Diverging margins and unit economics in AI compute infrastructure
AI summary card
Diverging margins and unit economics in AI compute infrastructure
Bernstein compares neoclouds, colocation REITs, and emerging AI infrastructure, arguing that CIFR/WULF stand out in lease margins, financing costs, and hyperscaler customers; IREN has high margins but still lags in scale and enterprise business, while colocation REITs such as DLR/EQIX combine high-credit customers, revenue visibility, and moderate capital expenditure.
- On margins, CRWV project-level adjusted EBITDA margins are about 70-80%, and IREN’s steady-state adjusted EBITDA margin on its Microsoft contract is about 85%, but GPU and data center depreciation reduce IREN’s project-level EBIT margin to about 26%.
- On revenue/MW, neocloud providers have higher revenue intensity than traditional colocation because they sell real estate, hardware, and compute services together; EQIX remains strong in colocation MW pricing, while cloud pricing is driven by tight supply-demand conditions, enterprise and on-demand workloads, chip upgrades, and software services.
- On capital expenditure, IREN’s vertically integrated model requires about $43-45Mn per IT-MW, including about $29Mn for GPUs and $14-16Mn for data centers; emerging AI infrastructure only builds powered shells, typically about $8-12Mn/IT-MW.
- On counterparties, hyperscalers are both high-quality anchor customers and long-term competitors; Google’s backstop on Fluidstack leases can convert neocloud tenant credit into hyperscaler credit, but it does not cover construction-period risk.
Report interpretation
Overview
This report compares traditional data center REITs (EQIX, DLR, CoreSite), neocloud providers (CRWV, IREN, NBIS), and emerging AI infrastructure/bitcoin miners (CIFR, WULF, CORZ, etc.) within the same AI compute ecosystem. The core question is how different business models convert per-MW capacity into profit, revenue, financing capacity, and tolerable risk.
Core views
The report’s core view is that AI infrastructure demand is strong, but value distribution depends on who bears GPU, data center construction, customer credit, and technology delivery risk. Neocloud providers have higher revenue/MW, but face greater capital expenditure, GPU depreciation, and customer concentration risk; emerging AI infrastructure has lower capital expenditure, strong lease margins, and can reduce financing costs through hyperscaler customers or backstops; colocation REITs have stronger revenue visibility and customer credit, and do not bear GPU risk, making them a relatively balanced model between the two.
Analysis framework
The report compares company groups across five dimensions—operating margins, revenue/MW, financing structure, capital expenditure intensity, and counterparty quality—and combines AFFO multiples, EV/adjusted EBIT multiples, and project financing costs to explain valuation and risk differences across assets.
Methodology notes
Compare neoclouds, colocation, and emerging AI infrastructure by layer within the AI infrastructure stack
Neocloud providers sell more complete compute solutions and therefore generate higher revenue/MW; colocation models have more stable rents; emerging AI infrastructure improves project margins through lease structures and cost pass-through.
Measure capital intensity and depreciation burden across different models using capital expenditure per IT-MW
Vertically integrated neoclouds must invest in both data centers and GPUs, resulting in the highest capital expenditure; leased-facility neoclouds mainly bear GPU costs; colocation and emerging AI infrastructure usually do not bear the GPU layer.
A stronger credit party such as Google covers rent or termination fees if the tenant defaults, improving the credit quality of project debt
This mechanism can convert the credit risk of neocloud tenants such as Fluidstack into hyperscaler-grade collateral, helping CIFR/WULF obtain better project financing terms, but construction-period and pre-completion risks remain uncovered.
Data center REITs are mainly valued on AFFO multiples, while CRWV is valued on EV/adjusted EBIT multiples
DLR’s target price is based on 27x 2027E AFFO, EQIX on 25x 2027E AFFO, AMT on 18x 2027E AFFO, and CRWV on 28.4x 2027E adjusted EBIT.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- CIFR/WULFEmerging AI infrastructure and data center suppliers transformed from bitcoin miners
- Strengths
- High lease margins, relatively low capital expenditure, declining financing costs, and gradually gaining hyperscaler customers or credit support from Google, AWS, and others.
- Weaknesses
- The business depends more on a small number of large projects and has high requirements for construction delivery, technical specifications, and customer concentration.
- Comparison
- Compared with neocloud providers, they do not bear GPU-layer capital expenditure or technology iteration risk; compared with traditional colocation, they offer greater growth elasticity and more notable improvements in project financing.
- Risks
- Pre-completion construction risk, stringent hyperscaler delivery requirements, lease concentration, and credit support that does not cover all risks.
- IRENVertically integrated AI cloud and data center operator
- Strengths
- Steady-state adjusted EBITDA margin on the Microsoft contract is about 85%, and landlord economics plus owned infrastructure deliver strong unit economics.
- Weaknesses
- Scale and enterprise business still lag behind cloud providers such as CRWV, and capital expenditure per IT-MW is as high as about $43-45Mn.
- Comparison
- EBITDA margins are higher than CRWV’s, but GPU and data center depreciation narrow the EBIT margin advantage; capital expenditure is significantly higher than colocation and emerging AI infrastructure.
- Risks
- GPU depreciation and technological obsolescence, financing pressure, utilization volatility, and hyperscaler customer concentration and competition.
- CRWV/CoreWeaveNeocloud/AI cloud compute platform
- Strengths
- Has large-scale contracts with hyperscalers and AI labs, uses a leased-facility model to reduce self-built data center capex, and has high revenue/MW.
