$3.5 Billion Senior Secured: The Nebius Facility Is a Collateral Contract, Not a Growth Story

Fact: Nebius Group N.V. has secured a $3.5 billion senior secured financing facility, arranged by Citibank. The disclosure runs approximately one hundred words. It contains no date. No author. No original source attribution. No interest rate. No tenor. No collateral schedule. No covenant thresholds. No drawdown conditions. No confirmation of whether the facility refinances existing debt or constitutes net-new capital. I have audited financing announcements for a decade. A senior secured facility described in a hundred words is not a financing announcement. It is a headline wearing the costume of disclosure. The phrase "senior secured" carries legal weight on its own: assets are pledged, the lender sits first in the repayment waterfall, and financial covenants almost certainly attach. None of that is quantified. None of that is disclosed. Protocol integrity is binary; trust is a variable. The facility exists. Everything that determines whether it accelerates Nebius or strangles it does not appear in the source material. That gap is the subject of this analysis. Nebius is not a normal startup. It is the surviving shell of Yandex N.V. — the Dutch-domiciled parent of the Russian search giant — after a two-year, multi-jurisdiction divestiture forced by sanctions exposure. The 2024 restructuring left the entity with roughly $2.5 billion in inherited cash, a European data center footprint in Finland (Mäntsälä) and France (Paris), an engineering cadre drawn from Yandex's search infrastructure teams, and four operating business lines: the GPU cloud (Nebius AI Cloud), the inference and fine-tuning platform (AI Studio), the data-labeling and RLHF operation (Toloka), and the autonomous driving unit (Avride). In December 2024, Nvidia, Accel, and Orbis led a $700 million equity private placement into the company. That equity raise matters here for a specific reason. It establishes the prior capital structure: cash-heavy, debt-light, story-driven. The $3.5 billion senior secured facility changes the character of the balance sheet. Equity capital tolerates losses. Secured debt does not. A company that borrows against assets has told the market something it cannot retract: that it believes its cash flows are bankable, or that the equity window has closed, or both. The neocloud sector — GPU cloud specialists as distinct from hyperscalers — is in a capital, power, and chip-allocation elimination tournament. CoreWeave sits in the first tier, publicly listed, with scale and backlog leadership. Nebius, Lambda, Crusoe, Together, and Fluidstack occupy the second tier. Bitcoin-miner conversions like IREN, Applied Digital, and Galaxy form the third. Nebius is the strongest European entrant and the head of the global second tier, but the gap to CoreWeave is not cosmetic. It is measured in contracted megawatts and in the cost of capital. Volatility is the tax on uncertainty. This facility is a bet that the tax is payable. The first number that matters is the interest burden. Industry comparables for asset-backed GPU debt — CoreWeave's DDTL facilities being the most cited — price in the SOFR plus 250 to 400 basis point range. Assume a 7% all-in cost. On $3.5 billion, that is $245 million in annual interest, before any principal amortization. To service that at a 50% GPU cloud gross margin, Nebius must generate roughly $490 million in new annual revenue just to cover the coupon. Amortization raises the requirement further. A five-year tenor with straight-line principal implies another $700 million per year of debt service — bringing the annual cash requirement north of $900 million. This is not a rounding error. It is a threshold. Every dollar of that requirement must come from contracted revenue, and the source material discloses none of it. The coverage ratio — EBITDA to interest expense — is the only metric that matters to the lender and the only metric absent from the announcement. The word in the announcement is "facility," not "funding." A facility is a commitment — typically a term loan or revolving credit line — with drawdown conditions attached. Those conditions routinely include lease execution milestones, data center delivery dates, leverage ratio tests, and minimum liquidity covenants. Reading "$3.5 billion" as "$3.5 billion in the bank" is the most common misreading of this genre of disclosure. The actual cash may be zero at announcement and released in tranches as operational milestones are met. A facility that is signed but undrawn is a promise, and promises carry conditions. The announcement does not say what is pledged. This is the single most consequential omission in the document. Three possibilities exist, each with a different risk profile. If the collateral is GPU inventory, the loan is exposed to the classic maturity mismatch. NVIDIA's product