The Reserve Report Nobody Files

Executive Summary

From data to compute

We once said data was the new oil. But the economics of AI suggest that the more important scarce resource may be compute.

Oil derives its value from a scarce physical resource and the infrastructure required to extract, transport, and refine it. AI compute has created a similar economic model, scarce capacity, large upfront investments, long-term commitments, and a growing ecosystem of suppliers, infrastructure owners, and users.

The difference is that the financial reporting framework was built for traditional businesses. It was not designed to show the full network of contracts, financing arrangements, and ownership interests now emerging around AI infrastructure.

This publication follows that money and asks a simple question:

What do the financial statements show, and what remains outside the picture?

This publication is a companion to The AI Infrastructure Reckoning. That publication examined the balance sheets funding the AI buildout; this one examines the contracts, entities, and counterparties behind it.

Four questions frame the analysis:

Sections 1–7 trace the economics from compute scarcity through the offtake arrangements, ownership structures, and allocation of risk.

The reporting question

At the center of the AI infrastructure model is one important assumption:

How long will today’s compute remain economically useful?

Chip technology is evolving rapidly. Yet the useful life assigned to the underlying infrastructure and equipment remains largely a management estimate rather than a market-priced measure.

That estimate affects depreciation, asset values, returns on investment, and, ultimately, the economics of the infrastructure being financed.

1. From Oil to Compute: The Analogy That Finally Fits

Exhibit 1a — How the economics moved from oil to compute

Why the Oil analogy works

The oil industry organized itself around three activities: extract, move, and refine. Compute reproduces the sequence almost exactly, makes the chips, builds the capacity, delivers the intelligence, and the tiers beneath line up just as neatly.

Exhibit 1b — The AI infrastructure chain, and its oil equivalent

The key difference is where customers ultimately derive value. Businesses do not generally buy crude oil simply to own it; they pay for the products and services that oil enables. AI is increasingly following a similar model. Businesses may not need to own GPUs or data center infrastructure themselves. Instead, they pay for the outcomes enabled by computing capacity, from generating AI responses and training models to analyzing data, automating processes, and developing new products.

Scarcity is what drives economics

The comparison with oil becomes more relevant when looking at scarcity. Compute requires significant physical infrastructure, specialized technologies, and substantial upfront investment. Supply also cannot be expanded immediately when demand increases. These constraints, together with a concentrated supply chain and long lead times for new capacity, can allow companies controlling critical parts of the infrastructure chain to earn a premium when demand exceeds available supply.

This is also where the economics of compute differ from the earlier “data is the new oil” narrative. Data can generally be copied and reproduced at very low cost. Compute, by contrast, requires physical assets, electricity, specialized chips, and data center capacity. It is therefore both capital-intensive and capacity-constrained.

Data can be replicated. Compute requires capacity.

What happened to the oil premium?

The history of the oil industry provides another useful perspective. Energy accounted for approximately 28% of the S&P 500 in 1980, compared with approximately 4% today. This does not mean oil became less important. Over time, investment expanded energy supply, reduced its relative cost, and allowed more of the economic value to flow to the industries and products built on top of it.

AI compute could follow a similar path.

Exhibit 1c — Where the value sits today, and where it moves as supply expands

Today scarcity concentrates the premium upstream

As supply expands the premium erodes and value moves downstream

2. Where The Scarcity Sits

We have seen how capital is flowing into AI infrastructure. The next question is where the economics are actually being captured.

The answer starts with the scale of the infrastructure being built and then moves further up the supply chain to the chips, manufacturing capacity, and software ecosystem that make that infrastructure valuable.

2.1 The unit of account: gigawatts

A useful way to think about the scale of the AI buildout is in gigawatts (GW) of data center capacity.

Building 1 GW of AI data center capacity can cost approximately USD 50 billion to USD 55 billion, depending on the technology, location, and scope of the investment. A significant portion of that investment is tied to servers, graphics processing units (GPUs), memory, and other computing equipment.

This changes the way we should think about the underlying asset.

The value of an AI data center is not simply the land, building, or power connection. A large part of its economic value sits in the computing equipment inside it.

That leads to an important accounting question:

How long will the chips remain economically useful?

2.2 Where the scarcity begins

The raw material used to make a semiconductor is abundant. The constraints emerge further along the production chain.

Three important bottlenecks illustrate this:

Exhibit 2a — The three tollbooths: how free sand becomes a USD 40,000+ chip

Source: Uniqus analysis. ASML describes its EUV technology as unique to ASML; TrendForce puts TSMC at 70.4% of the top ten foundry revenue, and no public figure quantifies its share of leading-edge nodes specifically. CUDA was first released in November 2006.

Exhibit 2b — From raw material to AI chip: where the scarcity is manufactured

The important point is that the scarcity does not come from the raw material itself. It is created by the specialized technology, capacity, expertise, and ecosystem required to turn that material into a high-performance AI chip.

This is different from oil.

In oil, the scarcity exists in the resource. In semiconductors, much of the scarcity sits in the process of converting an abundant resource into a highly specialized product.

That scarcity can translate into significant margins for companies controlling critical parts of the supply chain.

Natural scarcity vs. engineered scarcity

Governments are also becoming more involved in these bottlenecks through export controls, subsidies, and industrial policy. These measures can influence where advanced chips and manufacturing equipment can be produced and who can access them.

2.3 How long can the advantage last?

• Engineered scarcity has one important difference from natural scarcity: it can change.
• Oil reserves cannot be reproduced. Semiconductor capacity can. New factories can be built, alternative suppliers can emerge, and customers can develop their own chips.
• This is particularly relevant to Nvidia. Some of its largest customers are investing in custom AI accelerators, reducing their reliance on general-purpose GPUs over time.
• The competitive question is therefore not simply whether another company can build the same GPU. It is whether customers can develop alternatives that meet their specific computing requirements at a lower cost.
• Nvidia’s margins also illustrate how the economics can change as supply constraints move through the value chain. Nvidia reported a 75% gross margin in Q2 FY2027, while indicating that margins would decline in subsequent quarters, in part because of higher memory costs. Management expects margins to improve again as pricing and supply conditions change.

The AI buildout is increasing demand not only for GPUs, but also for memory and other components. As a result, cost pressures can move up and down the supply chain rather than remaining with a single participant.

