The loudest argument about Nebius Group is not about whether AI demand is real. Both sides agree it is. Michael Burry, who disclosed a short on NBIS on 6 August at roughly $212 and added to it on 12 August at $247, is not claiming that customers will stop buying GPU hours. He is claiming that the depreciation schedules underneath the reported profits are wrong. That distinction is the whole story, and it is why this stock can trade at $248.43 after a quarter in which revenue grew 454% year over year. The bulls are underwriting contracted revenue per megawatt. The bears are underwriting the useful life of a GPU. Those are different arguments about the same fleet, and only one of them resolves on the income statement. Our working bracket into 2027 is a $390 bull case against a $130 bear case, with spot at $248.43 as of the 18 August close.

Here is what almost every piece of NBIS coverage gets wrong: it treats the Q2 print as evidence for one side. It was evidence for both. The same release that produced a 50% adjusted EBITDA margin in the AI cloud business also disclosed a $5.66bn capex quarter, and capex is precisely the input that becomes the depreciation charge Burry is disputing. A blowout quarter in a capital-intensive business does not settle a depreciation argument — it enlarges it. Every incremental gigawatt Nebius commissions makes the accounting question bigger in absolute dollars, not smaller. That is the synthesis the bull-versus-bear framing usually misses, and it is the reason a $390 outcome and a $130 outcome can both be defended off the same set of audited numbers.

Key facts

  • Spot price $248.43 at the 18 August 2026 close, down 7.6% on the session; 52-week range $62.01 to $299.86 — stockanalysis.com, 19 August 2026
  • Q2 2026 revenue $582.3m, up 454% year over year; AI cloud revenue $575m, up 514% — Nebius Q2 2026 results, 11 August 2026
  • AI cloud adjusted EBITDA margin reached 50%; ARR moved from $1.9bn to $3bn inside 90 days — Nebius Q2 2026 results
  • Capex of $5.66bn in the quarter — the line that drives the depreciation dispute — Nebius Q2 2026 results
  • Michael Burry shorted at ~$212 on 6 August, added at $247 on 12 August: “Nebius is what the top of a boom looks like” — Cassandra Unchained Substack, via Seeking Alpha
  • Vineland, NJ planning board approved Phase 2 8-1 on 18 August, clearing a gating risk on a $17.4bn Microsoft agreement — WHYY, 18 August 2026
  • Standard contract economics run $20m-$25m of revenue per megawatt; management says it has signed a first deal in the $40m-$50m per MW band — Nebius Q2 2026 earnings call

What is actually happening, and why the tape looks confused

Nebius sells AI compute capacity. The mechanism is closer to commercial property than to software: the company secures power and land, builds data centre shells, fills them with GPUs, and signs multi-year contracts against that capacity. Revenue per megawatt is the rent. The GPUs are the building. Depreciation is how you expense the building over its assumed life.

That framing explains the apparently contradictory tape. Revenue growth of 454% tells you the leasing side is working. A 7.6% drawdown on 18 August tells you the market is repricing something else — in this case the cost of financing the build and the assumptions underneath the reported margin. The stock peaked near $300 in mid-June, fell back toward the high $100s by late July, and has round-tripped to $248. That is not the price action of a settled thesis.

The funding mix matters more than most coverage allows. On the Q2 call the structure disclosed was roughly 40% of operating cash flow covered by customer prepayments, with asset-backed debt covering most of the remainder and the at-the-market equity programme acting as the marginal source of additional funding. Prepayments are the cheapest money in that stack and the most flattering to the equity story, because they arrive before the depreciation does. Asset-backed debt is priced against the same assets whose useful life is in dispute. If the depreciation assumption is wrong, the financing cost assumption is wrong in the same direction.

Nebius is not an isolated case here. The same argument runs through every neocloud balance sheet, which is why we mapped the sector-wide version of it in our AI data centre bear case on $99bn of backlog against $50bn of debt. What separates Nebius from its peers is not the accounting question — it is the quality of the counterparties and the Nvidia relationship sitting behind the equity.

What the company, the Street and the short seller are each actually saying

The three constituencies are arguing past each other, and reading them side by side is more useful than picking one.

