NBIS After a 34% Day: 22x EV/ARR Against a Peer at 7x

The headline results (Nebius's 2026-08-12 release and the same-day conference call):

String the quarterly revenue prints together and the shape of the curve argues better than any single print does: $105.1M in Q2 2025, $399.0M in Q1 2026, $582.3M this quarter — five and a half times in a year — while adjusted EBITDA travelled from −$21.0M to +$236.2M over the same stretch. The turn in margin is worth more than the slope of revenue, because it says the unit economics of this business hold once scale arrives, rather than revenue being bought with subsidy.

The guidance requires one piece of arithmetic, and that arithmetic is the foundation under every judgment later in this note. Full-year revenue guidance stands at $3.0–3.4B; the first half actually delivered $981.3M, which leaves $2.02–2.42B for the second. Converted to sequential growth, Q3 and Q4 each have to put up +43% (low end of guidance) to +60% (high end) — and the quarter just reported grew +46% sequentially. The low end amounts to running at exactly the current pace for two more quarters; the high end requires accelerating from here.

The ARR line sets a steeper bar. March-end ARR was $1.90B (backing out the +58% sequential move), June-end was $3.0B, an increment of $1.10B in the quarter. The year-end target is $7–9B, which means the two remaining quarters have to add $4.0–6.0B, an average of $2.0–3.0B each. There is exactly one number to watch in the next report: whether the quarterly ARR increment can jump from $1.1B to above $2B. Revenue is a report card on the last three months; the ARR increment is a deposit on the next twelve. If the latter stalls, the former's curve caves in two quarters later.

The demand-side evidence is unusually hard. Management said as much on the call, about as bluntly as it gets: "we are sold out on capacity, because whatever we can build, we can sell." Long-term contracts — one to three years — are priced at $20–25M of revenue per megawatt per year, with customer prepayments covering 50%–60% of the matching capital expenditure. Short contracts of six months or less are being negotiated at $40–50M per megawatt, sometimes higher. Same capacity: the customer signing for three years pays one price, the customer renting for six months pays more than double. That spread is the most honest reading available of how tight supply is. Management put the payback period on recent projects at under two years.

The $8.04B of cash on the balance sheet sounds like a thick cushion until you set it against this company's burn rhythm: $5.66B of capital expenditure in a single quarter against $2.25B of operating cash flow, a $3.41B gap, with full-year capex guidance of $20–25B. On balance-sheet cash alone, that runs out in a little over two quarters.

The funding sources look like this (company disclosures and the 2026-08-12 call):

Put together, that stack broadly covers the low end of the capex guidance; the high end requires more financing. What readers should actually retain is the cost. Against roughly 271 million shares outstanding, 12.7 million shares diluted existing holders by 4.7% in a single quarter, and that is already a completed fact; the remaining 12.3 million, if fully used, adds another 4.5%. Any per-share value derived from the current market cap has to net that layer out first.

There is one more adjustment to unpack. Adjusted EBITDA is reported at $236.2M on a 41% margin, while stock-based compensation in the same quarter ran $102.5M against $14.7M a year earlier — nearly sixfold in a year, and 17.6% of revenue. Stock comp is added back in the adjusted figures, but the shares it hands out are real, irreversible dilution. Deduct it again and the 41% margin lands back around 23%. Both numbers are correct; which one you use depends on whether you think stock handed to employees counts as a cost. Anyone who thinks it doesn't is, in practice, paying part of that payroll themselves.

A note on GAAP while we're here. This company reported GAAP net income of $621M in Q1, then −$0.12 per share in Q2. The same set of books producing a profit one quarter and a loss the next tells you the GAAP net income line currently has little to do with how the business is running at Nebius; read operating cash flow and ARR instead.

Start upstream with silicon. The company's strategic alliance with NVIDIA is written into Item 1 of the 10-K, and NVIDIA is also a shareholder. In an environment of tight GPU supply and hyperscalers fighting over allocation, that relationship does buy a steadier delivery cadence. But it isn't exclusive — CoreWeave is an equally senior NVIDIA partner, and IREN holds a five-year, $3.4B cloud contract with NVIDIA of its own (per DCD and MLQ reporting). It is closer to a ticket into the first tier than an advantage over the others already inside it.

The real bottleneck is power and land. Management raised the year-end contracted power target to 5GW and described the company as one of a small handful globally "capable of building more than 1GW of new capacity per year." Read that the other way round: the ceiling on capacity expansion is set by substations, grid interconnection permits and local approvals, not by orders. The execution risk named on the call was the regulatory process at the Vineland, New Jersey site. Chips can be bought with money and relationships; queue time for grid interconnection cannot. That is the most brittle link in the chain.

