The AI Boom Is Becoming a Balance-Sheet Test
For much of the artificial intelligence boom, capital expenditure was interpreted as a signal of strength.
The companies buying the most advanced chips, constructing the largest data centres and securing the greatest computing capacity were assumed to be building an unassailable lead. Higher spending implied greater ambition. Greater ambition implied future dominance.
That relationship is becoming less automatic.
The sharp semiconductor sell-off on 28 July 2026 suggests that investors are beginning to ask a different set of questions. Asian chip stocks fell heavily as concerns intensified over AI infrastructure financing, elevated valuations and growing competition from China. South Korea’s KOSPI declined 10.8%, while Samsung Electronics and SK Hynix fell 13.4% and 14.7% respectively.
This was more than a routine correction in an overcrowded trade.
It reflected a change in what the market wants AI companies to prove.
The first stage of the AI cycle rewarded access to capital, computing power and technological capability. The next stage will test whether those advantages can produce economic returns before depreciation, financing costs and competition begin eroding them.
The central question is moving from “Who can build the most?” to “Who can earn enough from what has been built?”
AI spending is no longer automatically bullish
Investment in productive capacity can create enormous long-term value. It can also destroy capital when expenditure is driven more by competitive fear than by measurable demand.
The difficulty is that the two can appear identical during the early years of an investment cycle.
A company building additional AI infrastructure may be responding to genuine customer demand. It may also be spending defensively because management fears being left behind. Both companies will report higher capital expenditure. Both may announce new data centres, partnerships and computing capacity.
Only one may eventually earn an acceptable return.
This is why AI capital expenditure can no longer be assessed in isolation. Investors must determine whether spending is:
- supporting contracted or visible demand;
- improving the company’s competitive economics;
- producing recurring revenue;
- creating pricing power;
- lowering operating costs; or
- merely preserving market position in an increasingly expensive race.
For several years, companies were rewarded simply for demonstrating commitment to AI. That period encouraged a form of competitive escalation. Once one hyperscaler increased its investment budget, rivals faced pressure to respond, regardless of whether near-term monetisation had been fully established.
The result is a classic capital-cycle problem.
High expected returns attract investment. Investment expands capacity. Expanded capacity intensifies competition. Competition places pressure on utilisation, pricing and ultimately returns.
AI demand may continue growing rapidly throughout this process. Yet strong demand does not guarantee that every provider of AI infrastructure will earn attractive returns.
Reuters reported that UBS expected hyperscaler capital expenditure to rise 76% in 2026 to approximately US$673 billion, before growth slows to 25% in 2027 and 6% in 2028. The same report noted that 82% of respondents in Bank of America’s July fund manager survey viewed semiconductors as the market’s most crowded trade.
The risk is therefore not necessarily that AI spending collapses.
The subtler risk is that spending continues, but stops accelerating fast enough to support the revenue expectations and valuations built around it.
Revenue growth does not equal economic returns
AI-related revenue can grow while shareholder economics deteriorate.
That may appear contradictory, but revenue is only the beginning of the calculation. Investors must also assess the assets required to generate that revenue and the cost of continually replacing, upgrading and financing those assets.
Five measures are becoming increasingly important.
1. Operating margins
AI revenue must be evaluated after computing, energy, staffing and customer-acquisition costs.
A rapidly growing AI business may still create limited value if intense competition pushes down prices while the cost of serving customers remains high.
The relevant question is not simply whether AI revenue is increasing. It is whether each additional dollar of revenue produces an attractive incremental margin.
2. Depreciation
Data centres, servers and advanced chips are long-lived accounting assets, but their economic usefulness can change quickly.
When newer chips deliver significantly better performance or energy efficiency, older infrastructure may become less competitive before it is fully depreciated. Accounting depreciation may therefore understate the speed of economic obsolescence.
Companies can report healthy earnings while the cash required to maintain competitive infrastructure continues rising.
3. Financing costs
The first phase of the AI buildout was largely supported by the substantial internal cash generation of the world’s largest technology companies.
As infrastructure requirements expand, external financing becomes more important. That shifts part of the AI debate from equity-market optimism to credit-market capacity.
