Q2 did not settle the $NBIS debate. It killed the easy version of it. Revenue grew 454% year over year. AI Cloud ARR exited June at $3.0B. Adjusted EBITDA margin reached 49.7%. Those numbers prove the business works. They do not tell investors whether a share of NBIS ultimately captures the economics after hundreds of billions of dollars of infrastructure, GPUs, leases, replacement capital, debt and dilution move through the system. We rebuilt our Nebius model (first comment) from the ground up to answer that question. THE STREET’S 2027 NUMBER IS WRONG @RealJimChanos recently pointed to a real disconnect. Street expects roughly $11.5B of 2027 revenue. Nebius should operate around 1.5 GW of average connected capacity that year. Divide one by the other and you get less than $10M of revenue per connected MW, nowhere near the $20M to $25M per MW midterm contracts or $40M to $50M short-duration contracts Nebius is now signing. The numbers do not reconcile because the comparison uses the wrong denominator and the wrong timing convention. Connected facility power is not billable IT power. Year-end capacity is not average capacity. Exit ARR is not recognized annual revenue. New-contract ACV per MW is not fleet-wide revenue per MW. Our 2027 Base case models roughly: 1.50 GW of average connected power 1.059 GW of average billable IT power 1.596 GW of year-end billable IT power 1.074 GW of new billable capacity at $19.87M of ARR per MW $31.56B of AI platform exit ARR $20.01B of recognized group revenue That is approximately 74% above the Street number Chanos cited. We are not trying to reconcile our work down to consensus. We think consensus is materially too low because most sell-side models still do not properly model commissioning, PUE, customer acceptance, contract cohorts and partial-year revenue recognition. The contradiction is not inside Nebius’s contract disclosures. It is inside Street’s model. WHAT THE BUILD ACTUALLY COSTS Every megawatt in our forecast now traces to a named facility or sits inside an explicit undisclosed capacity bucket. We do not assign unidentified capacity to Oklahoma, Spain, Estonia or any other geography just because the company has employees or operating signals there. In our 2030 Base case, 4,086 MW maps to named sites. Another 2,114 MW remains undisclosed or unannounced. That uncertainty is visible instead of being disguised as fake geographic precision. Cost depends on the structure of each site. Owned greenfield requires the most sponsor capital. Build-to-suit reduces upfront cash requirements but creates lease claims. Colocation relies more heavily on partner infrastructure. The Base 2030 cost stack is approximately: $18.89M of physical infrastructure cost per connected MW $37.69M of compute, networking and storage per incremental active IT MW $59.30M of total cash build cost per incremental active IT MW Management has described the current capital stack as roughly 20% data-center implementation and 80% GPU deployment. Our 2026 bottom-up model lands at approximately 17.4% physical infrastructure and 82.6% compute. The resulting cumulative Base growth CapEx from 2026 through 2030 is roughly $285B. That is the number investors have to confront. Not because demand is weak, but because extraordinary demand still has to be physically delivered. MARGINAL CONTRACT PRICING IS NOT FLEET PRICING The bull-side shortcut is just as important to evaluate as the bear side. (and we address some very well formed analyses from @JonahLupton and @meeijer in the report) You cannot take the best contract Nebius signs today and apply it to every MW operating in 2030. Our 2027 Base cohort is composed of long-duration investment-grade contracts, core midterm contracts and short-duration scarcity capacity. The weighted headline economics are roughly $21.6M per MW. After realization adjustments, the cohort enters at $19.87M per billable MW. But each cohort expires and reprices on its own schedule. The installed fleet contains different hardware generations, contract durations, customer types and renewal economics. That is why Base realized fleet revenue reaches roughly $19.19M per average billable MW in 2030 while the newest cohort enters above $21M. Both figures are correct. They measure different things. Our Base operating model reaches: 2026 revenue: $3.2B 2027 revenue: $20.0B 2028 revenue: $45.9B 2029 revenue: $73.9B 2030 revenue: $103.2B 2030 adjusted EBITDA reaches approximately $56.1B at a 54.4% margin. EBITDA IS NOT WHAT THE OWNER KEEPS Michael Burry’s @michaeljburry depreciation criticism gets butchered by both sides. Accounting life, physical life, commercial life and economic productivity are not the same