Could Nebius Become a Trillion-Dollar AI Hyperscaler?

01/ Verdict

The thesis is strong. A trillion-dollar outcome is not a conclusion, it is a tail scenario that demands exceptional execution

Most of the bull case for Nebius survives contact with the filings, which is more than can be said for the average viral investment thesis. What does not survive is the leap from well positioned to trillion dollars, and the distance between those two phrases is where the actual work sits.

Nebius’s greatest strength is that it does not have to make a one-way bet on which AI model family wins. If closed models retain their lead, Nebius can benefit from demand for large centralised clusters; if a hybrid structure emerges, from multi-model orchestration and inference; and if open-weight models reach the frontier, from neutral cloud, sovereign deployments and specialised inference demand. In Morgan Stanley’s publicly available summary, Nebius appears, contrary to the viral claim, as a secondary beneficiary in all three scenarios, not just two.[1]

Strongest part of the thesisModel agnosticism, scarce power and compute capacity, and long-term customer contracts.

Most important correctionThe “approximately $50B backlog” is not the same accounting measure as $33.6B of RPO.

Biggest overstatementThe Meta backstop does not eliminate demand risk across the entire buildout.

Key variable to monitorNot revenue growth alone: ROIC after depreciation, customer concentration and software mix.

72Thesis quality
out of 100

Core conclusion of this report Nebius’s chance of becoming a “major AI-native hyperscaler” is now more than a speculative idea. Yet moving from an approximately $54 billion market capitalisation to $1 trillion would require an approximately 18.5× increase in value. That implies an extremely demanding journey towards roughly $83 billion to $167 billion of annual revenue by 2035, together with durable platform margins.

02/ Fact check

The core story is directionally right, but several figures and conclusions have been conflated

A viral post can be broadly correct and still be wrong in the places that decide whether you should act on it. Eleven claims went round; the table below puts each one next to what the filings, the official releases and the research actually say.

A Correct label means the factual statement is supported by the sources. It does not mean the investment outcome is certain.

Read the middle column as a verdict on the **statement**, not on Nebius.
Claim	Verdict	Verified or corrected reading
Morgan Stanley identifies Nebius as a beneficiary in two of three scenarios.	Incorrect	In the publicly available summary, Nebius and CoreWeave are shown as secondary beneficiaries in all three outcomes. NVIDIA is the leading beneficiary in each.[1]
Q1 2026 revenue was $399M, up 684% YoY.	Correct	Group revenue was $399.0M; core Nebius AI Cloud revenue was $389.7M, or roughly 98% of group revenue. Quarter-end ARR was $1.92B.[2]
Contracted backlog is approaching $50B.	Partly	Adding the headline maximum values of the Microsoft and Meta agreements produces approximately $46B to $49B. However, the accounting anchor as of 31 March was RPO $33.585B. Maximum contract value, RPO and future recognised revenue are not the same concept.[3]
Meta’s $15B backstop removes demand risk from the entire buildout.	Overstated	The backstop applies to unsold capacity in specified future clusters defined in the agreement. Construction, power delivery, financing, deployment timing, pricing, technology obsolescence and demand risk at other facilities remain.[4]
More than 60% of enterprises use open-weight models.	Directionally right	In McKinsey’s survey, 63% of 703 AI-experienced technology leaders and senior developers reported regular use of open-source AI at the model layer. That does not mean 63% of the entire enterprise universe; the sample and the definition of open source also include open-weight and partially open solutions.[5]
vLLM and TensorRT-LLM are Nebius’s own software layer.	Misleading	Nebius can integrate and optimise them in production inference, but vLLM is an open-source framework and TensorRT-LLM is part of NVIDIA’s ecosystem. Nebius’s defensible value lies not in owning the frameworks, but in integration, scheduling, reliability and unit economics.[6]
Nebius still owns about 28% of ClickHouse; the stake is worth $4B or more.	Outdated	The latest filing does not disclose an exact percentage. The official carrying value as of 31 March, following the January funding-round revaluation, was $1.518B. Multiplying an old ownership percentage directly by a $15B private valuation is unreliable because of dilution and preferred rights.[3]
Avride raised $375M of external capital.	Partly	The “up to $375M” total combines strategic investment and commercial commitments from Uber and Nebius. It is not all third-party capital.[10]
Nebius has a minority economic stake in Toloka.	Incomplete	Nebius lost majority voting control; according to the filing, its voting interest fell to 49%. The company now refers to a “significant economic interest”, but the exact economic percentage is not clearly disclosed in public materials.[19]
Tavily was acquired for about $275M.	Accounting difference	Media headlines used higher figures; under the SEC purchase accounting, the fair value of consideration transferred was $189.7M, with additional contingent and compensatory elements linked to performance and employment.[3]
The Eigen AI deal was worth about $643M.	Correct	Nebius announced a transaction with a signing-date value of approximately $643M, comprising cash and Class A shares; the deal closed in June 2026.[9]