Debt may be entirely appropriate when it finances assets supported by durable cash flow. It becomes more problematic when repayment depends on continuing valuation expansion, optimistic utilisation assumptions or future pricing that has not yet been demonstrated.
Reuters reported that hyperscalers have increasingly turned towards external financing and that weaker-than-expected returns could produce a sudden withdrawal of funding.
The AI boom is therefore becoming a balance-sheet test not only for infrastructure owners, but also for the lenders and capital markets supporting them.
4. Free cash flow
Earnings can rise even as free cash flow falls.
That distinction is becoming central to the AI investment case. Reuters calculated that Microsoft, Alphabet, Amazon, Meta and Oracle could collectively spend more on capital expenditure than they generate in free cash flow by 2027, based on LSEG consensus estimates. Their combined capital expenditure forecasts for 2026 had risen from approximately US$485 billion in January to around US$730 billion by July.
The issue is not that capital expenditure is inherently negative. The issue is whether the future cash flows created by that expenditure justify its scale and timing.
Investors should increasingly focus on:
- free cash flow after AI-related investment;
- cash returns relative to total capital deployed;
- the period required to recover infrastructure costs;
- reliance on debt or equity issuance;
- utilisation rates; and
- the sensitivity of returns to lower AI pricing.
These indicators reveal more about economic value than headline revenue growth alone.
5. Return on invested capital
Return on invested capital, or ROIC, compares the operating profit generated by a business with the capital required to produce it.
For the AI investment cycle, this may become the defining metric.
A company can dominate AI headlines, report strong revenue growth and operate world-class infrastructure, yet still generate disappointing returns if the capital base grows faster than its operating profit.
The most valuable AI companies may not be those with the largest assets. They may be those capable of producing the greatest amount of recurring cash flow from each dollar invested.
The strongest companies may not be the biggest spenders
The popular assumption is that the largest AI budgets will create the largest competitive advantages.
That may be true for a small number of businesses with sufficient scale, distribution and proprietary demand. It will not necessarily be true across the industry.
The strongest companies may instead be those that can monetise AI while allowing someone else to carry much of the infrastructure burden.
These could include software providers embedding AI into existing products, cybersecurity firms securing expanding digital systems, businesses using AI to improve operating margins, or specialised platforms controlling valuable customer relationships and proprietary data.
Their advantage may come from distribution rather than computing capacity.
This distinction resembles earlier technology cycles. Infrastructure builders make the transformation possible, but application-layer companies can sometimes capture more of the long-term economics because they require less capital and own the customer relationship.
The future AI value chain may therefore separate into three broad groups:
Infrastructure owners will carry the greatest capital requirements and must maintain high utilisation.
Technology suppliers will benefit from the buildout but face cyclical demand, competition and the risk of eventual overcapacity.
AI adopters and application businesses may generate smaller headline revenues initially, but could achieve superior returns if AI improves productivity without materially increasing their capital intensity.
The market has spent several years identifying who benefits from more AI spending.
The next opportunity may lie in identifying who benefits when AI becomes cheaper, more widely available and less differentiated.
AI may support economic growth while disappointing shareholders
A technology can transform an economy without rewarding every company financing that transformation.
Railways expanded commerce but produced uneven returns for railway investors. Telecommunications infrastructure improved connectivity while periods of excessive network investment destroyed capital. The internet reshaped almost every industry, even though many companies funding the original buildout failed to earn acceptable returns.
AI may follow the same pattern.
Its economic benefits could be broad and enduring. AI investment is already supporting manufacturing, construction, semiconductor exports, energy demand and technology employment across several economies.
A Reuters poll of nearly 500 economists found that AI-related investment was helping to maintain global growth forecasts despite persistent inflation and weaker expectations across many individual economies. Median global growth forecasts remained at 2.9% for 2026 and 3.1% for 2027, while South Korea and Taiwan received significant upgrades partly because of AI-related activity.
The IMF has similarly observed that demand for AI and other technologies has helped offset weakness elsewhere in the global economy, while also identifying a correction in AI market expectations as a downside risk.
This creates an important distinction between economic productivity and investment profitability.
AI may raise output, improve efficiency and create entirely new services. Customers, workers and economies may capture much of that value through lower prices, better products and higher productivity.
