The pattern we see again and again in compute-heavy AI project financing is instructive: the conversation that generates the most friction among sponsors is not about large language models or foundation model architecture – it is about capital structure. Specifically, it is about why so many compute-heavy AI projects are stalling at the point where institutional debt should be flowing, and what sponsors are getting wrong before they ever reach a credit committee.
The conventional map says AI companies raise venture capital, burn through it on GPUs, and hope the revenue follows. The actual route – the one that works for compute-heavy projects at scale – runs through a very different terrain: project finance debt, infrastructure equity, power purchase agreements, and the kind of financial model that a credit committee can actually stress-test. Most sponsors arrive at that terrain with the wrong documentation, the wrong framing, and the wrong sequence. What follows is a grounded map of how AI infrastructure projects actually get financed.
Why AI Infrastructure Is a Capital-Intensive Asset Class
The comparison that matters here is not AI versus software. It is AI compute infrastructure versus a power plant or a toll road. Both require large upfront capital expenditure, generate revenue as a function of utilisation over time, carry long-lead procurement risks, and depend on physical infrastructure – land, power, cooling, interconnect – that cannot be scaled with a sprint cycle.
A large-scale GPU cluster – the kind required to run serious model training or inference workloads at commercial scale – involves procurement costs that routinely run into hundreds of millions of dollars for hardware alone. Nvidia’s H100 and H200 chips, the current workhorses for frontier model training, carry significant per-unit costs at market rates, and a meaningful training cluster requires thousands of them. Add facility costs, power infrastructure, cooling systems, networking, and the land or long-term lease required to house the build, and the total capex for a serious compute project looks less like a startup funding round and more like a mid-sized energy project.
Power is not a secondary consideration. It is a first-order constraint. A hyperscale AI data centre drawing 100 megawatts is, from a utility and grid perspective, a significant industrial load – comparable to a large mining operation or a mid-sized desalination plant. Sponsors who frame their project in those terms unlock a different category of lender: the infrastructure fund, the development finance institution, the energy-focused project finance bank. Those institutions have capital pools and risk mandates that dwarf what the venture capital market can deploy into physical assets. The reframing is not cosmetic – it is structural, and it changes both the investor shortlist and the documentation requirements. Sponsors who have worked with experienced capital raising consulting firms understand that this reframing must happen at the structuring stage, not after the first lender conversation.
Put simply, capital and structure must align before the first institutional conversation begins. Sponsors who arrive without that alignment typically discover it the hard way, after months of meetings that go nowhere.
The Main Financing Structures for Compute-Heavy Projects
There is no single capital structure that fits every AI infrastructure build. The right configuration depends on project scale, the sponsor’s balance sheet, whether anchor customers have committed, and what stage of de-risking the project has reached. In practice, most large compute builds draw from a combination of the following:
- Senior secured project finance debt – typically a majority of total project cost, provided by infrastructure banks, specialist lenders, or development finance institutions against contracted revenue and physical assets. This is the backbone of the capital stack for projects with demonstrable offtake or signed customer contracts.
- Equipment financing and sale-leaseback structures – particularly relevant for GPU procurement, where specialised lenders or leasing vehicles can finance hardware separately from the facility, often at better advance rates than a blended project loan. A number of GPU financings completed in recent years have used sale-leaseback structures where the hardware is sold to a financial investor and leased back to the operator.
- Mezzanine or subordinated debt – sits between senior debt and equity in the waterfall, carries higher yield, and is often used to bridge a gap where senior leverage limits leave a portion of the capital stack unfilled. Infrastructure mezzanine funds are active in this space.
- Strategic equity from hyperscalers – Microsoft, Google, and Amazon have all made strategic investments or committed compute credit programs to AI ventures. These are not passive financial investments; they come with platform commitments, distribution rights, and supply chain implications that need to be modelled carefully before accepting the term sheet.
- Infrastructure equity funds – Brookfield, Macquarie, DigitalBridge, and comparable vehicles have published mandates for digital infrastructure that explicitly include AI compute. Their return expectations and hold period preferences shape how the equity tranche of the capital stack should be sized and structured.
- Government grants, incentives, and concessional finance – particularly relevant in jurisdictions that have framed domestic compute capacity as a strategic priority. The US CHIPS and Science Act, the EU’s AI Factories initiative, and several Southeast Asian government programs are actively providing grant funding, low-cost loans, or tax incentives for qualifying compute infrastructure projects.
The sequencing of these layers – and the covenant math that holds the stack together – is where most projects either gain momentum or get stuck. The parallel experience of how a national well retirement platform structured its capital raise illustrates how layered capital structures require the same disciplined sequencing regardless of industry.