03/ Scenarios

Three possible futures for the AI model market, and three different revenue paths for Nebius

The reason this thesis has legs is not that Nebius has picked the winning side of the model war. It is that the company does not have to pick at all, and the diagram below is the clearest way to see why.

What follows is not a copy of Morgan Stanley’s proprietary research chart. It is an independent analytical reconstruction based on the publicly available news summary dated 3 August 2026.[1]

The three futures, side by side

1 · Closed wins

Workloads concentrate in hyperscalers and frontier labs. Training clusters, optical networks and gigawatt-scale power become critical.

Nebius: dedicated capacity → secondary beneficiary

2 · Hybrid market

Closed models handle complex reasoning; open models handle high-volume and domain workloads. The need for routing and orchestration increases.

Nebius: Token Factory and managed inference → secondary beneficiary

3 · Open reaches frontier

Inference spreads across private data centres, sovereign cloud and the edge. Local capacity and efficient serving become critical.

Nebius: neutral open-model platform → secondary beneficiary

The takeaway: Nebius appears on the winning side of all three branches, but never as the primary winner. Important nuance: in the open-weight scenario, on-premises and enterprise hardware vendors such as Dell, HPE and NetApp may benefit more directly. Nebius does not win automatically; it must genuinely differentiate its neutral cloud and software layer.

How Nebius wins in each world

Closed world

Wins as a capacity provider

Even though customers such as Microsoft and Meta are hyperscalers themselves, power and GPU scarcity can force them to buy third-party AI cloud capacity. That is Nebius’s most visible revenue engine today.

Hybrid world

Wins as a platform

In a world where every request does not go to the same model, model selection, evaluation, routing, fine-tuning, guardrails and observability become economically valuable.

Open world

Wins as neutral infrastructure

Companies using open-weight models often want production-grade compute and serving without being locked into a proprietary model vendor. That is Nebius’s natural lane.

Why the Jevons mechanism is critical The logic in Morgan Stanley’s public summary is that falling inference costs do not necessarily reduce aggregate compute spending. As long as the economic value of an AI task remains far above its token cost, cheaper inference can create more workflows, more agent steps and broader deployment. Put differently: cost per task ↓ → task volume ↑ → aggregate compute demand ↑.[1]

04/ Open weights

Open models are becoming less an alternative to closed models and more a second engine within the same enterprise stack

The open versus closed framing is the wrong one. Enterprises are not choosing sides, they are running both, and the survey data shows a stack where open components sit underneath proprietary ones at almost every layer.

The research by McKinsey, the Mozilla Foundation and the Patrick J. McGovern Foundation surveyed 703 AI-experienced technology leaders and senior developers across 41 countries. Regular use of open-source AI was 63% at the model layer, 59% at the tools layer and 56% at the data layer.[5]

This matters for Nebius because “free weights” do not make production inference free. GPUs, memory, networking, cache management, autoscaling, reliability, security, evaluation and model optimisation still carry costs. That operational layer is precisely the value pool Nebius is targeting.

