Capital markets are now taking a more cautious approach to equity and bond investments in the artificial intelligence sector. Technological progress, productivity gains, and value creation often do not move in lockstep, and this has become a critical period of observation.

Over the past few years, AI has been primarily a valuation story in capital markets. Investors focused on model capabilities, computing scale, and future market potential, driving up valuations for chipmakers, cloud computing firms, and a host of AI companies. But as 2026 gets underway, a shift is increasingly worth noting: AI infrastructure investment is relying more heavily on bond issuance and other credit instruments for funding.

The Bank of England's July Financial Stability Report shows that as of early May, five AI "hyperscalers"—Meta, Alphabet, Amazon, Microsoft, and Oracle—accounted for more than 15% of year-to-date issuance in the U.S. dollar investment-grade corporate bond market. Their bond issuance in the first half of 2026 has already surpassed the full-year total for 2025. Citing data from BNP Paribas, Reuters reported that AI-related bond issuance had reached roughly $220 billion year-to-date by August 10, compared with only $125 billion in the same period last year. Meanwhile, credit spreads on bonds issued by some major tech firms have begun to widen, with investors demanding higher risk compensation.

This should not be simplistically read as a sign that the "AI bubble is about to burst." More notable is that the constraints on AI investment are evolving: capital markets once tested technological imagination and growth potential, but they will increasingly scrutinize returns on invested capital.

From a Valuation Story to Capital Deepening

Understanding this wave of AI investment requires viewing it not just as technological innovation but as a large-scale exercise in capital deepening. Major general-purpose technologies typically need substantial complementary investment before they translate into genuine productivity gains. Electrification required power plants and grids, automobiles required highways, and the internet required communication networks and servers. AI, similarly, depends on chips, data centers, electricity, and network infrastructure.

The "productivity J-curve" in economics captures this phenomenon: after a new technology emerges, firms must first invest capital, reorganize operations, and restructure processes, while productivity benefits may only appear much later. This also sets the current AI wave apart from the earlier internet expansion driven by search, social media, and digital commerce. Generative AI is far more capital-intensive; as training and inference scales grow, demand for chips, servers, and electricity typically rises, even though improvements in algorithm and hardware efficiency offset some of that demand. The result is a disconnect: markets still price AI according to the high-growth logic of the software industry, yet corporate balance sheets increasingly resemble those of infrastructure businesses. Beyond revenue growth, depreciation, capacity utilization, free cash flow, and financing costs are all gaining importance.

The Real Test Is Return on Capital

From an economic standpoint, AI investment ultimately must answer a classic question: can the returns generated by invested capital sustainably exceed the cost of that capital? If AI significantly boosts labor productivity, today's massive capital expenditures could be entirely justified. Companies building data centers and purchasing chips are essentially trading current resources for future output capacity. As long as these investments ultimately translate into higher revenue, lower costs, and sufficient profits, capital deepening becomes part of productivity growth.

The challenge lies in the timing gap: investment happens today, but returns can only be verified in the future. This gap was easy to overlook in a low-interest-rate environment, but with U.S. real interest rates now significantly higher than they were for most of the 2010s, waiting has become more expensive. The higher the cost of capital, the lower the present value of distant cash flows, and the higher the minimum return a project must achieve.

Thus, judging whether AI investment is rational requires more than looking at how fast AI revenue is growing; it also demands examining how much new capital was spent to generate that revenue. If each additional unit of revenue requires ever-expanding server, data center, and electricity investments, revenue growth does not automatically translate into shareholder value.

AI infrastructure also carries a distinctive form of economic depreciation. Buildings, grid connections, and land for data centers can last many years, but the computing equipment at their core may quickly face technological obsolescence. A chip that still functions perfectly can lose economic value the moment newer generations deliver significantly lower unit computing costs. If algorithmic efficiency improves rapidly, future tasks will require less computing power, and the scarcity value of existing equipment may decline. This creates an intriguing paradox: firms are willing to invest huge sums because they believe technology will advance quickly, but the faster the technological progress, the faster some existing capital loses economic value.

When AI Financing Enters the Credit Market

Until now, AI infrastructure investment abroad has largely relied on the strong operating cash flows of the big tech companies themselves. As capital needs expand, the role of corporate bonds, private credit, asset securitization, and various off-balance-sheet financing methods is growing. The Bank of England has already noted that AI financing is extending into the broader credit market, increasing the financial system's exposure to this investment cycle.

Credit markets assess assets differently from equity markets. Stock investors can price in distant futures, but creditors care more about cash flows, debt maturities, and refinancing capacity. This is particularly important when some AI assets tied to long-term debt have shorter economic lives, making maturity mismatches and asset value fluctuations highly relevant. Since 2026, the European Central Bank has also noticed major U.S. tech companies significantly increasing euro-denominated bond issuance. Credit spreads on some hyperscaler euro bonds have widened, showing that investors are demanding higher risk compensation.

Still, it is too early to conclude that AI financing is broadly "crowding out" other corporate funding. More accurately, as tech firms take a larger share of the credit market, this potential impact has entered the watchlist of regulators. This is why comparing tech stock price-to-earnings ratios alone is no longer sufficient to assess whether there is an "AI bubble." What deserves closer attention is the relationship between capital expenditures and revenue, free cash flow, return on invested capital, and credit spreads. If AI investment ultimately keeps returns on capital persistently above the cost of capital, today's investment boom can be understood as the capital formation required for a technological revolution. Conversely, if operating cash flows are chronically insufficient to cover interest, capital replacement, and new investment, markets will eventually shift from debating "whether computing power is scarce" to "whether computing power is in surplus."

From Computing Power Investment to Investment Efficiency

Historical experience reminds us that technological value and investment returns are not the same thing. The expansions of railroads, telecommunications, and the internet all went through phases of concentrated capital inflows followed by corrections. The bankruptcies of many railroad companies in the 19th century did not negate the economic value of rail, and the bursting of the internet bubble did not end the digital economy. A technology can generate enormous social benefits without every investing firm earning commensurate private returns.

For AI, the real question is not simply "is there a bubble" but how large the social productivity gains will ultimately be, and how much of those gains can translate into cash flows that investors can actually capture. This matters equally for China's capital markets. In recent years, models, chips, and computing infrastructure have been central to the country's AI industry buildout. In 2026, alongside continued deployment of hyperscale intelligent computing clusters and computing-power coordination initiatives, there is also a clear push to accelerate commercialized and scaled applications of AI across key industries. This means that as AI moves from technological breakthroughs into industrial application, the evaluation criteria should gradually shift from "how much has been invested" to "how much has been produced."

For companies, computing power is first and foremost a production input, not a measure of final economic output. Building data centers and purchasing servers creates capital, but it does not guarantee productivity or cash flow returns that match the scale of investment. The questions capital markets will ultimately need to answer are not new: how much capital did companies deploy, how much sustainable cash flow did they generate, and can return on invested capital consistently exceed the cost of capital over the long term?

Technological revolutions need imagination to get started, but they cannot be funded by imagination forever. As AI moves from the laboratory onto balance sheets, and from equity valuations into bond pricing, it begins to face the test that all large-scale investment ultimately cannot avoid: the rate of return on capital.