The AI Infrastructure Reckoning

Executive Summary

Generative AI has triggered an unprecedented corporate investment cycle. When ChatGPT launched in November 2022, Amazon, Microsoft, Alphabet, and Meta collectively spent approximately USD 158 billion per year on capital expenditures, comfortably funded by free cash flow. Four years later, the same companies have guided to approximately USD 760 billion for 2026 and consensus for 2027 now clusters at USD 1.0–1.3 trillion.

Against this backdrop, this publication examines the AI infrastructure investment cycle and its implications for hyperscalers, investors, and financial reporting.

1. The Three Eras of Hyperscaler Capital

Era I — The cloud buildout (2015–2022): disciplined and self-funded

• The first phase of hyperscaler investment was largely about building the cloud infrastructure needed to support a growing digital economy. In 2015, the Big-4 hyperscalers invested USD 23.8 billion in data center infrastructure. Through 2020, annual capex growth averaged roughly 32%, funded entirely from operating cash flow. By 2022, combined spend had reached approximately USD 158 billion, with capex-to-revenue ratios in the low-to-mid teens.
• For investors, this period was defined by capital discipline. Spending was expected to support clear growth opportunities and remain within the company’s ability to generate cash. When Meta increased its capital spending plans by two-thirds in 2022, at a time when the advertising market was weakening, the market reacted sharply.

Era II — The inflection (2023–2025): AI changes the scale of investment

• The launch of ChatGPT brought a new dimension to the infrastructure equation. Generative AI showed how much computing power is required not only to train large models, but also to run them at scale. Hyperscalers responded with a rapid increase in spending.
• Combined capex across the five largest players has grown at an average of roughly 72% per year since the release of GPT-4. On a Big-4 basis, annual spending increased from USD 152 billion in 2023 to USD 251 billion in 2024 and USD 416 billion in 2025. That represents two consecutive years of approximately 66% growth.
• All four companies reached record levels of capital intensity during this period. At Amazon and Meta, free cash flow also began to decline, providing the first clear indication that the traditional model of funding the buildout from internally generated cash was coming under pressure.

Era III — The next phase (2026–2030): spending moves toward a trillion dollars

• The scale of investment has increased again in 2026.
• Amazon has raised its planned capital spending to USD 220 billion. Alphabet has raised its outlook twice, most recently to USD 195–205 billion. Microsoft has maintained its capital spending outlook at USD 175 billion, down from an earlier USD 190 billion following a shift in the treatment of certain leases from finance to operating. Meta has raised the lower end of its range to USD 130–145 billion.
• Memory and component inflation has also become a factor, with three of the four companies citing higher costs. Looking ahead, current consensus estimates put 2027 hyperscaler capex at around USD 1.0–1.3 trillion.

The changes across these three eras are significant. What began as a relatively disciplined, internally funded cloud investment cycle has evolved into a much larger infrastructure buildout, with funding, depreciation, and returns increasingly important to the investment story.

Source: Company filings and guidance (Amazon, Microsoft, Alphabet, Meta); includes finance leases where disclosed; calendar-year basis. 2026 reflects guidance as raised through the 29–30 July 2026 earnings reports. 2027–2030 is an illustrative base case consistent with Street 2027 consensus of $1.0–1.3 trillion.

2. Why Simple Comparisons Can Be Misleading

• Comparing hyperscaler capex is not as simple as putting the numbers side by side. Each company uses slightly different definitions and reporting practices. For example, Microsoft includes finance lease right-of-use assets and accrued but unpaid property, plant, and equipment (PP&E) in its reported capex. Meta includes principal payments on finance leases, while Alphabet reports cash purchases of PP&E.
• These differences can have a meaningful impact on the headline numbers. For 2024, estimates of combined Big-4 capex range from USD 226 billion to USD 251 billion, largely depending on how finance leases are treated.
• The gap becomes more important as spending increases. For 2026, estimates of Big-4 capex range from around USD 730 billion on a cash basis to USD 760 billion or more when finance leases are included. Microsoft also disclosed in July that a change in accounting classification will affect how it reports capex going forward.
• The bottom line is that reported capex is not always a like-for-like measure of investment. As AI infrastructure spending reaches unprecedented levels, differences in accounting and reporting can materially alter the comparison. The numbers, therefore, need to be normalized before drawing conclusions about the scale of investment.