The downstream structure deserves more careful handling. Contracted revenue stands above $40B, mostly from investment-grade customers, with the Microsoft contract at $17.4B and the Meta contract widely quoted in the press at $27B. There is a definitional point that has to be stated plainly: within the Meta contract, the firm commitment for dedicated GPU capacity is $12B, with up to a further $15B representing options on incremental capacity. Only the sum of the two produces the headline $27B. Between the committed obligation and the ceiling figure sits a decision the customer has not made, and every valuation model that treats the ceiling as a revenue base is overstating the case.

Customer concentration is the other side of the same coin. The bulk of that $40B comes from Microsoft and Meta, which is both why the asset-backed loan priced at a mid-single-digit rate and why those two companies' capex cadences are effectively this company's destiny curve. The four new deals signed in the quarter brought in a different class of counterparty — model companies like Reflection and Cohere, plus a quantitative firm. They can afford prepayments, but none of them carries an investment-grade credit. Putting billion-dollar TCV on model companies that aren't yet profitable makes risk and return two faces of one object: contract quality determines both whether these agreements can be pledged at a bank and which contracts break first when the AI capex tide goes out.

Three comparable companies reported in the same week, and the readings only mean something side by side (each company's August 2026 releases and calls):

The direction is identical across all three — every dollar they can borrow is being converted into megawatts, all three raised capex guidance, all three are using customer prepayments to push risk off their own balance sheets. The differences sit in two places. On scale, CoreWeave is more than four times Nebius, and the contracted-revenue gap is wider still at two and a half times. On earnings quality, Nebius is the only one of the three whose AI cloud segment adjusted margin has touched 50%.

IREN's numbers carry a separate lesson. Its AI cloud revenue is less than a third of its own bitcoin mining line, yet it carries a 2026 ARR target above $4B, roughly 85% of it contracted (DCD). In this one sector, "power already connected," "money already contracted" and "revenue already recognized" are three things separated by one to two years each, and anyone who compares them in a single table will systematically overrate whichever company sits furthest from recognized revenue. Nebius's position across those three boxes is middle-to-front: the most aggressive power target, mid-pack contracted value, the fastest revenue recognition.

P/E is a useless yardstick here — GAAP earnings per share are still −$0.12, so the denominator is negative. EV/EBITDA is barely better, because adjusted EBITDA adds back stock compensation worth 17.6% of revenue. The only measures that support a cross-sectional comparison right now are EV/ARR and EV per megawatt of contracted power.

Enterprise value assembles like this (2026-08-13 closing data and the June-end balance sheet):

Take three different denominators and the answers diverge sharply:

For comparison, CoreWeave: enterprise value of roughly $89.9B (GuruFocus, August 2026), about 7.3x 2026 revenue, with an end-July EV/Revenue reading of 8.2x (Alpha Spread).

The threefold distance between 22.9x and 7.3x is the premium the market paid after this print. That premium only looks reasonable under one assumption: that year-end ARR of $7–9B has already happened. Because once you use the year-end target, Nebius at 7.6–9.8x and CoreWeave at 7.3x sit right on top of each other — which is another way of saying the current price has booked a report card that won't be published until next February, and booked it at full marks.

The $13.8M-per-megawatt reading looks cheap, and the trap sits in the same place: contracted power is not built power, built power is not contracted revenue, and turning all 5GW into revenue-producing halls requires more than $20B of capital expenditure a year for several consecutive years, plus the timeline hostage to grid approvals described above. What the market is prepaying for is execution, not demand — the demand link has already been demonstrated by sold-out capacity and a doubling in short-contract pricing.

Compute rental is fundamentally a commodity business, and three sets of readings locate it.

Prices are still pointing up. H100 hourly rents have moved from under $2 at the start of the year to above $2 and approaching $3; B200 has gone from just under $5 in January to $5.50–5.80 (aimultiple's GPU price index and IntuitionLabs' 2026 compilation). Last-generation hardware still commanding rising rents after the next generation ships in volume is itself a refutation of the intuition that GPUs depreciate the moment a new part arrives. Nebius's own contract pricing points the same way and more extremely: $20–25M per megawatt on long contracts, $40–50M on short ones, spot at double the term price.

Utilization is full. All three comparables say the same thing — everything built gets sold, Nebius's phrasing being that capacity is sold out, while CoreWeave signed more than $25B of additional customer commitments in the six weeks after quarter-end.