- Weaknesses
- Capital expenditure still depends heavily on GPU procurement, financing costs are high, and the on-demand and spot compute business is still in the early stage of expansion.
- Comparison
- Lower capital expenditure intensity than IREN; higher revenue/MW than colocation REITs, but also higher risk.
- Risks
- Customer concentration, GPU depreciation, hyperscalers being both customers and competitors, and dynamic demand being difficult to use as collateral for structured debt.
- EQIX/DLR/CoreSiteTraditional data center colocation REITs and colocation platforms
- Strengths
- High customer credit quality, strong visibility of contracted revenue, no GPU risk, and DLR/EQIX colocation EBIT margins of about 50-65%.
- Weaknesses
- AI-ready data center design pushes up capital expenditure; DLR is more exposed to hyperscalers and wholesale customers, while EQIX has smaller deployment scale and shorter average contract duration.
- Comparison
- Capital expenditure is below neoclouds and above emerging AI infrastructure; revenue/MW is lower than cloud providers, but cash flow stability is stronger.
- Risks
- Construction cost inflation, technical specification upgrades, hyperscaler customer concentration, and wholesale colocation pricing pressure.
- NBIS/NebiusNeocloud/partially vertically integrated AI infrastructure platform
- Strengths
- A more diversified customer base including AI startups, model developers, and enterprise users, with stronger infrastructure control.
- Weaknesses
- FY25 capital expenditure per MW is about $41Mn, higher than CRWV’s, and average customer credit quality is lower than in a hyperscaler-anchored model.
- Comparison
- Capital expenditure falls between CRWV at about $21Mn/MW and IREN at about $43-45Mn/IT-MW, making the model more hybrid.
- Risks
- Volatility in on-demand revenue, shorter contract duration, weaker financeability as collateral, and uncertain utilization.
Key data
- CRWV project-level adjusted EBITDA margin70-80%Represents the high adjusted EBITDA margin of neocloud projects at the adjusted EBITDA level.
- IREN Microsoft contract steady-state adjusted EBITDA margin85%Landlord economics and owned infrastructure drive a higher EBITDA margin.
- IREN project-level EBIT margin26%Data center and GPU depreciation dilute GAAP operating margins.
- CRWV project-level EBIT margin25-30%The gap versus IREN is relatively small at the EBIT level.
- DLR/EQIX colocation business EBIT margin50-65%A higher mix of retail and interconnection business supports colocation margins.
- Emerging AI infrastructure lease gross marginAbout 85%, and can approach 100% under triple-net leasesLease structures are shifting from adjusted gross leases toward triple-net leases.
- IREN Microsoft project capital expenditure$43-45Mn/IT-MWIncludes about $29Mn/IT-MW for GPUs and $14-16Mn/IT-MW for data center capital expenditure.
- Emerging AI infrastructure capital expenditure benchmark$8-12Mn/IT-MWTypically only the powered shell is built, with tenants bringing their own GPUs.
- DLR/EQIX/CoreSite capital expenditureAbout $15.6M/$16.2M/$13.6M per IT-MWFalls between neoclouds and emerging AI infrastructure.
- Contracted AI/colocation capacity signed by miners6GW, 17 deals, over $110BnRepresents about 20% of crypto miners’ planned 30GW power portfolio.
- Change in CIFR project financing cost7.125% down to 6.0%Coupon improved from the first Fluidstack-Google Barber Lake deal to the latest AWS Stingray transaction.
Impact & implications
The investment implication is that AI infrastructure is not a single-demand trade, but a repricing of different ways of bearing risk. If investors prefer high growth and high revenue/MW, neoclouds offer more upside but also higher risk; if they prefer credit quality, lease visibility, and no GPU risk, colocation REITs such as DLR/EQIX are more defensive; if they prefer improving project financing and low GPU exposure, emerging AI infrastructure such as CIFR/WULF stands out more.
Risks
- Hyperscalers are both the highest-quality customers and the most credible long-term competitors, which may weaken the long-term pricing power of neocloud providers.
- GPU capital expenditure, depreciation, and technology iteration risk may prevent high EBITDA margins from fully translating into EBIT margins.
- Neocloud tenant credit is weaker than hyperscaler credit; backstops can improve credit but usually do not cover construction-period or pre-completion risk.
- Large projects above 100MW impose higher requirements on delivery timing, technical specifications, and execution capability, and delays can affect repeat contracts and financing.
- On-demand and spot GPU compute can improve utilization and pricing flexibility, but they are highly cyclical and difficult to use as collateral for structured debt.
- Labor and equipment cost inflation may increase data center capital expenditure, and if cost pass-through clauses are lacking, project returns may be compressed.
What to watch
- Whether CIFR/WULF continue to win repeat contracts from hyperscalers such as AWS and Google, and whether backstop, lease, and financing terms improve further.
- Progress on the 200 IT-MW liquid-cooled facility IREN is building for Microsoft, and whether existing infrastructure in British Columbia and Childress can reduce follow-on GPU deployment costs.
- Changes in CRWV’s customer mix across hyperscalers, AI labs, enterprises, and on-demand customers, and whether GPU financing spreads decline.
- Changes in the customer mix of DLR, EQIX, and CoreSite between hyperscaler wholesale demand and enterprise interconnection business.
- How much of crypto miners’ planned 30GW power portfolio can continue to convert into contracted AI/colocation capacity.
- Whether next-generation chip migration and value-added layers such as software orchestration and managed services continue to lift revenue/MW for cloud providers.