Open weights are not the same as fully open source Model weights may be accessible while training data, training code, full methodology or commercial usage rights remain less open. Adoption data is therefore misleading unless licence terms and degrees of openness are separated.

Regular open-source use across the AI stack

Adoption is highest where the model itself sits and falls away sharply at the layers that cost money to run: hosting, inference and model modification.

Models

63%

Tools

59%

Data

56%

UX / applications

49%

APIs / prompting

49%

Hosting / inference

32%

Model modifications

25%

McKinsey survey of respondent organisations, 2025. Percentages do not exclude using open and proprietary tools together across different layers.

Why an enterprise reaches for an open model

Speed

For high-frequency workloads, per-token economics, latency and throughput may favour a smaller model or an optimised open model.

Control

Regulated industries and sovereign workloads may demand greater control over model behaviour, data locality and deployment architecture.

Customisation

For domain-specific use cases, fine-tuning, distillation and model routing may be more economical than relying on a single frontier API.

05/ Full stack

Nebius’s real ambition is not merely to rent GPUs; it is to own the economic control points of the AI production stack

Renting GPUs is a commodity business with a bad ending, and Nebius knows it. Everything the company has bought or built in the past year points at the layers where margin survives a price war.

By its own description, Nebius combines a proprietary cloud software architecture with in-house server, rack and data-centre design, which is the part of an AI factory buildout that most operators specify rather than design. Token Factory provides pay-per-token and managed deployment for open models, while Tavily’s agentic search, Eigen AI and Clarifai inference IP move the stack further upwards.[6]

The five layers, from land to tokens

1

Power and data centre

Land, grid access, greenfield construction, cooling and high-density deployment.

Owned / leased / partner

2

Servers, racks, network

In-house hardware design, supply chain, high-speed fabric and cluster operations.

Nebius design

3

AI cloud platform

Compute, storage, Kubernetes, observability, security, scheduling and developer services.

Proprietary layer

4

Token Factory

60+ open models, OpenAI-compatible API, fine-tuning, autoscaling and optimised endpoints.

Managed product

5

Agentic and optimisation

Tavily search, Eigen model and system optimisation, licensed Clarifai inference IP.

Acquired / licensed

Long rows of illuminated server racks receding down a data centre hall. — The layer everyone sees. The argument is about the four layers above and below it. Illustration, AI generated.

Where could the moat emerge?

The moat may come not only from GPU access, but from operational learning that delivers more tokens per second, higher utilisation, lower latency, fewer incidents and faster model onboarding on the same hardware. If those gains improve customer unit economics, Nebius can move away from bare-metal price competition.

What remains unproven?

The company does not yet regularly disclose Token Factory’s standalone revenue, gross profit, customer count or ARR contribution. The claim that the “software layer creates high margins” is therefore a plausible strategic hypothesis, not a fact fully demonstrated by segment financials.

Framework ownership is not the same as platform value vLLM is a community-led open-source framework; TensorRT-LLM belongs to NVIDIA. Nebius’s potential advantage is not owning these tools, but running them more efficiently within its own hardware, scheduling, cloud control plane and customer-support system.

06/ Financials

Growth is extraordinary, and so is the capital requirement

Q1 2026 was the first genuinely large-scale quarter in Nebius’s ramp-up, and it is the quarter where the two halves of this company become visible at once: revenue compounding at a rate almost nothing else in public markets can match, and a capital bill six times larger than the revenue it produced.

Revenue and adjusted EBITDA rose rapidly, while capex, depreciation, debt and lease commitments expanded at the same time. Valuing the company solely on revenue growth or adjusted EBITDA would therefore miss a substantial part of the economics.

$399.0M

Group revenue, up 684% YoY

$1.92B

Quarter-end ARR, last month × 12

$129.5M

Adjusted EBITDA, group, non-GAAP

$212.0M

D&A, 53% of revenue

−$128.0M

Operating result, revenue less operating costs

≈$2.5B

Q1 capex, GPU and data centre

Why those two profit lines do not meet Adjusted EBITDA is a non-GAAP measure that adds back share-based compensation and other items, so it does not bridge to the GAAP operating result by subtracting D&A alone. $129.5M of adjusted EBITDA less $212.0M of D&A gives −$82.5M, while the stated operating result is −$128.0M. The roughly $45.5M difference is the size of those addbacks.

