Discuss a Project
Economic Perspective · AI and the Economy

AI Investment and the Returns Behind the Build-Out

Alex Lampton
Anglo Siam Advisory · 2 October 2026
Series AI and the Economy, Issue 2 Draws on company filings, BIS research, competition authority findings and investment theory Data company figures are for the quarter to June 2026 Approach argued economic analysis, not a primary-data report

AI does not have to fail for companies to build more capacity than its returns can justify. Our hypothesis is that spending on the next wave of data centres, chips and power is running ahead of what that capacity will earn. Whether firms keep building depends on the returns they expect over the assets' working lives, set against the full cost of building, powering and financing them. If those expectations weaken, new projects are likely to slow. This report examines the evidence for that hypothesis and what a slowdown would mean for businesses that rely on AI providers for their prices, capacity and terms.

The difficulty is that firms must decide what to build before they know how much customers will use it. Company accounts show large spending and growing cloud sales. Yet cloud revenue includes services that predate the current AI build-out. The same infrastructure can also serve several products.1 Those figures cannot tell us what one more AI data centre will earn. To assess the hypothesis, we examine the public clues we can observe: paying customers, contract commitments, reported prices, and reported financing and operating costs.

This report asks one question: whether the next round of AI capacity will earn its full cost. It does not forecast outside shocks such as export controls, energy prices, interest rates, regulation or a sudden fall in the cost of AI models. Each would work through the same comparison, changing either expected returns or costs, and each could strengthen or weaken the case for further building.

For a company with an established cloud business, new AI capacity may protect existing customers as well as win new revenue. Its return could therefore be real even where AI sales alone look modest. Rivals may face the same incentive, and some of what one firm gains could be market share lost by another. Building can then make sense for each firm while their combined investment exceeds what the wider economic gains justify.2 That possibility does not settle our central question: whether the next round of capacity will earn enough to cover each firm’s full costs.

Scale

What is being invested in?

The scale of investment depends on what we count. The BIS estimates that by mid-2025, spending on US data centres, including the equipment inside them, and on IT manufacturing facilities was equivalent to about 1% of GDP. Its broader measure, which adds other IT equipment and software, reached 5%, above its previous peak during the dot-com boom.3 The 5% figure is total IT investment, not AI spending alone. Even the narrower measure includes chip plants encouraged partly by US industrial policy. The BIS makes a separate historical comparison using the rise in data-centre and IT-manufacturing construction. On that measure, the increase is similar to the US shale boom of the 2010s and about half the rise in IT investment during the dot-com era. Neither comparison tells us whether firms are earning a return on their AI capacity.

Theory

Why firms keep building

Investment decisions turn on the next project. In Jorgenson's framework, a firm compares what an additional asset is expected to earn with the yearly cost of owning it. This user cost reflects financing, depreciation, taxes and expected changes in the asset's price.4 For AI hardware, depreciation and expected price changes matter: chips can become economically obsolete before they stop working, and cheaper successors can further erode their value.5 The comparison shapes how much capital a firm wants to hold. If expected returns weaken, or financing becomes more expensive, it can scale back new orders without switching off the centres it has already built. The construction cost is already spent. A centre can remain worth operating even if it would not be worth building today. It still needs to earn more than its avoidable running costs and the value of any alternative use.6 A slowdown could therefore appear in plans, contracts and construction starts while customers continue using existing AI services.

A project can look worthwhile on today's forecast and still be worth postponing. Building a data centre commits money to a site and equipment, some of which may be specialised and hard to recover. Waiting may reveal more about paying demand, costs and the performance of newer chips. Economists call the value of keeping that choice open an option to wait.7 But waiting has costs of its own. A rival may secure customers or scarce power capacity first; in parts of Europe grid connections can take years, and advanced chip supplies have tightened.8 If the pressure to move early dominates, firms can rationally build ahead of demand. If waiting is more valuable, they may hold back until returns are clearer. Theory identifies both forces. Evidence must tell us which is stronger today.

Evidence

What can company accounts tell us?

Microsoft offers a useful test case. In the quarter ending June 2026, Microsoft reported $41 billion in capital expenditure, including finance leases, and more than 30 million paid Microsoft 365 Copilot seats. Copilot also forms part of its wider E7 software bundle.9 The seats show paid adoption; the expenditure shows the scale of the commitment. But the capital figure covers both AI and other cloud services. The seat count does not reveal Copilot's revenue or cost to serve. Nor can we divide a quarter's spending by a seat total built up over time to estimate a return. The useful question is whether each new block of capacity will earn enough through extra revenue and retained customers to cover its full cost. Microsoft does not publish the figures needed to answer it.

