Founder of Northwise. Trying to understand technology, incentives, and where the value goes.
Joined September 2026
- Tweets135
- Following20
- Followers928
- Likes569
Pinned Tweet
I remain bullish on $IREN after working through its latest earnings. The strongest evidence is coming from the terms customers are signing, especially the combination of higher prices and more money paid before service begins.
IREN disclosed a new five-year contract at about $22 million of annual revenue per IT megawatt, with the customer prepaying 90% of the associated GPU cost. It also signed a short-duration contract near $40 million per IT megawatt. Different commitments, different economics, but both help explain why delivering capacity into this market can be so valuable.
The five-year deal is particularly useful. A customer is committing to a price and supplying much of the money needed to acquire the hardware. That can reduce the financing IREN needs to arrange elsewhere and give it more room to choose how quickly it expands. The advance is then credited against future service, so it cannot be counted again as additional revenue or cash collected later.
This is where the existing contract book and the next wave of capacity need to be considered separately. Older agreements give IREN revenue visibility and help finance the build. New capacity can be sold into a stronger pricing environment. A rising share of those newer deals can lift revenue per megawatt across the fleet without assuming that Microsoft or another existing customer suddenly pays a higher price.
Our 2027 forecast illustrates how gradually that change reaches the income statement. New deals account for about 43% of billing capacity at year-end, but only 20% of revenue earned during the year in our Base case. Much of the capacity arrives late. The business can finish December with much stronger economics than its full-year results initially suggest.
I also think IREN deserves to be assessed on the business it is actually building. Its ownership of land, power and data centers gives it control over physical infrastructure and potential financing options that a pure hardware renter does not have. Mirantis adds real orchestration capabilities, including inside the NVIDIA agreement. A broader cloud offering still has to earn customer adoption and show what it contributes to pricing, reliability and retention. I have not assumed a separate software premium simply because IREN now owns the software.
That distinction runs through our work on $NBIS and $CRWV too. The sector's pricing evidence is useful, but the same rental quote cannot tell us what three different companies will earn after their contracts, operating costs and financing.
For IREN, the funding requirement remains substantial even with strong customer demand. Our Base case includes $27.15 billion of calendar 2027 capex and $5.07 billion of new equity. We allow for that dilution without relying on a share-price recovery to make the financing work. I still see an attractive opportunity, with considerable execution risk.
The first test is the initial program's customer acceptance, now targeted for January and potentially February. Later in the year, Horizon 5 and 6, Sweetwater and the Canadian conversions need to turn construction progress into paid service. Permitting and power conditions need to be followed by phase, particularly in Texas.
Our new IREN 2027 report is now live at Northwise. The financial model summary is attached below. The operating analysis is public, including the contract mix, site timing and financing. Premium includes the valuation, sensitivities and fully editable workbook.
Report link in the first reply.
ALT Northwise IREN calendar 2027 financial model summary comparing Bear, Base and Bull. Revenue: $4.93B, $5.92B and $7.51B. EBITDA before stock compensation: $3.31B, $4.21B and $5.64B. Year-end billing capacity: 526.7, 748.3 and 876 IT MW. Exit annualized revenue: $7.67B, $13.41B and $18.17B. Capex: $17.25B, $27.15B and $30.21B. Gross equity: $4.50B, $5.07B and $5.70B. The table also shows contract pricing, funding, customer credits, debt, cash and ordinary share counts. All scenarios depend on financing; exit run rate is not full-year revenue. Premium price targets are excluded.
I’m going to be around this evening aside from a run.
Ask me anything about $NBIS $IREN or $CRWV earnings.
Open to discuss any results and/or our models and plans.
I think it’s extremely important to unpack what’s been an extremely volatile couple of months.
Fundamentals across all three names remain very strong, despite prevailing narratives and macro disruptions.
Tyler | Northwise retweeted
$NBIS Our Q3 update deserves more than a revenue beat headline or a reaction to the capex number.
Customer terms improved, financing options expanded and the platform broadened.
Delivery slipped, the 2026 outlook weakened and management changed how it wants investors to measure the business.
Here is how we are working through it.
1. Start with what was actually earned.
Group revenue reached $980.8M, up 571% year over year and 68% sequentially. About $325M came from a nonrecurring initial phase of an asset-light partnership. Subtracting that item leaves approximately $655.8M, but that calculation is not a separately disclosed recurring-revenue figure. We would not annualize the entire headline quarter as compute rental income.
2. The asset-light model is beginning to produce revenue.
Nebius earns money when it enables its architecture and platform on partner-funded capacity, and as that capacity comes online. Management expects a small contribution to available capacity in 2027. The initial fee demonstrates a commercial path, but it does not tell us the eventual recurring fee yield, margin or amount of partner capacity. Those need separate assumptions.