Q1 scale comparison

US$ billions; visual scale normalised to capex.

Revenue

$0.399B

D&A

$0.212B

Capex

$2.473B

Interpretation: Q1 capex was approximately 6.2 times revenue for the same quarter. That is not inherently negative, because the capacity is being built for future revenue. But capital productivity must sit at the centre of the investment thesis.

RPO recognition schedule

$33.585B of unsatisfied RPO as of 31 March 2026; includes estimates of variable consideration and may change.

$33.6BGAAP RPO

29% next 24 months

39% months 25 to 48

32% later periods

Most of the backlog lands beyond the next two years, so near-term revenue depends on delivery, not on the headline.

Headline maximum contract values

≈$46B to $49B

Microsoft up to $19.4B[12] plus new Meta up to $27B; varies depending on whether the earlier approximately $3B Meta agreement is included.

Accounting anchor, 31 Mar 2026

$33.585B

Unsatisfied remaining performance obligations. Subject to timing, performance constraints and estimates of variable consideration.

Do not mistake a capacity headline for revenue-generating capacity. Every stage between contracted power and billable utilisation carries construction, grid, supply-chain and commissioning risk.

According to the Q1 shareholder letter, contracted power exceeded 3.5GW and year-end guidance was above 4GW. By contrast, expected connected power was only 800MW to 1GW by the end of 2026. The economic funnel is: contracted power → connected data centre → active IT load → installed GPUs → customer acceptance → billable utilisation.[20]

Why does cash flow look deceptively strong?

Q1 operating cash flow was $2.258B, driven mainly by a $3.198B increase in deferred revenue, which is prepayment received for services to be delivered in the future. This can be an excellent financing advantage, but it is also a future delivery obligation.[3]

Why does net income not represent operating performance?

Q1 net income from continuing operations was $621.2M, but included a $780.6M non-cash revaluation gain on the ClickHouse investment. Trailing P/E or net income is therefore not a useful way to assess the core cloud economics.[3]

Accounting watch From Q1 2026, Nebius extended the useful life of server and network equipment from four years to five. The change reduced quarterly depreciation expense by $43.1M and increased net income by $41.6M. It is a lawful and disclosed change in estimate, but investors should monitor whether the accounting life remains aligned with the equipment replacement cycle.[3]

The balance sheet in three numbers

$9.3B cash

Cash and cash equivalents were $9.298B as of 31 March. Liquidity is strong, but a significant portion should be assessed alongside customer advances and new financing.

$8.45B debt

Current and non-current debt totalled approximately $8.45B. In July, the company also announced a $775M secured facility backed by GPU assets and contracted cash flows.[13]

$9.9B future leases

Undiscounted future payments under data-centre leases that had not yet commenced were approximately $9.905B, with an average lease term of 11 years.[3]

07/ Meta backstop

The $27 billion Meta agreement reduces demand risk. It does not eliminate all risk

This is the single claim that did most of the work in making the thesis go viral, and it is the one that needs the most care. A backstop is a real thing with a real perimeter, and the perimeter is narrower than the retelling.

The five-year structure has two parts: $12B of dedicated NVIDIA Vera Rubin capacity, due to begin delivery in early 2027; and an obligation for Meta to buy up to $15B of capacity in specified new clusters that cannot be sold to third parties.[4]

Risks it genuinely reduces

The risk that specified clusters remain idle
Difficulty securing debt or asset-backed financing for the buildout
Initial demand visibility for new Vera Rubin deployments
Potential customer diversification when capacity is sold to third parties

Risks it does not eliminate

Grid connection, permitting, construction and commissioning
GPU procurement, networking, cooling and on-time delivery
Financing costs, leverage and refinancing
Technology obsolescence and future pricing compression
Utilisation of clusters outside the agreement
Customer concentration around Meta and Microsoft

The strategic elegance of the structure Nebius plans to sell the capacity first to higher-margin or more diversified third-party customers; Meta then takes specified residual capacity. This can create downside protection similar to a take-or-pay arrangement. But the scope and performance obligations in the contract are narrower than the social-media claim that the “entire buildout is guaranteed”.