CoreWeave makes the financing question more visible. In the quarter ending June 2026, the specialist AI cloud provider reported $2.6 billion in revenue and a $626 million net loss. Before interest, CoreWeave reported a $49 million operating loss after depreciation. That result uses a six-year estimated life for its technology equipment. If the equipment loses economic value sooner, the cost of providing capacity is higher than its reported depreciation suggests. Net interest expense was $640 million, more than the reported net loss. CoreWeave also had about $104 billion in revenue under contract but not yet recognised. Around 40% is expected within two years; the rest is spread across the following four and a half years.10 Those commitments are evidence of demand, though three customers supplied 72% of the quarter's revenue. The largest supplied 36%, compared with 71% for the largest customer a year earlier. CoreWeave must deliver the promised capacity to recognise the contracted revenue. The backlog does not show what that delivery will cost. The test for the next project is whether its revenue will cover power and other running costs, financing and depreciation. If chips need replacing before the contracts end, those costs matter too.

Outlook

What could slow the build-out?

Firms have reasons to continue building, at least for now. Microsoft said in late July that Azure demand still exceeded its available capacity. In September, CoreWeave said it was signing new compute contracts at higher prices. Some run for only three to six months.11 These are company statements, but both point to strong demand for capacity those firms can supply today. Short contracts cut both ways. They do not lock in prices for the years over which a new centre must recover its costs. Those prices can reset quickly if supply catches up. The pipeline is expanding too. CoreWeave's contracted power rose by about half a gigawatt in six weeks, although that is not capacity already operating. If paying use keeps rising as more centres open, utilisation and prices could justify further construction. If capacity grows faster than paying use, lower utilisation or weaker prices would reduce the expected return on the next project. The question is how much paying demand there will be when the capacity ordered today comes online.

Finance could slow the next round of building even if customers still want capacity. A specialist that relies on borrowing may face higher rates or find that lenders will fund fewer projects. A large platform has more room to draw on cash from established products. But the BIS says rising capital spending is testing that source of finance too, and expects debt to play a greater part. Research by Fazzari, Hubbard and Petersen explains why internal cash flow can matter when outside funding is costly. It does not tell us which AI firms face that constraint.12

A financing-led slowdown would show up in tighter loan terms or delayed projects; a fall in share prices alone would not establish one. The BIS found much the same private-credit loan spreads and maturities for AI-related firms as for other borrowers through 2025. That comparison does not show whether lenders had tightened terms over time. The BIS also contrasts those loan terms with equity valuations that imply exceptional returns. The loans cover broadly defined AI-related borrowers, while the valuations concern listed AI firms. They are not the same companies or the same financial claims. If loan spreads reflect underlying risk, the BIS suggests lenders may be understating it or equity investors overstating future cash flows. In January 2026, the bulletin's authors judged broader macroeconomic and financial stability risks to be moderate.13

Transmission

How would a slowdown reach other businesses?

Slower investment would first mean fewer new orders. If a planned data centre is postponed, builders, electrical contractors and chip suppliers may miss out on work they expected. If construction has begun and then stops, existing jobs may be lost and contracts may need renegotiating. Lower wages and profits can reduce spending elsewhere. This is the chain behind the multiplier developed by Kahn and Keynes.14

We cannot apply a fixed multiplier to AI spending. The BIS uses a rule of thumb that equipment accounts for about three-quarters of a new data centre's cost, and notes that some of it is imported. Construction work is more closely tied to the site. The wider effect depends on where suppliers and workers spend their income, and whether they find other work. Past booms give a sense of scale: the BIS finds that GDP growth in the year after six investment booms averaged more than one percentage point below growth during the booms. The largest fall followed the dot-com boom.15 Other forces affected growth too, so this is context rather than a forecast for AI. The first effect falls on firms involved in building capacity. Businesses buying AI would be affected through their provider's prices, capacity or service.

Businesses using AI could face quite different outcomes. Spare capacity may lower prices if providers compete for customers. A provider unable to finance expansion may limit new capacity or seek higher prices at renewal. Failure could force customers to move data and workflows. None follows automatically from slower construction. The effect can travel through firms the buyer never pays directly: a software vendor may use another company's model on a third company's cloud. A disruption at either layer could reach the buyer. Terms can change without a crisis. Microsoft said in July that it had added usage-based billing to Copilot Cowork and GitHub Copilot alongside seat licences.16

Switching matters in each case. The CMA's 2025 cloud investigation found that data transfer fees and interoperability barriers limited switching, and that Amazon and Microsoft held significant market power. Both announced steps to address those barriers in March 2026. The CMA's ongoing investigation into Microsoft's business software, including Copilot, is examining whether rival AI tools can integrate with it.17 This does not measure the cost of leaving any particular AI application.

Conclusion

What can we conclude?