The Q&A adds something important: Ophir Nave said the 2027 revenue guidance excludes potential asset-light and other partnership revenue. That makes the $12–13B outlook a different scope from a model that already includes those activities. Successful new partnerships could add to the outlook, but we should not extrapolate $325M every quarter or treat the upside as contracted revenue.
3. The earnings improvement needs a cash-flow check.
Group adjusted EBITDA was $465.8M. The AI cloud segment generated $518.2M at a 53.2% margin. Group continuing net income remained negative $37.5M, while adjusted net income was positive $90.9M. Operating cash flow was negative $273M and cash capex was $7.3B. These are different measures of progress, and one strong metric does not replace the others.
Dado specifically warned against annualizing Q3’s margin. The quarter benefited from point-in-time revenue, while new sites incur rent, power and payroll before billing starts. He expects owned capacity to improve the margin profile as it arrives in the second half of 2027. That supports a quarterly ramp, not a flat extrapolation of today’s 47.5% group margin.
4. Customers are accepting stronger terms.
Nebius reports more than $12B of contract value signed since the beginning of Q3. Weighted annual contract value per MW exceeded $25M on one-to-five-year deals, while average GPU-hour pricing rose more than 20% sequentially. Reserve auctions cleared over 15% above the reserve price, and short-term clusters priced above twice the reserved rate. The duration and hardware mix still determine what those prices mean for the overall fleet.
5. Upfront customer cash is increasingly important.
More than 90% of Q3 signed contract value carried a prepayment. For eligible agreements of at least a year, prepayments covered more than 100% of associated capex, including construction and networking. That is powerful evidence on those deals, not a statement that customers fund every dollar Nebius spends. Comparable one-to-three-year deal payback improved from 22 months to 19 months, excluding the timing benefit of prepayments.
6. Backlog and bookings need different clocks.
September RPO was $38.7B, versus $37.5B in June. The $12B-plus signing figure runs from the start of Q3 through the update, while RPO is a quarter-end balance affected by revenue recognition and contract scope. Marc clarified this directly: several agreements were signed after September and should appear in the following quarter’s commitments.
The high majority of the new TCV came from AI-native customers, largely existing platform customers scaling up. We cannot treat the difference between bookings and RPO as cancellations or claim enterprise already drives most new business.
7. Vineland is a real execution setback.
Nebius now expects more than half of the New Jersey Microsoft capacity delivered by year-end, with the remainder scheduled for Q2 2027. Some temporary generation was retired before the transition to permanent power; Bloom generation is operating and being expanded.
Other third-party colocation phases also moved into the first half of 2027. Contracts remaining in place protects future demand, but later billing still changes cash flow and financing costs.
8. Owned sites become more central.
Management says its owned projects remain on time or ahead and expects them to supply most future capacity. Missouri, Lappeenranta, a second Mäntsälä site and Alabama are expected to begin deployments in 2027. Their more than 3GW of eventual capacity is a multi-year program, not next year’s billable fleet. The more useful near-term question is how much becomes accepted, paid service in each quarter.
Andrey identified power and connection timing as the principal constraint. He said all 2027 equipment supply had been confirmed and technical integration work was advanced. Those are management’s assurances, not proof that every system is installed and billing. His comments on Pennsylvania were less precise about revenue timing than the four other named deployments. We should not insert a firm Pennsylvania revenue start from this answer.
9. The power pipeline continues to grow.
Nebius has more than 4GW contracted and retains its 5GW year-end target. Missouri expanded by 300MW to 1.1GW. The pipeline also includes planned capacity in Alberta, a 50MW South Carolina colocation and two India colocations. These widen the opportunity, but power rights, completed buildings and revenue-producing hardware remain separate milestones.
10. There is more to the platform than GPU rental.
Token Factory is approaching a multi-hundred-million-dollar run rate, with contract value up 17-fold sequentially. Databricks and Datadog add enterprise validation; Tavily extends Nebius into agent search and data. Initial Vera Rubin systems are live and being integrated, with CPU and inference infrastructure also expanding. The investment question is how these products improve retained customer spending and returns on capacity. We cannot count the same compute revenue again as an incremental software business.
11. The 2026 outlook moved lower despite the quarterly beat.
Revenue is now expected near the bottom of the $3–3.4B range. The call transcript records Dado narrowing revenue to $3–3.2B, while the written letter says the low end of $3–3.4B. We are preserving that distinction rather than silently treating the two formulations as identical.
The approximately 40% adjusted EBITDA margin is unchanged, but capex rises to approximately $27B. The old December ARR target was $7–9B; the new explicit milestone is above $7B in February. We would not describe that as the whole previous range simply moving two months.
12. The 2027 guide needs to be read as a connected plan.
Management expects $12–13B of revenue, $40–45B of capex and pretax breakeven in Q4, after financing costs and using five-year compute depreciation. Breakeven in the final quarter is not profitability for the full year. The capital spending also supports capacity with a revenue life beyond 2027.
