Nebius Sees Rising AI Computing Prices Despite Neocloud Debt Worries

Have you ever watched a market where everyone frets about debt loads yet the actual product keeps getting more expensive? That is exactly the strange tension playing out right now in the world of AI infrastructure. While share prices of several specialized cloud providers have taken a beating over the past month, the people actually selling the computing power are saying demand is still outrunning supply and prices continue to climb.

Why AI Computing Prices Keep Climbing Even as Debt Concerns Mount

I have been following this space closely for a while, and the contrast feels almost jarring. On one side you have investors suddenly nervous about how much leverage some of these newer cloud players are carrying. On the other side, the executives who negotiate deals every day report that customers are still lining up and bidding higher for capacity. It is the kind of disconnect that makes you lean in and pay attention.

One of the clearer voices this week came from a senior revenue leader at a major AI infrastructure firm. He noted that prices for computing power, measured in dollars per megawatt, are running ahead of the company’s own internal forecasts. That is not the language of a market that is cooling off. It is the language of a market still heating up.

The Growing Scrutiny Around Neocloud Balance Sheets

Let’s be honest for a moment. The term neocloud itself still feels new to a lot of people. These are the specialized operators who buy large quantities of advanced graphics processors, build out data centers, and then rent that capacity to companies training or running large AI models. The business model looks straightforward on paper: secure the hardware, lock in long-term power and cooling, then sell access at a premium because demand is so intense.

Yet the capital intensity is enormous. Buying the latest generation of chips requires billions. Building the facilities that can house them requires still more. Many of these companies have turned to debt markets to fund the expansion. When borrowing costs for one of the larger players recently spiked, the market reaction was swift. Shares across the group dropped sharply as investors started asking harder questions about leverage, interest coverage, and whether the economics would hold once newer, more powerful chips arrive.

In my view, that scrutiny is healthy. Any industry growing this fast needs adults in the room asking about the durability of cash flows. At the same time, it is possible for balance-sheet concerns and strong underlying demand to coexist for a while. That seems to be the situation today.

What Recent Capacity Auctions Are Revealing

One practical way these firms gauge the market is through auctions of available computing capacity. A recent auction at one of the larger providers showed continued upward pressure on pricing. The revenue chief described it as “continued upward mobility” in the price per megawatt. That single metric has become a key indicator for the entire sector.

Think of price per megawatt as the effective rent a customer pays for a standardized block of power that supports a certain number of high-end processors. When that number keeps rising, it tells you two things at once: customers still need the capacity badly enough to pay more, and the available supply has not yet caught up.

I find this particularly interesting because it cuts against the narrative that older generations of chips will quickly become worthless. If the market is still willing to pay up for capacity built around previous chip generations, the depreciation schedules many analysts have been using may prove too aggressive.

Older GPUs Are Holding Their Value Better Than Expected

Here is a detail that caught my attention. Another executive in the same broader industry noted that rental rates for earlier high-performance chips are actually higher today than they were when those chips first came to market a few years ago. The same pattern appears for the generation that debuted more recently. In both cases, the market is assigning greater value now than at the original launch.

That observation challenges a common assumption. Many people expected that each new wave of processors would immediately render the previous wave economically obsolete. Reality looks messier and more favorable for the infrastructure providers. As AI models grow more sophisticated, the volume of useful work that can be extracted from existing hardware keeps expanding. Companies are finding ways to generate more value per unit of compute, which supports higher willingness to pay even for slightly older silicon.

There’s tremendous demand and a lot of it is because the underlying model capabilities have gotten so much more advanced that people are just creating so much more value out of the compute infrastructure.

That comment from an industry CEO lands with particular force. It suggests the demand curve is shifting outward, not merely moving along a static curve. When the tools themselves become more productive, the infrastructure that powers them becomes more valuable.

Capacity Outlook for 2027 and Beyond

Looking further ahead, the same revenue executive offered a strikingly confident view of future availability. If the company brings its planned 2027 capacity online, he expects it will sell out. He went further and suggested a similar dynamic could appear in 2028. In other words, the multi-year supply pipeline still looks insufficient relative to the demand pipeline.

That kind of forward visibility is rare in most technology markets. It implies that large enterprise and AI-focused customers are already locking in commitments years in advance. When buyers are willing to contract that far out, pricing power tends to remain with the sellers for longer than many models assume.

Of course, nothing is guaranteed. Power availability, construction timelines, and chip supply can all shift. Still, the current signal from the people closest to the deals is one of persistent tightness.

How Different Providers Are Positioning Their Balance Sheets

Not every company carries the same debt profile. Some have been more aggressive in using leverage to accelerate growth. Others have taken a more measured approach. One executive was careful to distinguish his firm’s position, noting that it does not carry the same kind of debt load that has drawn recent attention. That distinction matters. Markets tend to punish the most leveraged names first when sentiment turns cautious, even if the underlying demand story remains intact.

In my experience watching capital-intensive industries, the companies that survive periods of heightened scrutiny are usually the ones that can demonstrate both strong unit economics and reasonable leverage. The current environment is testing exactly that combination.

