Nebius or CoreWeave? The Real Risk Profile for Legal AI Infrastructure

Financial figures last verified: July 30, 2026, using FY 2025 and Q1 2026 data where those periods are specified. This is not legal, investment, or tax advice. Daloopa’s comparative analysis is used as the main cross-company financial data set because its figures are traceable to company filings through its MCP tool; where company or market materials are available, they are cited separately.

A reader searching for a Nebius vs CoreWeave AI cloud stocks comparison 2025 may expect a better-buy argument. That is not the most useful question for a law firm buying legal AI software that may sit on top of one of these clouds. The more practical question is which infrastructure provider is less likely to become a continuity, outage, renegotiation, or vendor-instability problem after the legal AI tool has already been approved.

The compact answer for Q3 2026: Nebius currently presents lower counterparty risk for legal-AI dependency. CoreWeave is larger by revenue, but that scale is financed through a much heavier debt and lease structure. Nebius has net cash, owns more of its capacity, and shows a stronger underlying AI cloud margin trajectory. CoreWeave’s reported adjusted EBITDA margin looks impressive until depreciation, amortization, and stock-based compensation are put back into the operating-risk discussion.

Side-by-side risk view, before the story gets too neat

That table is the part of the comparison a legal-technology buyer should want before the product demo, not after it. A legal AI vendor can be accurate, well designed, and contractually polished while still depending on an infrastructure provider whose economics are becoming more brittle underneath the application layer.

The stock chart explains the timing; it should not drive the diligence

CoreWeave priced its IPO at $40 per share on March 28, 2025, later reached a June 2025 all-time high near $187, and was around $74 by late July 2026 in the price tracking cited by Daloopa and market commentary [2][1]. Daloopa’s comparison also found that Nebius’s market capitalization surpassed CoreWeave’s in late March 2026, even though CoreWeave had roughly five times the revenue; from CoreWeave’s IPO date, Nebius returned +886% versus CoreWeave’s +168% in that data set [1].

Those moves make the comparison searchable. They do not make it procurement-grade. A law firm is not holding a stock certificate in the ordinary course when it buys a legal AI tool; it is holding an operational dependency. The relevant failure mode is not simply that a share price falls. It is that financing costs, leases, capacity commitments, or customer concentration change the cloud provider’s incentives and the legal AI vendor’s reliability.

That is why infrastructure durability belongs in the same conversation as model performance and data-retention clauses. A firm that treats cloud capacity as someone else’s background problem may discover that the background is exactly where the disruption begins. The same budget and vendor-durability issue appears in why the AI chip sell-off matters for law firm AI budgets, where the legal risk is less about enthusiasm for AI and more about who absorbs cost when the capital stack tightens.

CoreWeave’s 56% adjusted EBITDA margin is not the comfort it first appears to be

The most important CoreWeave number is not its revenue scale. It is the gap between adjusted EBITDA and operating profit once the cost of the capacity model moves through the statements. Daloopa’s analysis shows CoreWeave with a 56% adjusted EBITDA margin, while also showing adjusted operating margin falling from 19% in FY 2024 to 1% in Q1 2026 as depreciation and amortization on colocated property and equipment ramped. Q1 2026 GAAP operating loss was $144M, and the adjusted EBITDA figure excluded about $1.1B per quarter in depreciation, amortization, and stock-based compensation [1].

For a software company with modest physical assets, EBITDA can sometimes be a useful shorthand. For an AI cloud provider built around expensive GPU capacity, depreciation is closer to the main character. It is not a stray accounting pebble in the shoe. It is the way the cost of the infrastructure arrives in the income statement after the procurement team has already admired the revenue growth.

The legal-procurement consequence is straightforward. If the cloud layer’s economics depend on excluding a very large quarterly depreciation, amortization, and compensation burden, the downstream legal AI vendor may face future pressure through price increases, stricter capacity allocation, renegotiated terms, weaker service levels, or a reduced appetite for smaller customers. None of those outcomes requires the infrastructure provider to fail. A vendor can remain technically successful while becoming a more difficult counterparty.