Asked about the assumptions, Dado said the guidance builds in conservative deployment timing, with a larger owned-site contribution giving management more control. He sees customer demand extending into 2028. That gives us identifiable upside variables: earlier acceptance and excluded partnership revenue. It does not justify treating every favorable scenario as the expected outcome.
13. This is why we challenged the ARR communication.
We understand management’s concern that annualizing a single month can distort a growing business, especially with lumpy partnership revenue. But giving the full capex number without an exit recurring-revenue outlook leaves investors less able to connect spending to year-end scale.
A deployment beginning to bill late in the year contributes little to annual revenue while entering the following year with much greater earning capacity. Our request is for a clearly defined bridge between capacity, recurring revenue and the Q4 profitability milestone, with one-time fees separated.
14. Funding is improving, but the obligations do not disappear.
Nebius expects $17–20B of customer prepayments next year. It expects prepayments and financing against investment-grade contracts to cover 75–85% of capex.
The July $775M facility at SOFR plus 2.5% provides a concrete financing example. But expected financing is not the same as cash already drawn, and prepayments create future service obligations. They must be reconciled through cash flow, debt and future collections without being counted twice.
The Q&A also changes how we should model the contract mix. When external funding is expensive, management can allocate more capacity to core and longer contracts with larger prepayments.
When financing becomes attractive, it can retain more short-term and spot exposure. Financing costs therefore affect the revenue opportunity as well as interest expense. Holding a premium short-duration mix fixed while independently assuming unlimited cheap funding would miss this tradeoff.
Prepayments on existing capacity can fund subsequent construction too. That helps liquidity, but it does not eliminate the future service owed to customers.
15. The balance sheet and share count still matter.
Quarter-end cash was approximately $7.4B against approximately $14.2B of debt, with additional restricted cash that cannot simply be treated as freely spendable. Nebius also raised approximately $1B net through its ATM in Q3, alongside August’s convertible financing.
Dado disclosed 4.6M ATM shares issued at a weighted average $229.10, with 7.6M shares still available under that program. Available authorization is not a commitment to issue them. More capital can create value when returns justify it, but equity issuance and conversion terms belong in the per-share model.
16. The other businesses remain separate sources of value and cash needs.
Avride is expanding its autonomous fleet and delivery footprint. TripleTen reduced consumer cash burn. Toloka reported strong growth, while ClickHouse remains a minority investment. Their potential matters, but an equity stake is not cash available for construction until it can be monetized on acceptable terms.
17. Enterprise and open models broaden the opportunity, but the mix is still emerging.
Management does not publish an enterprise-versus-AI-native revenue split. It sees partnerships with Databricks, Datadog, Snowflake and Palantir as leading indicators, alongside direct enterprise adoption.
Roman sees more open and specialized models expanding the customer base that Nebius can serve. We would test that through paid workloads and expansion, not import an ecosystem token-volume statistic as Nebius market share.
18. The software explanation goes beyond a generic margin premium.
Token Factory is a per-token managed service with different service commitments from GPU-hour infrastructure. Routing, caching, optimization and post-training are intended to help customers move across the platform.
Eigen and Clarifai are integrated; Inferize brings faster model loading intended to improve utilization on spiky workloads. Roman did not give standalone product margins. An owned-infrastructure cost advantage is a thesis to model, not permission to add an unsupported software margin on top of compute economics already counted.
19. Community and permitting risk remains site-specific.
Management said it has no Texas operations and does not currently anticipate material schedule effects from the political shift in its operating states. It emphasized local engagement and answering concerns from both parties. That is management’s assessment. It should inform a site-by-site risk ledger, not become a blanket assumption that permitting cannot delay projects.
Overall, we believe $NBIS continues to execute at an extremely fast paced elite level on tasks within its control.
We do believe that releasing capex guidance and not exit 2027 was an overlooked stumble that we hope to see rectified in the coming days.
We will be releasing our updated 2027 model shortly, with new models for 2028, 2029, & 2030 coming soon as well.
We are still incredibly bullish on the company and are taking advantage of this post earnings dip.
However, we do admit that the lack of clarity here will likely take some time for the market to digest with the current negative environment in place.
Tyler | Northwise retweeted
$NBIS: Agreed.
Market does not like giving Capex but not giving ARR
just my opinion of course
@romanchernin I understand the shift to providing revenue and earnings guidance to allow traditional investors to understand $NBIS .
However, providing the full capex number, while no longer providing the full ARR guide significantly impairs most from understanding the scale associated who aren’t deeply knowledgable on the Nebius buildout and financials.
Most investors are not going to understand what that capex translates into and are going to assume it requires 45B plus to create 13B in revenue.
Please reconsider releasing the 2027 ARR guide on the call.
Many of us work on X work hard to bridge the communication gap, but that metric would make things significantly easier in tandem with your Q4 2027 breakeven announcement.
