Demand for generative-ai computing is pushing specialized cloud operators into a capital-intensive race for power, data-center space and accelerator capacity, with CoreWeave, Nebius and Cerebras each reporting customer commitments that exceed the infrastructure they have already brought online. Their latest quarterly results show a common trade-off: contracts and backlogs are rising rapidly, but the companies must spend billions on facilities and hardware long before those commitments become cash-generating revenue.
The buildout is increasingly placing electricity access alongside chips as a central constraint on ai expansion. Operators with sites near available grid capacity, including some companies with roots in cryptocurrency mining, are seeking to convert that power access into long-term ai hosting agreements.
CoreWeave reported that 98% of its second-quarter revenue came from long-term contracts, leaving only 2% from on-demand consumption. The company’s revenue rose 112% from a year earlier to $2.6 billion, while its backlog reached $104 billion, up 246% year on year. It also added $25 billion in customer commitments at the beginning of the third quarter.
Those figures point to demand that has already been contracted rather than merely projected. Yet the cost of making capacity available is weighing on CoreWeave’s financial results. Gross margin declined eight percentage points to 66% as data-center rent, electricity and expansion expenses increased. The company reported a $49 million operating loss and a $626 million net loss, including $640 million in interest expense related to its gpu-backed debt financing.
CoreWeave spent $9.4 billion on capital expenditures during the quarter, more than three times its quarterly revenue. It raised its 2026 capital-expenditure forecast to between $35 billion and $39 billion and aims to exceed 1.85 gigawatts of active power capacity by the end of the year. That scale would put its planned electrical footprint in the range traditionally associated with major hyperscale data-center projects.
Nebius reports faster expected payback on cloud contracts
Nebius is pursuing a similar expansion strategy, though its quarterly numbers show stronger reported margins as it scales its ai cloud business. The company, created from the breakup of Yandex’s international assets following the $5.4 billion sale of its Russia business in 2024, reported second-quarter revenue of $582 million, up 454% from a year earlier.
Ai cloud accounted for $575 million, or 98%, of that revenue. Nebius reported gross margin of 77%, up six percentage points from the previous year, and adjusted ebitda of $236 million, representing a 41% margin. Its operating result remained negative, with a $176 million operating loss as new infrastructure began adding $260 million of depreciation and amortization costs.
Nebius said it signed four new ai cloud contracts with average contract values above $1 billion. The company reported contract values of $20 million to $25 million per megawatt, with shorter agreements priced as high as $40 million to $50 million per megawatt.
The company expects contracts signed in the second quarter to recover their associated capital expenditure and operating costs in roughly 22 months, compared with its earlier estimate of two to three years. That projected recovery period is important for a business spending heavily before revenue is recognized: Nebius invested $5.7 billion in capital expenditures during the quarter, almost 10 times its revenue, while maintaining full-year capex guidance of $20 billion to $25 billion.
Nebius also expects more than $9 billion in customer prepayments in 2026, which it said would cover about 50% to 60% of related capital expenditures. Prepayments can reduce the amount of external financing needed for new capacity, although they do not eliminate the execution risk of building and energizing data centers on schedule.
Cerebras shifts toward cloud services
Cerebras, best known for designing wafer-scale ai processors, is moving more of its business toward recurring cloud services as customers seek access to deployed computing capacity. The company reported second-quarter revenue of $180 million, up 74% from a year earlier.
Cloud and other services revenue rose 281% to $126 million, while hardware revenue fell 23% to $54 million. The shift gives Cerebras a larger share of recurring service revenue but also requires the company to maintain more infrastructure itself.
Cerebras reported a $477 million operating loss. It also disclosed a “core” operating loss of $34 million after excluding items including stock-based compensation linked to its May listing. Reported gross margin was 14%, while its core gross margin reached 41%, up roughly nine percentage points from a year earlier but below 46.5% in the first quarter.
The quarterly decline reflected the cost of leasing back systems that Cerebras had previously sold, allowing the company to meet demand through Cerebras cloud. It reported $25.4 billion in remaining performance obligations and said OpenAI remained its principal customer. Capacity build timing has limited how quickly orders can be converted into revenue, according to the company.
Cerebras said it had more than 600 megawatts of data-center capacity deployed or contracted through 2027 and raised fiscal 2026 core revenue guidance to between $880 million and $890 million.
Power access is becoming an ai infrastructure asset
The competition for ai capacity extends beyond specialist cloud companies. Meta is expanding custom-chip development and planning multi-gigawatt-scale computing deployments, while SpaceX has sought to sell access to its Colossus cluster. These projects compete for the same practical inputs: substations, transmission connections, construction capacity, cooling systems and high-density server space.
That demand is creating a potential opening for operators that already control power-heavy facilities. Riot Platforms said last week that it had signed a 20-year lease agreement to host advanced computing workloads rather than dedicate the relevant capacity to cryptocurrency mining. The company expects the agreement to generate roughly $9 billion of revenue over its term and said initial capacity is scheduled for delivery later this year.
The economic appeal is straightforward. A mining operation can be curtailed, expanded or relocated based on electricity prices and network economics, while an ai hosting agreement can attach contracted payments to a site with scarce grid access. The transition is not automatic: ai tenants require data-center-grade redundancy, network connectivity, cooling and construction standards that many mining sites were not originally designed to provide.
Claims that mining capacity will necessarily fall or that block production will slow as sites are repurposed go too far. Proof-of-work networks automatically adjust mining difficulty over time, and mining companies can deploy equipment elsewhere, replace older machines or maintain separate operations. The more immediate development is a contest for energized infrastructure, where a secured power connection can now support either digital-asset mining or high-value ai computing services.
As AI reshapes finance and infrastructure, explore automated strategies with our AI copy trading guide tailored for next-gen traders.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