2.4 The useful life question

The economics of the AI infrastructure buildout ultimately depend on how long the underlying equipment remains productive. When a GPU is sold, the supplier receives the sale proceeds. The buyer, however, records the equipment as an asset and depreciates it over its estimated useful life.

If the equipment remains productive for six years, the investment is recovered over a longer period.

If the technology becomes economically obsolete after three years, the same investment needs to be recovered much faster, and the buyer may need to replace the equipment sooner than originally planned.

The same investment, different replacement cycle

The difference is significant. A shorter useful life does not simply increase depreciation expense. It can also increase the amount of capital required to maintain the same level of computing capacity.

One assumption sits at the center

There is no single market standard that determines how long an AI chip will remain economically useful. The estimate depends on factors such as technology changes, performance requirements, utilization, power efficiency, and the availability of newer chips.

3. The Offtake Wave: The Contract Is the Collateral

3.1 The Offtake Web: Who Is Committed to Whom

AI infrastructure is increasingly being built against contracted future demand.

The flow is increasingly circular.

Exhibit 3a — The circular flow: the contract that funds the capacity it pays for

Nvidia’s August 2026 announcement illustrates this shift: it partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on financing platforms intended to mobilize more than USD 500 billion of third-party capital for AI infrastructure.

The scale of the commitments

The comparison is intended to show the scale and duration of the commitments, not to suggest that all amounts represent legally binding obligations of the same type.

3.2 The Lessees: Revenue vs. Compute Commitments

The important point is not simply that AI companies are committing to capacity. They are committing to capacity well before the related infrastructure is built and revenue is generated.

A simple way to frame this:

For the infrastructure provider, a long-term contract provides visibility into future revenue, but the capital has to be deployed today.

CoreWeave is a useful example. Its Q2 2026 revenue backlog was approximately USD 104 billion, excluding more than USD 25 billion of additional commitments added in early Q3. At the same time, CoreWeave had approximately USD 35bn of debt at Q2 2026, up from less than USD 8bn in 2024.

3.3 The lessors and the supplier: contracted revenue against their own financing

The table above sets out what each party has contracted and how it pays for it. Underneath the numbers, all five are doing a version of the same thing, committing capital now against revenue that arrives later. What differs between them is how large the gap is, how it is funded, and who is left holding the asset if the revenue does not arrive on the expected schedule.

Oracle

Oracle has a large commitment from an AI company to supply computing capacity. It cannot supply that capacity until it has built it, and it does not earn most of the revenue until the capacity is running. The money, therefore, goes out well before it comes back in. Because the build is large relative to Oracle’s existing business, a meaningful part of it is funded by issuing debt and equity rather than from operations.

The risk sits at the front of the sequence: Oracle carries the capital cost for years before the contract pays for it.

Microsoft, Amazon, Google, and Meta

These four are in the same position, but with very different balance sheets behind them. They generate substantial cash from established businesses and can fund a large share of the buildout from that cash rather than from borrowing.

That changes the financing risk, not the underlying exposure. The capital still goes in before the capacity earns anything, and the assets still have to be used heavily enough, long enough, to justify what was spent. Their question is less whether the buildout can be funded than whether it will be used, which is a different risk from the one a smaller operator such as CoreWeave faces.

CoreWeave

CoreWeave is the clearest illustration of the pattern. It has a very large, contracted backlog that customers have committed to pay in future years. To earn it, CoreWeave has to build and operate the infrastructure now, which means heavy capital spending today, funded largely by debt.

The whole model rests on one question: can the future contracted revenue service the debt taken on to create it? And underneath that sits a second question about the asset itself. A GPU financed today against a six-year life is one set of economics. If it is commercially superseded after three years, it has to be replaced sooner than the financing assumed, and the revenue it was meant to earn has to come from somewhere else.

Five risks therefore run together:

• Financing risk — can the business service the debt?
• Customer risk — will customers continue to take the capacity they have contracted for?
• Utilization risk — will the infrastructure be fully used?
• Technology risk — will today’s GPUs remain economically useful?
• Depreciation risk — Is the useful life assigned to them appropriate?

TeraWulf, Riot, and Fluidstack

These developers are using long-term AI contracts to raise the money to build. The distinguishing feature is that much of the financing is tied directly to a particular project, or secured on the GPUs themselves, rather than to the borrower as a whole. A lender is therefore underwriting three things at once.

A long-term contract + A physical asset + Expected future cash flow

That combination can provide a sound basis for financing. It does not remove the questions the lender still has to ask: what happens if the customer does not perform, if the project is delayed, if the GPUs lose value, or if the capacity is never fully used?

Nvidia

Nvidia is the most interesting of the five, because the conventional picture no longer describes it. In that picture, Nvidia sells a GPU, the customer pays, and the transaction ends. What has developed instead is a set of overlapping roles held simultaneously in the same ecosystem.

Each role is unremarkable on its own. Held together, they mean Nvidia’s exposure does not end when the chip ships. It continues through the equity it holds in the buyer, the capacity it has agreed to purchase if nobody else does, and the debt it has helped make financeable.

4. Financialization, in the Wrong Order

The financialization of compute is no longer hypothetical. It is happening quickly. In March 2026, a supplier disclosed an order book of roughly USD 1 trillion stretching through 2027. In June, the largest initial public offering in US history involved a company that is both a compute landlord and a compute customer. In August, six major asset managers agreed to mobilize more than USD 500 billion for AI infrastructure. And from October 2026, GPU compute futures are scheduled to begin trading on a major exchange.

Oil followed a different sequence. Investors first needed to know what a company owned, so reserve accounting came first. Standardized reserve reporting followed, and only then did oil futures become a major market. Compute is moving in the opposite direction. Capital and long-term contracts came first. Trading is arriving before there is a standard way to measure the useful life, capacity, or economic value of the underlying assets.

Exhibit 4a — Disclosure first, trading second. Compute is reversing the sequence

Oil — disclosure first, trading second

Investors could not value a resource they could not see, so the reserve report came first.

Five years of mandatory reserve reporting preceded oil’s first futures contract. Compute is getting the futures contract first.

The difference is not the size of the market. It is the sequence.