08/ Capital model

The key to the trillion-dollar thesis may be the shift towards an asset-light model announced in July 2026

Capital-heavy lessors do not get platform multiples, and no amount of revenue growth changes that. The July announcements are the first serious attempt to move Nebius out of that category, which makes them more consequential than any single contract.

Partner-owned infrastructure

Under Nebius’s new model, the partner finances and owns the facility and hardware, and operates the data centre. Nebius supplies the systems architecture, hardware design, cloud software, service levels and go-to-market. Economics may include revenue share, licensing fees, commissions or committed capacity.[14]

Why this matters: if the company can extend the same software and control plane into more regions with less incremental capital, its valuation multiple could move away from that of a capital-heavy lessor and towards a platform business.

Asset-backed financing

The approximately $775M senior secured facility announced in July is backed by deployed GPU infrastructure and contracted cash flows from investment-grade customers; pricing is SOFR plus 2.50%, with maturity in October 2030. The company says customer cash flow, together with the financing, covers more than 100% of the underlying deployment capex.[13]

Why this matters: if a cycle of contract, then asset financing, then freed corporate capital, then new capacity can be established, pressure from equity dilution may decline.

Post-Q1 customer-diversification signal In July 2026, Reflection AI announced that it had signed more than $1B of compute capacity with Nebius, including access to NVIDIA’s newest chips. Reflection’s work on open-source frontier models provides a concrete customer example directly aligned with Morgan Stanley’s open-weight scenario. Even so, concentration risk remains material because of the scale of the Meta and Microsoft contracts.[21]

The critical test The asset-light model’s promises of “minimal incremental capital” and “high-margin revenue” have not yet been demonstrated in segment financials. Partner-capacity revenue, gross margin and service-level performance should be monitored separately in 2027 and 2028.

09/ Portfolio

The ancillary assets carry real value, but they are not a standalone bridge to $1T

Every bull case for Nebius eventually reaches for the side assets, and they do carry real value. They also add up to a rounding error against the number this article is testing.

ClickHouse

$1.518B

31 Mar 2026 carrying value

Open-source real-time analytics database. Following the January 2026 funding transaction, Nebius revalued its investment upwards by $780.6M. This is not a liquid market value; exit timing and preferred rights matter.

Avride

Up to $375M

Uber and Nebius commitments

Robotaxis and delivery robotics. Commercial rides launched through Uber in Dallas; a monitoring specialist was present in the vehicle during the initial rollout. Significant optionality, but high execution and regulatory risk.

TripleTen

$11.6M

Q1 2026 revenue

Technology-reskilling edtech platform. Small relative to the core AI cloud; cash burn, strategic fit and potential monetisation paths should be monitored.

Toloka

49%

Voting interest after deconsolidation

AI data and human-validation platform. Nebius lost control but retained a significant economic interest. The exact economic percentage and future dilution should be monitored explicitly.

Up-stack M&A across the AI stack

February 2026

Tavily

Agentic search infrastructure. SEC fair value of consideration transferred: $189.7M; additional performance and retention components apply.

May 2026

Clarifai IP and core team

Token Factory’s optimisation capabilities were strengthened through inference IP licensing and the transfer of the core engineering team.

May to June 2026

Eigen AI

Approximately $643M signing value. Model, systems and kernel-level inference optimisation.

July 2026

Asset-light platform

Licensing and operating Nebius architecture and software stack on third-party data centres.

Strategic reading The Tavily, Eigen and Clarifai transactions suggest that Nebius is positioning for a world in which inference workloads overtake training. But for the acquisitions to create value, Token Factory revenue, customer retention and gross-profit contribution must become visible.