The public evidence does not establish our hypothesis that firms are building the next wave of AI capacity faster than it can earn back its costs. The risk is credible: competition can reward building early, and the BIS expects debt to finance more of the expansion. The BIS also highlights a tension: AI-related and other private-credit loans carry similar spreads, while listed AI companies have valuations that imply exceptional returns. The groups and financial claims differ, so the comparison raises a question rather than measuring a project's expected return. Paid seats, contracted revenue and company reports of capacity shortages point to demand too. We cannot see what an additional centre will earn over its working life after its full costs. Nor can we infer from these accounts that overbuilding has occurred across the industry.

We cannot see what an additional centre will earn over its working life after its full costs.

Evidence would sharpen the picture in different ways. Fewer new orders and construction starts would show whether the build-out is slowing, not why. Renewal prices and paying use as new centres open would test whether demand supports the capacity. Separate AI revenue and cost disclosures, including depreciation and financing, would bring us closer to measuring returns. Tighter lending terms would show financing pressure; customer concentration would show which specialists depend heavily on a few buyers. For a business buying AI now, the useful questions are when its price can change, who supplies the model and computing behind its service, what capacity is promised, and how long it would take to switch.

Notes

Notes and sources

  1. 1. Microsoft, “FY26 Q4 Investor Metrics,” definition of Microsoft Cloud revenue, microsoft.com; and FY26 Q4 earnings call, on AI and non-AI infrastructure, microsoft.com. Microsoft Cloud includes Microsoft 365 Commercial cloud, Azure and other cloud services, the commercial portion of LinkedIn, and Dynamics 365. These disclosures do not isolate AI-only revenue or a return on an additional data centre. ↑
  2. 2. Phurichai Rungcharoenkitkul, “The AI investment race,” BIS Working Paper 1367 (July 2026), bis.org. The paper models competition for a few dominant positions and estimates investment above a socially efficient benchmark. Its result is model dependent and does not establish whether individual firms will recover their costs. ↑
  3. 3. Iñaki Aldasoro, Sebastian Doerr and Daniel Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin 120 (January 2026), Graphs 1 and 4 and accompanying discussion, bis.org. The mid-2025 spending measures include data-centre equipment estimated at three times construction spending. Graph 4 compares changes in investment as a share of GDP across selected booms; its AI category is defined as data-centre and IT manufacturing construction. ↑
  4. 4. Dale W. Jorgenson, “Capital Theory and Investment Behavior,” American Economic Review 53, no. 2 (1963), pp. 247–259, especially pp. 248–251, aeaweb.org. Jorgenson distinguishes desired capital from the timing of new projects and replacement investment. The application to AI data centres is this report’s interpretation. ↑
  5. 5. Microsoft, FY26 Q4 earnings call, microsoft.com, describes roughly two-thirds of that quarter’s capital expenditure as short-lived assets, primarily CPUs and GPUs. This supports the relevance of economic depreciation but does not establish a universal lifespan for AI chips. The effect of an expected fall in asset prices on user cost follows Jorgenson (1963), cited above. ↑
  6. 6. Avinash K. Dixit and Robert S. Pindyck, Investment Under Uncertainty (Princeton University Press, 1994), ch. 7, “Entry, Exit, Lay-Up, and Scrapping”, jstor.org. The decision to keep an existing asset operating depends on expected future cash flows and avoidable costs, not on recovering expenditure that is already sunk. This does not imply that every existing site should stay open. ↑
  7. 7. Avinash K. Dixit and Robert S. Pindyck, Investment Under Uncertainty (Princeton University Press, 1994), jstor.org. The value of waiting applies when an investment is at least partly irreversible and new information may arrive. The competing incentive to commit early is modelled for AI by Rungcharoenkitkul (2026), cited above; neither source establishes which force currently dominates. ↑
  8. 8. International Energy Agency (IEA), “Overcoming energy constraints is key to delivering on Europe's data centre goals” (November 2025), iea.org, reports EU grid connection waits of two to ten years, depending on country. IEA, “Data centre electricity use surged in 2025” (April 2026), iea.org, reports tightened supply chains for advanced chips and IT components. These are location- and period-specific constraints, not universal project lead times. ↑
  9. 9. Microsoft, “Fiscal Year 2026 Fourth Quarter Earnings Conference Call” (29 July 2026), microsoft.com. The $41 billion capital expenditure figure included $5.6 billion in finance leases; Microsoft separately reported $35.8 billion in cash paid for property, plant and equipment. It described CPU and GPU investment as serving both AI and other infrastructure. Microsoft also described E7 as a suite combining Copilot, E5, Entra and Agent 365; it did not say how many of the 30 million paid Copilot seats came through bundles. ↑