Oil developed a disclosure framework before a deep futures market developed. Compute is attracting capital and creating financial markets, while important questions around capacity, useful life, utilization, and economic value are still being addressed.

Why the futures market matters

A futures market could provide something the compute market has not had at scale: an observable forward price for compute. The potential chain is:

Exhibit 4b— The transmission chain: from installed capacity to financing decisions

This does not mean that a futures price automatically determines the accounting value or useful life of a GPU. But over time, a sufficiently active market could provide an external data point against which management assumptions can be assessed.

For example:

Practical Perspective

A market price is not automatically an accounting benchmark

A newly established futures market may initially have limited trading volume and history. Early prices may therefore provide useful market information without necessarily being sufficient, on their own, to support an accounting conclusion.

The key indicators to watch:

• Trading volume and open interest
• Bid-ask spreads and liquidity
• Length of observable price history
• Stability of settlement prices
• Relationship between futures prices and actual compute pricing
• Whether the contracts represent the same type of compute being used in the underlying infrastructure

5. Follow One Dollar: The Loop Every Party Books in Full

A simple example illustrates how interconnected the AI infrastructure ecosystem is becoming.

Exhibit 5a — The circuit: each step, who takes it, and what it books as

Consider a single dollar moving through the AI supply chain. Nvidia invests in an AI lab. The lab uses that money to commit to computing capacity. The cloud provider buys Nvidia chips to build that capacity. Nvidia then earns revenue from selling the chips.

Why the framework cannot see the loop

• The transactions may be accounted for correctly, but the connections can be difficult to see. An investment, capacity commitment, revenue, backlog, and financing may sit across different financial statements, making the broader economic relationship difficult to reconstruct.
• Disclosure can remain fragmented. Materiality thresholds and different disclosure requirements may result in pieces of the relationship appearing in different parts of the financial statements rather than as one connected picture.
• Roles are increasingly overlapping. A supplier may also be an investor, a financing partner, or a capacity provider in the same ecosystem. This creates a network of interrelated transactions rather than a simple supplier–customer relationship.

The challenge is therefore not only how each transaction is accounted for, but whether the financial statements allow investors to see how the transactions connect.

Why is this different

Vendor financing is not new. Similar structures have appeared in earlier technology cycles, including periods when equipment suppliers helped customers finance purchases.

What is different in today’s AI infrastructure model is the number of roles that can sit with the same participants.

A supplier may simultaneously be:

• a seller of GPUs;
• an investor in an AI customer;
• an investor in an infrastructure provider;
• a provider of financing or credit support; and
• a participant in arrangements that support future capacity.

This creates a network rather than a simple buyer-seller relationship.

6. Connecting the Accounting to the Economics

Where does the accounting framework stop giving a clear answer?

The arrangements below illustrate some of the questions emerging as AI infrastructure becomes more interconnected. For each arrangement, we look at three things:

1. What is recorded today

2. What the arrangement raises from an economic perspective

3. Where the current framework may not provide a complete answer

One example is included where the framework works as intended. That contrast is important: the issue is not that the accounting framework cannot address these transactions. It is that some newer structures do not fit neatly within existing requirements.

A technology supplier invests in an AI customer while also generating significant revenue from that customer. In some cases, suppliers have invested billions of dollars in AI labs, alongside investments by hyperscalers.

For example:

An AI chip supplier sells USD 5 billion of GPUs to an AI company over several years. At the same time, the supplier invests USD 2 billion to acquire a 10% stake in that AI company.

The supplier, therefore has two relationships with the same customer:

Supplier → invests USD 2bn → AI company

Supplier ← receives USD 5bn of GPU purchases ← AI company

Why it matters

Both transactions may be accounted for separately. The investment may be recorded as an investment, while the GPU sales are recorded as revenue.

The broader question is whether the investment and the customer’s purchases are economically connected. For example, is the investment helping secure future purchases from the customer?

What each party books today: The stake is carried at fair value and remeasured on the next funding round an “observable transaction” even where the supplier participated in that round. Chip and cloud revenue from the same investee is recognized at 100%.

The question it raises: The investment and the commercial relationship are economically connected. This raises questions such as:

• Is any part of the investment, in substance, consideration provided to the customer and therefore a reduction of revenue?
• Does the accounting need to distinguish between revenue from a customer that has received funding from the supplier and revenue from other customers?
• Should the investment and the related purchase commitments be viewed together?

Where the framework stops: Existing requirements focus on ownership, significant influence and related-party relationships. These thresholds do not necessarily require disclosure of the broader relationship between an investment, customer revenue and purchase commitments when the investment falls below the relevant threshold.

The result is that the individual transactions may each be accounted for appropriately, while the broader commercial relationship may be difficult to see from the financial statements.

What each party books today: Where the arrangement falls within the applicable customer consideration requirements, the value of the instrument is treated as consideration payable to the customer and reduces revenue as the related goods or services are transferred.

The question it raises: In this case, the accounting framework provides a clear answer. The accounting reflects the economics of the arrangement.

Where the framework stops: The requirements address equity instruments granted to a customer. They do not necessarily address the broader question of what happens when a supplier instead purchases an interest in the customer’s equity.

This distinction becomes important as supplier investments in AI customers become more common.

A data center campus may be owned through a separate vehicle, financed with significant third-party debt, and leased to an AI company or hyperscaler. The tenant may also have an equity interest in the vehicle and provide a residual value guarantee.

For example:

An AI company needs a USD 10 billion data center but does not want to build and finance the entire facility itself.

A separate infrastructure company is created:

Infrastructure company → owns and finances a data center

AI company → leases the data center and uses most or all of its capacity

Lenders → lend USD 8 billion to the infrastructure company

AI company → provides a long-term lease and may also provide a guarantee

The AI company may own 20% of the infrastructure company, while another investor owns 80% and retains the formal control rights.

Why it matters

The AI company may recognize its investment, lease asset, and lease liability, and disclose its guarantee. The infrastructure company’s USD 8 billion debt may remain outside the AI company’s balance sheet if the AI company does not control the entity.

A reader, therefore, sees several separate pieces:

Investment + lease + guarantee + infrastructure company debt

But the economic relationship is much more connected.

What each party books today: The tenant may record:

• an investment in the joint venture;
• a right-of-use asset and lease liability for the contractual lease term; and
• a guarantee or other exposure in the notes, where required.