Tyler | Northwise retweeted
@romanchernin I understand the shift to providing revenue and earnings guidance to allow traditional investors to understand $NBIS .
However, providing the full capex number, while no longer providing the full ARR guide significantly impairs most from understanding the scale associated who aren’t deeply knowledgable on the Nebius buildout and financials.
Most investors are not going to understand what that capex translates into and are going to assume it requires 45B plus to create 13B in revenue.
Please reconsider releasing the 2027 ARR guide on the call.
Many of us work on X work hard to bridge the communication gap, but that metric would make things significantly easier in tandem with your Q4 2027 breakeven announcement.
The market has become deeply skeptical of AI infrastructure at a time when customers are making increasingly substantial commitments to use it.
$IREN’s latest five-year contract paired higher pricing with a customer prepayment covering 90% of the associated GPU cost. $CRWV now expects to meaningfully exceed its 2027 ARR target. Customers need the capacity badly enough to help fund its arrival, while shareholders have been selling down the companies supplying it.
That tension sits at the center of my view on the sector.
These businesses can change enormously between now and the end of next year. A facility that starts billing in October contributes three months to annual revenue, but enters the following year with twelve months of earning capacity. With several deployments arriving in succession, the business shareholders own in December can be much larger than the annual results suggest.
There is also more being built around that capacity. $NBIS and CoreWeave are developing the software, reliability and customer workflows that make a cloud useful beyond access to a particular GPU. IREN is bringing its power and infrastructure position into that same opportunity.
My conviction comes from the combination: scarce capacity, customers willing to commit capital, and cloud platforms gaining more ways to earn their business. At equity values below the revenue run rates shown here, successful delivery could leave investors looking at very different companies a year from now.
The selloff has pushed $NBIS, $IREN and $CRWV below 1x the annual revenue run rates we expect them to reach by December 2027.
At yesterday’s close, Nebius stood at 0.82x, IREN at 0.62x and CoreWeave below 0.90x, using our Northwise Base forecasts for Nebius and IREN and CoreWeave’s $30bn management target. Last night, CoreWeave said it expects to meaningfully exceed that target.
Those valuations sit alongside customers committing to years of capacity, paying higher prices and advancing substantial sums to secure equipment. The demand is showing up in contracts and customer money.
We think the market is underestimating how quickly that demand can become a much larger operating business. As new capacity starts billing, the revenue base can change dramatically within a few quarters.
The animation shows how far equity valuations have fallen relative to that opportunity.
CoreWeave $CRWV reports Tuesday and Nebius $NBIS on Thursday. Our new Northwise article proposes nine categories for AI infrastructure and a narrower definition of neocloud.
Between the October 6 and November 6 closes, Nebius lost 48.2%, CoreWeave 49.4%, IREN $IREN 47.6% and Cipher $CIFR 57.8%. The Nasdaq-100 ETF $QQQ fell 13.6% over the same period. Those companies have different customers, assets and obligations, yet for much of that period their shares moved like one trade.
Musk's October consolidation warning and last week's reports about Microsoft $MSFT tightening oversight of suppliers have kept the debate focused on whether the neocloud sector can survive its financing and delivery pressures. A project arriving late, a loan becoming more expensive and a customer finding cheaper computing elsewhere reach shareholders through different parts of the business.
We reserve “neocloud” for providers whose principal product is GPU capacity. They secure equipment from suppliers such as NVIDIA $NVDA and AMD $AMD, arrange the power and financing, and sell access. Fluidstack's managed-capacity business and Nscale's infrastructure contracts fit here. Procurement, utilization, contract pricing and the cost of capital explain much of what they earn.
When equivalent capacity becomes easier to buy, customers gain bargaining power. A merchant still owes its lenders and eventually has to replace the fleet. A protected contract or an enduring cost advantage can make that a good business, but the economics are highly sensitive to what the customer pays for access.
An integrated cloud has more ways to earn that customer's business. Storage, networking, scheduling, recovery, deployment and inference can improve how much useful work the hardware produces. Customers may stay because moving means rebuilding a production workflow, migrating data and accepting operational risk. Software can support better utilization and retention even when it is bundled into the compute bill.
That is why I place Nebius and CoreWeave among AI-native Hyperscalers. Both control infrastructure, operate substantial software stacks and deliver across major world regions. Their strategies still deserve separate treatment.
CoreWeave has built around large customer commitments that support financing and deployment. Weights & Biases, Forge and SUNK Anywhere extend it further into developer workflows and infrastructure beyond its own fleet. Nebius also uses anchor contracts, while building Token Factory, enterprise integrations and a broader customer cloud. Its partner-infrastructure model creates another route to expansion, with partners funding assets and Nebius supplying the operating platform and customer relationship.