3. What the four companies actually reported in July 2026

In July 2026, all four hyperscalers reported their latest results. All four also raised or extended their spending plans. Their results show a common theme: AI investment is growing rapidly, but the impact on growth, cash flow, and profitability is not the same across the four companies.

The key numbers from their latest reported results are set out below:

Sources: Amazon Form 8-K and press release, 30 July 2026 (sec.gov); Microsoft financial year 2026 fourth quarter release, 29 July 2026 (microsoft.com/investor); Alphabet second quarter 2026 release, 22 July 2026; Meta second quarter 2026 release, 29 July 2026 (investor.atmeta.com). Microsoft is reporting a June financial year-end, so its figures are for the fourth quarter and the full year.

The numbers show why the current investment cycle deserves a closer look. Revenue is growing strongly across the group, and cloud growth is particularly strong at Microsoft and Alphabet. At the same time, the cash cost of building AI infrastructure is becoming increasingly visible. Amazon and Alphabet reported negative free cash flow on the measures highlighted above, while Meta’s free cash flow was only USD 784 million after USD 31.1 billion of capital spending.

Amazon’s quarter also included an unusual item that is important when looking at its reported earnings. Net income of USD 62.6 billion included USD 53.4 billion of pre-tax gains recorded outside operating income, arising mainly from its investment in Anthropic. This is a valuation gain on a private investment rather than cash generated from selling an asset.

The broader point is the change in capital intensity. Capex-to-revenue ratios of 31% to 62% across the Big-4 (refer to the chart below) are a long way from the asset-light model that defined these businesses for much of the cloud era.

As the balance sheets become more capital-intensive, growth and margins are no longer enough to understand the performance of these companies. Return on invested capital, asset utilization, depreciation, and funding will increasingly matter alongside revenue growth and operating margins.

Source: Uniqus estimates, based on companies’ latest 2026 CapEx guidance and consensus revenue estimates; Microsoft presented on a fiscal-2026 basis.

4. Funding the AI Buildout

4.1 The funding model has flipped

• For most of the last decade, hyperscalers funded their investment largely from operating cash flow. That model is now under pressure.
• J.P. Morgan Asset Management estimates that bonds issued by these companies accounted for about 2% of total US dollar investment-grade issuance between 2022 and 2024. It expects that share to reach about 9% in 2026, with roughly USD 300 billion of investment-grade debt issuance from the group over the coming year. It also expects free cash flow across the group to turn negative in the near term.
• External capital is therefore becoming a much larger part of the funding mix. Alphabet priced an USD 84.75 billion equity offering in June 2026 to support its AI infrastructure and compute investments. Morgan Stanley estimates that global data center investment could face a financing gap of around USD 1.5 trillion through 2028, with private credit, corporate debt, and securitized markets expected to help fill the gap. Goldman Sachs also notes that infrastructure funds raised a record USD 221 billion in 2025 and had around USD 400 billion of available capital as of September 2025.
• The pressure is already visible in the latest results. Amazon’s trailing free cash flow had moved to an outflow of USD 7.6 billion. Meta’s quarterly free cash flow fell 91% to USD 784 million, while Alphabet reported its first quarter of negative free cash flow.
• As more outside capital enters the market, the way AI infrastructure is financed is also changing. Capacity is increasingly being developed through build-to-suit arrangements, leases, joint ventures, and special-purpose vehicles. Meta’s USD 14 billion El Paso data center, for example, sits in a vehicle in which funds managed by BlackRock hold an 80% interest. About USD 12.5 billion of the funding is provided through debt, with Meta remaining the sole tenant.