The companies paying for the infrastructure will earn attractive returns only if they can retain a sufficient portion of those benefits.
That outcome depends on market structure.
When a small number of companies control scarce computing resources, they may enjoy strong pricing power. When models become more efficient, chips become more available and competition increases, more of the economic value may pass to users instead.
The technology can therefore succeed even as the economics of providing it become less attractive.
AI is becoming a credit question as well as an equity question
The financing structure behind the AI boom deserves far more attention.
Equity investors tend to focus on growth, market share and long-term addressable markets. Credit investors focus on contractual cash flows, debt service, asset values and downside protection.
As AI infrastructure becomes more capital-intensive, the credit perspective becomes increasingly relevant.
Lenders and investors will need to examine:
- who ultimately pays for the computing capacity;
- whether customer commitments are binding;
- how concentrated the customer base is;
- whether contracts extend beyond the financing period;
- who bears electricity and construction cost overruns;
- how quickly equipment may become obsolete;
- whether assets have meaningful resale value; and
- whether refinancing is required before projects become cash-generative.
A data centre may appear to be a tangible infrastructure asset. Its economic value, however, depends heavily on power access, connectivity, equipment configuration, customer demand and the pace of technological change.
Not every AI infrastructure project should receive infrastructure-like financing terms.
Some projects may have stable, contracted cash flows and strong sponsors. Others may contain far more technology, utilisation and refinancing risk than their physical appearance suggests.
The dividing line will become increasingly important as the buildout expands beyond the balance sheets of the largest technology companies.
What investors should watch next
The next phase of the AI cycle will be determined less by announcements and more by financial conversion.
Investors should watch whether AI-related revenue begins growing faster than the capital required to support it. They should also monitor whether depreciation assumptions remain realistic, whether financing costs rise, and whether companies can sustain shareholder distributions without increasing leverage.
Three signals will be particularly important.
First, capital expenditure growth must eventually moderate without causing revenue expectations to collapse.
Second, AI adoption must translate into measurable margin improvement for customers, not merely higher technology expenditure.
Third, infrastructure owners must demonstrate that utilisation, pricing and contractual demand can support acceptable returns across the full life of the assets.
Companies that pass these tests may emerge with durable competitive advantages.
Those that fail may discover that technological importance is not the same as economic profitability.
Frequently asked questions
Is the AI boom ending?
Not necessarily. AI investment and adoption may continue expanding for years. The change is that investors are becoming more selective about which companies can convert that growth into sustainable free cash flow and attractive returns on capital.
Why is AI spending becoming a balance-sheet issue?
AI infrastructure requires substantial upfront investment in chips, data centres, networking equipment and energy capacity. As spending rises, companies must rely more heavily on operating cash flow, debt or external capital, increasing the importance of leverage, depreciation and financing costs.
Which financial metrics matter most for AI companies?
Investors should examine operating margins, capital expenditure, depreciation, free cash flow, leverage, utilisation rates and return on invested capital. Revenue growth alone does not show whether AI investment is creating economic value.
Could AI transform the economy but disappoint investors?
Yes. AI can improve productivity and economic growth while competition, falling prices and excessive infrastructure investment reduce the returns earned by individual companies.
Who could benefit most from the next phase of AI adoption?
Potential beneficiaries may include companies that use AI to increase productivity, strengthen existing products or improve margins without carrying the full cost of building and maintaining large-scale computing infrastructure.
The AI boom is not becoming less important.
It is becoming more financially demanding.
Technological leadership will continue to matter, but it will no longer be sufficient. The winners will need to combine innovation with capital discipline, durable monetisation and balance sheets capable of supporting the investment required.
The next stage of the AI cycle will reveal which companies have built productive assets and which have simply accumulated expensive capacity.
The question is no longer whether AI will transform the economy. It is who will earn an acceptable return from paying for that transformation.
Tigris Asset Management Pte Ltd holds Capital Markets Services Licence CMS101520 issued by the Monetary Authority of Singapore. This material is for informational purposes only and does not constitute an offer or solicitation to buy or sell any investment product. Past performance is not indicative of future results. Investment products are available only to accredited and institutional investors as defined under the Securities and Futures Act 2001 of Singapore.