My view of Nebius's long-term resilience rests on that combination. It can compete on the cost and reliability of completing a workload, and it can distribute its platform through infrastructure it does not have to fund entirely itself. The opportunity grows if customers keep choosing the service and partners earn enough to keep supplying capacity. Its own hardware commitments, financing and delivery still have to produce an attractive return.
IREN belongs in Emerging AI-native Hyperscalers. It starts with control of land, power and data centers, and Mirantis adds real software capability. The next step is integrating that capability into a cloud customers repeatedly use. Its latest five-year deal, with prepayments covering 90% of associated GPU capex, also shows how much contract structure can change the amount shareholders need to finance. Acquisition, customer adoption and funding are separate parts of that transition.
Lightning, Together and Crusoe Cloud sit in Full-stack AI Clouds. Our technical cutoff requires operating control of compute, core cloud services, managed workloads, integrated AI workflows and enterprise delivery. AI-native Hyperscalers add delivered geographic reach, repeated large deployments and customer breadth. A large GPU count alone tells us little about those capabilities.
We retain Hyperscalers for Amazon $AMZN, Microsoft, Alphabet $GOOGL and Oracle $ORCL, and use GP Cloud Providers for smaller or more regional general-purpose clouds such as DigitalOcean $DOCN. Fireworks, Baseten, Modal and Replicate within Cloudflare $NET sit among AI Software and Inference Specialists. Compute Marketplaces form another category because matching buyers with third-party capacity creates different inventory and operating obligations.
HPC Infrastructure and Hosting covers businesses such as $WULF, $CIFR, $CORZ, $APLD, $HUT, $RIOT and $KEEL, with Galaxy $GLXY's Helios and mixed businesses such as Bitdeer $BTDR assessed by activity. Their economics run through power, construction, rent, tenant credit and the next lease. A tenant's GPU rental price can fall while the host continues collecting contracted rent; that pressure may reach the host later through tenant health or renewal terms.
SpaceX $SPCX illustrates why the business line comes first. External GPU contracts, Grok, Cursor, captive training, Starlink and launch belong in different parts of the analysis. One corporate name can contain several economic models.
A cheaper GPU hour can squeeze a capacity merchant, lower an inference specialist's input cost and create room for a cloud to deliver more useful work at a better price. The same change can arrive much later in a landlord's cash flow. Those differences should shape what we expect from earnings and what we are willing to pay for the business.
When CoreWeave and Nebius report, newly delivered capacity, customer adoption and funding terms will tell us more about their future earnings than the neocloud label. The valuation should follow what each company can retain after paying for the infrastructure and capital its customers need.
Full article in the first reply.
ALT Northwise map of nine AI infrastructure categories. Broad cloud ecosystems: Hyperscalers and GP Cloud Providers. Integrated AI services: AI-native Hyperscalers, Full-stack AI Clouds and Emerging AI-native Hyperscalers. Capacity access: Neoclouds and Compute Marketplaces. Specialized functions: AI Software and Inference Specialists, and HPC Infrastructure and Hosting. Each category includes its customer product and company examples, separating cloud platforms, GPU capacity, software execution and physical hosting.
Nebius $NBIS can use partnerships to reach power and sites that would take time to recreate itself. CLT1 shows why that can help a cloud expand, and why the economics need to be examined on both sides of the agreement.
This is a former Bitcoin hosting property in South Carolina, where AIB plans to deliver two 25 MW AI halls for Nebius. Spartanburg County has a data-center moratorium, but CLT1 has an electricity agreement and a proposed development agreement that has passed second reading. Final approval remains ahead. That combination gives the project a more concrete route forward than a power announcement alone.
The map helps explain what is being assembled. The building has to fit alongside electrical equipment, cooling infrastructure, neighboring homes, stream buffers and drainage. The usable asset emerges when the site, power, engineering and local permission all work together. Recreating that combination elsewhere requires more than buying another parcel.
For Nebius, working with AIB means another developer can advance that physical infrastructure while Nebius builds elsewhere. It can pursue several routes to capacity at once, drawing on power positions and development work it did not have to originate entirely itself.
But the economic cost of the facility still reaches Nebius through the hosting agreement and customer prepayments. Nebius also supplies the servers and GPUs. AIB must earn a return on the infrastructure it develops, while Nebius must earn enough from customers to cover hardware, hosting, operations and financing. The value created by a scarce power location is shared across those businesses.
I would judge the arrangement by whether the time gained and revenue opportunity justify those commitments. Owning a campus gives Nebius more control and the potential benefit of later expansion. A hosting agreement can provide another route to growth, with a different allocation of costs and responsibilities. Either can produce attractive returns if the economics and delivery work.
Financing connects the two companies before the first GPU is installed. Nebius's commitment helps AIB raise project capital; Nebius then depends on AIB delivering the facility. AIB's October presentation still showed financing incomplete and estimated Q4 2027 delivery. Slippage at the host can move Nebius's revenue into a later period while its demand opportunity remains intact.