The debt of the separate vehicle may remain outside the tenant’s balance sheet if the tenant does not control the vehicle. About USD 27.3bn of vehicle debt appears nowhere on its balance sheet.

The question it raises: The structure can create a different economic picture from the individual accounting entries.

For example:

• The tenant may be the sole or primary user of the campus.
• The campus may have been built specifically for that tenant.
• The tenant may provide a significant residual value guarantee.
• The vehicle’s debt may ultimately depend on the tenant’s lease payments.

This raises a practical question: How much of the underlying infrastructure and financing exposure is borne by the tenant, even when the vehicle is not consolidated?

Where the framework stops: Consolidation depends on the applicable control model, including power and exposure to variable returns. Lease accounting separately considers the contractual lease term and reasonably certain extension options.

As a result, the financial statements may show the investment, lease, and guarantee separately, while the vehicle’s debt appears elsewhere.

The accounting can therefore be correct for each individual element without giving the reader a single view of the overall arrangement.

Exhibit 6a — The campus vehicle: what the tenant reports, and what it does not

Source: Uniqus analysis of a representative campus joint venture as described in company announcements and analyst commentary, October 2025 to March 2026. Illustrative of the structure rather than of any particular transaction; individual terms vary.

A supplier may agree to purchase unused capacity from an independent cloud provider or provide other support that helps the provider finance its infrastructure.

For example:

A cloud company spends USD 6 billion building a new AI data center. It expects an AI customer to use most of the capacity, but lenders want additional protection if the customer does not use all of it.

The AI chip supplier agrees:

The supplier, therefore, becomes a potential buyer of the cloud company’s unused capacity.

Why it matters

The supplier originally sells GPUs to the cloud company. It now also supports the cloud company’s future revenue.

The arrangement could therefore look like:

Supplier sells GPUs → Cloud company

Cloud company builds capacity → AI customer

Supplier agrees to buy unused capacity → Cloud company

The supplier is no longer simply a hardware vendor. Its commitment can also help finance the infrastructure.

What each party books today: Depending on the terms, the arrangement may be recorded as:

• a future purchase commitment; or
• a guarantee or other financial instrument.

Revenue from the supplier’s sale of chips to the same customer is generally accounted for separately.

The question it raises: the economic relationship can extend beyond a simple supplier-customer transaction.

For example:

• If lenders rely on the supplier’s commitment, does the arrangement effectively provide credit support for the customer’s financing?
• Does a supplier’s commitment to purchase the customer’s future output affect how the underlying sale and financing arrangements should be viewed together?

Where the framework stops: The distinction between a purchase commitment and a guarantee can depend on the specific terms of the arrangement.

The accounting framework also considers control separately from financial support. As a result, supplier support may affect the economics of the financing without necessarily resulting in consolidation or a single disclosure of the overall relationship.

A hyperscaler may provide credit support for a data center landlord’s long-term lease or debt financing. In certain arrangements, the hyperscaler receives warrants or other consideration in exchange for providing support.

For example:

A data center landlord wants to borrow USD 5 billion to build a facility. A hyperscaler agrees to lease the facility for 20 years and provides a guarantee or other credit support for the landlord’s financing.

The structure looks like:

Lender → USD 5bn financing → Data center landlord

Landlord → data center → Hyperscaler

Hyperscaler → long-term lease/guarantee → supports landlord’s financing

Because the hyperscaler has a strong credit profile, lenders may be more comfortable providing the financing.

Why it matters

The hyperscaler may not own the data center or the landlord’s debt. However, its lease commitment and financial support may be central to the project’s economics.

This raises a broader question: How much economic exposure does the hyperscaler have to an asset that it does not own?

What each party books today: The guarantor may recognize the guarantee at fair value and account for any warrants received as financial assets, subject to the applicable guidance.

The landlord and lender may also reflect the guarantee in their respective accounting and credit assessments.

The question it raises is: If the landlord’s primary source of cash flow is the lease with the hyperscaler, and the hyperscaler also provides credit support, how should the overall economic exposure be viewed?

The question becomes particularly relevant when the financing depends heavily on the tenant’s credit quality.

Where the framework stops: The accounting for guarantees, leases, and consolidation is based on separate requirements.

A guarantee does not, by itself, determine who controls the underlying entity. As a result, a company may provide significant economic support to an infrastructure asset without recognizing the asset or related debt on its balance sheet.

Long-term compute agreements can include large minimum commitments while also providing termination or cancellation rights. This creates an important distinction between the headline value of a contract and the amount that is actually enforceable.

For example:

An AI company signs a five-year agreement with a data center operator for USD 1 billion of compute capacity.

At first glance:

Five-year contract × USD 1bn per year = USD 5bn of contracted revenue

The operator, therefore, describes the arrangement as USD 5 billion of backlog.

But the contract also allows the AI company to terminate the arrangement with 90 days’ notice.

The economics are therefore different from a firm five-year commitment.

Why it matters

The headline backlog may be USD 5 billion, but the amount that is actually enforceable could be significantly lower if the customer can cancel the contract.

This creates an important distinction between:

Contract value

and

Enforceable future revenue

It also raises questions about the customer’s ability to pay and whether the contract remains attractive if compute prices fall.

What each party books today: Remaining performance obligations generally reflect the enforceable period of the contract. Other measures, such as supplier “backlog” or landlord “contracted revenue,” may be presented outside the financial statements and may not have a consistent definition.

The question it raises is: How much of the reported backlog is:

• legally enforceable?
• cancellable at short notice?
• dependent on future pricing?
• concentrated with private or unrated customers?

A large headline backlog may therefore provide limited information about the amount of revenue ultimately expected to be realized.

Where the framework stops: There is no single accounting definition of “backlog.”

Companies can therefore use different measures and definitions. Remaining performance obligation disclosures also do not necessarily provide a complete view of cancellation rights or the customer’s credit profile.

This makes it important for investors to understand what sits behind the reported backlog, rather than relying only on its headline value.

Another structure involves a supplier providing equity or financing to a special purpose vehicle that owns the supplier’s equipment and leases it to another customer. In some arrangements, the customer may also provide equipment or prepayments to the vehicle.

