Summary: Haverhill price target $205 (12 months)
Nebius is one of the best-run AI cloud builders in the market, but at about $221 the stock already pays for more than the good news. Our 12-month price target of $205 (through October 2027) sits about 7% below the current price, so we rate the shares Sell.
Company filings through Aug 2026; Haverhill Financial Research model; price as of Nasdaq close Oct 9, 2026
The business is working. The AI cloud segment earned a 50% adjusted EBITDA margin in Q2, and new deals are priced at $20–25M of yearly revenue per megawatt.
The cash is not there yet. Nebius spent $8.1B on equipment and buildings in the first half against $4.5B of operating cash, and plans $20–25B for the full year. Shareholders fund the gap through new shares and convertible notes.
Our view: we like the company more than the stock. We would buy closer to $160, or on proof that deals keep pricing above $20M per MW while dilution slows.
Investment thesis in brief
Our thesis: Nebius has won the right to grow, but growth here costs so much capital that the per-share payoff is smaller than the headline numbers suggest.
What we like. Demand for AI computing still outruns supply. Nebius says it could sell all of its 2027 capacity today on current terms and is choosing not to. Pricing is rising, not falling: Q2 deals averaged above $20M of yearly revenue per megawatt, versus roughly $12M for its 2026 base. Customers now prepay for capacity, which cut the expected payback on new deals to under two years.
What the market may be underrating. Nebius writes its own cloud software rather than renting out bare chips. That software layer (Token Factory for running AI models, Tavily for AI search) is small today but is what separates a cloud from a landlord. The ClickHouse stake, carried at $1.5B, is a second source of value most investors ignore.
What holds us back. Every new megawatt needs billions of dollars up front, and Nebius has paid for much of it by selling stock and issuing notes that turn into stock. The fully diluted share count is now roughly 369M against 254M shares outstanding at the end of 2025. Unless the company starts funding growth mostly from customer cash and secured debt, shareholders will keep sharing the upside with new investors.
How the business actually makes money
Nebius rents AI computing power by the hour, the month or the year, and 98% of revenue ($574.9M of $582.3M in Q2) comes from that AI cloud. Customers pay for time on clusters of NVIDIA GPUs plus the storage, networking and software that make them usable.
Who pays, and for what. Big tech companies such as Microsoft and Meta buy dedicated capacity on multi-year contracts with investment-grade credit, which Nebius uses to back secured debt. AI labs and model builders, including Reflection, Cohere and a US AI lab, buy large training and inference clusters on one- to three-year contracts; about 70% of Q2 deals included prepayments. Enterprises and startups such as Revolut and monday.com pay as they go for managed inference (Token Factory), on-demand GPUs and AI search (Tavily). New in Q3, urgent buyers take short three- to six-month blocks of capacity at premium prices, which Nebius puts at $40–50M per MW.
Why customers keep paying. Capacity is scarce, so switching means waiting in line elsewhere. Most revenue sits under signed contracts: $37.5B of remaining performance obligations, of which 36% is due within 24 months and 40% in the two years after. Workloads also deepen over time; one early customer has grown its usage more than 20x.
The concentration catch. Three customers made up 59% of Q2 revenue (24%, 21% and 14%), and Nebius does not name them. The large Microsoft and Meta contracts are the most likely candidates, so a single renewal or dispute matters.
The other pieces. Avride (self-driving cars and delivery robots, $1.0M Q2 revenue, $40.1M adjusted EBITDA loss) and TripleTen (tech education, $10.0M revenue, $9.4M loss) are small and lose money. Nebius also owns minority stakes in ClickHouse, carried at $1.52B, and Toloka.
What the numbers say
Revenue is compounding fast and the AI cloud is profitable before depreciation, but the company still loses money on a GAAP basis and spends far more cash than it brings in.
Source: company quarterly results releases and 2025 annual results; Q3 2025 derived from full-year 2025 less the other three quarters
Q2 2026 at a glance. Revenue was $582.3M, up 454% from a year earlier, and the AI cloud earned a 49.7% adjusted EBITDA margin. The annualized run-rate reached $3.0B in June, and contracted backlog stood at $37.5B. Nebius spent $8.1B on equipment and buildings in the first half against $4.5B of operating cash flow, and ended June with $8.0B of cash plus $1.1B restricted (from the Q2 results release and interim financial statements).
Three things investors should know.
Cash flow is real but pre-paid. Of $4.5B of first-half operating cash flow, $4.4B came from the rise in deferred revenue, meaning customers paid ahead for capacity Nebius has not delivered yet. That cash is a gift today and an obligation tomorrow.
The depreciation change helps reported results. Moving servers from a 4-year to a 5-year life cut first-half depreciation by $86.1M. GPUs may lose value faster than five years if new chip generations keep arriving every 12–18 months.