  10. 10. CoreWeave, Form 10-Q for the quarter ended 30 June 2026, sec.gov, condensed consolidated statement of operations and Note 2. The filing reports $103.7 billion in unsatisfied remaining performance obligations, including billed and unbilled amounts, net of estimated variable consideration such as service credits and delivery delays. It expects 41% to be recognised by June 2028, 39% between months 25 and 48, and the rest between months 49 and 78. Customer shares of quarterly revenue were 36%, 26% and 10%; the filing does not name the customers. Its largest customer supplied 71% in Q2 2025, versus 36% for the largest in Q2 2026; customer labels may differ across periods. CoreWeave's Form 10-K for 2025, sec.gov, gives technology equipment a six-year useful life, raised from five years in 2023. The June 2026 10-Q reports no material change to those accounting policies or critical estimates. ↑
  11. 11. Microsoft, “Fiscal Year 2026 Fourth Quarter Earnings Conference Call” (29 July 2026), microsoft.com; CoreWeave, “CoreWeave Continues to Contract New Compute Capacity at Higher Prices” (17 September 2026), coreweave.com. CoreWeave reports pricing of about $40 million per megawatt, measured as annualised revenue divided by power required to serve the clusters, and contracted power of about 4.2 gigawatts at 11 August, up from 3.7 gigawatts at 30 June. The release does not state the comparison behind “higher prices”, and its contracts do not establish a market-wide price trend. Contracted power is not a measure of operating capacity. ↑
  12. 12. Steven M. Fazzari, R. Glenn Hubbard and Bruce C. Petersen, “Financing Constraints and Corporate Investment,” Brookings Papers on Economic Activity 1988, no. 1, brookings.edu; Iñaki Aldasoro, Sebastian Doerr and Daniel Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin 120 (January 2026), Graphs 2 and 3 and accompanying discussion, bis.org. Cash-flow sensitivity alone does not prove that a particular firm cannot raise finance. BIS Graph 2 covers five large US technology firms; Graph 3's private-credit comparison uses broad PitchBook categories for AI, big data and cloud technology, with data through 2025. The equity valuations discussed by the BIS concern listed companies, so the comparison does not match loan and equity pricing for the same borrowers. ↑
  13. 13. Iñaki Aldasoro, Sebastian Doerr and Daniel Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin 120 (January 2026), Key takeaways and concluding discussion, bis.org. The authors say macroeconomic and financial stability risks appear moderate; the bulletin states that its views do not necessarily reflect those of the BIS or its member central banks. ↑
  14. 14. R. F. Kahn, “The Relation of Home Investment to Unemployment,” The Economic Journal 41, no. 162 (1931), pp. 173–198, jstor.org; John Maynard Keynes, The General Theory of Employment, Interest and Money (1936), ch. 10, “The Marginal Propensity to Consume and the Multiplier,” cambridge.org. Kahn developed an employment multiplier; Keynes related changes in investment to income and consumption. Neither supplies a numerical multiplier for AI projects. ↑
  15. 15. Iñaki Aldasoro, Sebastian Doerr and Daniel Rees, “Financing the AI boom: from cash flows to debt,” BIS Bulletin 120 (January 2026), Graphs 1 and 4 and accompanying discussion, bis.org. The roughly three-quarter equipment share is a rule of thumb used by the BIS, not a measured chip share; the BIS notes that some equipment is imported. Graph 4 compares GDP growth in the year after six historical investment booms with average growth during each boom. The association does not isolate the effect of investment from other causes. ↑
  16. 16. Microsoft, “Fiscal Year 2026 Fourth Quarter Earnings Conference Call” (29 July 2026), microsoft.com. Management said it had added usage-based billing to Cowork earlier in July and introduced usage-based billing for GitHub Copilot during the quarter. This does not mean all Microsoft 365 Copilot licences moved away from per-seat pricing. ↑
  17. 17. UK Competition and Markets Authority, “Cloud services market investigation,” final decision (31 July 2025), gov.uk; “CMA announces package of actions on business software and cloud services” (31 March 2026), gov.uk; and “CMA launches strategic market status investigation into Microsoft’s business software ecosystem” (14 May 2026), gov.uk. The 2025 findings concern cloud infrastructure. The separate business software investigation remains open and does not establish the cost of switching a particular AI application. ↑
Disclosure. This article was developed with AI assistance for source discovery, research organisation and drafting. I set the research question, discussed the evidence and directed the interpretation. Material claims were checked against the cited sources during drafting. Responsibility for the final argument and any remaining errors is mine. Nothing in this article is investment advice.
Get started

Choosing an AI provider for your business?

ASA can help you check when your prices can change, who runs the service behind it, and how long a switch would take, before you commit.

Discuss a Project More Insights