Payback periods are shortening, with management putting recent projects at under two years. In a capital-intensive business, a shrinking payback period is a classic marker of a cycle still in its upswing, because it means pricing is running ahead of build costs.

The call: mid-upswing — not late, but past the easiest money of the early phase. The reasoning is on the supply side. CoreWeave raised 2026 capex to $35–39B, Nebius has to build more than 1GW every year, and IREN is sitting on 2.91GW of already grid-connected power. Those megawatts turn into sellable capacity over the next six to eighteen months. Industry analysis broadly expects H100 rents to soften another 10%–20% as the Blackwell family rolls out at scale (IntuitionLabs).

Two trackable leading indicators of a top:

This chapter is about weekly-timeframe signals and governs entry and exit timing over the next few weeks; readers holding for a year or more can skip straight to the next chapter.

Start with where this stock sits. The AI cloud compute subsector shows net inflow in the 2026-08-13 sector money flow reading, its second consecutive day. Look further back: net flow over the last 5 sessions is +$13.8M, while the last 20 sessions run −$345.6M. A month of money walking out, and only the last five days bringing a sliver back. Over the same window, the 5-day price change is +9.5%, of which August 12 alone contributed 34%. Price is running ahead of money, and by a long way — a rally with that shape is driven by event buying on earnings day, not by sustained inflow.

Now the parent sector. Technology sits in the weakest tier of the eleven majors (2026-08-13), and the AI cloud compute subsector itself sits only in the middle tier, having climbed out of the weakest tier just on August 12, with its RRG quadrant in Weakening — relative strength is deteriorating even while absolute momentum persists. A subsector that just escaped the weakest queue while its parent sector remains in it needs more days to prove upside persistence. Two is not enough.

Where money is going matters more than where it's leaving. The three strongest inflow destinations on August 13 were gold, silver and silver miners; the heaviest outflows were industrial REITs, retail REITs and electric utilities. The rotation families already formed are materials and healthcare, with industrials forming; gold miners, silver miners and copper miners each climbed into the strongest tier in succession in early August. This stock's sector belongs to none of the formed rotation families. The three strongest money-migration paths run utilities to consumer discretionary, utilities to materials, and consumer staples to materials — not one of them ends in technology.

On the macro side, the overall market state is constructive (reading +43) and VIX is low (z-score −1.64), but Federal Reserve net liquidity of $5.84T is up only +0.3% week over week and essentially flat, with the 10-year Treasury at 4.7%, so the liquidity environment carries a caution label. Beta is available; the incremental water level is not visibly rising. In that setup money rotates from existing positions rather than lifting everything, and whatever isn't in the leadership sees its bounces stay bounces.

Put into a trader's language: the demand story got confirmed in the earnings report, but sector-level money has been net negative for the past month, the last two days of reflow (+$13.8M) come nowhere near offsetting it (−$345.6M), and the price has already moved 34%. This is not a rally being pushed along by buying; it is a repricing driven by short covering plus an event.

Long term (quarters to years): worth following, but two costs come with it.

The supporting case is hard. Revenue +454% year over year and still +46% sequentially, an AI cloud segment adjusted margin already at 50%, capacity sold out, short-contract pricing at double long-contract pricing, payback under two years, and most of the $40B-plus of contracted revenue coming from investment-grade customers. This is a business whose unit economics have been demonstrated and whose demand is running ahead of supply, in an industry cycle still in mid-upswing.

The first cost to accept is dilution. 4.7% in one quarter has already happened, 4.5% of authorization remains, and $20–25B of capital expenditure for the year means the financing does not stop. The second is valuation: 22.9x EV/ARR is three times the peer's 7.3x, and the entire basis for that premium is ARR going from $3.0B to $7–9B by year-end. There is one verification point, and it is whether the quarterly ARR increment can clear $2B (this quarter it was $1.10B). Clear it, and growth digests the multiple; miss it, and 22.9x travels toward the peer's 7x, which is not a gentle stretch of road. So the way to hold it is in tranches, following the quarterly reports, building position after the verification rather than before it — and accepting in advance that 30%-scale drawdowns are normal in this kind of name. It fell 18% in the month before earnings (as of 2026-08-05).

Short term (days to weeks): don't chase.

Price has completed its repricing ahead of the money: +34% in a day, +9.5% over five days, while the sector's trailing 20-day money flow is still −$345.6M net out and the recent reflow amounts to two days and $13.8M. The parent sector is in the weakest tier, the subsector was promoted to the middle tier only two days ago, and the RRG position is still Weakening. Leadership money is in gold, silver and materials; technology has no place in any of the formed rotation families; macro liquidity is flat and flagged caution. Chasing a 34% candle in that combination is a bet that someone else keeps buying higher, and the flow evidence supporting that assumption does not currently exist. What to wait for is specific: either sector money genuinely turns, or price returns to a level with something behind it. Either one showing up is where the risk-reward starts to make sense.

