Glossary

Contracted-power guidance, connected capacity, and the gap to capacity that actually earns.

Contracted power

Land and grid capacity Nebius has secured — not capacity that is running. Contracted power becomes revenue only after construction, energization, and sale. It is the most quoted number about this company and the least connected to current earnings: outside models value energized capacity, which is a small fraction of this.

Connected capacity

Power that is physically energized and able to run hardware. This, not contracted power, is the quantity every serious model multiplies by a price per megawatt. The gap between the two is the buildout still to come — and the buildout risk.

Sold capacity

Energized capacity under customer contract. Nebius has never disclosed it, and dividing revenue by an assumed contract price to infer it produces an estimate that depends entirely on the assumed contract mix. This site does not publish that estimate as though it were an observation.

Capacity implied to be earning

Nebius never discloses how many megawatts are actually under contract and generating revenue. Dividing ARR by the blended contract rate gives the closest honest estimate — and the answer lands far below connected power. That gap is the single most important thing to understand here: energized capacity is not earning capacity, and management's own guidance implies the same. This is a derivation, not a disclosure, and it moves with the assumed contract mix.

Connected against earning

Management guides to 800MW–1GW connected and $7–9B of ARR at the same date. At the blended contract rate those two figures only reconcile if well under half of connected power is earning — consistent with management saying the year-end capacity becomes fully active during the first half of 2027. Anyone multiplying connected megawatts by a headline contract rate will overstate revenue by roughly two times.

Contracted-power guidance ladder

The company has raised its year-end contracted-power target four times in twelve months (>1 GW → >2.5 → >3 → >4 → 5 GW). Watch the interval between raises rather than the level: shortening intervals mean the buildout is outrunning its own plan.

ARR target, year-end 2026

Management's guided exit rate for 2026, against $3.0B at mid-year. Reaching it requires the second half to add roughly as much run-rate as the entire prior history of the business. That is the single clearest test of whether the buildout is converting on schedule.

Pace against the year-end ARR target

Whether revenue is keeping up with the run-rate the year-end target requires. It compares the share of the promised gain delivered against the share of the year elapsed, so 'behind pace' means behind schedule rather than failed. This is the only guided figure Nebius reports often enough to check mid-year; connected power and contracted power are only reported at year-end, so they cannot be tracked this way.

The price side: contract mix, the blended rate per megawatt, operating leverage and margin.

Blended annual contract value per megawatt

The fleet-weighted price across the whole contract book, and the number that actually multiplies against capacity to produce ARR. It is far below the headline rate on new deals, because most installed capacity is still on older, cheaper long-term contracts. Models that use the newest deal price as the fleet rate overstate revenue substantially.

Rate on newly signed capacity

What Nebius charges on deals signed now, rather than what the installed base earns. This is the leading indicator: today's marginal rate becomes tomorrow's blended rate as old contracts roll. Rising marginal rates against flat capital costs is the cleanest evidence of pricing power that exists in this business.

Contract mix

Management splits the book into three types that earn very differently: long-term investment-grade deals near $12M per MW signed to finance the buildout, mid-term deals with AI labs at $20–25M, and short-duration deals at $40–50M or more. The tier rates are disclosed; the share of capacity in each is not, so the mix shown here is an assumption and the blended rate moves with it.

Auction price discovery

A pilot auction cleared 15% above any price the company had previously seen and 20% above its own pipeline for the same hardware. It covers only a small slice of capacity, but as direct market evidence it suggests capacity has been underpriced rather than fully valued.

Annual contract value per megawatt

The yearly price charged per megawatt of capacity sold. 2026 base contracts were signed near $12M/MW; deals closed in Q2 2026 ran $20–25M; short-term and auction deals exceed $40M. Rising ACV/MW is pricing power — the same megawatt earning more.

Operating leverage

Each cost line as a share of revenue. All four falling while revenue grows means costs are not scaling with the business. Cost of revenue 29→23%, product development 41→33%, SG&A 65→30%, D&A 72→45% between Q2 2025 and Q2 2026.

Cost of revenue / revenue

Direct cost of delivering compute as a share of revenue. Falling from 29% to 23% over four quarters means each dollar of revenue costs less to serve as the fleet utilizes.