Bridging the gap between business and capital
From concept to investor-ready
We ensure your project resonates with the market, delivering the confidence investors need to move forward.
How Project Finance Debt Applies to AI Infrastructure
Project finance, at its core, is a lending structure where repayment depends on the cash flows generated by a specific asset rather than on the sponsor’s corporate balance sheet. It is the structure that finances pipelines, power plants, toll roads, and LNG terminals. It works for AI infrastructure for exactly the same reasons it works for those assets: there is a discrete, identifiable revenue stream tied to physical capacity, and that revenue stream can be modelled, stress-tested, and pledged as security.
What a credit committee will actually examine when underwriting an AI infrastructure project finance loan is not the technology thesis. Credit committees underwrite cash flow. The questions they ask are specific: What is the total installed compute capacity, expressed in petaflops or GPU-hours? What utilisation rate has been assumed in the base case, and what is the floor utilisation at which the project still services its debt? Who are the customers, what is the tenor and structure of their contracts, and what is the customer concentration risk? What is the power cost assumption, and is it fixed or floating? What is the hardware refresh schedule, and has the capex for that refresh been included in the model?
Sponsors who answer those questions fluently – with a model that shows the full cash flow waterfall from revenue through operating costs, debt service reserve, senior debt repayment, and equity distribution – move through credit committees at a materially different speed than those who arrive with a narrative and a spreadsheet assembled in the week before the meeting.
It is important to remember that AI infrastructure project finance is not yet fully standardised. Lenders are still developing their underwriting frameworks for GPU depreciation curves – hardware that was state-of-the-art in 2023 may be economically obsolete by 2027 – for customer concentration in markets where a small number of hyperscalers or frontier AI labs represent the bulk of demand, and for the power agreement structures that underpin the physical viability of the project. This ambiguity creates both risk and opportunity. The sponsor who arrives with a well-structured model and a clear capex-opex decomposition educates the lender while simultaneously differentiating the project from the crowd. In each case, independent assessment of the technical assumptions strengthens that position considerably.
Building the Financial Model Before the Pitch
The single most persistent mistake in AI infrastructure financing is the sequence: pitch first, model later. The correct sequence inverts that entirely. The financial model is the single point of truth from which everything else flows – the pitch narrative, the term sheet conversation, the information memorandum, and the lender roadshow. Building it after investor pushback means retrofitting analysis to a narrative, and sophisticated institutional capital reads that immediately.
For a compute-heavy AI project, the model architecture should include, at minimum:
- A utilisation-based revenue build – not a top-down market size estimate, but a bottom-up calculation from total installed capacity, expected utilisation by customer tier (enterprise versus hyperscaler versus research), blended price per GPU-hour or compute unit, and a ramp curve that reflects realistic sales cycles.
- A capex-opex decomposition that clearly separates the one-time capital expenditure items (GPU procurement, facility construction, power infrastructure buildout, networking) from recurring operating costs (power, cooling, maintenance, software licensing, personnel). Lenders size debt against the asset base; they need to see these clearly separated.
- A hardware refresh schedule – GPU generations typically have commercial relevance windows of several years before the next generation forces a pricing recalibration. A model that ignores refresh capex is incomplete.
- A debt service coverage ratio (DSCR) analysis run across at least three scenarios: a base case, a downside case (utilisation falls materially, power costs rise, a key customer delays ramp), and an upside case. Lenders want the downside; equity investors want the upside; the base case is what everyone publicly agrees to.
- A cash flow waterfall that reflects the actual priority of payments: operating costs first, then debt service reserve funding, then senior debt repayment, then mezzanine, then equity distributions. This waterfall is what the project finance lender is actually buying when they extend debt.
The sensitivity analysis is the credibility test. A model that shows only the base case is not analysis – it is a forecast. Sponsors who arrive with a model that has been genuinely stress-tested, where the downside assumptions are defensible and the DSCR still clears a lender-acceptable floor, are the ones who earn trust at the table. That trust, in practice, is what separates a fundable project from a stalled one. The broader lessons from infrastructure projects that stalled because financial preparation was inadequate are a useful reference for any sponsor building this discipline for the first time.
What Institutional Lenders Look for in an AI Infrastructure Deal
The misconception worth addressing directly is that a compelling AI use case is sufficient to attract institutional debt. It is not. Technology thesis is not a credit underwriting criterion. Lenders underwrite cash flow, counterparty quality, and asset security.