10/ The maths

Can Nebius reach $1 trillion? The maths is far harder than the story

This is where the thesis stops being a narrative and becomes arithmetic, and the arithmetic is unforgiving. Every route to $1 trillion demands a revenue figure that would put Nebius among the largest infrastructure businesses on earth.

$1T

Value increase required to move from today’s approximately $54.0B market capitalisation to $1 trillion.

≈18.5×

This is not merely a rise in the share price. Share count, future dilution, warrants and M&A consideration will also affect the ultimate shareholder return.

Revenue required for a $1T valuation

Illustrative terminal sales multiple; not a forecast.

Terminal value / sales (simplified)	Required annual revenue	vs 2026 guidance midpoint
6×	$166.7B	52.1×
8×	$125.0B	39.1×
10×	$100.0B	31.3×
12×	$83.3B	26.0×

Even the most generous multiple still asks Nebius to grow revenue 26 times over from the 2026 guidance midpoint.

CAGR required from 2026 midpoint to 2035

Starting point: $3.2B revenue-guidance midpoint; nine-year period.

12× case

43.6%

10× case

46.6%

8× case

50.3%

6× case

55.1%

The lower the multiple the market is willing to pay, the more brutal the growth requirement becomes. Nothing here is impossible; all of it is rare.

Three routes to the same number

Capacity case

Nebius becomes a multi-gigawatt global AI infrastructure supplier, but low differentiation and capital intensity keep the terminal multiple constrained. This scenario requires the highest revenue to reach $1T.

Hyperscaler case

Compute, storage, networking, managed services and enterprise reliability scale; the customer mix diversifies. A higher multiple becomes possible, though revenue of around $100B may still be required.

Platform case

Token Factory and partner-operated Nebius cloud generate capital-light software economics. This is the scenario most likely to justify the highest multiple, and the one with the least financial evidence today.

The real bridge to $1T Not merely announcing 4+ GW of contracted power, but repeatedly converting connected, then energised, then GPU-installed, then customer-accepted, then high-utilisation capacity into recurring platform revenue across hundreds of deployments.

11/ Bull vs bear

The strongest investment thesis, and the fault lines that could break it

Both columns below are drawn from the same filings. That is the honest position on Nebius: the bull case and the bear case are not competing readings of the evidence, they are the same evidence weighted differently.

Bull case

↑

Model-agnostic demand

Closed, hybrid or open-weight: every path consumes compute, network and power.

↑

Scarcity moat

Grid-ready land, power, GPU supply and operating capability are not easily replicated in the near term.

↑

Contract visibility

Microsoft, Meta and new AI-native customers provide revenue visibility and financing collateral.

↑

Full-stack economics

Hardware design plus cloud control plane plus inference optimisation could create a margin pool above bare metal.

↑

Asset-light replication

Partner-owned sites could globalise the Nebius platform with less corporate capex.

↑

NVIDIA alignment

A $2B strategic investment is a strong signal for supply-chain and technical collaboration.

Bear case

!

Capital intensity

The $20B to $25B 2026 capex guidance is far above the $3.0B to $3.4B revenue guidance.

!

Customer concentration

A procurement, timing or strategy change at a small number of hyperscalers could have a major effect.

!

Power ≠ revenue

Contracted MW, connected MW, active GPUs and billable utilisation are not the same.

!

Pricing and obsolescence

New GPU generations, custom silicon and inference efficiency could weaken the economics of older capacity.

!

Software moat unproven

Token Factory’s standalone revenue and margin contribution are not yet disclosed.

!

Valuation already rich

At roughly 16.9× the 2026 guidance midpoint on market-cap/revenue and about 28× quarter-end ARR, the valuation already embeds high expectations.

12/ What to watch

Which KPIs will show whether the thesis is working?

Revenue growth is the least informative number Nebius reports, because at this stage of a buildout it is almost guaranteed. These nine tell you whether the capital is actually working.

01

RPO conversion

Speed of RPO conversion into revenue, changes in cancellation and variable consideration, and customer acceptance.