For example:

A chip supplier wants to sell USD 3 billion of GPUs to an AI company, but the AI company does not want to fund the entire purchase upfront.

The supplier, therefore, helps establish a separate financing vehicle.

The structure looks like:

Supplier → provides equity/financing → Vehicle

Vehicle → buys USD 3bn of GPUs → Supplier

Vehicle → leases GPUs → AI company

AI company → makes lease payments → Vehicle

Vehicle → uses payments to service debt

The supplier has effectively helped create the financing structure through which its own equipment is purchased.

Why it matters

The supplier may recognize revenue when it sells the GPUs to the vehicle, subject to the applicable accounting requirements.

But economically, the supplier has also provided financing or support that helped make the sale possible.

This raises questions about:

• whether the sale is truly independent;
• who controls the vehicle;
• who ultimately bears the risk of the GPUs; and
• whether the supplier’s financing, ownership, and sales relationships should be considered together.

What each party books today: The supplier may recognize revenue on sales of equipment to vehicles, subject to the applicable accounting requirements.

The vehicle may separately recognize the equipment and related financing, while the tenant recognizes its lease or service costs.

The question it raises: The structure raises broader questions:

• When does a sale to a vehicle represent a substantive sale?
• Who has the economic exposure to the equipment?
• Who controls the vehicle?
• How should the supplier’s financing or ownership interest be considered alongside its equipment sales?

Where the framework stops: The answers depend on the specific ownership, control, and financing arrangements.

Where the supplier’s ownership interest does not result in consolidation, transactions between the supplier and the vehicle may be accounted for separately. The resulting financial statements may therefore show individual transactions without making the full economic relationship easy to reconstruct.

Across these examples, the accounting framework generally answers the individual accounting question.

The more difficult question is often broader:

How should a reader understand the economic relationship when investment, financing, capacity commitments, revenue, and infrastructure ownership are all connected?

In six of the seven examples, the individual requirements may not fully answer that broader question. The example involving equity instruments granted to a customer shows that when the standard explicitly addresses the issue, the accounting can align closely with the underlying economics.

The emerging challenge is therefore not simply whether each transaction is accounted for correctly. It is whether the financial statements provide enough information for investors to connect the transactions and understand the overall exposure.

Practical Perspective: What the Financing Cost Tells Us

• Financing cost can provide insight into how the market views the structure.
• In one example, debt issued by the campus vehicle carried a coupon of approximately 6.6%, compared with around 5.4% on the sponsor’s own long-term unsecured debt.
• The vehicle financing carried a higher reported financing rate, which represents a significant additional financing cost for the campus vehicle.
• The sponsor had a strong credit profile, but the debt was raised through a separate vehicle rather than directly by the sponsor. The higher financing cost, therefore, provides an indication of the economic cost of keeping the debt outside the sponsor’s balance sheet.
• The individual elements of the arrangement are disclosed:
  • the sponsor’s investment;
  • lease obligations; and
  • residual value guarantee.
• Rating agencies may also make adjustments when assessing the sponsor’s overall exposure.
• However, a reader may still need to connect these disclosures to understand the bigger picture:
  • the campus is owned and financed through a separate vehicle;
  • the vehicle depends on the tenant’s lease payments; and
  • the tenant provides additional support through its guarantee.

Each piece is disclosed. The challenge is seeing how the pieces fit together.

• The broader lesson is not that special-purpose entities or non-consolidated structures are inappropriate.
• Rather, an arrangement can comply with the accounting requirements for each individual transaction while the overall economic relationship remains difficult for a reader to understand.

7. Who Is Paid First, Who Holds the Risk

The loop shows how the dollar travels. The question that matters when it stops traveling is a different one: who is paid first, and who is left holding the asset.

Exhibit 7a — Cash flows inward, risk flows outward

Risk flows outward — it settles at the edge, with the layer listed first.

Source: Uniqus analysis of the arrangements set out in Sections 3 to 5. Layers are ordered by priority of payment and residual exposure, not by size. Oracle is placed with the leveraged operators because it is borrowing and issuing equity ahead of the revenue its contracts will bring.

Who Carries the Risk?

• Cash flows through the chain, but risk does not. Suppliers are generally paid when equipment is sold, while the longer-term economic exposure sits with the parties that own, finance, or depend on the infrastructure.
• The outer layers carry greater dependence on future cash flows. Infrastructure providers and lenders depend on future customer payments, asset utilization, and the value of the underlying equipment. They also have less direct visibility into how quickly AI lab revenues will grow.
• Private credit is becoming an important source of funding. The Bank for International Settlements estimated private credit to AI-related borrowers at more than USD 200 billion in January 2026, with estimates pointing to USD 300–600 billion by 2030. This places a significant amount of capital at the outer end of the financing chain.
• Backstops can shift risk back toward the core participants. Investments, guarantees, capacity commitments, and other forms of support can reduce the exposure of infrastructure providers and lenders, while increasing the financial exposure of the companies providing that support.
• The key question is timing. Can AI lab revenues grow sufficiently to support their long-term compute commitments before infrastructure providers need to refinance their debt?
• Revenue growth provides some context. One major AI lab’s revenue run-rate increased roughly sevenfold over 18 months, while another’s increased roughly threefold. At the same time, significant equity capital has been raised to provide additional funding while revenues continue to scale.
• The outcome depends on how quickly revenue grows relative to fixed commitments. If growth continues as expected, the ability to support long-term commitments may improve over time. If growth slows materially, the same take-or-pay commitments could create pressure for both sides of the arrangement:
  • the AI lab may face higher fixed payment obligations; and
  • the infrastructure provider may face increased collection and utilization risk.
• Contract terms, therefore, matter. Termination rights, minimum commitments, payment obligations, and other contractual protections can become particularly important when revenue growth does not keep pace with infrastructure commitments.
• The broader point: understanding the risk requires looking beyond the individual transaction and considering the contract, underlying asset, financing, and expected customer cash flows together.

8. The Useful Life Question

The economics of the AI infrastructure buildout depend on one important accounting estimate: how long the underlying compute assets remain useful.

• The useful life of a GPU or accelerator affects:
  • the period over which the asset is depreciated;
  • annual depreciation expense;
  • the timing and amount of replacement capital expenditure; and
  • the economics of financing the infrastructure.
• Small changes in the assumed replacement cycle can have a significant impact when applied to an installed base running into hundreds of billions of dollars.