Product development / revenue

Engineering spend as a share of revenue, down from 41% to 33%. The platform is largely built; revenue is growing into a mostly fixed engineering base.

SG&A / revenue

Overhead as a share of revenue, down from 65% to 30% — the steepest of the four cost lines. Overhead barely grows while revenue compounds.

Depreciation & amortization / revenue

The accounting cost of the GPU fleet spread over its useful life, as a share of revenue. Falling D&A share means revenue is outgrowing the depreciation of the installed base — but see the useful-life note on this line before crediting all of it to operations.

D&A useful-life change

Effective 2026 Nebius extended server useful life from 4 to 5 years. This reduces reported depreciation independent of any operational improvement — part of the D&A ratio's fall is an accounting choice, not efficiency.

Operating margin against the peer benchmark

Every valuation of this company turns on an assumed EBIT margin, and the only real-world evidence comes from CoreWeave, which guides to roughly 8% for FY2026 and low teens by the fourth quarter. Management guides Nebius to 20–30% over the medium term. Assuming the top of that range means assuming several times the margin the largest operator in the industry earns at comparable scale — possible, given Nebius owns rather than leases part of its estate, but it is the least evidenced input in any model.

Capex payback period

How long a deal takes to repay the capex it required — now 1 year 10 months, down from 2–3 years. Shorter payback means capital recycles faster and less outside financing is needed per megawatt.

Capex per megawatt

What one megawatt of capacity costs to build, and the denominator of every return calculation here. It is not separately disclosed, so the figure shown is implied from payback and contract value. Component and memory cost inflation pushes it up through 2027, which erodes returns even if pricing holds.

Prepayment coverage of new-deal capex

Share of new-deal capex funded by customer cash paid up front rather than debt or equity. Currently 50–60%. Higher means less dilution per megawatt built.

Quarterly revenue

Revenue actually recognised in the quarter. This is not ARR: ARR annualises the most recent run-rate, while revenue accumulates over the whole period. Capacity that lands in December flatters ARR and barely touches that year's revenue, which is why the two diverge so widely in a fast buildout.

Quarterly capital expenditure

Cash spent on data centers and hardware in the quarter. Read it against prepayment coverage and payback — capex is only alarming when the capital it consumes isn't being recycled.

Customer commitments, who signed them, and how creditworthy those counterparties are.

Customer commitments

Contracted future revenue — money customers are obligated to spend over multi-year terms. Commitments convert to revenue only as capacity comes online, so this is a ceiling on near-term growth, not a forecast.

Annualized run-rate revenue

The most recent revenue run-rate multiplied out to a year. It moves before reported revenue does, which is why the company guides on it — but it annualizes a single point, so capacity energized in the final weeks of a period lifts ARR far more than it lifts that period's revenue.

ARR against recognised revenue

ARR is a snapshot of the exit rate; revenue is the area under the curve. In a business adding capacity this fast the two diverge enormously, and conflating them is the most common error in analysis of this company. Use ARR to judge momentum and revenue to judge what was actually earned.

Backlog coverage

Customer commitments divided by current annualized revenue: how many years of today's revenue is already signed. High coverage de-risks growth but concentrates it — check how much sits with a single customer and whether that customer could actually pay.

Customer roster

Named customers are disclosed; mystery entries are commitments the company references without naming the buyer. Concentration matters more than count — one hyperscaler can dominate the backlog.

Backlog by counterparty credit

Take-or-pay contracts are only worth the creditworthiness of whoever signed them. Investment-grade commitments from Microsoft and Meta are close to certain; billion-dollar commitments from venture-funded AI labs depend on those labs continuing to raise capital. A backlog that looks identical in dollars can carry very different risk depending on this split — and this is precisely what the market punished in comparable names.

Customer concentration

The share of disclosed commitments sitting with the largest one and three customers. High concentration means the loss or renegotiation of a single contract reprices the whole thesis. Only disclosed commitments are counted, so the true denominator is larger and these shares are an upper bound.

How the buildout is paid for, what it costs, and how much of it dilutes existing holders.