The Broader Confirmation Coming from Adjacent Players

Additional color arrived this week from an unexpected quarter. A large aerospace and technology company reported that revenue in its AI-related business had jumped dramatically. More importantly for the infrastructure discussion, its finance leadership indicated that pricing was tracking slightly above the upper end of a previously disclosed range measured in dollars per watt. That independent data point lines up with the auction results and executive commentary coming from the specialized cloud providers.

When multiple independent sources point in the same direction, the signal becomes harder to dismiss. Supply continues to lag demand. Pricing power remains real. The market for high-performance computing capacity is still a seller’s market for the time being.

What Rising Prices Mean for the Economics of the Business

Higher rental rates improve the return profile on the heavy capital outlays these companies are making. Recovering the cost of advanced processors and the associated data-center buildout becomes easier when the revenue per unit of capacity keeps rising. That dynamic also provides a buffer against the inevitable arrival of newer chip generations.

There is a second-order effect worth noting. As long as older chips continue to command attractive rates, the effective useful life of the installed base extends. That stretches the depreciation period and improves cash-flow visibility. Analysts who have been modeling rapid obsolescence may need to revisit those assumptions.

• Higher realized prices support faster payback on GPU investments
• Extended economic life of previous-generation chips improves overall returns
• Multi-year sell-out visibility reduces near-term utilization risk
• Strong demand from advancing AI models underpins willingness to pay

None of this eliminates the risks. Interest rates, construction delays, and shifts in customer budgets can still create turbulence. Yet the current pricing environment is providing a meaningful cushion.

Investor Sentiment Versus Operating Reality

Share prices across the group have been volatile. Debt concerns have dominated the narrative for weeks. That is understandable. Markets often focus on the most immediate risk factor, especially when leverage is involved. At the same time, the operating data continuing to emerge tells a different story about the health of the underlying market.

I have seen this pattern before in other infrastructure buildouts. Sentiment can swing hard on financing worries while the actual product continues to clear at higher prices. Eventually the two narratives have to reconcile. Either demand softens and prices fall, validating the caution, or pricing remains firm and the stronger operators demonstrate that their capital structures are manageable.

Right now the second path still looks more plausible based on the commentary coming from the people closest to the deals.

The Role of Model Progress in Sustaining Demand

One under-appreciated driver is the rapid improvement in the capabilities of the AI models themselves. Each advance in model performance tends to unlock new use cases and increase the volume of computation customers want to run. That creates a feedback loop: better models drive more demand for compute, which supports higher prices, which in turn funds more infrastructure, which enables still larger models.

As long as that loop remains intact, the pressure on capacity is likely to persist. The executives speaking this week appear to believe the loop is still accelerating rather than slowing.

Practical Implications for Watching the Sector

For anyone following these companies, a few signals seem worth tracking closely over the coming quarters. First, the trajectory of price-per-megawatt realizations in successive auctions or contract announcements. Second, commentary on the utilization and residual value of older chip generations. Third, any changes in the cost of capital or the willingness of lenders to finance further expansion. Fourth, concrete evidence on whether the multi-year capacity pipelines are in fact selling out as projected.

Those four data points together will tell a clearer story than any single share-price move or headline about leverage.

The table above is a simplified way of organizing the moving parts. Reality will of course be more nuanced, but frameworks like this help keep the analysis grounded when headlines swing from extreme optimism to extreme caution.

A Personal Take on the Current Moment

I have found that the most useful periods in any high-growth sector are the ones where the narrative and the numbers temporarily diverge. They force everyone to examine assumptions more carefully. Right now the narrative is dominated by leverage and financing risk. The numbers coming from the commercial side of the business are still pointing to scarce supply and rising prices.

Perhaps the most interesting aspect is how calmly some of the operators are responding. They acknowledge the market’s concerns about debt while continuing to report that their own capacity is in strong demand and that pricing is exceeding plan. That combination of transparency and confidence is worth watching.

None of this is investment advice, of course. Markets can remain skeptical longer than many expect, and capital structures can still matter a great deal if conditions tighten further. Yet the fundamental picture of AI computing demand does not appear to be cracking. If anything, the latest comments suggest it remains robust.

Looking Ahead Without Overconfidence

The next several quarters will test whether the pricing power observed today can persist as more capacity comes online and as the next generations of processors arrive. It will also test which balance sheets can comfortably support the ongoing capital requirements.

For now, the clearest message from the people running the auctions and negotiating the contracts is straightforward: demand is still running ahead of supply, and prices are still rising. That reality sits alongside legitimate questions about leverage. Both can be true at the same time. The companies that navigate the next phase successfully will likely be those that respect both sides of the equation.

In a market moving this quickly, staying close to the primary commercial signals rather than the secondary market noise often proves the more reliable guide. The latest round of executive commentary has provided a useful reminder of that principle.

The story of AI infrastructure is still being written in real time. Rising computing prices are one of the more concrete chapters available at the moment. How the debt and financing chapters resolve will shape the rest of the narrative. For anyone trying to understand the economics of this buildout, both parts of the story deserve equal attention.