AI Compute Buildout Model | Mid-Market Case

If you are a family office, a private equity fund, or another investor looking for AI compute infrastructure exposure, this model will help you understand if the investment case is legit.

If you are a new operator in the AI compute infrastructure business, this model will tell you what your investors expect from the project to make it live.

Schedule a call with us: calendar.app.google/pK5okJCf2Y4Pm1g36, or email stepan@cobursa.com, stani@cobursa.com with any questions. Follow us on X: @stepanVC, @amusingventures.

This model is for illustrative purposes only and does not constitute professional advice. By using this tool, you agree to the terms, disclaimers, and copyright notice at the bottom of this page.

Market context

The digital infrastructure market is verticalizing rapidly. As global AI power demand scales into tens of gigawatts, traditional boundary lines between real estate, utilities, and technology are blurring, driving intense consolidation alongside new entry points:

• Colocation developers: moving up the stack from passive real estate landlords into active technology operators to capture premium, bare-metal GPU rental margins.
• NeoClouds: moving down the stack by acquiring land, physical data hulls, and generation-scale behind-the-meter energy assets to secure vertical control over the physical layer and protect against structural power deficits.
• Mid-market entrants: private equity, family offices, and smaller operators entering the space, drawn by attractive compute yields but constrained by high execution risk and conservative private credit underwriting.

The technology stack is evolving rapidly. Each new NVIDIA generation, released on an annual cycle, delivers structurally superior performance per watt and per dollar of capital, causing the residual values of prior generations to reprice as soon as a next-generation platform ships. The newest architectures (Blackwell Ultra and Rubin-class) require 120-370 kW+ per rack with full direct-to-chip liquid cooling, heavy floor loading, specialized mechanical plumbing, and 480V or 800V power distribution.

Most existing colocation facilities are not ready for these power densities, as they were engineered for air-cooled racks in the 20-40 kW range. Converting them requires significant capital expenditure, extensive facility downtime, and massive engineering overhauls. Only a limited number of purpose-built or rapidly convertible sites can host the newest generation of GPUs today, and even fewer are prepared for the still-higher densities expected in subsequent generations. The highest-quality powered sites and Tier-1 colocation relationships are locked early by hyperscalers, neoclouds, sovereigns and mega-funds.

This structural deficit creates a significant opportunity for prefabricated modular data centers (MDCs). By deploying factory-integrated, liquid-cooled pods onto more available, secondary 1 to 10 MW power sites, operators can bypass 5-to-7-year utility grid queues and compress development timelines from 24 months down to 4-6 months, provided high-speed fiber is in proximity. SemiAnalysis estimates that modular builds will capture 30% market share of total deployed AI compute capacity by 2028.

New entrants face severe information asymmetry, scarce access to specialized liquid-cooled facilities, and extreme operational complexity. A lack of specialized AI-infrastructure expertise constrains the ability to source, diligence, and execute transactions at institutional speed and quality. At the same time, these barriers to entry also create opportunities for more agile capital where supported by the right partnerships addressing the capability gaps.

Build-out project structure

AI data center build-outs involve multiple parties contributing equity, debt leverage, power access, land, customer pipelines, specialized resources, and operational capabilities. Successful deployments are built on strategically aligned partnerships where every party clears its required financial return relative to the risk it carries.

A typical deployment coordinates a multi-party project structure. Asset boundaries dictate the combination of counterparties: a pure powered-land transaction may only involve a site owner and developer; an asset-light GPU cluster integrates an operator, colocation landlord, lender, and offtaker; while a full-stack deployment requires the entire ecosystem.

• Sponsor (deal lead). Controls the SPV, sets the capital structure. Historically a sole infrastructure fund, but increasingly structured as integrated JVs combining capital, technology, and power access.
• Equity provider. Ranges from institutional capital to private equity and family offices backing the transaction. Increasingly includes strategic capital (e.g., chip manufacturers/OEMs) providing hardware allocations or structural residual-value guarantees as part of the equity package.
• Lender. Commercial banks or private-credit funds providing senior construction debt, mini-perm facilities, mezzanine tranches, or specialized asset-backed GPU debt.
• Offtaker. Hyperscalers, NeoClouds, foundation model labs, AI-native startups, or enterprises. The creditworthiness of the counterparty dictates whether debt leverage can be raised at all.
• Land and power owner. Controls the physical site and secures the structural grid connection, substation rights, or behind-the-meter generation capacity.
• Developer and operator. Designs, builds, and runs the facility or GPU cluster.
• Colocation facility provider. Provides high-density powered shells equipped with direct-to-chip liquid cooling infrastructure for rent to GPU fleet operators under wholesale capacity contracts.
• OEM or chip vendor. Allocates scarce GPUs and, in some structures, provides residual-value or baseline capacity backstops.
• Utility or independent power producer. Delivers grid power or co-owns on-site generation infrastructure.
• GTM & orchestration partners. Used when the sponsor lacks direct AI compute sales, GPU cluster operations, or low-latency orchestration capabilities.

Deal equilibrium

For a build-out deal to close, every party must clear its specific risk-adjusted financial return and underwriting hurdles.

Institutional sponsors & equity providers

Allocate equity capital based on strict, risk-adjusted profiles, with levered IRR hurdles scaling directly with tenant technology and counterparty credit risk:

• Hyperscaler leases (AWS, Google, MSFT): 10%-15% levered IRR hurdles (investment-grade / low real estate risk).
• Major NeoCloud operators: 18%-25%+ levered IRR hurdles (blended infrastructure and venture technology risk).
• Less established operators / GPU fleet exposures: 25%-35%+ levered IRR hurdles (high execution, asset utilization, and credit risk).