4.2 How this looks from outside the companies

• The funding shift is not limited to the hyperscalers themselves. It is becoming a broader feature of the data center market.
• The Bank for International Settlements (BIS) examined the issue in January 2026. It concluded that the scale of planned investment will require a move away from funding primarily through operating cash flow toward debt, with private credit expected to play a growing role. It assessed the broader macroeconomic and financial stability risks as moderate, but noted that the sustainability of the investment cycle depends on AI companies meeting high earnings expectations. It also observed that equity valuations have moved ahead of what debt markets are pricing.
• The BIS estimated that private credit to AI-related borrowers had already exceeded USD 200 billion and could reach USD 300–600 billion by 2030.
• The scale of the underlying infrastructure market is also expanding quickly. Dell’Oro Group raised its forecast for worldwide data center capital expenditure in August 2026 to more than USD 3 trillion by 2030, almost twice its earlier February forecast of USD 1.7 trillion.

This means the funding question is becoming a reporting question as well. As more AI infrastructure is financed through debt, private capital, leases, and joint ventures, investors will need greater visibility into where the assets sit, who controls them, who carries the debt, and what commitments sit outside the reported capex number.

For companies, this makes the quality and transparency of financial reporting increasingly important—not just for accounting compliance, but for giving investors a clear picture of the risks and commitments behind the AI buildout.

5. Depreciation is the bill that arrives later

• Capital spending is a cash outflow, but it does not reduce reported profit when the cash is spent. The expense is recognized on the income statement later, through depreciation, spread over the asset’s useful life.
• This is why much of the earnings impact of the spending over the last three years is still ahead. Microsoft provides a clear example. Its depreciation expense increased from USD 11.0 billion in the financial year 2023 to USD 15.2 billion in 2024 and to USD 22.0 billion in 2025, according to its annual report. The increase reflects infrastructure that has already been built and placed in service, rather than spending that is still to come.
• Morgan Stanley estimates that Microsoft, Oracle, Meta, and Alphabet could record more than USD 680 billion of cumulative depreciation over four years and this will only increase if we factor Amazon into the mix, which had previously reduced the server lifespan from six years to five years. The estimate illustrates the potential scale of the earnings impact as today’s infrastructure moves through its useful lives.
• Alphabet provides another useful example. Its servers are depreciated over six years, which means assets being installed today could continue to affect earnings well into the early 2030s. Alphabet has also warned that higher infrastructure investment will increase depreciation and related data center operating costs, putting pressure on its profit and loss account.

5.1 The Useful-Life Estimate Matters More Than It Looks

One of the most important assumptions in depreciation accounting is the useful life assigned to an asset. It is a judgment, and even a relatively small change can have a meaningful effect on reported earnings.

The hyperscalers have changed their useful-life assumptions in both directions in recent years:

Sources: company Forms 10-K and 10-Q filed with the SEC; Amazon’s fourth quarter 2024 useful-life study and 2025 quarterly disclosures; Microsoft financial year 2026 fourth quarter earnings call.

5.2 Power and memory

The constraints on AI infrastructure are also changing.

GPUs remain constrained, but advance commitments are making them easier to secure. Power and the infrastructure needed to deliver it are becoming a more significant constraint. J.P. Morgan estimates that 122 GW of global data center capacity will be installed between 2026 and 2030, highlighting the scale of the power and construction investment required to support the AI buildout.

Input costs are also affecting the spending mix. Amazon attributed the increase in its 2026 capital spending outlook from USD 200 billion to USD 220 billion primarily to higher memory costs. The company also indicated that, despite the additional spending, AWS would remain short of capacity relative to demand through 2026 and 2027.

The result is a more complicated investment equation. Companies are spending more to build capacity, while also having to consider how quickly that capacity becomes obsolete, how long it will remain productive, and what economics it can generate over its useful life.

6. The Revenue Side: The Price of AI is falling

Every dollar of this capex ultimately produces one product: computed intelligence, sold by the token, the seat, or the impression. The capex side of the ledger is disclosed in exhaustive detail. The revenue side also needs more attention.

6.1 Demand under contract is clear

Today’s AI infrastructure is being deployed against identifiable customer demand, with major cloud providers reporting large contracted backlogs and persistent capacity constraints.