This is where my broader view of Nebius as a cloud platform meets the physical business. Software and customer relationships can help it earn more from the infrastructure it accesses. Shareholders benefit when those advantages leave an attractive return after paying everyone required to deliver the service. CLT1 gives us a specific project on which to follow that process, from local approvals through to the economics of serving customers.
A former Bitcoin hosting property in South Carolina could become two 25 MW AI halls for Nebius $NBIS. Its next county vote is on the tentative November 16 agenda, as the project works through a local data-center moratorium.
At CLT1 in Moore, AIB plans to replace the old container installation with a purpose-built facility under a 12-year Nebius hosting agreement.
Fifty megawatts can get lost beside the much larger campuses in Nebius's plans. Each completed hall would give the company another place to install GPUs and serve customers while the wider network expands. AIB develops the building, power infrastructure and cooling; Nebius supplies the computing equipment.
The county documents put that opportunity into focus. Second reading passed on November 3, and the new packet includes a proposed site plan and design standards. We see a credible route through the moratorium, with the final agreement and permits still outstanding.
AIB's October presentation estimates delivery of both halls in Q4 2027. Financing was still incomplete in that presentation, and active Nebius IT capacity at CLT1 has not been publicly confirmed. Construction, equipment installation and customer activation will determine when the contracted capacity starts earning revenue.
Our new standalone report follows the property from Bitcoin hosting to the proposed AI redevelopment, explains the power and approval path, and connects CLT1 to the wider Nebius footprint.
Full report in the first reply.
ALT Thomas & Hutton's proposed CLT1 site plan in Moore, South Carolina, dated November 6, 2026. The drawing shows the proposed building, electrical and mechanical yards, nearby homes, stream buffers and detention basins. Proposed design, not completed construction or final approval.
I recently said our $NBIS 2027 model was some of the strongest work the team had produced, yet one of our least-viewed Nebius reports.
With the fear and volatility across AI infrastructure and Nebius reporting Thursday, I wanted more people to have access to that work before earnings. We’ve opened the full report, including the analysis behind our $479.29 probability-weighted target for December 2027.
I’ve been writing about the problems with grouping these businesses together as “neoclouds.” The debate around cheaper compute is a good example of how much gets lost.
Our report works through how Rubin could allow Nebius to earn more revenue from each megawatt while customers pay less for the work their applications actually complete. The performance gains can be large enough for both sides to benefit.
That illustration uses NVIDIA’s workload-specific performance preview, so utilization and negotiated pricing will determine how much of the opportunity Nebius captures. But it gives us a much more useful way to examine the business than assuming falling compute costs automatically destroy its economics.
Nebius’s software and inference platform add another dimension: helping customers get more useful work out of that infrastructure, reducing idle capacity and giving them reasons to stay as hardware becomes more widely available.
This is the detail I want Northwise to bring to a sector where sentiment can move much faster than people can work through the underlying businesses. Financing and execution deserve scrutiny. So does the possibility that consensus is substantially underestimating what Nebius can earn.
After Thursday’s earnings, we’ll use the refreshed data to build separate Premium models for 2027, 2028, 2029 and 2030, following the expansion year by year.
For now, the report below is open to everyone. I hope it helps you go into earnings with a stronger understanding of the thesis. And to our members, thank you for supporting the research that makes releases like this possible.
With $NBIS reporting Thursday, November 12, and so much fear and volatility across AI infrastructure, we’ve made our latest Nebius 2027 report free to read in full, including the valuation and every price target our Premium members received in October.
We think the market is underestimating how much business Nebius can deliver next year. In our October 26 report, we forecast $21.86B of 2027 revenue against the $12.24B Street estimate we reviewed, a gap of 78.5%. Our EBITDA forecast is $12.49B against $6.53B, or 91.3% higher.
Most of that difference comes from the amount of capacity we expect to earn revenue during 2027. We worked through the sites, hardware deliveries, customer commitments and commissioning schedules quarter by quarter. At our expected fleet yield, the Street revenue estimate would require roughly 513 average billable MW. Our model reaches about 938 MW.
Our targets for December 31, 2027:
Bear: $340.31
Base: $471.44
Bull: $628.75
Probability-weighted: $479.29
We weight those cases 30%, 40% and 30%. Even our Bear case produces $19.07B of revenue, more than 50% above the Street estimate in the report. Our disagreement with consensus runs through the operating forecast, well before we put a valuation multiple on the business.
The build accelerates sharply toward the end of 2027. Our Base case closes the year with 1,613 MW of accepted IT capacity, but much of it earns revenue for only part of the year. Pennsylvania and Independence arrive late in our schedule. We expect a much larger business by December, while recognizing that a delayed commissioning date would remove months of revenue from the forecast.