Exhibit 8a — What the life assumption does to the replacement bill

A shorter useful life can therefore double the annual replacement requirement.

Source: Uniqus computation, illustrative. Assumes a constant USD 1.8 trillion installed accelerator base, unchanged replacement prices, and evenly distributed replacement. Annual replacement expenditure equals the assumed base divided by the replacement interval.

Why the estimate matters

The useful-life assumption affects different participants in different ways:

The important point is that the useful-life estimate is not neutral to the economics of the ecosystem. Different participants may have different economic interests in how quickly assets are replaced. The evidence is beginning to develop. There are arguments on both sides of the useful-life question:

Evidence that could support longer useful lives:

• Rental rates for some older generation accelerators have remained resilient.
• Older chips can be redeployed to inference, fine-tuning, and other workloads as newer chips take on the most demanding training workloads.
• Software optimization can improve the effective utilization of existing hardware.
• Demand for compute may continue to grow even as the cost of individual workloads falls.
• Rapid growth in AI revenues could support the large infrastructure commitments being made today.

Evidence that could support shorter useful lives:

• New generations of accelerators can deliver significantly better performance and power efficiency.
• The cost of new compute continues to decline on a per-unit basis.
• Older equipment may become less attractive for demanding workloads as newer chips are deployed.
• Rising infrastructure and financing costs increase the cost of carrying older assets.
• Faster technology cycles could result in more frequent replacement than current depreciation periods assume.

The conclusion is not that current useful lives are wrong. The more important point is that the assumption is becoming increasingly testable as market data develops.

Practical Perspective

The Case for Longer Useful Lives

There are several reasons why current useful-life assumptions may prove reasonable:

• Older GPUs are still generating value. Rental rates for some older generation accelerators have increased, rather than declined. This suggests that older equipment can continue to generate meaningful revenue even as newer chips enter the market.
• Older chips can move to different workloads. The newest chips may be used for demanding AI training, while older generations can continue to support inference, fine-tuning, and other workloads. This can extend the economic usefulness of existing equipment.
• Lower compute costs can increase demand. As the cost of running AI workloads falls, more applications may become economically viable. If demand for compute grows faster than efficiency improves, overall demand for infrastructure could remain strong.
• AI revenues are growing rapidly. The revenue growth of leading AI labs provides additional support for the large computing commitments being made today. For example, one major AI lab’s revenue run-rate increased from approximately USD 4 billion to more than USD 65 billion in fourteen months.

What does this mean for useful lives?

These factors provide a reasonable basis for expecting some AI infrastructure to remain economically useful for longer than the current technology cycle might suggest.

The question is therefore not whether a three-year or six-year useful life is correct in isolation. The more important question is whether the assumed useful life continues to be supported by

• rental rates and utilization;
• demand for older generation chips;
• technology and performance improvements;
• replacement patterns; and
• growth in AI workloads and revenues.

As more market data becomes available, these assumptions will become easier to assess. The focus should shift from the assumption itself to the evidence supporting it.

The case for longer useful lives is credible, but the supporting evidence is still developing. As compute pricing, secondary market data, and utilization of information become more observable, companies and investors will have more evidence to assess whether current useful life assumptions remain appropriate. This makes the next phase of the market particularly important for financial reporting.

8.1 What drives the economics: seven variables, six set by markets, one set by management

The useful-life assumption should not be viewed in isolation. The economics of compute are being shaped by several variables simultaneously.

Six market variables: One management estimate

Six of the seven key variables are market-driven, and most are moving against the existing asset base: output prices are falling, input costs are rising, capital is becoming more expensive, and new technology is reducing the value of older generations.

Useful life is the exception. It is management’s key discretionary estimate, which is why it is likely to face the greatest scrutiny.

And while depreciation is noncash, it is not inconsequential. Useful lives affect quarterly earnings, borrowing arrangements, replacement capex requirements, and the potential for impairment.

9. Key Accounting, Reporting, and Control Considerations

Most discussions about AI infrastructure focus on chips, power, and capital spending. But behind these investments are accounting and control issues that will increasingly matter to CFOs, controllers, audit committees, and auditors, including estimates, consolidation, guarantees, revenue visibility, and internal controls.

Five areas stand out

One important point: if your exposure to compute comes through a multi-year purchase or capacity commitment rather than owning the accelerators, these issues do not disappear.

A capacity contract may contain an embedded lease, potentially bringing another party’s hardware onto your balance sheet. The supplier’s pricing and ability to fulfill a long-term contract may also depend on the useful-life assumptions it has made. And a take-or-pay commitment can shift the impact of the supplier’s assumptions to the buyer.

These questions matter not only to owners of compute infrastructure, but also to those buying compute capacity.

9.1 Useful lives and residual values

Even a small change in the useful life assigned to accelerators can have a material effect on reported earnings, and the estimate is about to acquire an external benchmark for the first time. Boards should expect an evidence-based process rather than an annual policy formality.

• Economic life, not physical life. The question is not how long the equipment can run, but how long it is expected to remain economically productive as successor generations ship and output prices fall.
• Evidence, not assertion. Technology roadmaps, workload migration patterns, utilization data, secondary-market pricing, and once liquid, forward curves are all relevant support for the estimate.
• Residual value is no longer automatically zero. A developing secondary market for accelerators makes a zero residual assumption something that must be supported rather than assumed.
• Changes are changes in the estimate. Revisions are applied prospectively and disclosed where material, which places a premium on documenting the evidence at the date of the change.

Implementation matters

• Bring the right data together. Important evidence may sit outside the finance function, including technology roadmaps, equipment utilization, deployment data, and secondary market pricing. Finance should establish a regular process to collect and review this information with the technology and procurement teams.
• Group assets based on how they are used. Useful lives may differ by chip generation, workload, and location. For example, equipment used for AI training may have different economics from equipment used for inference. Applying one useful life to the entire fleet may therefore not reflect the underlying economics.
• Keep a clear record of the assessment. Document the information used to determine the useful life, including technology roadmaps, pricing data, utilization information, and the date of each source. This will allow the assessment to be understood and reproduced in future periods.
• Start before market data becomes available. Establish the data collection and asset segmentation process before new market pricing becomes available. This will allow companies to compare market evidence with an already documented view of their own asset base.