cadence is roughly annual — Blackwell gives way to Rubin, Rubin to the next generation. The economic life of a GPU in production is three to four years at best. A five- to seven-year loan secured against depreciating silicon is undercollateralized by year three. The residual value curve is steep, the secondary market is thin, and the generational switch is unpredictable. These are the three defects of GPU collateral, and they conflict directly with the risk assumptions of conventional credit. If the collateral includes the parent guarantee or equity pledges, the risk transmits to the entire group. A default at the facility level becomes a group-level event, and the inherited cash balance is effectively pledged as a backstop. If the collateral includes long-term power purchase agreements and data center leases, the bank is effectively lending against a power position. PPAs and land leases are more tradable than silicon and have longer useful lives. This is the cleanest structure for the lender and the least discussed in the coverage. Nebius's visible demand depends on a small number of hyperscale buyers — Microsoft and Meta are the names in public reporting. Take-or-pay contracts of this type are precisely why a bank underwrites the facility at all. Banks do not lend against model capability. They lend against backlog. The corollary is that Nebius's repayment capacity is now bound to the capital expenditure rhythm of two or three counterparties. If a single hyperscaler trims AI capex by 20%, the facility's coverage ratio moves against the borrower immediately. Concentration is not a footnote in this deal. It is the underwriting. NVIDIA is simultaneously a shareholder in Nebius and its GPU supplier. Nebius's largest customers are also NVIDIA's largest customers. This closed loop is now standard across AI infrastructure: upstream vendor equity, downstream customer commitments, and bank debt jointly financing a single company's expansion. The structure works until any one node revises expectations. Then the repricing is chain-linked. This is the most underappreciated systemic risk in the AI infrastructure complex, and it is invisible in any single company's disclosure. When vendor equity, customer contracts, and bank debt all depend on the same demand forecast, there is no diversification — only correlated exposure dressed as a capital stack. The facility is arranged by Citibank — a major commercial bank, not a private credit fund. In 2025, the dominant financiers of AI infrastructure have been Apollo, Blue Owl, and Blackstone. Commercial bank leadership signals that the project cleared internal credit approval and industry exposure limits. It also signals probable syndication: the bank will likely distribute portions of the facility to other lenders. If syndication stalls, the cost of capital rises passively. Citibank serves multiple neoclouds and hyperscalers; its participation is not an exclusive endorsement, and it should not be read as strategic validation. Bank participation is a credit decision, not a strategic one. Companies of this type typically borrow at the parent level with guarantees, then establish project-level SPVs for each data center. If Nebius follows this pattern, each future site could carry its own facility. Consolidated leverage would then be systematically understated on the balance sheet that investors actually read. The $3.5 billion may be the first tranche of a series, not a standalone event. The correct question is not "how much did they borrow" but "how much will they borrow through structures that do not consolidate." If Nebius depreciates GPU assets over six years while the economic life is three to four, reported profit is systematically overstated and the interest coverage ratio is flattered. Investors should demand the depreciation schedule and the contracted per-GPU-hour rate. Without those two numbers, no unit-economics claim survives stress-testing. Depreciation is where capital-intensive businesses hide their problems, and it is the first assumption an auditor should interrogate. At current GB200 NVL72-class deployment costs — industry convention runs $30-50 million per megawatt including facility retrofit, networking, and power infrastructure — $3.5 billion corresponds to roughly 700 to 1,100 MW of new capacity potential, excluding interest and working capital. That is an order-of-magnitude jump. It also means Nebius must solve site selection, grid interconnection, supply chain, and operations hiring simultaneously within two to three years. Money is not the bottleneck. Delivery capacity is. The transformer and power distribution supply chain is the tightest link in the entire AI buildout, with lead times of two to three years for some components. A dollar committed today buys megawatts two years from now, if it buys them at all. Finland's Mäntsälä site draws on Nordic low-carbon power — nuclear