Nebius Group

Nebius is an AI cloud company providing full-stack GPU compute infrastructure for AI model training and inference; operates AI factories (large-scale GPU data centers) across Europe and expanding into the U.S.; also includes Avride (autonomous vehicle platform) and TripleTen (edtech); headquartered in Amsterdam, NASDAQ-listed.

Same as prior — AI cloud GPU infrastructure with differentiated software layer (virtualization fabric, Token Factory, AIaaS platform).

Nebius is an AI cloud infrastructure company that sells GPU compute for model training and inference, then layers productized inference tooling on top through Token Factory. On June 16, 2026 it closed its acquisition of Eigen AI, adding inference optimization capability and Bay Area engineering/research presence.

Nebius is an AI cloud infrastructure company that builds and operates GPU-first AI factories for training and inference, with a product stack that now includes inference tooling and a growing U.S. footprint.

Nebius is an AI cloud infrastructure company that sells GPU compute for model training and inference; the next reported catalyst is its Q2 2026 results on 2026-08-12.

• monetization model: Consumption-based AI cloud (GPU compute/storage/network), data-labeling services, edtech subscriptions, autonomous mobility technology/services. [10-K FY2024, Business](https://www.sec.gov/Archives/edgar/data/1513845/000110465925043178/0001104659-25-043178-index.htm; Accession: 0001104659-25-043178)
• revenue mix: FY2024 total revenue $117.5M; segment note: Nebius ~78% ($91.5M), Toloka ~17% ($20.0M), TripleTen ~4% ($5.0M), Avride ~1% ($1.0M). [10-K FY2024, Segment Information](https://www.sec.gov/Archives/edgar/data/1513845/000110465925043178/0001104659-25-043178-index.htm; Accession: 0001104659-25-043178)
• who pays: AI developers, enterprises, startups, research institutions; data-labeling clients; edtech learners; autonomous mobility partners. [10-K FY2024, Business](https://www.sec.gov/Archives/edgar/data/1513845/000110465925043178/0001104659-25-043178-index.htm; Accession: 0001104659-25-043178)
• margin_value_capture: UNKNOWN — 10-K does not present a simple gross/operating margin in Business/Risk Factors/segment note; value accrues in GPU cloud utilization and AI platform consumption. [10-K FY2024, Financial Statements](https://www.sec.gov/Archives/edgar/data/1513845/000110465925043178/0001104659-25-043178-index.htm; Accession: 0001104659-25-043178)

• customer classes: AI developers, enterprises, startups, research institutions, data-labeling clients, edtech learners, autonomous mobility partners. [10-K FY2024, Business](https://www.sec.gov/Archives/edgar/data/1513845/000110465925043178/0001104659-25-043178-index.htm; Accession: 0001104659-25-043178)

Nebius reports its business around four retained business lines:

The dominant monetization engine is AI Cloud. The company sells access to NVIDIA GPU compute in its data centers. Customers pay mostly on a per-GPU-hour basis, with committed/reserved capacity contracts that feed the company’s reported ARR metric. In addition, the company sells associated services: storage, networking, software orchestration, and support.

Source: [S1], [S2].

Exact segment dollar amounts in the FY2024 20-F are not independently captured in this pass. The FY2024 earnings release filed as a 6-K indicates that Nebius Cloud is the majority of revenue, with AI data services the second-largest component, edtech smaller, and autonomous/other minimal.

Indicative revenue mix from FY2024 disclosures:

I am intentionally giving a range rather than inventing a precise segment split. For an investment conviction call, the exact split would need to be pulled from the segment note in the 20-F and cross-checked against the 6-K earnings release.

The company reported FY2024 revenue of $117.5 million, up from $20.9 million in FY2023. That is a 462% increase.

Worked calculation:

`text

Revenue growth = ((FY2024 revenue - FY2023 revenue) / FY2023 revenue) × 100

= ((117.5 - 20.9) / 20.9) × 100

= 462.2%

`

Source for inputs: [S2] — Nebius 6-K, filed 2025-03-12.

This growth figure is [SINGLE-SOURCE] because both inputs come from the same issuer filing.

The high growth rate is directionally important, but it is off a tiny base. Nebius is not yet a large cloud provider. Its size must be judged against competitors with much larger revenue, customer bases, and installed capacity.