The second half of this note turns those two verdicts into something executable: what the conditions look like, which numbers the levels sit on, and what to check each day.

Entry conditions (two of them; one satisfied before position size is discussed, both before adding)

Condition one, the flow side, readable in the daily review every day. The AI cloud compute subsector's cumulative 20-day net flow turns from negative to positive (the 2026-08-13 reading is −$345.6M), while the technology parent sector exits the weakest tier. Right now only the 5-day net flow has turned positive (+$13.8M) and the subsector has just moved into the middle tier — both are still short. The point of this condition is that it separates "a one-off, event-driven repricing" from "money genuinely starting to allocate to this direction." The first is a place to sell; only the second is a place to buy.

Condition two, the numbers in the next quarterly report, derived straight from the arithmetic in Chapter 1. Next-quarter revenue growth of ≥ +43% sequentially, meaning an absolute figure of ≥ $833M ($582.3M × 1.43, the pace required by the low end of full-year guidance); and simultaneously a quarterly ARR increment of ≥ $2.0B (June-end $3.0B, year-end target $7–9B, $4.0–6.0B to be added across two quarters). This quarter's actual readings were +46% sequential and a $1.10B ARR increment — the revenue leg already clears, the ARR leg is short by half. Of the two, the ARR increment is primary and sequential revenue is secondary, because revenue is already-signed contracts being recognized while ARR is the volume of new signing.

Three reference levels

Two numbers daily. The AI cloud compute subsector's daily money flow direction and its consecutive-day count (currently: net inflow, 2 consecutive days, 2026-08-13), plus the technology parent sector's relative position. If the consecutive-inflow streak goes from 2 days to more than 5 and the parent sector leaves the weakest tier, condition one begins to hold. Conversely, if the inflow breaks within 3 days and the subsector drops back to the weakest tier, that promotion was nothing more than a one-off shock from earnings day, and condition one is void.

Three numbers weekly. GPU spot rents (H100 hourly currently in the $2–3 range, B200 at $5.50–5.80), peer contract announcements (CoreWeave signing $25B+ in the six weeks after quarter-end, IREN raising its ARR target from $3.7B to above $4B), and macro liquidity (net liquidity $5.84T, 10-year Treasury 4.7%, 2026-08-13). If H100 rents break below $2 and stay weak for two consecutive weeks, the cycle call in Chapter 6 moves back a notch. If the 10-year crosses above 5%, every high-multiple name resting on discounted forward cash flows comes under pressure together, and this stock's beta ranks in the highest tier of that group.

Three numbers quarterly, all in the earnings report. The quarterly ARR increment (threshold $2.0B, last quarter $1.10B), sequential revenue growth (threshold +43%, last quarter +46%), and prepayment coverage on new contracts (currently 50%–60%). The third is the easiest to overlook and the earliest leading indicator — falling coverage means the company is trading looser terms for orders, which would hit both the financing structure of Chapter 2 and the cycle call of Chapter 6 at once. Also check the share count each quarter: if quarter-end shares outstanding grow more than 5% over the prior quarter, dilution has outrun ARR growth.

Three situations trigger an update: either the ARR increment or sequential revenue coming in materially below threshold; the per-megawatt spread between short and long contracts compressing inside 50%; or price breaking the $193 early-warning level. If any of those appears, we will carry the update in that day's daily review.

Catalyst calendar

The numbers on that tracking checklist are ones we watch every day — the subsector's flow direction and streak, technology's position among the majors, GPU spot rents, the 10-year Treasury. If any of them crosses the thresholds written above, we will publish an update alongside that day's daily review, so readers don't have to sit refreshing the data pages themselves. Every figure in this note has a source: the financials come from Nebius's August 12, 2026 release, the same-day call and the June-end balance sheet; the peer readings come from CoreWeave's and IREN's own announcements; the money flow and rotation readings come from the corresponding data pages on the site, with dates marked in parentheses throughout so each one can be checked individually. However well a company's story is told, it still comes down to those two threshold numbers — the quarterly ARR increment and sequential revenue growth. November will answer them.

This article was produced on the NextPick research workbench; every money-flow, tier and valuation figure it cites can be verified on the corresponding data surfaces of the site, and the moment any tracking condition listed in the report triggers, an addendum will go out together with that day's daily review.