We also follow who is paying for that capacity and on what terms. Longer contracts bring customer advances that help fund construction and equipment. Shorter contracts give Nebius more exposure to the prices customers are willing to pay for scarce compute. We expect that mix to support strong growth, with the financing burden increasing if prices weaken or customers contribute less cash upfront.
Our Base case includes $58.77B of cash capex and $12.46B of new equity during 2027. Those new shares are already included in our targets. We remain bullish after allowing for that financing because we expect the additional capacity to support far more revenue and earnings than the Street estimates in our report.
Thursday’s disclosures will help us judge whether the sites, contracts and financing are progressing as we expect. After earnings, we’ll refresh the data and publish separate Premium models for each individual year: 2027, 2028, 2029 and 2030. Members will be able to follow how the expansion develops year by year, including the revenue, funding needs and dilution along the way.
Our members received this research first, and their support allows us to open selected reports to the wider community. We hope this one gives people something useful to work through ahead of Thursday. The full report is linked in the first reply. Thanks again for supporting independent research.
ALT Northwise Nebius 2027 price targets in US dollars per share for December 31, 2027: Bear $340.31 with 30% weight, Base $471.44 with 40% weight, Bull $628.75 with 30% weight, and probability-weighted target $479.29. Full report now free to read at northwiseproject.com.
We are working on a Northwise proposal to reclassify the companies being grouped together as “neoclouds,” separating their business models, financing needs and exposure to commoditization.
Watching $NBIS get sold alongside the rest of the neocloud trade has reinforced something I think our sector needs to address: we are using one label for businesses with very different ways of making money.
The financing fears are real. Expensive infrastructure, depreciating GPUs, large customers with bargaining power, and capital commitments that arrive before revenue have contributed to selling across Nebius $NBIS, CoreWeave $CRWV, IREN $IREN and other AI infrastructure names. Those pressures affect a GPU renter, a cloud platform and a landlord differently.
A company that acquires scarce GPUs from suppliers such as NVIDIA $NVDA and AMD $AMD, finances them and rents them to customers benefits from the supply shortage. If supply catches up, customers gain alternatives and rental pricing comes under pressure. The debt, equipment costs and eventual replacement bill remain.
For Nebius, a significant part of my conviction comes from the software platform, managed inference, enterprise relationships and distribution it is building around the hardware. Token Factory already serves production workloads. Its infrastructure partnership model creates a route for Nebius to supply the cloud platform and customer relationship while partners finance and own the facilities and GPUs.
If customers choose Nebius because it helps them deploy models, run them reliably and complete useful work at a lower cost, its value can extend beyond the scarcity of a particular GPU. If partners can earn acceptable returns while paying Nebius to operate and distribute their capacity, Nebius can expand without funding every increment of hardware itself.
I see Nebius developing into an AI-native hyperscaler, with more ways to retain customers and earn revenue as raw computing becomes more competitive. That is the basis for my view that its long-term commoditization risk is meaningfully different from that of a merchant GPU renter.
Nebius still has substantial infrastructure commitments, large dedicated-capacity customers, financing requirements and delivery risks. Software must translate into adoption and better economics. Partners must make money too. Recent concerns about Microsoft $MSFT’s oversight of projects need to be examined company by company, including for Nebius. We should be able to recognize the platform opportunity while being specific about the obligations required to build it.
CoreWeave $CRWV also has substantial software and developer tools, alongside its strategy of using large customer commitments to support infrastructure financing. IREN $IREN’s physical infrastructure and Mirantis acquisition give it a different starting point. Acquiring established software brings capabilities, but integrating those capabilities into a cloud offering that customers adopt takes further work.
TeraWulf $WULF and Cipher $CIFR sell hosting infrastructure under a different set of obligations. Fireworks and Baseten compete around the software and operations required to serve models. Their customers are paying for different things, and their shareholders are exposed to different costs.
A decline in GPU rental prices can squeeze a merchant’s spread, lower an inference provider’s input costs, and leave an existing fixed hosting lease largely unchanged in the near term. Competition, contract terms and customer health determine how those effects eventually reach shareholders. Calling all of that “neocloud risk” obscures the differences we need to model.
The proposal will span established clouds such as Amazon $AMZN, Microsoft $MSFT, Alphabet $GOOG and Oracle $ORCL; AI-native platforms; specialist clouds; merchant GPU providers; inference software; and physical infrastructure. It will also distinguish the business lines and stages of development within names such as $CORZ, $APLD, $HUT, $BTDR and $KEEL.
The work will distinguish what each company delivers today from what it hopes to become. We’ll examine who funds the assets, who bears their replacement cost, who controls the customer relationship and what keeps that relationship valuable as supply expands.
We should be able to explain why Nebius and CoreWeave belong in a broader platform category while pursuing different strategies, why IREN’s software ambitions deserve attention without being treated as a finished product, and why a hosting company can have a defensible business without owning an inference stack.