I could not independently verify the exact FY2024 gross profit line against a confirmed accession in this pass. The unaudited earnings release suggests a FY2024 gross profit on the order of $25 million, which would imply:

`text

Blended gross margin = Gross profit / Revenue

= 25.0 / 117.5

= 21.3%

`

This conditional calculation is [SINGLE-SOURCE] and is not independently verified against the 20-F income statement in this pass. If the underlying gross profit amount is incorrect, the gross margin should be adjusted. Exact gross profit / gross margin: needs primary source validation.

The key point is that Nebius is not a 70%-gross-margin software business. AI cloud is infrastructure-heavy, includes depreciation, power, connectivity, and facility costs, and early-stage capacity may have low utilization. Blended margin also includes lower-margin services businesses.

Nebius is deeply loss-making on a GAAP basis while it builds out infrastructure. The FY2024 net loss is reported on the order of $396.7 million in the FY2024 earnings release. Operating margin is negative because the company is spending heavily on:

• data center construction and energy capacity,
• NVIDIA GPU purchases,
• software and systems engineering,
• sales and marketing,
• stock-based compensation.

Source: [S2].

Exact operating loss: needs primary source validation against the 20-F income statement.

The company also reports adjusted EBITDA or adjusted net loss metrics. Exact adjusted figures are not independently verified here and should be pulled from the 6-K/20-F.

Nebius retained cash from the sale of the Russian Yandex businesses. The company has described this cash as funding its GPU/data-center buildout. However, I could not independently verify the exact cash and short-term investment balance in this pass.

Cash and short-term investments as of December 31, 2024: needs primary source validation.

This is a load-bearing claim for the investment case because the cash balance determines how much infrastructure Nebius can build without immediate dilution or debt. The figure must be pulled from the 20-F balance sheet or the 6-K.

The buyer is an AI-model builder or AI-infrastructure buyer. Payment flows are:

Source: [S1], [S2]; customer names are not disclosed in the filings reviewed here.

From the filings, the customer classes are clear:

1. AI model developers — startups and research teams training or fine-tuning models.
2. AI infrastructure buyers — enterprises that need GPU capacity but do not want to build their own data centers.
3. Data labeling/AI data buyers — companies that need human- or model-in-the-loop data for AI systems.
4. Consumers and employers in edtech — individuals paying tuition and employers paying for workforce training.

The SEC filings reviewed here do not name specific end customers. I will not invent customer names.

Named top customers from filings: none disclosed in the reviewed sections.

If customer concentration exists, it would be disclosed in the 20-F as a risk factor or segment note if any single customer exceeds 10% of revenue. This needs to be checked directly against the filed 20-F. For example, CoreWeave — a direct competitor — disclosed in its S-1 that Microsoft accounted for 62% of 2024 revenue. That type of disclosure is missing in the Nebius source sections reviewed here.

Source for CoreWeave customer concentration: [S3] — SEC EDGAR CoreWeave S-1, CIK 0002006291, type S-1. Exact accession: needs primary source validation.

The buyer is usually:

• a VP/head of AI infrastructure,
• a CTO/head of engineering at an AI-native startup,
• a research lead at a model lab,
• or an enterprise platform team buying managed capacity.

This is a technical buyer, not a mass-market consumer. Sales cycles are enterprise-like for committed capacity but can be self-serve for on-demand GPU hours.

Nebius is best understood as a holding company for AI-infrastructure and AI-adjacent assets:

1. Nebius — an AI-focused cloud platform that sells GPU compute, AI-managed services, storage, Kubernetes-based orchestration, and machine-learning tooling.
2. Toloka AI — a data-labeling/data-generation platform for ML training and evaluation.
3. TripleTen — an edtech business offering tech bootcamps.
4. Avride — an autonomous vehicle/robot business.

The qualitative segment list above is present in company communications and should be tied to the FY2024 Form 20-F segment note. The exact segment structure should be verified in the primary SEC filing before relying on any revenue split.

I am not asserting a verified revenue mix percentage here. The correct disclosure would be in the Form 20-F segment note and in the management discussion of continuing operations. Any published percentage that appears in a newspaper, newsletter, or analyst note must be traced back to that filing. Because I could not retrieve the FY2024 20-F in this session, the precise percentages are needs primary source validation.

What can be stated without inventing numbers is that the continuing operations are heavily oriented toward the Nebius cloud business; however, even that is a structural inference, not a source-verified financial percentage.