Customer prepayments are the cheapest money Nebius has. Deferred revenue rose from $1.58B to $5.98B in six months. Management expects more than $9B of prepayments in 2026.
Balance sheet and debt (June 30, 2026, with known changes since)
Source: Nebius Q2 2026 interim financial statements (notes 8, 11, 12, 16) and Aug 24, 2026 notes closing release
The company is not over-borrowed today: cash roughly matches convertible debt. The pressure point is the pipeline of commitments, which is far larger than the cash on hand.
Competition and moat
Nebius competes in a crowded field, and its edge is engineering depth and speed, not a lock on customers.
Where the moat is real. Nebius designs its own servers and racks, runs its own cloud software, and owns more than 75% of its contracted power. It holds NVIDIA's top performance designation on GB300 systems and is among the first to receive Vera Rubin systems. NVIDIA also invested $2B in Nebius in March 2026, which helps with chip allocation.
Where it is thin. Most of what customers buy is still GPU time, and GPU time is a commodity once supply catches up. Switching costs for training customers are measured in months, not years. If chip supply loosens in 2027–2028, pricing per megawatt is the first thing that would come under pressure.
Management, governance and dilution
Management has executed well on building capacity, but shareholders have paid for it with a fast-growing share count.
Leadership. Founder and CEO Arkady Volozh previously founded Yandex and controls voting through Class B shares carrying ten votes each (33.5M Class B shares outstanding). The company is Dutch, headquartered in Amsterdam, and files with the SEC as a foreign private issuer, so it reports quarterly on Form 6-K rather than 10-Q.
How growth has been funded since mid-2025. Nebius has paid for its buildout with convertible notes, at-the-market share sales, NVIDIA's $2B investment in March 2026 and, in July 2026, a secured debt facility backed by customer contracts. The largest single raise came in August 2026: $3.45B and $2.3B of convertible notes that turn into shares above about $313 and $325.
Adding outstanding shares, NVIDIA's warrants, notes convertible below today's price and unvested awards, we count roughly 369M fully diluted shares. That excludes the August 2026 notes, which add about 18M more if the stock passes ~$313.
Insider selling. Form 4 filings dated October 5, 2026 show the COO sold 500,000 shares and the CEO, CTO and infrastructure chief sold a further ~81,000 shares. Some of this was reported as tax-related, but the size of the COO sale is worth watching.
Accounting judgment calls. Management lengthened server lives to five years in 2026 and excludes stock pay and acquisition costs from adjusted EBITDA. Both are common in the industry, but they make adjusted figures look better than cash reality.
What is it worth
We value Nebius at $205 per share over the next 12 months (to October 2027), a probability-weighted blend of three scenarios. That is about 7% below today's price, which supports a Sell.
| Assumption | Bear | Base | Bull |
|---|---|---|---|
| Revenue | |||
2026 revenue Guidance $3.0–3.4B | $3.0B | $3.2B | $3.4B |
2027 revenue | $7.5B | $11.0B | $14.0B |
2028 revenue | $10.5B | $17.5B | $24.0B |
| Profit and spending | |||
2028 adjusted EBITDA margin AI cloud was 49.7% in Q2 | 45% | 52% | 56% |
2026 capital spending Guidance $20–25B | $20B | $22.5B | $25B |
2027 capital spending | $22B | $32B | $38B |
| Funding and valuation | |||
2027 gap funded with new shares Rest from debt and prepayments | 50% | 50% | 50% |
Price of new shares sold Drives dilution | $170 | $240 | $300 |
EV / 2028 EBITDA Discount to software for asset wear | 8x | 10x | 12x |
Diluted shares, end-2027 Includes all convertibles | ~417M | ~417M | ~411M |
| Value per share | $75 | $192 | $357 |
| Probability weight | 25% | 50% | 25% |
| Probability-weighted value $204 | Price on Oct 9 ~$221 | Haverhill price target $205 –7% |
Our method. Nebius has no stable profits to put a P/E on, and its cash flow is distorted by prepayments and heavy building, so a standard discounted cash flow swings wildly with small changes. We use a forward EBITDA method instead:
Forecast 2028 revenue and adjusted EBITDA, the year the 2026–2027 buildout should be fully earning.
Apply an EV/EBITDA multiple that a buyer in October 2027 would pay for that year.
Subtract the net debt we expect at the end of 2027 after funding the buildout, and add the ClickHouse stake at its $1.52B carrying value.
Divide by fully diluted shares, including convertible notes, NVIDIA's warrants and the new shares we expect Nebius to sell to fund 2027.
We treat convertible notes as shares rather than debt, which is conservative on dilution. Avride and TripleTen get no separate value; their losses are already in group EBITDA.