Management made one claim on the Q2 call that dominates the bull case: it could pre-sell its entire 2027 capacity today and is deliberately choosing not to, holding capacity back to capture higher yields from shorter-duration, premium demand. That is a statement about pricing power, and it is falsifiable — if 2027 capacity ends up contracted at standard rates, the claim was wrong. The company also disclosed it sees $40m-$50m per megawatt opportunities against a $20m-$25m standard, and has signed the first deal in that band.

The sell side moved up but not to euphoria. Baird raised its target to $340 from $250. Bank of America’s Tal Liani went to $310 from $280 with a Buy rating. The 12-month average target across covering analysts sits around $272.56 — only 9.7% above spot. That is a Street that likes the business and is not underwriting a melt-up.

Burry is making a narrower and more technical claim than “AI is a bubble”. His argument is that AI infrastructure operators are assuming useful lives for accelerators that are longer than the two-to-three-year real replacement cycle, which understates depreciation and overstates earnings. Note what he is not disputing: the revenue, the contracts, or the demand. Note also the structural caveat — Scion Asset Management is closed, so there are no 13F filings behind these positions. The disclosures come through his Cassandra Unchained Substack, which means the size of the position is not independently verifiable.

“Nebius is what the top of a boom looks like.” — Michael Burry, Cassandra Unchained, 12 August 2026

There is a genuine counter to Burry from the operating side, and it is worth stating fairly: if A100s are still running economically into 2029, then a five-year schedule on newer H100 and B200 fleet is arguably conservative rather than aggressive. Second-hand utilisation of prior-generation accelerators is the empirical test, and it has so far run in the operators’ favour. This is the crux, and it will be resolved by resale markets, not by argument.

The numbers: where $390 and $130 actually come from

NBIS over 12 months to 18 August 2026. Spot $248.43, with the $390 bull and $130 bear scenarios marked. Price data: stockanalysis.com.

The $390 bull case is a revenue-per-megawatt argument. Nebius exited Q2 with $3bn of ARR, having added $1.1bn of it in a single quarter. If the premium band it has begun signing — $40m to $50m per MW against a $20m to $25m standard — becomes a meaningful share of the 2027 book rather than a single deal, the revenue per unit of commissioned capacity roughly doubles without any additional gigawatts. Layer that onto an AI cloud business already running a 50% adjusted EBITDA margin, and you reach a number above the highest published Street target of $340. That is $390, or 57% above spot. It requires the pricing claim to be true at scale and the capacity to actually come online.

The $130 bear case is the depreciation argument carried to its conclusion. If the economic life of the fleet is two to three years rather than five to six, the annual depreciation charge against a $5.66bn quarterly capex run-rate roughly doubles. A 50% adjusted EBITDA margin survives that — adjusted EBITDA sits above depreciation — but reported earnings and free cash flow do not, and the multiple the market is willing to pay compresses accordingly. The level is not arbitrary: NBIS traded through $130 as recently as March and April 2026. The bear case is not a crash to an unprecedented level. It is a round trip to where the stock sat five months ago, which the chart above makes plain.

Both brackets sit either side of spot, and that is deliberate. A prediction page whose bull case is below the current price is not a prediction — it is a stale artefact. Readers should treat any NBIS target published before the 11 August Q2 release as pre-dating the single most important disclosure in the company’s history as a listed AI cloud operator.

For context on how the same maths lands on comparable balance sheets, our CoreWeave CRWV bull and bear analysis and our IREN $84 bull versus $22 bear breakdown run the identical framework across two peers with materially different funding structures.

The regulatory and permitting tension nobody priced until this week

On the night of 18 August, Vineland’s planning board approved DataOne’s 587,980 square foot Phase 2 expansion by a vote of 8-1, after public comment ran past 11pm. Phase 1 is already under construction. The approval matters financially because dedicated capacity from that New Jersey site underpins a multi-year Microsoft agreement valued at $17.4bn — the approval removed a genuine execution risk rather than merely adding a site.

It also showed where the binding constraint on this entire sector now sits. The vote drew boos from the room and a shout of “You sold us out!” from residents who said they had been blindsided by a plan amended several times. The objections were specific: noise, water consumption, emissions, LNG storage on site, and the sheer scale of the build.

This is the part of the AI infrastructure trade that does not appear in any DCF. Nebius has secured up to 1.2 gigawatts of power and land for a planned Pennsylvania facility and talks about a pipeline running toward 5 gigawatts. Every one of those gigawatts has to clear a local planning board somewhere, in front of residents who bear the water and emissions cost and capture none of the compute upside. An 8-1 vote is comfortable. It is not a guarantee that the next one goes the same way, and the political economy is moving against permissive approvals, not toward them.