thing. $CRWV is recontracting A100 capacity into 2029 despite the architecture launching in 2020. That is strong evidence that older GPUs do not become commercially worthless after two or three years. It does not mean an old GPU retains frontier pricing forever. It also does not answer the power-opportunity-cost problem. A functioning accelerator can still deserve replacement if newer hardware produces several times more value from the same scarce, permitted and energized MW. So our model separates GAAP depreciation from normalized replacement capital. In Base: Revenue: $103.2B Adjusted EBITDA: $56.1B GAAP depreciation: approximately $40.0B Interest expense: approximately $5.0B Normalized replacement reserve: approximately $33.1B Normalized owner free cash flow: approximately $16.0B In Bear, Nebius still reaches approximately $62.8B of revenue and $29.5B of adjusted EBITDA. Normalized owner free cash flow is negative $10.3B. That is the point. A company can become enormous and still be a poor investment if maintaining the machine consumes more capital than the machine produces. WHO FUNDS THE BUILD DECIDES THE STOCK OUTCOME Our financing waterfall runs through customer prepayments, internal operating cash, secured debt, strategic-asset monetization and common equity, in that order. Across the Base forecast, cumulative funding includes approximately: $96.5B of customer prepayments $98.1B of secured and project debt $2.3B of strategic-asset monetization $12.0B of common equity Customer prepayments are not free money. They may require lower pricing, longer duration, priority access or other commercial concessions. The exact counterfactual cost is not publicly disclosed, so we stress it rather than inventing a precise answer. The share-count dispersion is where the model becomes violent. 2030 fully diluted shares: Bear: approximately 686.8M Base: approximately 445.0M Bull: approximately 402.6M Bear builds less infrastructure than Bull but issues dramatically more stock because weaker contract quality reduces prepayments and debt capacity exactly when capital becomes most expensive. That is why the same business can support radically different shareholder outcomes without requiring AI demand to disappear. HYPERSCALERS ARE A FINANCING BRIDGE, NOT THE END STATE Arkady’s position is the right one: hyperscalers are friends today and competitors tomorrow. The large Microsoft and Meta contracts provide cash flow, prepayments, investment-grade collateral, utilization and proof that Nebius can deliver at scale. Nebius is using those contracts to finance the infrastructure and platform it needs to broaden beyond them. The long-term thesis is not that Microsoft rents Nebius GPUs forever. It is that Nebius uses today’s hyperscaler economics to build Token Factory, Aether, managed inference, open-model support, enterprise relationships and an asset-light distribution layer before those customers internalize more capacity. The Base case includes 1.5 GW of partner-financed capacity by 2030, producing approximately $9.75B of revenue at a 70% margin. Delivering the same revenue through owned infrastructure would require roughly 490 MW of billable IT capacity and close to $29B of additional CapEx. That is why the asset-light model could matter so much. It is also why we refuse to value it as proven software economics before it scales. WHAT WE PUBLISHED The operating model is public. The site-by-site capacity schedule, connected-to-billable conversion, build-cost engine, contract cohorts, replacement-capital logic, financing waterfall, debt treatment, dilution mechanics and principal risks are all laid out in full. Premium members receive the Bear, Base and Bull per-share valuations, scenario probabilities, probability-weighted target, present value, required-return framework, action bands and the downloadable 38-tab workbook behind the research. Memberships are Northwise’s only revenue source. No ads, affiliate links, sponsored coverage or paid placements. We also launched the rebuilt Northwise site at It is no longer a chronological pile of articles. Research now connects through company pages, models, related theses, portfolio activity and structured filters. Premium members can access live valuation outputs and downloadable workbooks, while free accounts can follow companies, save research and receive alerts. The core Nebius question is no longer whether AI demand exists. It is whether the company can convert an unprecedented physical build into durable fleet economics without allowing debt, leases, replacement capital and dilution to absorb the value before it reaches common shareholders. That is the problem our rebuilt model is designed to solve.