H100 card is $2.93 an hour, under the old $3.85 list. On the new B300 card it is $6.18, under $7.85. “Up to” is not “everyone gets.” The headline increase is the unreserved rate.
Spot is a third price. Preemptible rates can change every 15 minutes, and the ceiling sits one cent under on-demand, but the printed floor is $0.79 on H100 and H200 and $0.99 on B200 and B300. An interruptible job is not paying $4.50. The October card is about firm capacity. Most revenue is priced as annual contract value per megawatt, not off this page.
CoreWeave priced the same shortage in megawatts
CoreWeave did not publish an H100 card. On September 17 it published a contract print. The release, carried on Nasdaq, says that since June 30, 2026, CoreWeave has continued to contract new compute at higher prices. In the third quarter it has signed short-dated contracts of about three to six months, “with pricing of approximately $40 million per megawatt, calculated as annualized revenue divided by power required to service the related clusters.”
The same paragraph points back to the August 11 earnings call. CoreWeave said it added more than $25 billion of net new customer commitments early in the third quarter, not yet in the revenue backlog as of June 30, and that total contracted power rose to approximately 4.2 gigawatts as of August 11 from approximately 3.7 gigawatts as of June 30. That 4.2 gigawatt figure is a level already reached by August 11, not a forecast that power was heading toward 4.2. The release does not quote a named executive. There is no sentence from Michael Intrator in it.
FinanceFeeds has already separated that $40 million print from what the older fleet earns, in CoreWeave’s $40 million per megawatt is not what the fleet earns. The units do not match. Nebius is quoting a GPU-hour. CoreWeave is quoting annualized revenue per megawatt. Dividing $40 million by 8,760 hours gives about $4,566 per megawatt-hour, which is still not a GPU-hour. Neither company publishes how many accelerators sit in a megawatt after networking and cooling, so a CoreWeave “H100 equivalent” hour derived from this release would be an invented number.
Nebius has been using the megawatt language since August 12, before the October list change. In the shareholder letter, founder and chief executive Arkady Volozh wrote that second-quarter deals yielded more than $20 million per megawatt on average, and that four landmark AI-cloud contracts averaged more than $1 billion of total contract value each, at $20 million to $25 million per megawatt. He also wrote: “We see a price opportunity in the $40-50 million per MW range and we signed our first one this week.” CoreWeave’s third-quarter short deals, at about $40 million, sit on the bottom edge of the band Nebius said it was already clearing on short-term capacity in August.
Volozh also wrote that core deals were “driven by increasing prices for new-generation GPUs and more than 30% higher pricing on older-generation GPUs versus Q1.” That 30% is a contract comparison with the first quarter, not the 16.9% move from $3.85 to $4.50. Adding the two percentages would double-count. Signed deals had already repriced older GPUs, and the walk-up card moved again on October 1.
Two sentences from the letter are the company’s own words, not a paraphrase. “We closed our largest AI Cloud deals on our strongest terms to date, at prices that represent a step-change in the economics of our business.” And, a few lines later: “We could sell our entire 2027 capacity on these terms today.” “These terms” refers to the block he had just described, the $20 million-plus deals and the $40 million to $50 million short-term opportunity. It is not a disclosure that every 2027 megawatt is pre-sold at $40 million. He says Nebius is deliberately not selling the whole book forward, because some capacity is worth more kept open. The text is the Q2 2026 shareholder letter.
Amazon’s reserved hour is the expensive one
The third public card is Amazon Web Services capacity blocks. They are reservations, not on-demand. Checked on October 2, the AWS tables price an eight-GPU instance and, in parentheses, the rate per accelerator.
In US East the reserved accelerator rates checked on October 2 are $5.191 for an H100 (p5.48xlarge, eight GPUs at $41.528 an instance-hour, Northern Virginia and Ohio), $5.97 for an H200 on p5e in Ohio, $6.865 for an H200 on p5en in Northern Virginia, $12.355 for a B200, and $14.04 for a B300. Tokyo’s p5 H100 rate is $4.720, under Nebius’s new card, so the geography matters. Against the US East reservation, Nebius’s October 1 card is still cheaper: $4.50 versus $5.191, $5.40 versus $5.97 or $6.865, $8.50 versus $12.355, and $9.50 versus $14.04. CoreWeave published no GPU-hour, so it cannot be ranked here. The table is AWS capacity-block pricing.
Effective October 7, AWS says capacity-block rates per accelerator across available regions will be $5.970 on P5, $6.866 on P5e, $7.895 on P5en, $14.208 on P6-B200, and $16.146 on P6-B300, higher in GovCloud. On the US East H100 line, $5.191 to $5.970 is 15.0% on that reservation rate only. The same notice leaves on-demand and Savings Plan prices unchanged.
The products are not the same hour. Nebius on-demand can be started and stopped if the GPUs are free. An AWS capacity block is reserved in advance and paid up front at the rate in force when it is bought. Even with that difference, the hyperscaler’s reserved H100 in US East sits above Nebius’s new list, and Amazon has scheduled the reserved rate higher on October 7 while leaving on-demand unchanged.
The price is rising while the buildout is, too
A glut runs the other way. The owner of empty halls cuts the rate to fill them. Nebius is adding halls and raising the walk-up rate in the same season.
In the second quarter, purchases of property and equipment and intangible assets were $5.657 billion, against group revenue of $582.3 million. That is 9.7 times the quarter’s revenue going into equipment and buildings. The letter rounds capital expenditure to about $5.7 billion, driven by GPUs and data-center expansion. First-half purchases were $8.130 billion. AI-cloud revenue was $574.9 million, up 514% from a year earlier. Group revenue was up 454%. Annualized run-rate revenue was $3.0 billion at June 30, up from $1.9 billion at March 31. AI-cloud adjusted EBITDA was $285.7 million, a 49.7% margin.