In practice, the deal-specific factors that move an AI infrastructure project from interesting to fundable are relatively consistent across lender types:
| Factor | What Lenders Examine | Why It Matters |
|---|---|---|
| Contracted revenue | Signed offtake agreement, LOIs, or reservation agreements | Predictable cash flow reduces underwriting risk |
| Customer quality | Credit rating or financial standing of anchor customers | Counterparty risk is a primary underwriting variable |
| Power agreement | Fixed-price or indexed PPA with a creditworthy utility | Power is the largest opex; floating cost is a key risk |
| Site control | Owned or long-term leased land with permitting clarity | Physical security for the loan; regulatory risk signal |
| Capex certainty | Fixed-price EPC or procurement contracts | Cost overrun risk is a common project finance killer |
| Sponsor experience | Track record in infrastructure or data centre development | Lenders back people as much as they back projects |
Similarly, the information memorandum (IM) for an AI infrastructure project must be structured to defend the model under scrutiny, not to sell a vision. The IM should lead with capacity assumptions, power agreements, and customer concentration analysis – not with AI market size slides. Document architecture matters. An IM that opens with a TAM chart signals to an infrastructure lender that the sponsor does not yet understand what the lender actually needs to see. Long-term strategic investors in this asset class have seen that pattern before. They recognise it quickly, and they move on.
Bridging the gap between business and capital
From concept to investor-ready
We ensure your project resonates with the market, delivering the confidence investors need to move forward.
Capital Stack Design: Layering Grants, Debt, and Equity
The optimal capital stack for an AI infrastructure project is not determined by what the sponsor would prefer – it is determined by the project’s cash flow profile, the covenant math that senior debt imposes, and the return expectations of each capital tier. Getting this wrong does not just make the project harder to close; it can render it structurally unfundable, because a stack with too much leverage will breach DSCR covenants at base case, while a stack with too much equity will generate returns too low to attract institutional capital.
A representative capital stack for a mid-scale AI compute facility – one with contracted compute capacity and a signed anchor customer – might layer government grants or incentive programs as a first tranche of effectively free capital that reduces the equity requirement, followed by senior secured debt priced at a spread reflecting the asset’s risk profile, a mezzanine layer to bridge remaining gaps, and an equity tranche sized to meet institutional return thresholds. The exact ratios shift depending on the jurisdiction, the power agreement structure, and the quality of the offtake.
What is equally important to understand is that government incentive programs – whether US federal programs tied to domestic compute sovereignty, Australian government digital infrastructure grants, or LATAM development bank concessional loans for technology infrastructure – are not afterthoughts. They are structural elements that change the leverage math, and they should be identified and sized in the model before the private capital raise begins, not discovered midway through a lender roadshow. Healthy financial discipline at the structuring stage avoids that particular embarrassment. Engaging capital raising consultants who specialise in infrastructure capital stacks early in this process ensures the grant and incentive landscape is mapped before the private raise commences.
Milestone Sequencing: Matching Your Raise to De-Risking Events
Institutional capital does not flow in a single tranche at financial close. It flows in stages, and each stage corresponds to a specific de-risking milestone that changes the project’s risk profile and therefore the cost and availability of capital. Sponsors who understand this sequencing raise capital more efficiently and on better terms.
The de-risking milestones that typically correspond to capital unlock events in AI infrastructure projects are:
- Site control and permitting – the moment the project transitions from concept to located asset, unlocking early equity and development finance.
- Power agreement execution – a signed PPA or grid connection agreement dramatically reduces the single largest operational risk variable, typically enabling the first serious project finance conversations.
- Anchor customer LOI or reservation agreement – demonstrated demand from a creditworthy counterparty is often the trigger for a credit committee to approve indicative terms.
- Equipment procurement contract – a fixed-price GPU and hardware procurement agreement reduces capex uncertainty, a key project finance underwriting variable.
- Full project finance close – senior debt drawn, equity committed, and construction commenced.
In each case, the sponsor who arrives at the next conversation with the milestone clearly evidenced in the model – and with the resulting improvement in DSCR and risk profile translated into updated documentation – is the one who moves forward. The sponsor who arrives with a narrative update and no revised model is the one who waits. That is not a commentary on the quality of the project. It is a commentary on the quality of the preparation.
Cross-Border Capital Sources for AI Infrastructure Sponsors
Cross-border capital for AI infrastructure is structurally underreported, and the gap between what sponsors know about and what is actually available is significant. APAC sovereign funds – GIC in Singapore, Temasek, the Abu Dhabi Investment Authority – have active digital infrastructure mandates. Middle Eastern sovereign wealth vehicles have been publicly allocating to compute infrastructure in jurisdictions where power is cheap and permitting is faster than in North America or Western Europe. LATAM development banks, including CAF (the Development Bank of Latin America) and BNDES in Brazil, have published programs targeting digital infrastructure as a development priority.