This is also where disclosure and financing structure matter. The concern is not that debt automatically makes a company unsuitable; many infrastructure businesses are financed heavily. The concern is whether the business model can show durable operating margins after the cost of the infrastructure is recognized. For a related securities-risk angle involving CoreWeave and circular AI financing concerns, see Securities Fraud Risk in Nvidia’s Circular AI Financing.

Owned capacity changes the risk profile

Nebius owns more than 75% of its contracted capacity, while CoreWeave sources essentially all capacity through colocation leases, according to Daloopa’s comparison [1]. That distinction is not cosmetic. It changes who has control over facilities, renewal exposure, operating leverage, and the timing of capacity economics.

A provider that owns a large share of its contracted capacity may still face execution risk, power constraints, hardware obsolescence, customer concentration, and construction delays. Ownership is not immunity. But it gives the provider a different set of tools than a model that must continually translate customer demand into lease commitments, financing obligations, and third-party facility dependence.

The margin data lines up with that distinction. Nebius’s Core AI EBITDA margin doubled to 45% in Q1 2026, a trajectory Daloopa attributes to owned-facility economics [1]. That is the more important Nebius point than the cleaner-looking GAAP net income headline. The operating model appears to be gaining margin as capacity scales, rather than watching margin compress as the cost base catches up.

For law firms, this belongs in vendor diligence even when the direct contract is with a legal AI application rather than with Nebius or CoreWeave. The application vendor’s uptime, pricing, product roadmap, and support commitments can all be affected by infrastructure-layer control. The same logic applies beyond cloud contracts to hardware-specific dependencies; Etched’s $10.3B valuation flags a hidden legal AI risk is a useful parallel because it treats the stack below the software interface as part of legal AI reliability.

Nebius’s GAAP profit needs a footnote, not a victory lap

Nebius reported $101.7M in FY 2025 GAAP net income, while CoreWeave reported about a $1.2B GAAP net loss [1]. That sentence is true, but it is too tidy if left alone. Nebius’s FY 2025 GAAP net income included a one-time revaluation gain tied to its ClickHouse equity securities, so it should not be used as a simple proof that Nebius’s AI cloud operations were fully mature on a GAAP basis [1].

The better Nebius argument is narrower and stronger: it had a net-cash balance sheet, significant cash, a mostly owned contracted-capacity base, and improving Core AI EBITDA margin. Daloopa also reports $129.5M of adjusted EBITDA in Q1 2026 excluding the ClickHouse gain [1]. That combination says more about counterparty resilience than the isolated GAAP net-income comparison.

For legal buyers, the footnote matters. If a procurement memo simply says “Nebius is profitable and CoreWeave is not,” it overstates the clean contrast. If it says “Nebius currently shows lower counterparty risk because its liquidity, ownership model, and AI cloud margin trajectory are stronger, while its FY 2025 GAAP profit included a one-time gain,” the memo is more likely to survive contact with a finance partner.

Customer concentration is a shared risk, with different consequences

CoreWeave’s customer concentration is severe in the cited Q1 2026 data: Microsoft and OpenAI represented about 65% of revenue, and Microsoft alone accounted for 45% to 62% depending on the metric used [1]. That creates a familiar legal-risk pattern. A provider can look stable because its customers are large, then become exposed because the customers are large enough to dictate terms, shift workloads, delay expansion, or build competing capacity.

Nebius is not diversified in the comforting sense either. Its two largest contracts total more than $46B, including a Microsoft commitment reported in the $17.4B to $19.4B range and a Meta arrangement reported at up to $27B [1][3][4]. Daloopa also reports that Nebius’s mid-market tail was growing 3.5x quarter over quarter, but that does not erase the strategic weight of the largest hyperscaler contracts [1].