02

Connected and active MW

Not contracted power, but energised, GPU-installed and revenue-generating capacity.

03

Revenue per MW and per GPU

Pricing, mix, utilisation and productivity of new-generation hardware.

04

GAAP margin after D&A

Operating margin, free cash flow and ROIC alongside adjusted EBITDA.

05

Customer concentration

Top-two customer share and diversification across enterprise and AI-native customers.

06

Token Factory attach

Inference revenue, tokens served, customer count, retention and gross-profit contribution.

07

Asset-light mix

Revenue from partner capacity, margin, capex avoidance and SLA performance.

08

Debt and dilution

Interest burden, secured-financing terms, convertible dilution, warrants and share-based M&A.

09

M&A integration

Product adoption, performance and monetisation from Tavily, Eigen and Clarifai IP.

13/ Conclusion

Nebius has a credible path to becoming a major AI hyperscaler; $1 trillion is not yet the base case

Strip out the excitement and one question decides this: whether the inference layer earns a genuine software margin, or settles into the same commodity economics that already govern where it makes sense to run AI workloads at all. Everything else is execution.

The evidence today suggests that Nebius is trying to build more than a conventional GPU lessor: in-house infrastructure design, a proprietary cloud control plane, open-model inference, agentic search, optimisation IP, long-term hyperscaler contracts and a new asset-light partnership model all connect to the same strategy. Its ability to benefit across all three Morgan Stanley scenarios is the strongest part of the thesis.

But to reach $1T, Nebius would need not only enormous capacity in the 2030s, but hyperscaler-quality economics. That means reducing customer concentration, approaching roughly $80B to $170B of annual revenue, generating high returns on invested capital after depreciation and financing costs, and proving that the software and inference layer constitutes a genuine margin moat.

The accurate formulation is therefore: not “Nebius will definitely become a trillion-dollar company”, but “Nebius is unusually well positioned for multiple directions in the model market, yet still faces enormous capital, execution and platform-monetisation tests before a trillion-dollar outcome becomes credible”.

Base caseMajor AI-native cloud; strong growth, but capital-heavy economics and a lower terminal multiple.

Bull caseGlobal hyperscaler; diversified customers, a multi-region platform and durable managed-service margins.

Moonshot caseAI operating platform; partner-owned global capacity plus high-margin inference software, and $1T becomes possible.

FAQ/ Questions people actually ask

Questions people actually ask about Nebius

What is Nebius?

Nebius Group is a NASDAQ-listed AI infrastructure company that builds and rents large-scale GPU clusters, and sells inference and platform services on top of them. It reported $399.0M of group revenue in the first quarter of 2026, roughly 98% of which came from its AI cloud business, and it ended the quarter with $1.92B of annualised run-rate revenue.

How much revenue would Nebius need to justify a $1 trillion valuation?

Between roughly $83B and $167B a year, depending on the sales multiple applied. At 12 times sales the figure is about $83.3B, at 6 times sales it is about $166.7B. Starting from the $3.2B midpoint of 2026 guidance, reaching those levels by 2035 would require a compound annual growth rate of between 43.6% and 55.1%.

Does the $27 billion Meta agreement remove Nebius's demand risk?

No. The five year structure has two parts: about $12B of dedicated NVIDIA Vera Rubin capacity due to begin delivery in early 2027, and an obligation for Meta to buy up to $15B of capacity in specified new clusters that cannot be sold to third parties. It reduces demand risk for that named capacity. It does not underwrite the rest of the buildout, and it concentrates counterparty risk on a single customer.

Is Nebius's contracted backlog $50 billion or $33.6 billion?

The accounting anchor is $33.585B of remaining performance obligations as of 31 March 2026. The figure close to $50B comes from adding the headline maximum values of the Microsoft and Meta agreements, which produces roughly $46B to $49B. The two are different measures and should not be used interchangeably.

14/ Method

Research approach and limitations

An article that spends its second section correcting someone else’s numbers owes the reader its own working. Here is where each figure came from and where the reconstruction stops.