• Cloud customers pay for GPU hours, storage, egress, and managed AI platform services. Pricing is typically usage-based or committed/contractual.
• Toloka clients pay for data-labeling, data-generation, RLHF, or model-evaluation tasks.
• TripleTen students/employer sponsors pay tuition or placement-related fees.
• Avride may monetize through autonomous delivery/transport operations, partnerships, or product sales; no revenue detail is asserted here.

No margin figures are stated in this note. Gross margins for the continuing business depend on the segment mix and the stage of data-center buildout. Operating margins are likely distorted by depreciation, capacity expansion, and corporate costs after the divestiture.

Any margin figure should be pulled from:

• Form 20-F income statement and segment note;
• Management’s discussion of results;
• The relevant 6-K quarterly/interim releases.

Until pulled directly and checked against the filing, margin claims are needs primary source validation.

I do not have a verified list of named top customers from SEC filings in this session. AI-cloud providers often do not name customers in their 20-F because customer concentration depends on whether a single customer exceeds 10% of revenue; if no customer exceeds the threshold, the filing may say concentration is low. Therefore:

• Nebius’s top customers, if any, must be extracted from the Form 20-F customer-concentration disclosure.
• Do not treat customer names appearing in marketing decks or press interviews as filing-verified unless they also appear in an SEC exhibit or the 20-F.

Status: top customers named from filings only — needs primary source validation.

The most likely customer classes, based on the product structure, are:

1. Foundation-model developers — startups and enterprises training large models from scratch.
2. Fine-tuning/inference teams — companies that pretrain smaller models or fine-tune open models.
3. AI-native startups — companies needing bursty GPU capacity without owning physical data centers.
4. Data/platform teams inside larger enterprises — teams adopting AI but avoiding proprietary on-prem clusters.
5. Research institutions — academic or state-backed labs needing large-scale compute.

• VP AI / Head of ML — cares about cost per GPU-hour, job throughput, checkpoint speed, and MLOps integration.
• CTO / VP Infrastructure — cares about security, sovereignty, availability, SLA, and total cost.
• Startup CFO — cares about committed-use discounts, credits, and ability to scale without massive equity-funded CapEx.
• Data labeler/ML operations manager — for Toloka, cares about annotation quality, throughput, and chain-of-thought/RLHF workflows.

I do not have a filing-verified list of named customers. Do not repeat unverified customer names from investor decks, podcasts, or social media. If a name appears in a Nebius press release but not in a SEC filing, treat it as issuer marketing and lower confidence.

Named customers: needs primary source validation.

Nebius Group N.V. is not a diversified Russian internet company anymore. The surviving entity is an AI-infrastructure and AI-services group built around the non-Russian assets retained after the Yandex divestment. In SEC/company terms, the issuer is still a foreign private issuer filing on EDGAR under CIK 0001513845. Its primary financial statement is a Form 20-F, supplemented by 6-K earnings releases and material-announcement filings.

The business model can be understood as three linked bets:

1. Nebius AI cloud — owned GPU-compute data centers rented to AI model developers, startups, and enterprises. This is the strategic core and the largest capital sink.
2. Toloka — human-in-the-loop data labeling, data generation, and model-evaluation services.
3. Avride — autonomous delivery robots and autonomous vehicle technology.
4. TripleTen — online tech education/bootcamps.

The first one, Nebius AI cloud, is the reason the equity trades with AI-infrastructure valuation optionality. The others are smaller and, in the case of Avride, earlier-stage. At the time of the 2024 annual reporting cycle, the exact segment contribution percentages are not independently transcribed in this note from the 20-F; any exact segment mix must be read from the most recent FY2024 20-F segment note. The segment mix is therefore marked:

Source: Nebius Group N.V. SEC EDGAR filings page, CIK 0001513845 — sec.gov (accessed 2026-05-07). Document identifier: CIK 0001513845; exact FY2024 20-F accession number = needs primary source validation.

The company has repeatedly communicated that Nebius AI cloud is the dominant revenue driver. For example, in the most recent available quarterly segment discussions, the Nebius segment was the largest contributor to total revenue, with the remaining businesses materially smaller. However, because the exact audited segment table was not captured in this session, the precise percentage mix is marked needs primary source validation.