The demand evidence strengthened materially in July. Microsoft’s commercial RPO rose 84% to USD 678 billion; Google Cloud’s backlog reached USD 514 billion and AWS reported USD 496 billion together representing approximately USD 1.69 trillion of contracted future revenue. All three companies indicate demand still outpaces supply. If that revenue curve materializes, today’s infrastructure spending would represent capacity built against substantial contracted and emerging demand, not speculative capacity for its own sake.

Sources: Microsoft financial year 2026 fourth quarter release; Alphabet second quarter 2026 release results. The AWS figure of USD 496 billion is drawn from secondary reporting of Amazon’s second quarter.

6.2 But the prices are falling

AI intelligence is becoming a commodity. As with other commodities, prices tend to move toward the cost of the lowest-cost provider that can deliver adequate quality.

Today, the cost of one million tokens varies dramatically: about USD 56 for premium frontier models, USD 26 for mid-tier models, around USD 1–USD 1.50 for efficient models from companies such as Meta and Google, and as low as USD 0.50 for leading Chinese open-weight models. That means a 112x price difference for models whose quality is increasingly similar on public benchmarks. As model performance converges, price and cost efficiency are likely to become increasingly important.

Source: Published list API prices (blended input/output), company pricing pages, Artificial Analysis, mid-2026. Actual negotiated and committed-use prices could be lower. Approximate; tiers and models change frequently.

6.3 But it is showing up inside customers’ own costs

The decline in AI prices is not limited to the companies building the models. For enterprises, AI costs can still add up quickly, particularly when teams use expensive frontier models for tasks that could be handled by lower-cost models.

This is already becoming a bigger focus for companies managing their cloud spend. The FinOps Foundation’s 2026 survey of 1,192 practitioners, representing more than USD 83 billion of annual cloud spend, found that 98% are now managing AI spend. This compares with 63% a year earlier and 31% two years earlier.

As AI use expands, companies will need to pay closer attention to the cost of each use case. This could lead CFOs and technology teams to introduce tighter controls over AI spending, including monitoring token usage, selecting the right model for each task, and understanding the cost of AI across different applications.

The issue is therefore not simply how much AI a company is using, but whether it is using the right level of AI capability for the job at the right cost.

6.4 The scissor analysis: 24x volume against 96% deflation

• Goldman Sachs estimates that AI token usage could grow by about 24x by 2030.
• But higher usage does not automatically mean higher revenue. The price paid for each token is also falling rapidly, with some estimates showing around a 96% decline across comparable model generations.
• The chart shows what happens when these two trends move together:
  • If prices fall by 90%, the 24x increase in usage more than offsets the price decline, and token revenue grows to about 2.4x.
  • If prices fall by 96%, the increase in usage is almost exactly offset by the price decline, leaving revenue broadly flat.
  • If prices fall by 98%, the price decline is faster than the growth in usage, and revenue falls by about half.
• This creates a “scissor” effect: the same growth in AI usage can produce very different revenue outcomes depending on how quickly prices fall.
• The key question is therefore not only how much AI will be used, but how quickly the price of AI will decline.
• For the companies investing billions in AI infrastructure, this is an important economic risk. Demand may continue to grow strongly, but if AI prices fall faster than usage increases, more and more volume will be needed just to maintain revenue.

Source: Uniqus illustration. Volume path: approximately 24x growth to 2030 (Goldman Sachs projection); price paths show cumulative declines of 90%, 96% and 98% over the same period. Indexed to 2025 = 100. Illustrative — realized prices, mix, and committed use discounts will differ.

7. Key Accounting, Reporting, and Control Considerations

Most of the discussion around AI infrastructure focuses on chips, data centers, and power. But behind the investment is another set of issues that will increasingly matter to CFOs, finance teams, audit committees, and auditors. These include accounting estimates, asset valuation, financing arrangements, revenue visibility, and internal controls. Seven areas stand out.

7.1 Useful lives

As discussed in Section 5 above, even a small change in the useful life of servers and other AI infrastructure can have a meaningful impact on reported earnings. This is likely to attract greater attention from investors, auditors, and regulators as the asset base grows.

