For the past two years, people in the compute-power space have gotten used to a familiar script: more and more cards keep coming out from Nvidia, so rents will inevitably drift downward. This week, Nebius went the other way. It increased prices for on-demand GPUs, moving from the previous-generation cards straight to the latest Blackwell lineup—prices that have already risen once earlier this year. When the news broke, the stock jumped that night; peers moved along with it. But by the time U.S. markets opened, the gain had been largely given back—more than half of it.
What the price increase signals is that Neocloud truly has pricing power. The question is why investors only gave it a half-day good reaction. That’s what I’ve been trying to figure out. $NBISB is currently around $216, still far below the mid-August peak.
According to Reuters, the new pricing schedule takes effect on October 1, with Nvidia GPUs across all generations raised by 17% to 21%. For H100, the per-card price goes from $3.85 to $4.50. The latest B300 saw the biggest increase, and CPU instances and memory also rose. The previous round was in May; this time, they added another layer at a higher level. CoreWeave previously hinted at price increases too, and in recent days capacity has basically been sold out.
What I care about most is the H100 segment. This card has been shipping for more than three years; by logic, prices should have fallen to clear inventory by now, yet rents are still moving upward. In its Q2 shareholder letter, Nebius also noted that the prices for new contracts signed for older-generation cards were more than 30% higher than in Q1. This ties into the largest controversy in AI infrastructure: depreciation. Starting this year, Nebius extended the depreciation life of servers and networking equipment from 4 years to 5 years. Michael Burry publicly criticized tech companies last year for lengthening server depreciation and making profits look better. Older cards become more expensive to lease, which is exactly the strongest rebuttal in the hands of the bulls: if cards can still be rented out at higher prices, adding another year of depreciation still holds up.
So why didn’t the stock win investors over. After reading the shareholder letters, I believe the direct contribution of the price increase to this year’s revenue isn’t as big as the headline suggests. Nebius relies on mid-term contracts of one to three years to make its living, and long-term customers already get discounted pricing. This time, the changes target on-demand listed prices, covering only part of the revenue base. The large orders signed in Q2 mostly correspond to capacity that won’t come online until the end of this year, contributing mainly in 2027. What the price increase changes is expectations for renewals and new contracts; the quarterly books basically don’t move.
Nebius is also testing more aggressive pricing. The shareholder letter says that in Q3 it ran the first pilot of a capacity auction, securing the highest Blackwell deal price the company has achieved so far, and it also signed its first short-term “urgent” orders for three to six months at prices clearly higher than standard contracts. If you view the on-demand price increase as part of this package, it looks more like it’s setting a price anchor for next year’s new contracts.
The pressure on the other side of the ledger is even bigger. In Q2, Nebius revenue was $582 million, while capital expenditures were about $5.7 billion in the same period—its spending pace is close to ten times its cash-in. How is the gap funded? The shareholder letter lists three options: customer prepayments, asset-backed loans, and a share issuance. By the end of June, the company sold more than 10 million shares via market-priced issuance, with an average price of $223.6. Then in August it issued a sizable batch of convertible bonds. Now the stock trades below the average price of that issuance; in the past month it’s fallen by more than 20%. Dilution concerns are a major driver. The benefits of the price increase have to first offset the dilution from each round of financing before they finally flow back to existing shareholders.
Wall Street is clearly divided on this. Goldman analyst Alexander Duval raised his price target to $328 by late August, arguing that it’s getting power faster than expected and that prepayments make expansion easier to finance. Truist initiated coverage with a Buy at the beginning of September, also emphasizing the pricing power created by scarcity. The cautious camp is watching execution; DA Davidson previously cut its target price due to delays at the New Jersey Vineland data center. On the day of the price increase, CoreWeave didn’t rise—it fell instead. Some media analysis suggests investors worry about how much of its capacity has already been locked in at the old prices, and how much debt is being carried behind that.
My take is that the shift of bargaining power toward the supply side is real. The H100 pricing is hard evidence—that part is where I stand with Goldman and Truist. But I don’t agree that price increases can solve Nebius’s balance-sheet problems within one or two quarters. The company’s own model shortens the payback period for new contracts to 1 year and 10 months; previously it was two to three years. This holds only if prices can stay firm next year. When the next-generation Rubin ships at scale and new capacity floods the market, if older-card rents reverse downward, both assumptions—5 years of depreciation and payback in under two years—will be challenged at the same time. Then, looking back, today’s price increase may end up being evidence of a peak.
There’s another line in the shareholder letter that shows just how much confidence management has. Nebius said that by today’s terms, the 2027 capacity could already be sold out. The company intentionally reserved part of it for urgent orders, hoping to sell at even higher prices. That’s a bet on the rising-price trend: if they’re right, next year’s profits come out higher by a good margin; if they’re wrong, it means empty cabinets paired with interest that can’t be paid back.