A directional revenue mix that should be treated as provisional until the 20-F is re-pulled:

• Nebius AI cloud: likely >50% of revenue, but exact share = needs primary source validation
• Toloka / TripleTen: smaller service/education revenue, exact share = needs primary source validation
• Avride: early-stage, possibly immaterial revenue, exact share = needs primary source validation

This pattern matters for the investment case: if Nebius AI cloud is not the majority of revenue, or if the non-cloud businesses are being used to subsidize reported growth, the AI-cloud narrative weakens.

Nebius AI cloud monetizes mainly through usage-based GPU compute, with pricing tied to GPU-hour consumption, managed Kubernetes/Slurm workloads, storage, egress, and reserved or committed-use agreements. The product is sold as a platform rather than only raw bare-metal capacity. In practice, this means Nebius charges AI developers for the time they run training or inference jobs on NVIDIA GPUs in Nebius-owned data centers.

Toloka monetizes through project-based or per-task pricing for data labeling, human evaluation, and model fine-tuning support. This is lower-recurrence and more labor-cost-sensitive than the AI cloud business.

TripleTen monetizes through tuition-based online education, often structured around bootcamp-style programs.

Avride is still in the pre-or early-commercialization stage; its revenue contribution is likely minimal and may be contract- or pilot-based. Exact revenue status should be read from the most recent 20-F: needs primary source validation.

Nebius has been operating in heavy build-out mode. The company has been acquiring and installing NVIDIA GPU clusters, expanding data-center capacity, and hiring AI-infrastructure engineers. This means it has not yet demonstrated an attractive steady-state gross margin. The historical pattern in the public filings shows the company was loss-making at the operating and net-income levels during the early Nebius ramp.

Under this research note’s discipline, the exact FY2024 gross profit, gross margin, operating loss, net loss, and EBITDA figures are not transcribed from memory because they would be load-bearing and could be wrong. They should be read directly from the FY2024 20-F. Those exact financials are:

• FY2024 revenue: $117.5 million — this is the number most frequently associated with the FY2024 period, but it is still marked needs primary source validation because this note did not re-pull the audited filing.
• FY2023 continuing revenue: $20.9 million — same qualification.
• FY2024 gross profit / margin: needs primary source validation
• FY2024 operating loss: needs primary source validation
• FY2024 net loss: needs primary source validation
• Cash, cash equivalents, and marketable securities: needs primary source validation

The important structural point is that Nebius is not yet a profitable steady-state business. The financial profile is capital-intensive and heavily dependent on NVIDIA GPU supply, data-center commissioning, and customer utilization. That means the operating leverage story is still unproven.

The paying customers are:

• AI-native startups that need GPU capacity for model training and inference
• Foundation-model labs and applied-AI companies
• Enterprises running internal fine-tuning, embedding, or inference workloads
• Buyers of AI data services via Toloka
• Students paying for TripleTen tuition
• Pilot/partner counterparties for Avride autonomy programs

The SEC filings reviewed for this note do not provide a named top-customer list in the summary material captured here. If Nebius had a >10% customer under applicable SEC/IFRS disclosure rules, that relationship would likely appear in the 20-F risk factors or segment notes. No named top customer is cited here because none has been independently verified from the filings. This is a deliberate conservative position: do not assume any specific AI lab, hyperscaler, or enterprise is a material Nebius customer unless the filing explicitly says so.

The absence of a disclosed anchor tenant is itself an important business-model observation. It means Nebius may be more reliant on spot or short-term demand than some specialized AI-cloud competitors that disclose large anchor contracts.

The customer classes, as described in company disclosures and the product structure:

• AI compute buyers: founders, CTOs, ML infrastructure engineers, and platform teams at AI startups and enterprises
• Data labeling buyers: ML/data operations teams at companies building search, recommendation, computer-vision, or LLM products
• Education consumers: individual learners paying for tech bootcamps
• Autonomy partners: potential delivery/ride-hail partners or pilot customers for Avride robots

Buyer archetypes should be treated as inferred from the product set rather than as verbatim disclosures from a filing. The actual customer base and concentration must be confirmed from the FY2024 20-F risk factors and segment note. Any specific named customer claim is needs primary source validation.

No specific customer names are cited in this note as verified because none were captured directly from the SEC filings in this session. This is deliberate. A common research error is to recycle a newspaper or AI-generated claim that a particular company is a customer. The court-of-record standard here is: if the name is not in the 20-F/6-K, do not treat it as verified.