Lenders & credit providers

Underwriting standards restrict leverage based on asset type and contract bankability:

• Facility capex (real estate): senior construction or infrastructure debt is typically capped at a maximum 60% loan-to-cost (LTC) ratio.
• GPU hardware debt: 70%-80% loan-to-value (LTV), conditional on take-or-pay offtake contracts or upfront customer prepayments.
• Debt capacity is sized strictly to maintain a debt service coverage ratio (DSCR) above 1.30x, backed by a 6-month forward debt service reserve account (DSRA).

Interest rates usually vary by offtaker credit quality: 7-10% with an investment-grade offtaker; 10-12% with a strong offtake; 12-15% with weak or no offtake. Debt can be more expensive and restrictive for mid-market players.

Offtakers mandate credible operators capable of meeting strict infrastructure delivery service level agreements (SLAs).

Developers and operators require a development fee, an asset-management fee, and typically a profit share (the promote) structured via an equity waterfall that triggers once the sponsor's preferred return hurdles are met (in a JV structure).

If the project cannot model out an equilibrium that hits the required sponsor hurdles and bankability metrics under conservative assumptions, the sponsor will walk, reallocating the equity capital to a project with a superior risk/return profile. A strategically aligned partnership reduces deal risk, improves bankability, and unlocks the asset's true value.

Model context

The model abstracts away real-world operational complexities to isolate the core financial variables defining deal viability. A live transaction requires tailored consultation and validation with technical and financial experts. To discuss your specific mandate, contact us via stepan@cobursa.com.

The primary objectives of this model are to:

• Quantify project scale & returns: demonstrate the indicative IT MW capacity and levered return on capital achieved by a specific dollar amount of equity investment in mid-market AI compute infrastructure.
• Verify deal equilibrium: evaluate if the proposed deal terms simultaneously clear the sponsor's required IRR hurdle and the lender's debt service covenants under a set of custom capital and operational assumptions.
• Perform sensitivity analysis: stress-test project cash flows and financial outcomes against the most critical commercial variables, including hourly GPU rental rates, annual price decay, and GPU fleet utilization.

Model deployment scenarios

To match varying mid-market risk mandates, the model permits toggling between two distinct operational structures:

GPU & IT-hardware ownership. The special purpose vehicle holds title strictly to the revenue-generating silicon (including physical GPU clusters, servers, memory, and high-speed InfiniBand networking fabrics). All power allocation, high-density liquid cooling, and building infrastructure are rented from a third-party wholesale colocation facility. Cash flows are generated by leasing compute hours or reserved capacity under multi-year or flexible contracts, net of pass-through utility tariffs and facility operational expenses.

Full-stack center ownership. This mode serves as a direct proxy for modular data center campus builds. It captures the underlying real estate leasehold value alongside the premium GPU compute margins. This configuration maximizes potential project IRR and long-term terminal value, at the cost of higher upfront capital intensity, technical engineering complexity, and single-site concentration risk.

Model configuration

• Integrated asset SPV: a sponsor-backed special purpose vehicle that holds title to the GPU clusters or to the physical power access and electrified shell, depending on the chosen deployment model.
• Equity partner: funds the SPV's capital requirements and structures the debt financing facilities.
• Data center developer & operator: builds and manages the physical site in exchange for contractual development and management fees.
• GTM partner: routes and commercializes the raw compute capacity under a structured revenue-share agreement.

Baseline assumptions and simplifications

• Facility: a liquid-cooled AI factory utilizing the NVIDIA Blackwell Ultra (GB300 NVL72) platform, sized dynamically based on the input equity ticket.
• Asset timeline: a single five-year asset lifecycle with a Year 5 exit, modeled as if the real estate converts to a stabilized triple-net (NNN) lease and the GPU clusters are assigned a structural residual value.
• Development: construction is modeled in Year 0, with capacity going live at the start of Year 1.
• Debt optimization: the capital stack utilizes a streamlined debt service profile rather than multi-tiered commercial amortization schedules.

Refer to the Assumptions tab for the comprehensive list of underlying parameters.

Disclaimer

This financial model and any accompanying materials (collectively, the "Model") are provided solely for illustrative, informational, and educational purposes. The Model does not constitute, and shall not be construed as, investment, financial, legal, tax, or accounting advice.

The Model abstracts and simplifies complex operational, structural, and regulatory realities. No rendering of legal structures (including special purpose vehicles), tax treatments, or accounting methodologies is intended to be definitive. The author is not acting as an investment advisor, fiduciary, or legal/tax consultant. All projections, market trends, return thresholds, interest rates, utilization metrics, and asset valuations are indicative baseline assumptions and subject to rapid, material change based on macroeconomic conditions, technological obsolescence, and utility constraints. Actual project performance, financing terms, and returns may vary significantly from the scenarios modeled herein. No representation or warranty, express or implied, is made as to the accuracy, completeness, or reliability of the financial mechanics or outputs presented. Users are strictly advised to seek independent, tailored consultation and validation from qualified investment, legal, tax, accounting, and technical experts before executing any transaction. The author accepts no liability for any financial decisions, losses, or damages incurred in reliance on this Model.

Copyright © 2026 Cobursa. All rights reserved. This Model and all text, structures, logic, and intellectual property contained herein are proprietary property of Cobursa. No part of this material may be copied, reproduced, distributed, republished, modified, or used in any manner, electronically, mechanically, or otherwise, without the prior, express written consent of Cobursa. Unauthorized use, disclosure, or distribution is strictly prohibited.