So my view on this price increase is fairly neutral: it proves demand, but it doesn’t prove the financial model. #AI算力 can next look at the three-quarter report in November—whether the prices on newly signed contracts for older cards can still hold, and whether the cadence of share issuance and convertible bonds slows down. If both of these move in a good direction, then there will be a reason to reclaim the gains the stock has given back this time.
What the price increase signals is that Neocloud truly has pricing power. The question is why investors only gave it a half-day good reaction. That’s what I’ve been trying to figure out. $NBISB is currently around $216, still far below the mid-August peak.
According to Reuters, the new pricing schedule takes effect on October 1, with Nvidia GPUs across all generations raised by 17% to 21%. For H100, the per-card price goes from $3.85 to $4.50. The latest B300 saw the biggest increase, and CPU instances and memory also rose. The previous round was in May; this time, they added another layer at a higher level. CoreWeave previously hinted at price increases too, and in recent days capacity has basically been sold out.
What I care about most is the H100 segment. This card has been shipping for more than three years; by logic, prices should have fallen to clear inventory by now, yet rents are still moving upward. In its Q2 shareholder letter, Nebius also noted that the prices for new contracts signed for older-generation cards were more than 30% higher than in Q1. This ties into the largest controversy in AI infrastructure: depreciation. Starting this year, Nebius extended the depreciation life of servers and networking equipment from 4 years to 5 years. Michael Burry publicly criticized tech companies last year for lengthening server depreciation and making profits look better. Older cards become more expensive to lease, which is exactly the strongest rebuttal in the hands of the bulls: if cards can still be rented out at higher prices, adding another year of depreciation still holds up.
So why didn’t the stock win investors over. After reading the shareholder letters, I believe the direct contribution of the price increase to this year’s revenue isn’t as big as the headline suggests. Nebius relies on mid-term contracts of one to three years to make its living, and long-term customers already get discounted pricing. This time, the changes target on-demand listed prices, covering only part of the revenue base. The large orders signed in Q2 mostly correspond to capacity that won’t come online until the end of this year, contributing mainly in 2027. What the price increase changes is expectations for renewals and new contracts; the quarterly books basically don’t move.
Nebius is also testing more aggressive pricing. The shareholder letter says that in Q3 it ran the first pilot of a capacity auction, securing the highest Blackwell deal price the company has achieved so far, and it also signed its first short-term “urgent” orders for three to six months at prices clearly higher than standard contracts. If you view the on-demand price increase as part of this package, it looks more like it’s setting a price anchor for next year’s new contracts.
The pressure on the other side of the ledger is even bigger. In Q2, Nebius revenue was $582 million, while capital expenditures were about $5.7 billion in the same period—its spending pace is close to ten times its cash-in. How is the gap funded? The shareholder letter lists three options: customer prepayments, asset-backed loans, and a share issuance. By the end of June, the company sold more than 10 million shares via market-priced issuance, with an average price of $223.6. Then in August it issued a sizable batch of convertible bonds. Now the stock trades below the average price of that issuance; in the past month it’s fallen by more than 20%. Dilution concerns are a major driver. The benefits of the price increase have to first offset the dilution from each round of financing before they finally flow back to existing shareholders.
Wall Street is clearly divided on this. Goldman analyst Alexander Duval raised his price target to $328 by late August, arguing that it’s getting power faster than expected and that prepayments make expansion easier to finance. Truist initiated coverage with a Buy at the beginning of September, also emphasizing the pricing power created by scarcity. The cautious camp is watching execution; DA Davidson previously cut its target price due to delays at the New Jersey Vineland data center. On the day of the price increase, CoreWeave didn’t rise—it fell instead. Some media analysis suggests investors worry about how much of its capacity has already been locked in at the old prices, and how much debt is being carried behind that.
My take is that the shift of bargaining power toward the supply side is real. The H100 pricing is hard evidence—that part is where I stand with Goldman and Truist. But I don’t agree that price increases can solve Nebius’s balance-sheet problems within one or two quarters. The company’s own model shortens the payback period for new contracts to 1 year and 10 months; previously it was two to three years. This holds only if prices can stay firm next year. When the next-generation Rubin ships at scale and new capacity floods the market, if older-card rents reverse downward, both assumptions—5 years of depreciation and payback in under two years—will be challenged at the same time. Then, looking back, today’s price increase may end up being evidence of a peak.
There’s another line in the shareholder letter that shows just how much confidence management has. Nebius said that by today’s terms, the 2027 capacity could already be sold out. The company intentionally reserved part of it for urgent orders, hoping to sell at even higher prices. That’s a bet on the rising-price trend: if they’re right, next year’s profits come out higher by a good margin; if they’re wrong, it means empty cabinets paired with interest that can’t be paid back.
So my view on this price increase is fairly neutral: it proves demand, but it doesn’t prove the financial model. #AI算力 can next look at the three-quarter report in November—whether the prices on newly signed contracts for older cards can still hold, and whether the cadence of share issuance and convertible bonds slows down. If both of these move in a good direction, then there will be a reason to reclaim the gains the stock has given back this time.