Source priority

SEC filings and company press releases were treated as primary sources; McKinsey research for adoption data; Reuters for transaction values and the current capex context; and Blockspace for the public summary of Morgan Stanley’s 3 August 2026 research.

Morgan Stanley limitation

Because the full proprietary report and original chart are not public, the scenario diagram in this article is not a line-for-line reproduction. It is an analytical reconstruction from publicly available text.

Market capitalisation

The approximately $54.0B figure multiplies the $212.58 price on 4 August 2026 by 253.898M shares outstanding as of 31 March 2026. Prefunded warrants, options, RSUs and subsequent changes could produce a different fully diluted value.

Valuation scenarios

The 6× to 12× sales multiples and 2035 CAGR calculations used for the $1T scenarios are purely illustrative sensitivity analyses; they are not a price target or investment recommendation.

Data currency

Because Q2 2026 results are due on 12 August 2026, the latest financial period in this report is Q1 2026.[15]

15/ Sources

Primary and supporting sources

Twenty-one sources, filings first. Where a claim rests on a company statement rather than an audited number, the text says so.

Morgan Stanley says open-weight AI could expand compute and power demandBlockspace, 3 August 2026. Public summary of the three scenarios.
Nebius reports first quarter 2026 financial resultsSEC Exhibit 99.1. Revenue, adjusted EBITDA, cash flow, shares.
Unaudited condensed consolidated financial statements, Q1 2026SEC Exhibit 99.2. RPO, deferred revenue, debt, leases, ClickHouse, Tavily, accounting policies.
Nebius signs new AI infrastructure agreement with MetaOfficial release, 16 March 2026. $12B dedicated plus up to $15B unsold-capacity arrangement.
Open source technology in the age of AIMcKinsey, Mozilla Foundation and Patrick J. McGovern Foundation, April 2025.
AI model inference on Nebius: Token FactoryOfficial product page. 60+ models, OpenAI-compatible API, managed inference.
Nebius to triple capacity at its Finland data centre to 75 MWOfficial release. Site-level capacity expansion. Relabelled: the original document cited this URL as evidence of full-stack architecture and in-house hardware design, which it does not support.
Nebius announces agreement to acquire TavilyOfficial release, February 2026. Agentic-search strategy; see source 3 for the purchase accounting.
Nebius agrees to acquire Eigen AIOfficial release, 1 May 2026. Approximately $643M signing value.
Avride secures up to $375M backed by Uber and NebiusOfficial release, October 2025.
Uber and Avride launch robotaxi rides in DallasOfficial Uber release, December 2025.
Nebius signs $17.4B AI infrastructure deal with Microsoft, potential $19.4BReuters, September 2025.
Nebius raises $775M in first secured debt financingOfficial release, 17 July 2026.
Nebius introduces a business model to scale its AI cloud through infrastructure partnershipsOfficial release, 15 July 2026.
Nebius Group N.V. announces date of second quarter 2026 results and conference callOfficial release. Results due 12 August 2026.
Nebius reports higher quarterly capex on AI cloud expansionReuters, 13 May 2026. Context for the $20B to $25B 2026 capex guidance.
Nebius expands in the UK with more NVIDIA-powered infrastructureOfficial release, June 2026. £1.7B, 65MW and Revolut production use cases.
NBIS market data4 August 2026 price reference. The market capitalisation calculation is approximate and static.
Toloka deconsolidation and retained interestsSEC Q1 2026 financial statements, the same filing as source 3. 49% voting interest and retained securities.
Nebius Q1 2026 letter to shareholdersOfficial shareholder letter. 2026 guidance, the distinctions between contracted, connected and active power, and capacity expansion.
AI startup Reflection signs over $1B computing deal with NebiusReuters, 14 July 2026. NVIDIA-backed compute capacity for an open-model builder.

Legal notice This material is for general information and research purposes only; it is not investment, tax or legal advice. Forward-looking scenarios may not materialise. AI infrastructure companies carry high volatility, financing, customer-concentration, technology and execution risks.

Could Nebius become a trillion-dollar AI hyperscaler?