How the buildout is paid for

Capacity is funded four ways — customer prepayments, debt, convertible notes, and share issuance. Only the last dilutes existing holders. Watch the mix shift: more prepayment and debt per megawatt means less dilution.

Shares outstanding

The direct measure of dilution. Step marks show when new shares were sold through the at-the-market program and at what average price. A rising share count is the cost existing holders pay for growth not covered by prepayments and debt.

At-the-market issuance

Shares sold gradually into the open market under a standing program. Selling at higher prices raises more cash per share of dilution — compare the average sale price across issuance periods.

Convertible notes

Debt that can convert into shares above a set price. Cheap financing while the stock rises, but it becomes future dilution if converted — read principal alongside the conversion price.

Effective breakeven above conversion price

The stated conversion price is not the level that decides the outcome. These notes accrete — $1,000 borrowed becomes $1,100 or $1,250 at maturity — but holders forgo that accretion if they convert. That pushes the price at which conversion actually beats repayment well above the headline conversion price.

Cash interest against reported interest

Interest accrues only on original principal, so the cash leaving the business is far below the interest expense that appears in the accounts; the difference is non-cash accretion. Reported earnings therefore understate cash generation here — the mirror image of the depreciation useful-life change, and worth knowing in both directions.

Net debt

Gross debt less cash, and the deduction that turns any enterprise value into an equity value. Outside models assume tens of billions by the end of 2027 as the buildout is financed; that assumption is usually a plug rather than an observation, so the current reported position is worth tracking against it.

Gross debt

Everything owed before netting cash. The mix matters as much as the level: asset-backed facilities secured against contracted cash flows behave very differently from corporate debt, and Nebius has barely touched the latter.

Cash and equivalents

Dry powder against a capital programme measured in tens of billions. Read alongside prepayments, which fund construction without appearing here as borrowing.

Deferred revenue

Customer cash received ahead of delivery. It is simultaneously the cheapest financing available — no interest, no dilution — and a forward demand signal, because it is money customers have already committed for capacity that does not yet exist. A rising balance leads revenue.

Asset-backed facility

Debt secured against contracted customer revenue rather than equity. Financing capacity this way avoids dilution entirely, but adds fixed obligations that must be serviced whether or not new deals close. Management describes it as a repeatable framework against a backlog of more than $40B.

Customer prepayments

Cash customers have paid ahead of usage. It funds construction before revenue exists and signals customer conviction — a customer does not prepay billions for capacity it doubts will arrive.

Tracked disclosed data centers with their state and megawatts, plus possible sites inferred from hiring.

Data center footprint

Each node is a site. Size is megawatts; ring colour is state — announced, under construction, or live. Contracted power (5 GW) is much larger than live capacity, and the gap between them is the buildout risk. Contracted capacity not yet assigned to a disclosed site does not appear on the globe.

Site state

Announced means a signed location with no construction disclosed; under construction means work is underway; live means energized capacity is serving customers. Only live megawatts can earn revenue today.

Data center hiring by metro and month, and first detections of possible new sites.

Hiring radar

Nebius posts data center technician roles for a site months before announcing it. Each cell counts open site-operations roles in a metro and month, drawn from the public careers board. It is a signal, not confirmation — roles are occasionally posted for sites that never open.

First detection

The first data center operations posting ever seen in a metro. This is the earliest public trace of a possible new site, recorded permanently even after the posting closes.

Model catalog, launch lag against upstream releases, and availability zones.

Models in the catalog

Open-weight models available for inference on Token Factory, observed directly from the public API. Breadth signals how seriously the inference business is being run — a stale catalog means the effort has stalled.

Model launch lag

The hours between a model's public release and its availability on Nebius. Day-one support is the company's claim; this measures it. Rising lag on flagship releases is an early sign the inference platform is losing priority.

Availability

Incident history by availability zone, drawn from the public status page. Enterprise customers read this page before signing — so should you.

Availability zones

Independent failure domains on the status page. New AZs appear here before marketing announces them, which makes this list an early indicator of capacity coming online.

Public list prices per GPU-hour across comparable vendors, with append-only change history.

GPU price index

Public list prices per GPU-hour across vendors. Large customers negotiate below list, so treat this as relative positioning rather than what anyone actually pays. A vendor drifting above the peer band must justify it with reliability or scale.