For an investor, the correct read is that permitting has migrated from a background assumption to a live gating item on the growth schedule. That is a change in the risk profile even though this particular vote went the company’s way.

What happens next

First, the depreciation question gets settled empirically, not rhetorically, and probably during 2027. The test is whether prior-generation accelerators retain economic utility past the five-year mark. If A100-class hardware is still generating revenue in 2029, Burry’s thesis fails on its own terms and the bear case decays toward the low $200s rather than $130. If resale values break down first, reported earnings across the entire neocloud cohort reset simultaneously.

Second, watch the 2027 contract mix rather than the headline growth rate. Management has staked its credibility on being able to sell 2027 capacity at premium, shorter-duration rates. The disclosure to look for is realised revenue per megawatt, not total revenue. Revenue can grow 400% while unit economics deteriorate; that combination is what a genuine top would look like.

Third, expect permitting to become a disclosed risk factor rather than an operational footnote. After Vineland, the marginal gigawatt is a political question. We would expect the next several site announcements to arrive with more community-engagement framing attached, and we would treat any site that slips more than two quarters as a leading indicator on the 5 gigawatt ambition.

Our base case is that NBIS spends the next two quarters range-bound between the March lows and the June highs while the market waits for the 2027 contract mix to become visible. The bracket resolves toward $390 if premium pricing holds at scale, and toward $130 if depreciation schedules are revised — and, critically, those are the only two disclosures that matter. Everything else is noise around a fleet whose useful life nobody has yet proven.

Frequently asked questions

What is the Nebius NBIS stock prediction for 2027?
Our bracket is a $390 bull case against a $130 bear case, with spot at $248.43 as of the 18 August 2026 close. The bull case requires premium revenue-per-megawatt pricing to hold across the 2027 book. The bear case requires depreciation schedules on the GPU fleet to be revised toward a two-to-three-year economic life. Analyst consensus sits near $272.56, with Baird at $340 and Bank of America at $310.

Why is Michael Burry short Nebius?
Burry’s argument is an accounting one, not a demand one. He contends AI infrastructure operators assume useful lives for accelerators that exceed the real two-to-three-year replacement cycle, which understates depreciation and overstates earnings. He disclosed the short on 6 August at around $212 and added at $247 on 12 August. Because Scion Asset Management is closed, there are no 13F filings and position size cannot be independently verified.

How did Nebius perform in Q2 2026?
Revenue rose 454% year over year to $582.3m, with AI cloud revenue of $575m up 514%. The AI cloud segment posted a 50% adjusted EBITDA margin, and annualised run-rate revenue moved from $1.9bn to $3bn within the quarter. Capex was $5.66bn. Full detail is in our Nebius Q2 2026 results coverage.

What did the Vineland approval actually change?
The 18 August 8-1 planning board vote approved a 587,980 square foot Phase 2 expansion and removed a gating execution risk on a $17.4bn Microsoft agreement that depends on capacity from the site. It did not resolve the underlying community objections over water, emissions and LNG storage, which remain a template for opposition at future sites.

Is Nvidia’s stake in Nebius relevant to the valuation?
Nvidia holds a 9.3% beneficial ownership position, disclosed via regulatory filing, on top of an earlier $2bn investment. It is a meaningful signal about supply access and strategic alignment, but it is not a floor under the share price. We covered the stake and its lock-up mechanics separately in our analysis of the Nvidia position in Nebius.

What is the single number to watch from here?
Realised revenue per megawatt on 2027 contracts. Standard economics run $20m-$25m per MW and management claims it has signed a first deal in the $40m-$50m band. If that premium becomes the mix rather than the exception, the bull case is live. If 2027 contracts land at standard rates, the growth is real but the multiple is not.

Sources: stockanalysis.com for price and range data; Seeking Alpha on the Burry position; WHYY on the Vineland planning board vote; Nebius on the Vineland project.

This article is analysis and information only. It is not investment advice, and it is not a recommendation to buy or sell any security. The bull and bear figures are scenario brackets, not price targets. Figures are accurate as at the 18 August 2026 close and market conditions change.