Those differences should drive our revenue assumptions, financing models and valuations, even when the stocks move together.
Expect new $IREN $CRWV and $NBIS models soon from the team following respective earnings.
We are pushing hard to refresh and revamp our core AI infra research to help shine some detailed analysis on the sector.
There is still a lot of FUD and fear circling.
Stay patient, average in to the highest quality names in the space, and make sure you know the thesis backwards and forwards, whether its from our work or from other fantastic independent coverage names in the space.
I put more weight on the source of NVIDIA $NVDA's earnings growth than on the size of a buyback announcement.
In our central forecast, normalized diluted EPS rises from $8.82 in calendar 2026 to $33.77 in 2030. Revenue growth does most of that work. The margin changes are a small drag, while the lower net weighted-average share count adds to the result.
That tells me where to spend my research time. The big questions are whether customers can fund the equipment, whether NVIDIA can deliver it and how much value the company retains as competition improves.
Repurchases help the per-share outcome. They cannot carry this forecast if the operating expansion disappoints. These are forecasts, not reported results or stock-return estimates.
1
5
474
Be patient today with $NBIS and others.
Don’t over-leverage.
The fundamentals will eventually take charge of the sentiment and picture.
Earnings continue to explode around the market, and demand remains n a consistent upwards trend.
That does not mean the market has to rally back next week.
I see this as a possible bottom, but it could be a 2018 like scenario.
Yesterday definitely felt near capitulation, but that’s often where a last unexpected down move finishes.
The red is a bit blinding, but this is a phenomenal opportunity to pick up the highest names across tech.
Some names we like most here:
AI Infra:
$NBIS
$AMD
$NVDA
Fintech:
$SOFI
$NU
$HOOD
Hyperscalers:
$AMZN
$GOOG
$META
Emerging:
$MELI
$JMIA
$GRAB
Discretionary:
$CELH
$HNST
$HIMS
Bonus:
$RKLB
$NFLX
$MU
$SNDK
With the market officially in correction territory, I believe it’s prudent to keep a long term horizon and to slowly move risk on in lots.
NFA, but we are buyers across Northwise and personal allocations with some valuations hitting GFC levels.
I find the AST SpaceMobile $ASTS competitive debate more useful when I start with what the customer must do to use the service.
Broadband through a separate terminal, emergency messaging on a supported device and a carrier service reaching an ordinary phone solve different problems. Their capabilities can overlap over time, but their adoption requirements and economics are different.
The appeal of AST's approach is straightforward: keep the phone, number and carrier relationship, then extend coverage where the terrestrial network falls away. That could make adoption easier because the carrier already has the customer.
The tradeoff is that the carrier is also a sophisticated wholesale buyer. Distribution reach does not tell me what AST will earn per line, how much capacity each line will consume or how bargaining power will be shared.
I built those questions into the free operating model accompanying the report. They are more useful starting points than treating the entire satellite industry as one market.
northwiseproject.com/researc…
1
2
380
We’re going to be deep diving heavily back into the AI infra buildout here soon.
Core model updates coming to $NBIS $CRWV $IREN to start. Looking to build year by year through 2030 with an emphasis of separating fact from FUD regarding where these names stand today.
We remain very constructive on the sector as a whole with special credit to those with a healthy balance sheet, plenty of financing flexibility for the buildout ahead, and those creating full AI native integrated stacks.
Overall, we believe the current market dynamics are extremely sentiment driven and could flip “overnight” at any time.
We aren’t predictors of short term technicals, but we believe the flip will come now or over the next several months.
In the meantime, we see a multitude of opportunities across the sector and continue to average into the most high quality names.
We view this is a generational buying opportunity for the next leaders in Cloud.
The fundamental picture has not worsened and if anything valuations are more attractive across tech as a whole than they have been over many years.
We are buyers and will continue to be at these prices, unless the thesis is fundamentally proven broken.
Congrats to those who picked up $NBIS under $150.
Our smaller active internal fund is now 100% nebius for the time being.
We might not be completely out of the woods yet, but I do see a lot of narratives dissipating/reframing after today.
Don’t underestimate how much politics influences the markets at times.
Really good opportunity to return to the fundamentals with earnings coming up.
Highly recommend checking out our 2027 model to familiarise yourself at some differentiated perspectives.
We expect a very large beat of street estimates for 2027, but management will likely raise guidance for exit ARR and capex a few times before it’s all said and done to keep some stability.
An older GPU can still work perfectly and be worth replacing.
For a power-constrained operator such as Nebius $NBIS or IREN $IREN, the same power slot might support a newer NVIDIA $NVDA system that earns more after its costs.
I compare those two uses of power. Book depreciation, the old machine's remaining cash contribution and the cost of installing its replacement can point to different dates.
I see room for a replacement cycle, but its timing has to be earned by better economics.
2
5
1,212