Boards and audit committees should expect companies to have a clear, evidence-based process for setting and revisiting useful lives. This should be more than an annual policy exercise. The assessment should consider factors such as technology changes, workload migration, asset utilization, and the expected economic life of the equipment.

Key accounting implications:

• Useful life assumptions for servers, accelerators, networking equipment, and other AI infrastructure can have a significant impact on depreciation expense and reported earnings.
• Rapid technology changes and falling AI compute prices make it important to assess whether the estimated useful life continues to reflect the period over which an asset is expected to generate economic benefits.
• Changes in useful life are generally changes in accounting estimates and should be supported by evidence and applied prospectively under the applicable accounting framework.
• Management should maintain clear support for these estimates, including technology roadmaps, expected utilization, workload migration plans, replacement cycles, and evidence of technological obsolescence.
• The key question is not simply how long the equipment can physically operate, but how long it is expected to remain economically useful.

7.2 Componentization, residual values, and end-of-life obligations

AI infrastructure is not a single, homogeneous asset. Data centers typically comprise servers, accelerators, networking equipment, power systems, and other components that may have different useful lives and disposal profiles. As the market for used AI equipment develops and companies enter into more complex infrastructure and site arrangements, questions around componentization, residual values, and end-of-life obligations are becoming increasingly relevant. These considerations can affect both the amount and timing of depreciation and the recognition of related liabilities.

Key accounting implications:

7.3 Impairment

The rapid pace of change in AI technology and economics creates a greater need to assess whether AI infrastructure will continue to generate the expected economic benefits over its useful life. Falling compute prices, changes in customer demand, lower-than-expected utilization, and faster technology cycles could affect the future cash flows associated with these assets. As the scale of investment increases, companies will need to consider whether these factors indicate impairment and whether existing asset groupings remain appropriate.

Key accounting implications:

• Rapid changes in AI technology mean impairment assessments may need to consider more than physical obsolescence or underutilization.
• A sustained decline in the price of AI compute could reduce the future cash flows expected from infrastructure. If prices decline faster than usage increases, this could indicate impairment.
• Asset grouping is also important. Training clusters, inference infrastructure, and data centers may share assets but generate or support cash flows at different levels.
• Companies should ensure that impairment assessments reflect the economics of the underlying assets, including expected utilization, pricing, technology changes, and future demand.
• The recognition and measurement requirements differ under US GAAP, IFRS, and Ind AS, making framework-specific analysis important for companies operating across jurisdictions.

7.4 Leases, joint ventures and special purpose vehicles

The scale of AI infrastructure investment is also driving greater use of alternative financing and ownership structures, including leases, joint ventures, and special-purpose vehicles. These arrangements can change how assets, liabilities, and commitments are reflected in the financial statements, even where the underlying economic exposure remains significant. The accounting assessment, therefore, needs to look beyond legal ownership and consider the rights, obligations, control, and risks associated with each arrangement.

Key accounting implications:

• The way AI infrastructure is financed or structured can significantly affect what appears on the balance sheet.
• Build-to-suit arrangements, leases, joint ventures, and special-purpose vehicles may result in different accounting outcomes depending on the rights, obligations, and level of control involved.
• Under US GAAP, companies need to consider the variable interest entity requirements under ASC 810. Under IFRS, the analysis focuses on control under IFRS 10.
• Lease arrangements also require assessment under ASC 842 or IFRS 16, as applicable.
• Companies should look beyond ownership and consider residual guarantees, purchase options, take-or-pay commitments, and other contractual terms that may create significant economic exposure.
• A reduction in reported capex does not necessarily mean a reduction in the underlying investment or commitment. Investors may need to consider the broader financing structure to understand the economics of the buildout.

7.5 Construction accounting, capitalized interest, and Construction in progress (CIP)

The scale and duration of the AI infrastructure buildout are making construction accounting increasingly important. Data center projects can involve significant investments in land, buildings, power infrastructure, servers, and other equipment, often developed over multiple years and across different locations. Determining which costs qualify for capitalization, how long capitalization should continue, and when an asset is ready for its intended use can therefore have a meaningful impact on the timing of asset recognition and depreciation.