The distinction is not concentration versus diversification. It is concentration interacting with balance sheet, margin, and capacity control. A concentrated customer base is less threatening when the provider has more liquidity, better current operating-margin direction, and more ownership of the capacity it is selling. It is more threatening when the same customer concentration sits on top of heavy net debt, thin current liquidity, lease obligations, and rapidly compressed adjusted operating margin.

This is also why Meta matters on both sides of the comparison. Meta’s reported Nebius deal supports Nebius’s demand story, but market commentary in July 2026 also pointed to Meta’s internal cloud ambitions as one factor pressuring CoreWeave shares during a roughly 30% monthly drop [4][5]. That does not prove a near-term CoreWeave operating disruption. It does show how quickly a hyperscaler relationship can move from customer-validation signal to strategic-threat question.

Legal teams have seen the same pattern at the application layer: single-provider dependency often looks efficient until the provider becomes unavailable, deprioritizes the customer, or changes its terms. For the professional-responsibility version of that problem, Claude’s June 2026 outages expose law firm ethics risk is the more direct outage example.

What the law-firm diligence file should ask next

The legal AI vendor should not be allowed to answer infrastructure questions only with “we use enterprise-grade cloud.” That is a marketing phrase, not a continuity plan. A firm does not need to turn its lawyers into data-center analysts, but it does need enough detail to know whether the cloud layer is a hidden single point of failure.

• Identify the actual infrastructure providers supporting the legal AI tool, including whether workloads are split across providers or concentrated on one AI cloud.
• Ask whether the vendor can move workloads if its AI cloud provider faces an outage, capacity restriction, contract dispute, or pricing change.
• Request disclosure of material subcontractors and cloud dependencies in the contract or security addendum, not merely in a sales response.
• Tie service-level commitments to remedies that matter for legal work: notice, escalation, data access, export capability, and business-continuity obligations.
• Review whether the vendor’s pricing assumes a specific inference or training-cost structure that could change if the cloud provider reprices capacity.
• Set a reassessment trigger when major infrastructure providers report new quarterly results, announce large financing transactions, lose major customers, or materially change capacity strategy.

This is not just procurement fastidiousness. A lawyer’s duty of technological competence increasingly requires understanding the operational limits of tools used in legal work. The point is developed in ABA Model Rule 1.1 and AI: What Competence Requires of Attorneys Using AI Tools. If a tool is material to litigation, due diligence, contract review, regulatory response, or client advice, the firm should know enough about the tool’s infrastructure dependency to avoid being surprised by a preventable continuity problem.

The narrower answer for Q3 2026

If the question is which stock has more upside, this comparison is incomplete. A real investment analysis would need valuation, growth expectations, competitive positioning, dilution, management execution, and a view on AI infrastructure demand. Those are not side issues for investors.

If the question is which provider presents lower counterparty risk for legal-AI infrastructure dependency in Q3 2026, the current evidence favors Nebius. The reason is not that Nebius is risk-free or that CoreWeave lacks demand. The reason is that Nebius’s net-cash position, owned-capacity model, and improving Core AI EBITDA margin create a more resilient operating profile than CoreWeave’s larger but debt-heavy, lease-dependent model with sharply compressed adjusted operating margin.

That answer should be revisited when CoreWeave’s Q2 2026 results are available and when all stock-price, market-cap, debt, liquidity, customer-concentration, and capacity figures can be freshly verified. Until then, a law-firm risk file that treats Nebius and CoreWeave as interchangeable AI-cloud names is missing the clause most likely to become someone else’s problem later.

References

1. Nebius vs CoreWeave: Neoclouds Financial Analysis — Daloopa
2. CoreWeave Announces Pricing of Initial Public Offering — CoreWeave, March 28, 2025
3. CoreWeave vs. Nebius: Which Artificial Intelligence (AI) Infrastructure Stock Is a Better Buy in 2026? — The Motley Fool, May 22, 2026
4. Meta signs cloud computing deal with Nebius worth up to $27 billion — CNBC, March 16, 2026
5. Why CoreWeave Stock Fell 30% in Just 1 Month — The Motley Fool, July 26, 2026