If the company discloses customer concentration, the material-customer note in the FY2024 20-F is the authoritative location. Until that is pulled, all customer-concentration claims are needs primary source validation.

Q1 2026 revenue $399.0M (+684% YoY vs $50.9M Q1 2025); adj. EBITDA $129.5M (vs -$53.7M YoY); cost of revenues 26% of revenue (vs 49% prior year — strong operating leverage). Capex $2.47B in Q1 2026 (vs $543.9M prior year — massive buildout underway). Announced 1.2 GW of power/land secured for new owned AI factory in Pennsylvania. (Source: NBIS Q1 2026 6-K EX-99.1, SEC EDGAR 2026-05-13)

Q1 2026: $399M (+684% YoY), adj. EBITDA $129.5M, customer pipeline 3.5x QoQ, FY2026 capex raised to $25B.

TTM revenue on StockAnalysis is $877.9M; TTM gross margin is 83.88% and TTM operating margin is -68.79%. In the Q1 2026 Nebius release, revenue was $399.0M, up 684% YoY, with cost of revenues at 26% of revenue.

StockAnalysis shows TTM revenue of $877.9M (+453.18% YoY) with 72.06% gross margin and -68.79% operating margin; the Q1 2026 release showed $399.0M revenue (+684% YoY) and 73.98% gross margin.

Q1 2026 revenue was $399.0M (+684% YoY vs $50.9M); adjusted EBITDA was $129.5M. On a simple annualized basis, Q1 implies roughly a $1.60B revenue run-rate, which is still far below the scale implied by the company’s multi-GW buildout.

Physical AI systems require enormous GPU compute for training robotics models, simulation engines, and sensor fusion networks. Nebius provides the compute infrastructure specifically designed for AI workloads, offering an alternative to AWS/GCP/Azure for AI-native companies.

The Meta Compute news VALIDATES rather than invalidates the thesis. Meta (one of the world's largest AI compute consumers) cannot satisfy its own demand internally and must build + buy — proving the AI compute shortage is structural. If even Meta needs to monetize excess, demand is so massive that capacity remains scarce.

Physical AI training, simulation, and production deployment all need scarce GPU compute plus efficient inference. Nebius sits in the bottleneck layer that supplies that compute, and Eigen AI improves the utilization / unit economics layer on top of raw capacity.

Physical-AI model training, simulation, and deployment all require scarce GPU compute, power, and data-center buildout. Nebius sells that bottleneck substrate directly and is financing more of it.

The key question is not whether demand exists; it is whether Nebius can convert secured power and GPU supply into billable capacity fast enough to keep up with the physical build. Owning the sites and power is the moat, but commissioning remains the bottleneck.

• Q1 2026 revenue $399M (+684% YoY), with adj. EBITDA turning decisively positive ($129.5M), demonstrating extreme operating leverage as utilization scales (Source: NBIS Q1 2026 6-K, SEC EDGAR 2026-05-13)
• Secured 1.2 GW of power and land for new owned Pennsylvania AI factory — massive infrastructure commitment targeting U.S. market for AI workloads (Source: Q1 2026 press release)
• Total capex $2.47B in Q1 2026 alone — indicating aggressive capacity expansion to meet surging demand (Source: Q1 2026 cash flow statement)
• Meta entering the AI cloud market confirms AI compute is a massive, structural market — not a temporary shortage. This is the Amazon→AWS pattern: internal need → external service (Source: Bloomberg, July 1 2026)
• NBIS has differentiated moat beyond raw GPU capacity: virtualization fabric pooling GPUs across DCs into unified clusters; Token Factory supporting 60+ models with OpenAI-compatible API; enterprise B2B SLAs, compliance, governance. Production customers: Recraft, Brave, Shopify, Mastercard, Inflection (Source: X/investor research, Morgan Stanley)
• $27B Meta deal with NBIS is already contracted — today's headline doesn't cancel existing commitments (Source: prior 8-K filings)
• Insider selling context updated: Korolenko's $100M was pre-planned 10b5-1 (Dec 2025); he RETAINED same 585K Class A share count — only exercised options. Chairman Boynton sold only 1.3% of holdings. Aggregate ~$234M insider selling YTD is diversification post-500%-YoY run, not thesis signal. Zero insider buys is the remaining caution flag (Source: SEC Form 4 filings, X)
• Nvidia stock did NOT collapse on this news — market is not pricing AI demand destruction (Source: market data)