Volozh wrote that Nebius expects to end 2026 with 5 gigawatts of contracted power, up from the more-than-4-gigawatt figure given the prior quarter, and to deploy more than 1 gigawatt a year from 2027. CoreWeave’s 4.2 gigawatts is that company’s own definition of contracted power, not the same audited megawatt. Both figures are being paired with higher prices, not with discounting.
Customers are paying for a large piece of the build, which is the test of whether the higher price is collectible. The letter says Nebius expects more than $9 billion of customer prepayments in 2026, that about 70% of second-quarter deals included prepayments, and that those prepayments covered 50% to 60% of the associated capex. Payback on those deals shortened to 1 year and 10 months, on the company’s method, from a historical two to three years. Around that quarter Nebius also raised a $775 million secured facility at SOFR plus 2.50%, sold about $2.8 billion of stock through June 30, and in August closed roughly $5.75 billion of convertible notes. The capital is priced on the bet that the GPUs keep renting.
On the 271,855,218 shares outstanding at June 30, the October 1 close of $232.28 is about $63.2 billion of equity. That count is stale if more stock was sold later, and it is large next to a $3.0 billion run rate. The list hike does not earn the valuation. Most revenue is not the on-demand card. The card tests the August claim that 2027 capacity could be sold on strong terms and that Nebius is choosing to hold some back. A second Nebius price file should stay separate. Token Factory is not a rented GPU, and FinanceFeeds covered that move in Nebius raised inference prices 18.3%, not cut them. Adding 18.3% to 21% is not a company-wide increase.
What a dearer H100 says about Nvidia
Nvidia closed on October 1 at $230.86. Nasdaq’s premarket indication around 5:20 a.m. Eastern time on October 2 was $233.43, up $2.57, or 1.1%. The 52-week high on that Nasdaq page is $236.54, so Thursday’s close sat 2.4% under the high. Nebius closed at $232.28, 22.5% under its 52-week high of $299.86, and was indicated at $236.23 in the same window, up $3.95, or 1.7%. CoreWeave closed at $88.57, 42.2% under its 52-week high of $153.20, and was indicated at $89.42, up $0.85, or 1.0%. This is not a forecast or a recommendation to buy or sell NBIS, CRWV, or NVDA. The prints will move again before the opening auction. The figures are Nasdaq’s record of the October 1 close and of that premarket window.
The tapes diverge on how much of the shortage is already in the price. Nvidia is near its 52-week high. CoreWeave, with short deals near $40 million a megawatt, is still more than 40% off its high, and Nebius is about 23% off its own. The rental power has not been a straight line into the neocloud stocks. It is a straight line into the case that the chip vendor still owns the bottleneck.
Nebius and CoreWeave rent H100s. They do not make them. A higher hour in October 2026 means Hopper’s secondary market has not cleared, so a buyer walking away from Blackwell is not being offered a cheap older card. That sits downstream of Nvidia’s server pricing, which FinanceFeeds covered in Nvidia is raising AI server prices more than 15%. Memory was the August cost story. October’s rental story is demand.
Nvidia’s second-quarter revenue was $96.2 billion, and the stock traded the margin guide rather than the top line. A $4.50 hour does not move a quarter that size. It says whether the next order still comes from customers who are short of capacity. Volozh on 2027, and CoreWeave near $40 million a megawatt, are a yes for now.
A glut would show up first in the old chip’s posted hour, or in a paper substitute for that hour. GPU futures have not become a deep hedge. The contract talk has stalled on delivery windows and a long hold, which is what you get when the physical card is still the scarce object. The hyperscale version of the same bet is written as backlog rather than as a GPU-hour. Oracle’s backlog and the debt raised against data centers assume the halls will be rented. Cheaper H100 hours would make that backlog the problem. Dearer hours make it a queue. October’s cards say queue. They do not make the debt safe. Nebius’s August release lists financing, sites, power, and the supply chain as the ways the build can still miss.
What would retire the scarcity reading
The card is evidence only if buyers who do not get a discount actually pay it. If October on-demand hours bill at $3.85, the right-hand column was a notice, not a rate, and GPU Finder’s October 2 snapshot is a reason to wait for an invoice. If the revenue that matters is locked inside older commitments, the walk-up hike is a press release. CoreWeave’s $40 million is the short book, not the fleet. If cleared spot keeps falling while the list stays up, that is a discount under another name. And if AWS cuts the October 7 reservation rates before they take effect, or Nebius reprints its October 1 column lower, the shortage call was early. Until then the public record is higher in three units: Nebius’s announced GPU-hour, CoreWeave’s megawatt of annualized revenue, and Amazon’s reserved accelerator.
What the card does and does not answer
Did Nebius raise every GPU? No. H100, H200, B200, and B300 have higher October 1 figures. RTX PRO 6000 and the L40S “from” rates do not. GB200 and GB300 racks have no public hour.
Is $4.50 what an H100 customer pays? Not necessarily. The on-demand column still says $3.85, and GPU Finder still showed that H100 rate, plus B200 at $7.15 and B300 at $7.85, on the morning of October 2 UTC. It showed the H200 already at $5.40. Discounts of up to 35%, and dynamic spot, sit below either column.
Is CoreWeave’s $40 million a GPU price? No. It is annualized revenue per megawatt on contracts of about three to six months. Nebius’s August language was $20 million to $25 million per megawatt on the large second-quarter deals, and $40 million to $50 million on short-term capacity.
Do higher rents mean Nvidia shares must rise? No. Nvidia already trades close to its 52-week high. The card says the prior-generation chip is still scarce. CoreWeave and Nebius are well below their highs anyway. Capex, power, and dilution sit between a higher GPU-hour and a higher share price.