Key accounting implications:

• The size and duration of AI infrastructure projects increase the importance of project-level accounting and controls.
• Companies should have clear policies for identifying costs that qualify for capitalization, including construction costs, directly attributable costs, and, where applicable, capitalized interest.
• Debt-funded construction can result in significant capitalized interest under the applicable accounting requirements, including ASC 835-20 and IAS 23.
• The timing of when an asset is ready for its intended use is particularly important. Depreciation generally begins when the asset is placed in service, even if utilization takes time to ramp up.
• Canceled, delayed, or materially re-scoped projects should be reviewed for recoverability and potential write-downs.
• Strong project accounting is increasingly important as multiple data centers, power infrastructure, and technology assets are developed simultaneously across different locations.

7.6 Revenue, backlog, and disclosure quality

As AI infrastructure investment continues to grow, the focus is shifting from the scale of investment to the ability to convert that capacity into sustainable revenue. Large customer commitments and backlog can provide visibility into future demand, but the underlying economics may change as AI compute prices, usage patterns, and customer requirements evolve. Contract duration, usage-based pricing, cancellation rights, and potential repricing can therefore affect both the timing and economics of future revenue.

Key accounting implications:

• Backlog and remaining performance obligations can provide useful visibility into future revenue, but the headline amount does not necessarily translate into revenue at today’s pricing or margins.
• Contract duration, cancellation rights, usage commitments, pricing mechanisms, and potential repricing can affect both the amount and timing of revenue.
• As AI compute prices decline, changes in customer usage and contract pricing could affect the revenue profile of existing commitments.
• Companies should also consider whether arrangements with strategic investors, partners, or related parties create additional disclosure considerations.
• Greater visibility into depreciation and amortization through the income statement and disaggregated expense information can help investors understand the earnings impact of the infrastructure buildout.
• Fair value changes in private AI investments can also create significant earnings volatility. Where material, companies should provide sufficient information about valuation methods, key assumptions, and sensitivity to help investors understand the impact.

7.7 Controls over a large spending program

The scale and pace of AI infrastructure investment increase the complexity and risk of the underlying financial reporting process. With significant amounts being spent across multiple projects, locations, and contractual arrangements, companies need reliable processes to ensure that costs are appropriately captured, capitalized, and depreciated, and that key accounting judgments are consistently applied. The effectiveness of these controls becomes increasingly important as the investment program expands and the number of estimates, transactions, and reporting requirements increases.

Key accounting implications:

• The scale of AI infrastructure investment increases the risk of errors in capitalization, project accounting, depreciation, impairment, and financial statement disclosures.
• Companies should have effective controls over capital approvals, project costs, construction-in-progress, placed-in-service dates, and depreciation commencement.
• Controls should also cover the estimates and judgments that can materially affect reported results, including useful lives, residual values, impairment assessments, and lease or SPV conclusions.
• Management review controls become particularly important where significant accounting conclusions depend on forecasts, technology assumptions, or complex contractual arrangements.
• Companies should maintain clear documentation supporting key judgments so that conclusions can be reviewed by internal audit, external auditors, and audit committees.
• Controls should also address commitments, contingencies, and contractual obligations arising from the broader AI infrastructure program.
• Given the pace of investment, control frameworks should evolve alongside the program rather than being strengthened only after the complexity has increased.

Cross-border implications — US GAAP and IFRS

The buildout is global, but the accounting frameworks diverge in areas that matter directly to AI infrastructure economics.

• Component depreciation is required under IAS 16 and only permitted under US GAAP
• Impairment uses the discounted value in use under IAS 36, against the undiscounted screen in ASC 360
• Borrowing costs follow IAS 23 against ASC 835-20
• Consolidation follows IFRS 10 against ASC 810
• Lease classification follows the single model in IFRS 16 against the dual model in ASC 842

Groups reporting under both, and Indian groups applying Ind AS, need one set of judgments that holds up in every framework they report under.