GPU list-price history

Each point is recorded when a vendor's published price range changes. The chart uses the midpoint of that range only to compare direction across vendors; it is derived, not a quoted transaction price. Missing days mean no detected list-price change, not missing daily market data.

Insider trades from Form 4, and intent-to-sell notices from Form 144.

Insider activity

Purchases and sales by officers, directors, and large holders, from SEC Forms 4 and 144. Most insider selling at growth companies is scheduled compensation, not a signal. A cluster of open-market purchases is rarer and more informative than any sale.

Form 4 — completed trades

Form 4 reports a trade that already happened, filed within two business days. Code P is an open-market purchase, S a sale, A an award, F tax withholding, M an option exercise.

Form 144 — intent to sell

A notice of intent to sell, filed before the trade — the sale may never occur. A 144 with no matching Form 4 within about 90 days suggests the sale did not go through. Never read a 144 as a completed sale.

Where each GPU generation sits in deployment.

GPU generation pipeline

Where each GPU generation sits in deployment. Received means first units are on site; validating means internal bring-up; production means customers can buy it; scaled means it is a fleet workhorse. Nebius does not disclose fleet mix percentages, so this tracks state rather than share — any percentage you see elsewhere is an estimate.

Stakes in the businesses outside the AI cloud.

Non-cloud holdings

Nebius holds stakes in businesses outside the cloud — ClickHouse, Toloka, Avride, TripleTen, Tavily. These are excluded from AI-cloud revenue but carry value. Stake percentages are disclosed only partially and valuations are last-round marks, not current prices.

A sourced ledger of analyst targets, short interest, Schedule 13D/G beneficial owners, and quarter-over-quarter 13F changes.

Tracked analyst targets

The median uses only the latest corroborated target from each tracked sell-side firm issued within the last 12 months. It is a reviewed sample, not a complete Wall Street consensus or a forecast by this site.

Reported short interest

FINRA short interest is a twice-monthly snapshot of open short positions. It is not daily short-sale volume, and a rise can reflect hedging as well as a directional bearish position.

Tracked 13F ownership

Form 13F reports arrive up to 45 days after quarter end, cover only qualifying managers, and may include amendments. Changes describe the tracked filing sample, not real-time or total institutional ownership.

Beneficial owners above 5%

Schedule 13D and 13G filings disclose beneficial ownership around the 5% threshold. A filing can group joint reporting persons that describe the same position, so their shares are not added together. Amendments can report that a filer has fallen below 5%; those filings remain in the history but leave the current-holder list.

Tracked rating mix

How the tracked firms are split between buy, hold and sell, using each firm's latest call. Ratings vocabularies differ by firm, so Outperform and Overweight count as buy and Sector Perform counts as hold. This describes the reviewed sample on this page, not every analyst covering the stock.

Short interest trend

The last year of twice-monthly FINRA snapshots. The direction matters more than any single reading: a position built steadily over months is a different thing from a spike between two settlement dates, and neither tells you why the position exists.

Scheduled events that can move the picture.

Catalysts

Scheduled events that can move the picture — earnings, note maturities, stated energization targets. Company timelines are targets, not commitments; dates marked estimated are inferred from cadence, not announced.

The newest SEC filings found by the daily automated check.

SEC filings

Every SEC filing as it appears. 6-K is a current report, 20-F the annual report, 424B5 and FWP are offering documents, SC 13G a large-holder disclosure, Forms 3/4/144 are insider forms. Offering documents landing between earnings dates usually mean new financing.

General

Delayed share price

A delayed quote shown only to anchor the price-target ladder and insider trade prices. It is not live and this site draws no conclusions from it.

Implied annual contract value

Sold megawatts times price per megawatt — the yearly contract value your scenario implies. A scenario, not a forecast: it holds if and only if the inputs you chose hold.

Implied ARR

The annualized revenue run-rate your scenario implies once the chosen capacity is sold at the chosen price. Compare it against management's stated guidance, which is the default preset.

Implied EBITDA

Implied ARR times the margin you chose. The calculator stops here deliberately — no multiple, no share count, no target price. Valuation is your job, not this site's.