Accounting recommendation

• Establish a formal, useful-life assessment process. Review the useful lives of AI infrastructure regularly, with particular focus on technology refresh cycles and changes in asset utilization, rather than relying only on the annual reporting cycle.
• Use external market evidence. Compare the company’s assumptions with available secondary market prices, utilization data, and, when available, forward pricing for compute. This provides external evidence to support or challenge the existing assumptions.
• Consider the disclosure implications. Where AI infrastructure is material to the financial statements, consider whether the useful-life and depreciation assumptions represent a significant accounting estimate and whether additional disclosure, including sensitivity information where relevant, would help users understand the impact on reported results.

9.2 Consolidation, leases, and financing structures

Capacity is increasingly developed through build-to-suit arrangements, joint ventures, and special purpose vehicles. These change what appears on the balance sheet without changing the economic exposure, and the assessment has to look past legal ownership to rights, obligations, control, and risk.

• Power and economics are both required. Under US GAAP the variable interest entity analysis, and under IFRS the single control model, can reach different conclusions on the same structure. Groups reporting under both need one set of judgments that survives each.
• Lease term judgment carries the weight. Where a short-stated term sits alongside a long residual value guarantee, the lease term conclusion determines how much of the arrangement is recognized at all.
• Guarantees are the quiet exposure. Residual value guarantees, capacity backstops, and credit support given to landlords require maximum potential payment disclosure and are frequently drafted to sit just outside the recognition threshold.
• Reported capital expenditure is not the measure of commitment. A reduction in reported capex may reflect a change in structure rather than a change in investment. The commitments note, not the cash flow statement, is where the obligation appears.

Implementation matters

• Review the arrangements before contracts are signed. Obtain term sheets, side letters, guarantees, and put/call arrangements early in the process. Involve finance when the key terms are still being negotiated, rather than after the agreement is finalized.
• Maintain one central register for infrastructure entities. Track ownership, third-party debt, guarantees, control rights, and the group’s share of cash flows in one place. Changes to any of these items should trigger a reassessment of the consolidation conclusion.
• Document one clear assessment across reporting frameworks. For groups reporting under both US GAAP and IFRS, document the analysis for each framework in one place. This helps ensure that the conclusions remain consistent and differences are clearly understood.
• Track key lease dates. Include renewal, extension, and residual value option dates in the close calendar. Reassess the lease term when circumstances change or when an option becomes reasonably certain to be exercised.

Accounting recommendation

• Establish an early accounting review. Significant infrastructure arrangements should be reviewed by accounting, treasury, legal, and business teams before contracts are executed.
• Consider the accounting issues together. Assess leases, consolidation, guarantees, and contractual commitments as connected issues. The conclusion in one area may affect the analysis in another.
• Explain the overall exposure clearly. Where an infrastructure vehicle is not consolidated, consider presenting its third-party debt, the group’s obligations, and the financial support provided by the group together in one disclosure.

9.3 Revenue, backlog, and remaining performance obligations

Large contracted backlogs provide genuine visibility into demand. The headline figure, however, does not necessarily translate into revenue at today’s pricing, and the enforceable term governs what belongs in it.

• Cancellable periods are not performance obligations. Where either party can terminate on short notice, the enforceable term is short, however long the contract’s stated tenor. Practice varies, and the variation is not visible to readers.
• Backlog is not a defined term. Supplier “order book” and landlord “contracted revenue” are non-GAAP measures with no definition. Where they are used, the basis of preparation matters more than the number.
• Counterparty credit belongs in the assessment. Where a multi-year obligation rests with an unrated private counterparty whose revenues are an order of magnitude below its commitments, collectability is a critical judgment rather than a routine one.
• Falling output prices reach existing contracts. Usage-based pricing and repricing mechanics mean that declining compute prices can change the revenue profile of commitments already signed.

Implementation matters

• Capture the key contract terms. Maintain key terms such as termination rights, notice periods, minimum purchase commitments, and pricing changes at the contract level. Update these terms when contracts are amended so that the revenue assessment reflects the current agreement.
• Reconcile backlog to signed contracts. Before reporting backlog, reconcile it to the underlying executed contracts. Clearly distinguish between amounts supported by enforceable contracts and amounts based on commercial expectations or pipeline activity.
• Assess the customer’s ability to pay. For large, long-term commitments from private or unrated customers, establish a documented credit assessment and update it regularly. The size of the contract alone should not be viewed as evidence that the amounts will be collected.
• Consider pricing changes upfront. Where contracts allow pricing to decrease over time, assess the potential impact on future revenue when the contract is entered into. Do not wait until a decline in pricing or demand creates a separate accounting concern.

Accounting recommendation

• Disaggregate remaining performance obligations by enforceability and, where material, by counterparty credit standing.
• Where non-GAAP backlog is disclosed, publish the basis of preparation alongside it, including whether cancellable periods are included.
• Treat collectability on long-dated offtakes from private counterparties as a critical estimate, supported by a documented credit assessment rather than by the size of the contract.

9.4 Investments in customers, and customers’ investments in you

This is the area where practice has moved furthest ahead of the guidance. Equity stakes now run in every direction across the value chain, and the accounting for each leg is settled, while the accounting for the relationship is not.

• Fair value gains are earnings. Marks on private stakes in the same ecosystem that buys the product can represent a substantial share of reported profit, and depend on round prices in which the holder may itself have participated.
• Consideration payable to a customer. Where an investment is linked to a customer’s purchases, the question of whether part of it is in substance a reduction of revenue is a live one and is settled guidance where the instrument is granted rather than bought.
• Two thresholds, pointing opposite ways. Usually, a holding above 10% of voting interests makes the investor a principal owner and a related party of the investee; the obligation on the investor turns instead on significant influence, presumed at 20%. The practical result is that the disclosure falls on the private party. Boards should be briefed on the exposure and whether or not either threshold is met.
• Valuation discipline. Material private stakes require robust valuation processes, with methods, key assumptions, and sensitivities disclosed where the earnings impact is significant.