The constraint is not the availability of cross-border capital. It is the documentation readiness of the project. A sponsor who wants to access GIC or a Gulf sovereign fund needs an IM that includes proper currency risk analysis, jurisdiction-by-jurisdiction regulatory mapping, and a financial model built to international project finance standards – not a pitch deck translated from English into a second language. The documentation bar for cross-border institutional capital is materially higher than for domestic venture, and sovereign risk considerations are real in some of the most attractive power-cost jurisdictions. Experienced project finance advisors who work across these geographies – across APAC, the Gulf, and LATAM – understand that the model must reflect local regulatory and currency dynamics honestly if the due diligence process is to hold.
AI infrastructure, in geographies where power is cheap and government incentives are available, represents a genuinely interesting cross-border opportunity. But it is only accessible to sponsors who have done the structural work to reach it. The experience of raising capital for a major mineral exploration program across multiple jurisdictions offers a practical illustration of the documentation discipline that cross-border institutional mandates demand.
Frequently Asked Questions
What makes AI infrastructure different from a typical tech startup raise?
AI infrastructure projects – compute clusters, GPU facilities, dedicated data centres – require large upfront capital expenditure in physical assets, contracted power, and real estate. That makes them structurally similar to power plants or toll roads, not software companies. The financing structures that fit (project finance debt, infrastructure equity, PPAs) and the documentation required (financial models, IMs, offtake analysis) are drawn from the infrastructure capital markets, not from the venture capital playbook. Sponsors who treat an AI infrastructure raise as a standard tech fundraise typically find institutional lenders unwilling to engage on their terms.
Can AI infrastructure projects use project finance debt?
Yes, and increasingly they do. Project finance structures debt repayment against the cash flows of a specific asset rather than the sponsor’s balance sheet. For AI compute facilities with contracted customers, signed power agreements, and a modelled revenue path, the project finance framework applies directly. The key underwriting variables are utilisation rates, power costs, hardware depreciation, customer concentration, and debt service coverage. A project that can model these convincingly across base and downside scenarios is a genuine project finance candidate.
What is the role of a financial model in an AI infrastructure raise?
The financial model is the single point of truth for an AI infrastructure raise. It drives the capital structure, the term sheet conversation, the information memorandum, and the lender roadshow. A model built correctly for this asset class includes a utilisation-based revenue build, a capex-opex decomposition, a hardware refresh schedule, a debt service coverage ratio analysis across multiple scenarios, and a cash flow waterfall. Institutional lenders will interrogate the model directly. A pitch deck without a supporting model is not investment-ready documentation – it is a conversation starter, and nothing more.
What government incentives are available for AI infrastructure projects?
Incentive availability is jurisdiction-specific and changes frequently, so any sponsor should conduct current due diligence with local advisors. That said, the US CHIPS and Science Act, the EU’s AI Factories program, and several national compute sovereignty initiatives in APAC and the Gulf have all made concessional capital, grants, or tax incentives available for qualifying compute infrastructure. These incentives can materially reduce the equity requirement and improve DSCR math. Identifying and sizing them before the private capital raise begins is a structural decision, not a post-close optimisation.
What documents does an institutional investor require for an AI infrastructure deal?
At minimum: a detailed financial model (not a simplified summary), a formal information memorandum structured around capacity assumptions and contracted revenue, a capital structure memo showing the proposed debt-equity configuration and covenant analysis, a site and permitting status summary, copies of or term sheets for power agreements and anchor customer contracts, and a management team summary. For cross-border raises, the IM should also include currency risk analysis and jurisdiction-specific regulatory mapping. Sponsors who cannot produce this documentation are not yet investment-ready, regardless of how strong the underlying technology is.
It is clear that the deals clearing institutional credit committees in AI infrastructure are not necessarily the ones with the most impressive technology or the most recognisable founder names. They are the ones where capital and structure have been aligned before the first investor meeting – where the model has been built honestly, the capital stack has been designed with covenant math in mind, the offtake architecture is clearly evidenced, and the documentation can withstand genuine scrutiny rather than simply generate interest. The capital raising advisors who operate in this space consistently point to that structural discipline – not the quality of the technology – as the differentiating factor between projects that close and projects that do not.
Strong projects do not fail because of weak fundamentals. They fail because the structure is not there when the capital arrives – or because the capital never arrives at all, because the structure was never built to receive it. The terrain is navigable, but only with the right map, built before the journey begins.