and hydro. France draws on the nuclear grid. Both offer competitive industrial electricity within Europe, and the low-carbon attribute converts directly into procurement credit with ESG-sensitive and sovereign customers. Against US markets — Northern Virginia and Texas, where interconnection queues are long and power prices are rising — the European position has an advantage in electricity price certainty and a disadvantage in permitting and construction speed. The trade is real, and it favors the patient operator over the fast one. Nebius is not a model developer. Its technical route is integration engineering: making NVIDIA's reference architecture faster, more stable, and cheaper to sell. The competitive metrics are MFU (model FLOPs utilization), mean time between failures, cluster startup time, effective network bandwidth, and storage throughput. None of these are disclosed. The software layer — AI Studio — is inference-layer arbitrage in a category already validated by Together, Fireworks, and Replicate. Its only defensible differentiation is cost coupling to owned compute, not model capability. Toloka, the data-labeling and RLHF environment business, is the one genuinely hard-to-replicate asset in the stack, and it is entirely absent from the financing narrative. That absence is itself information: the capital is being raised against compute, not against the data business. High-end NVIDIA silicon deployment is constrained by US export controls. A "global expansion" touching sensitive jurisdictions triggers license review. Nebius's Russian-origin background places it under stricter scrutiny in that review, not looser. EU data residency obligations under GDPR and the AI Act require architectural isolation, not contractual promises — and architectural isolation raises unit compute cost. CSRD disclosure requirements make energy attributes a financing condition rather than a marketing bonus. If the facility carries sustainability-linked pricing, the coupon is now tied to PUE and renewable share, and a grid delay becomes an interest expense. The disclosure omits, itemized: the rate structure — fixed or floating; the tenor — three, five, or seven years; upfront and commitment fees; covenant thresholds for minimum liquidity, net leverage, and interest coverage; default triggers; the collateral schedule; whether negative pledge clauses restrict future refinancing; whether drawdown is conditioned on specific customer contracts; whether the facility refinances existing debt or is net-new; the currency, euro or dollar, which would reveal the true expansion focus; the allocation among capex, refinancing, and working capital; and whether funds flow to non-AI lines such as Avride, Toloka, or TripleTen. Recovery is not a phase; it is a reconstruction. The same applies to disclosure. A hundred-word announcement is not a reconstruction. It is a stub. The bulls are not wrong about everything, and pretending otherwise is a failure of rigor. Three of their claims survive scrutiny. First, the take-or-pay contracts are real. Banks do not underwrite $3.5 billion against a slide deck. The fact that Citibank's credit committee approved the facility is itself evidence that Nebius presented a backlog of meaningful scale. That is a fact the skeptics must absorb, not dismiss. Second, the European power position is a genuine moat. Low-carbon, competitively priced electricity with sovereign-customer appeal is not replicable by writing a check. It takes a decade to build. The Finnish and French sites are structural assets, and the sovereign cloud niche — European governments, financial institutions, healthcare — has a real, if bounded, ceiling. A bounded moat is still a moat. Third, debt is cheaper than dilution at current valuations. If management believes the equity is undervalued, borrowing against contracted cash flow is the rational choice. The bear case that "they could not raise equity" and the bull case that "they chose not to" are observationally identical from the outside. Only the missing announcement date would resolve which is true — and that date is absent. The bulls' error is not the thesis. It is the omission of the coupon. A moat that costs 7% a year to defend is a different asset than a moat that costs nothing. The contrarian who ignores the coupon is as careless as the skeptic who ignores the backlog. The next disclosure that matters is not another facility. It is the first quarterly filing that breaks out backlog, per-GPU-hour pricing, depreciation schedule, and interest coverage in the same document. Until those four numbers appear together, every claim about Nebius — bullish or bearish — is inference dressed as analysis. Code is law, but logic is the jury. The jury has not been seated. The facility is signed. The terms are not. The question for the reader is not whether Nebius survives. It is whether the collateral survives the silicon.