Why these numbers. Our base 2027 revenue of $11B assumes year-end 2026 run-rate lands near the $8B midpoint of guidance and roughly doubles through 2027. That is consistent with the backlog: 36% of the $37.5B in contracted revenue, about $13.5B, falls in the two years to June 2028. The 52% margin sits just above the AI cloud's 49.7% Q2 level, allowing for scale but also for the money-losing side businesses. A 10x multiple is a discount to software because the assets wear out in about five years and must be replaced.
How sensitive the value is. In our model, each $2.5B of 2028 revenue is worth about $31 a share at 10x, and each two turns of the multiple about $44 a share at our base revenue. Across that grid, value per share runs from about $98 (8x on $12.5B) to $310 (12x on $22.5B).
Today's $221 price roughly equals our base case with 2028 revenue near $20B, or a 12x multiple on $16B. The market is already pricing in execution near the top of a reasonable range.
Today's valuation, for reference. At $221, shares plus NVIDIA's warrants are worth about $65B. Against June 30 cash and debt, that is roughly 22x the $3.0B run-rate and 20x the midpoint of 2026 revenue guidance.
What are the real risks
The biggest risk is not that demand disappears but that Nebius keeps needing outside money faster than its cash flow grows. We rank the risks by how much they would move our value per share.
Funding and dilution. 2026 capex of $20–25B is several times operating cash flow. Every dollar raised as equity at a lower price cuts per-share value; our bear case assumes stock sales near $170.
GPU pricing falls as supply catches up. Deal economics rest on $20M+ of yearly revenue per MW. A return to the ~$12M 2026 base would roughly halve the return on new capacity.
Customer concentration. Three unnamed customers were 59% of Q2 revenue. A delay, cut or renegotiation from one would hit revenue and the debt backed by those contracts.
Build delays. Revenue starts only once a site is powered, built, filled with GPUs and handed over. The owned Missouri, Alabama and Pennsylvania sites are scheduled for 2027 and later.
Chip obsolescence. Servers are depreciated over five years, but NVIDIA ships a new generation every 12–18 months, so older GPUs may earn less than planned late in their lives.
Prepayments reverse. $6.0B of deferred revenue is service Nebius owes. If a prepaying customer fails, Nebius still owes its suppliers and landlords.
AI spending cycle turns. Neocloud stocks already trade well below their highs on bubble worries; a pullback in AI lab funding would hit the smaller labs Nebius serves first.
Governance and insider selling. Founder voting control through Class B shares limits outside influence, and large October insider sales may weigh on sentiment.
Side-business losses. Avride burned $40.1M of adjusted EBITDA in Q2 alone and could need more capital.
Risks 1 and 2 drive most of the gap between our bear and base cases. Both show up in the numbers within a quarter or two, which is why they lead our watch list in the next section.
What would change our mind
Our Sell rests on one idea: the business is strong but per-share returns are capped by funding needs. These are the specific signals that would prove that wrong in either direction.
Signals that would move us to Hold or Buy
Funding shifts away from shareholders. Two consecutive quarters with no at-the-market share sales and no new convertibles, with growth funded by prepayments and secured debt like the July facility.
Run-rate beats the top of guidance. Year-end 2026 run-rate above $9B, the top of the $7–9B range, without a cut to the 40% adjusted EBITDA margin target.
Pricing holds into 2027. New deals in Q3 and Q4 2026 still averaging $20M+ of yearly revenue per MW, with short-term deals near the $40M+ the company has cited.
2027 guidance implies $12B+ of revenue. That would move our base case closer to today's bull case.
The price falls below about $160 with no change in fundamentals.
Signals that would make us more bearish
Pricing slips. New deal yields fall back toward the ~$12M per MW 2026 base, or management stops disclosing yields.
A large customer wobbles. Any of the three customers above 10% of revenue delays capacity, cuts a contract or stops prepaying.
Capacity slips. Connected power at year-end 2026 below the 800MW low end of guidance, or the 2027 owned-site openings pushed back.
Dilution accelerates. Fully diluted shares above ~420M by the end of 2027 without a matching rise in the revenue outlook.
Cash flow turns negative before capex. Operating cash flow below zero in any quarter, which would mean prepayments have stopped covering the gap.
Catalysts to watch
The next big test is the Q3 2026 report, expected in mid-November (the company has not confirmed the date).
Source: Nebius Q1 and Q2 2026 shareholder letters; dates are Haverhill estimates where the company has not confirmed them
Sources and disclosures
Every figure in this report comes from Nebius Group's own filings and releases, linked below; the analysis, model and price target are Haverhill Financial Research's own work.
Disclosures. Haverhill Financial Research is a publisher of general financial commentary. This report is not personalized investment advice and does not consider any reader's financial situation; readers should do their own research. Price targets are estimates of value, not guarantees. The author of this post holds no position in NBIS. Data as of October 9, 2026.