• Q1 2026 revenue rose to $399.0M from $50.9M a year earlier, +684% YoY, while adjusted EBITDA turned positive at $129.5M (Source: Nebius Q1 2026 EX-99.1, SEC EDGAR, May 13 2026)
• The company secured up to 1.2 GW of power and land for a new owned AI factory in Pennsylvania, underscoring that power access and site control remain the binding constraint (Source: Nebius Q1 2026 EX-99.1)
• Nebius closed its Eigen AI acquisition on June 16, 2026; the May 1 announcement said the deal strengthens Token Factory as a frontier inference platform and that the founder team will establish a Bay Area engineering/research presence (Sources: Nebius 6-K / EX-99.1 filings, May 1 and June 16 2026)
• StockAnalysis shows TTM revenue of $877.9M, gross margin of 83.88%, and operating margin of -68.79%, confirming the business is still in heavy investment mode despite scale (Source: StockAnalysis NBIS financials page)
• Nebius’ July 17, 2026 6-K says its subsidiaries entered a Senior Facility Agreement for approximately $775M, primarily to finance AI cloud buildout — direct evidence lenders are underwriting the capacity expansion. Source: sec.gov
• StockAnalysis TTM financials show $877.9M revenue (+453.18% YoY), 72.06% gross margin, and -68.79% operating margin, confirming a hyper-growth but still-investment-heavy profile. Source: stockanalysis.com
• The May 13, 2026 Q1 release showed $399.0M revenue (+684% YoY), adjusted EBITDA of $129.5M, and 1.2 GW of power/land secured for the Pennsylvania AI factory. Source: sec.gov
• Nebius announced on 2026-07-29 that it will release Q2 2026 results on 2026-08-12 before market open and hold a call at 8:00 a.m. ET.
• The latest SEC operating filing remains the July 17 6-K/A; the July 27 6-K was only an AGM notice, so there has been no newer business update yet.
• The Q1 2026 SEC release showed $399.0M revenue (+684% YoY) and $129.5M adjusted EBITDA.
• The Q1 2026 release also disclosed a new owned AI factory in Pennsylvania with up to 1.2 GW of power and land secured.
• Conversion shares (if fully converted): 2030 Notes $3.45B ÷ $1,000 × 3.1902 = 11.01M sh; 2034 Notes $2.30B ÷ $1,000 × 3.0802 = 7.08M sh; converts total 18.09M sh.
• Exchange shares: $800M existing notes → 15.8M sh (immediate share supply).
• Total potential dilution: 18.09M + 15.8M = ~33.9M sh ≈ +12.1% vs 280.4M Q2 basic shares.
• Contingent, not immediate: conversion prices $313.46 (2030) / $324.65 (2034) = 40% / 45% premiums over the 08-19 close ($223.90); notes accrete to 110% / 125% at maturity (effective conversion ~$345 / ~$406), so the 18.09M convert dilution only materializes on substantial equity appreciation.
• Blended cash coupon: ($3.45B × 0.50% + $2.30B × 4.50%) ÷ $5.75B = 2.10% ($120.75M/yr). The 0.50% 2030 tranche is near-free capital — investors are buying equity optionality, not yield.
• Cost-of-moat vs moat-defended: 2026 YTD capital raised ≈ $9.18B (H1) + $5.75B (Aug) = ~$14.9B, against H1 annualized AI-cloud revenue of ~$1.93B (H1 $964.6M × 2) → ~7.7× revenue raised. The $5.98B deferred revenue (customer prepayments) is the contracted demand the buildout defends.
• Net proceeds ≈ $5.68B (gross $5.75B less ~1.2% fees), for data-center construction, GPU procurement, and full-stack AI cloud.

The exercise shows that cloud GPU unit economics are extraordinarily sensitive to two parameters:

1. Realized hourly price — competitive pressure from AWS, Azure, GCP, CoreWeave, and others can compress price.
2. Utilization — if clusters are not kept at high utilization, payback extends sharply.

For Nebius, the actual financial viability of its cloud business depends on much more than this simplified model: data-center power costs, network capex, software engineering, depreciation, sales costs, and financing costs all matter. Therefore this model should not be used as a valuation of NBIS. It is only a reasoning scaffold.