Implementation matters

• Map every equity relationship across the value chain. Maintain one register of equity stakes, convertible instruments and warrants held in customers, suppliers and infrastructure partners, and of any interests those parties hold in the group. For each, record the ownership percentage, board or observer rights, and the revenue, purchases, commitments and receivables with the same counterparty. This lets the investment and the commercial relationship be assessed together rather than by separate teams.
• Assess investment and commercial agreements at the same time. Where an equity investment and a supply, capacity or financing agreement with the same counterparty are negotiated together or close together, review them as a package before signing. Identify whether the investment price, purchase commitments, pricing terms or exclusivity provisions depend on one another.
• Look beyond the ownership percentage when assessing influence. Significant influence can exist below 20% where the investor has board representation, participates in policy decisions, supplies critical technology or has material transactions with the investee. In an ecosystem where the investor is also a key supplier, these indicators may be present even at low ownership levels. Document the assessment and revisit it when rights, ownership or the commercial relationship change.
• Establish a valuation process for private stakes that does not rely on the round alone. Where a stake is remeasured on a funding round in which the group participated, assess whether that round is an orderly transaction and whether the price reflects terms specific to the group, such as commercial commitments, preferences or side arrangements. Support each mark with documented methods and assumptions, and reconcile unexpected movements before the close.

Accounting recommendation

• Test the investment price against fair value. Where an investment is linked to a customer’s purchases, compare the price paid for the equity with its fair value. Any excess may, in substance, be consideration payable to the customer and a reduction of revenue rather than part of the investment’s cost.
• Present the relationship, not just its parts. Where material, disclose together the investment, the revenue earned from the investee, the related remaining performance obligations and receivables, and any commitments or financial support provided. Apply this at any ownership level, rather than only where the related party or equity method thresholds are met.
• Make the earnings effect of private stakes visible. Where fair value gains on stakes in customers or partners are significant, disclose them separately, with the valuation method, the key inputs and whether the group participated in the round that set the price. Apply the same treatment consistently in non-GAAP measures and in the discussion of results.

9.5 Controls over a program of this scale

The pace and size of the buildout raise the risk of error across capitalization, project accounting, depreciation, impairment, and disclosure, and concentrate that risk in exactly the judgments least amenable to routine testing.

• Management review controls carry the load. Where conclusions depend on forecasts, technology assumptions and complex contracts, the precision and documentation of the review control are what an inspector will examine first.
• Estimates need owners. Useful lives, residual values, impairment indicators, lease term, and consolidation conclusions should each have a named owner, an evidence standard, and a review cadence.
• Contract intake is a control point. Structures are determined at negotiation. A control that operates after signature can only report the answer, not influence it.
• Monitoring must extend outside the entity. Compute price indices, successor product launches and counterparty credit developments are now inputs to the estimates process and need a monitoring control.

Implementation matters

• Define what each review control should cover. For every management review, clearly document what is being reviewed, the relevant threshold, what action is required if the threshold is exceeded, and what evidence should be retained.
• Set clear rules for when costs are capitalized. Establish a consistent trigger for capitalization, such as the commissioning or in-service date, and apply it consistently across locations. This reduces the need for case-by-case judgments.
• Have a backup for key accounting estimates. Assign a deputy for each key estimate owner and ensure that the supporting analysis and evidence are transferred when responsibilities change.
• Align controls with key project milestones. Review controls when significant events occur, such as commissioning, technology refreshes, or major changes in utilization. Do not rely only on quarter-end reviews to identify changes.

Accounting recommendation

• Treat AI infrastructure as a significant financial reporting program. The control framework should cover the full asset lifecycle, including approval, construction, capitalization, depreciation, impairment, and disposal.
• Assign ownership for monitoring market and technology changes. Designate responsibility for monitoring compute pricing, utilization, technology refresh cycles, and successor products. Define how this information will feed into useful-life, depreciation, and impairment assessments.
• Keep the audit committee informed of the overall structure. Where AI infrastructure is significant, provide the audit committee with visibility into the key contracts, financing arrangements, accounting judgments, and related controls, not just the resulting balances.

9.6 What a reserve report for compute would contain

Oil received its disclosure architecture because investors could not value a resource they could not see. The four items below are the compute equivalent of what they currently cannot see. None requires a new standard; every one is a disclosure that a preparer could adopt voluntarily tomorrow, that a comment letter could require next year, and that could be codified later.

• Revenue and backlog from investees, at any ownership level. Revenue, remaining performance obligations, and receivables from customers in which the entity holds equity, convertible instruments, or warrants with the investment, the commitment, and any linkage between them are presented together.
• Financial support extended to customers and landlords. Guarantees, capacity backstops, prepayments, and vendor financing, with maximum exposure and the third-party debt each one supports.
• Unconsolidated compute vehicles, shown whole. For each material campus joint venture or financing vehicle: its third-party debt, the entity’s lease and guarantee obligations to it, and the share of its cash flows the entity itself provides the three facts that are individually disclosed today and never presented together.
• The useful-life estimate, with its evidence. Accelerator lives and residual values, the secondary market and forward curve data they reconcile to, and the sensitivity of earnings and borrowing bases to a one-year change.

9.7 Cross-border implications

The buildout is global, and the frameworks diverge in the places that matter most to compute economics.

• Impairment. US GAAP tests recoverability on undiscounted cash flows; IFRS and Ind AS compare carrying value to the higher of fair value less costs of disposal and value in use. An observable, declining price curve therefore bites earlier outside the United States.
• Componentization. Required under IFRS and Ind AS, where significant components have different lives; permitted but not required under US GAAP, a live difference for a data center combining accelerators, networking, and power systems.
• Useful lives. IFRS and Ind AS require annual review of lives and residual values, which makes an unchanged six-year assumption harder to sustain without fresh evidence.
• Consolidation and leases. The variable interest entity model and the single control model can reach different answers on the same vehicle; lease classification follows a single lessee model under IFRS, against a dual model under US GAAP.

For the Middle East, the exposure is sovereign: funds converting a depleting geological endowment into a manufactured one, with the reporting consequences landing in consolidation and structured-entity disclosure for compute vehicles, and in impairment of vintages displaced by each new generation. In India, abundant talent and constrained power mean the exposure arrives through global capability centers, data center developers, and listed technology balance sheets, where embedded lease analysis of capacity contracts and commitment disclosure under Schedule III will be the pressure points.