Neoclouds: The Backlog Quality Test
Why AI infrastructure durability now depends on monetizable MW, contract structure, and renewal discipline
Note: This report builds on our April framework separating the different business models emerging across neoclouds, GPU clouds, and power/infrastructure landlords. Later this week, we will publish a more technical SWOT across hyperscalers and neoclouds, focused on how competitive capabilities map to AI workloads, cloud architecture, and broader CSP trends.
We continue to watch the neocloud and AI infrastructure landlord landscape because these companies remain one of the better market proxies for hyperscaler urgency. The hyperscalers are the demand center, but the neoclouds, GPU clouds, and powered-shell landlords show us how that demand is being translated into contracts, financing structures, power commitments, and monetizable capacity. They also give us a cleaner way to track the dollar value being assigned to each constrained MW across the AI infrastructure stack.
There was a brief “thank goodness for the bitcoin miners” phase of this cycle. The hyperscalers were running into a simple problem: AI demand was moving faster than their ability to bring new power online. Bitcoin miners, for all the volatility of the legacy business model, had already done some of the hardest work. They had sites, grid relationships, substations, and access to large blocks of power. That made them useful as a time-to-power accelerant, and it helped explain why the first wave of investor enthusiasm focused so heavily on who had capacity ready to convert.
That phase rewarded access. Primary among them things like, capacity, power, sites, and deployment speed carried valuations because scarcity cleared the market. For a stretch of time, the size of a provider’s backlog was a reasonable proxy for the quality of its business. That proxy is becoming less useful. As more capacity comes online and more contracts are signed, the relevant question is moving away from the size of the commitment and toward the durable economics behind it. The input that constrains the industry is power, so backlog quality increasingly has to be measured against the economic value created per constrained MW.
That is what we mean by monetizable MW. The concept is simple enough: how much durable value does a contract produce per constrained MW, and who carries the power, financing, hardware, and renewal risk required to produce it? Headline backlog tells us demand exists. It says much less about how much of that demand turns into residual value after the asset is built, financed, depreciated, and renewed.
The spread is large enough to matter. Powered-shell landlords monetize roughly $1.2M to $2.3M of revenue per MW each year and generally transfer GPU ownership and obsolescence risk to the tenant. GPU clouds and full-stack neoclouds monetize several times more, often $7M to $13M per MW, while keeping more of the financing intensity, utilization risk, GPU depreciation, and renewal exposure that come with owning the compute. That is one of the more important takeaways from the work: the same constrained MW can support a roughly 5x revenue-density spread depending on where the company sits in the stack. Higher revenue density can be attractive, but it usually comes with more of the risk.
There is also a power-acquisition cost sitting beneath the entire discussion. Our research points to roughly $4M of generation capex per usable MW before transmission, distribution, shell, cooling, network, or GPU clusters are included. The exact figure moves by region, power source, interconnection status, and timing. The larger point is that the MW itself is expensive long before it earns anything.
That cost links the power debate directly to backlog quality. Every provider in this category is trying to turn constrained MW into durable revenue before one of three clocks runs out: the customer contract, the useful life of the asset, or the financing behind it. A model tracks well when revenue per MW, contract duration, and asset life stay aligned long enough to clear the capital stack. It gets harder when scarce power is tied to revenue that fades before the leases, GPUs, or debt are paid down.
We also think the distinction between firm MW and flexible MW will become more important. Some AI workloads can shift across time or geography in ways traditional enterprise workloads cannot. That creates room for demand response, interruptible operation, workload shifting, and dynamic token pricing. For operators that can run power-aware, flexibility becomes a potential operating advantage. Power stops being a passive input and becomes an asset that can be managed.
Applied across the major names, this does not produce one clean ranking. Power Landlords screen strongest on duration and hardware-risk transfer. IREN owns the scarcest asset in the cycle, deliverable power, and still has to prove a durable compute premium beyond its first hyperscaler-backed deployment. Nebius has the most interesting contract design among the platform names, with a disclosure gap that keeps its optionality from becoming a conclusion. CoreWeave has the strongest commercial validation in the group and the clearest mismatch between long data center leases, shorter customer contracts, and GPU depreciation.
The first expansion signals are encouraging. Several large customers have added capacity, extended commitments, or signed new multi-year anchor agreements, which supports the view that external AI infrastructure remains strategically useful while internal capacity catches up. We still separate expansion during scarcity from renewal discipline once customers have more choices. The current wave proves demand, urgency, and willingness to use external partners. The more important test is whether those customers renew at attractive pricing once internal capacity improves, custom silicon is more widely deployed, and prior-generation GPUs need a second economic life in inference.
One distinction runs through the whole analysis. Stronger contracts reduce business-model risk without necessarily improving the equity claim. A high-quality customer commitment can still sit inside a capital structure that leaves limited residual value for shareholders. That is why we separate the quality of the customer commitment from the quality of the shareholder claim throughout.
The full report turns this into a company-by-company assessment of CoreWeave, Nebius, IREN, and the powered-shell landlord cohort, including Core Scientific, Cipher, TeraWulf, Hut 8, and Galaxy’s Helios campus. We map who owns the risk across power delivery, GPU depreciation, utilization, refinancing, and renewal exposure, and we quantify the CoreWeave duration math, the Nebius disclosure gap, the landlord revenue-per-MW tradeoff, and the renewal-discipline signals we will monitor from here.
The goal is to make the diligence work more precise. AI infrastructure demand is visible, but the next layer of separation will come from how well each model aligns key variables like monetizable MW with customer duration, asset life, financing terms, and risk ownership. That alignment is the more important directional signal than the headline backlog number, and it is what we will be watching from here.
Inside the Full Report
The full report turns this framework into a company-by-company underwriting, backed by the model and a full set of exhibits.
• The monetizable-MW framework in full, with revenue-per-MW and capex-per-MW ranges for landlords, GPU clouds, and full-stack neoclouds, and the UBS power-acquisition layer that sits beneath them.
• Company deep dives on CoreWeave, Nebius, IREN, and the powered-shell landlord cohort, including Core Scientific, Cipher, TeraWulf, Hut 8, and Galaxy’s Helios campus, each underwritten on duration, risk ownership, and renewal exposure.
• The CoreWeave duration math, including why revenue per MW reads near $18.5M on active power but compresses to about $5.3M across contracted power, and what that swing means for the renewal case.
• The Nebius disclosure gap quantified, with company-level ARR implying roughly $0.5M per MW against an estimated $16M per MW on the Microsoft hosting agreement.
• A side-by-side map of who owns the risk across full-stack platforms, bare-metal GPU clouds, and Power Landlords, layer by layer from power delivery to refinancing.
• The firm-versus-flexible MW distinction as a new operating advantage, and how power-aware operation opens a second revenue line.
• A risk scorecard ranking the models on power control, contract duration, hardware-risk transfer, financing support, disclosure quality, and customer concentration.
• The hyperscaler paradox, and what the next renewal wave will reveal about true dependence as internal capacity and custom silicon arrive.
• A renewal-discipline tracker naming the specific contract signals we will monitor, from prepayments and step-down pricing to GPU residual values and project-finance spreads.
• Where we could be wrong, and the data caveats behind the model.