For most of the generative artificial intelligence boom, investors have been encouraged to think about AI primarily as a technology capable of making other things cheaper.
Developers can write code faster. Financial institutions can accelerate research, compliance and documentation. Customer-service departments can automate interactions. Manufacturers can optimize supply chains. If artificial intelligence eventually generates the productivity improvements its proponents anticipate, its ultimate economic legacy may indeed be profoundly deflationary. Getting there, however, is proving extraordinarily expensive.
This summer, the economics of artificial intelligence have become an increasingly important part of the AI conversation. The debate is no longer simply about whether models will become smarter or whether OpenAI, Anthropic, Google, Meta or their Chinese competitors will produce the most capable systems. An increasingly important question is how much intelligence costs to produce, distribute and consume—and who will pay for it.
That question has enormous implications for wealth managers because the AI investment cycle is evolving from an equity-market phenomenon into a capital-markets phenomenon. Goldman Sachs Research estimated in August that it was tracking nearly $500 billion of AI-related debt issuance during 2026. Morgan Stanley, using a broader definition, forecast nearly $570 billion. The Bank of England noted in its July Financial Stability Report that issuance from five AI hyperscalers had reached a scale comparable to U.K. government gilt issuance over the same period. AI began this decade looking like another software revolution. Increasingly, it resembles an infrastructure boom.
Intelligence Has a Physical Plant
Artificial intelligence may appear immaterial to someone typing a question into a chatbot, but underneath that interface is an enormous industrial system. AI requires semiconductors, servers, networking equipment, data centers, cooling systems, transmission lines, substations and huge quantities of electricity. Training models requires computing power, while serving those models to hundreds of millions of people and businesses requires still more.
The resulting investment requirements are difficult to overstate. The biggest technology companies spent approximately $380 billion on capital expenditures in 2025, according to an analysis published by the University of Chicago Booth School of Business, and were expected to spend roughly twice that amount this year.
This creates one of the defining paradoxes of AI economics: the cost of an individual unit of artificial intelligence can fall dramatically while total spending on artificial intelligence rises dramatically.
Better chips perform calculations more efficiently. Smaller models can accomplish tasks that once required larger systems. Competition pushes model prices lower. Open-source models create additional pricing pressure. But falling costs stimulate demand. Businesses put AI into more workflows. Millions of workers become persistent users. Software companies embed AI into products. Agentic systems perform lengthy sequences of operations instead of answering individual prompts.
Intelligence per dollar improves while dollars spent on intelligence multiply.
The paradox is particularly important for agentic AI. A chatbot may process one question and produce one answer. An agent can reason through a problem, search databases, invoke software tools, evaluate results, retry failed approaches and coordinate with other agents. Business value can increase enormously—but so can computational consumption.
That expense is beginning to reach ordinary enterprises. An August Harvard Business Review article described an emerging “AI cost shock” as companies move from experimentation toward large-scale deployment. Early AI adoption was partly encouraged by vendors bundling or subsidizing expensive capabilities. Consumption-based pricing can expose a very different economic reality.
For financial services, the distinction between pilot and production matters. A wealth manager might begin by using AI to summarize meetings or draft communications. A mature agentic system might prepare for meetings, retrieve CRM and portfolio information, identify planning opportunities, monitor accounts, initiate workflows and conduct follow-up. Each additional layer of autonomy creates additional computational activity.
AI consequently introduces something resembling a variable production expense into activities historically dominated by fixed software licenses and human labor. Enterprise AI may therefore require as much attention to FinOps as computer science: measuring inference, compute and business outcomes closely enough to determine whether an AI system creates more value than it consumes.
When Cash-Rich Companies Start Borrowing
Those costs eventually lead to the capital markets.
The largest technology companies are not financially distressed. Alphabet, Amazon, Microsoft and Meta remain extraordinarily profitable and possess some of the strongest corporate balance sheets in the world. The significant development is that AI has become expensive enough that borrowing makes sense even for extraordinarily wealthy corporations.
Through 2024, according to the Chicago Booth analysis, the largest technology companies financed most aggregate capital spending through operating cash flow. Hyperscaler corporate bond issuance subsequently increased to approximately $120 billion in 2025, compared with an average of about $28 billion annually from 2020 through 2024. By the first half of 2026, issuance had already exceeded the total for all of 2025.
Debt allows companies to preserve liquidity and match long-term financing with infrastructure expected to generate revenue for years. Goldman Sachs expects roughly one-third of hyperscaler capital expenditures to be debt financed in 2026, increasing to about 35% in 2027.
What is unusual is not the financing technique. It is the scale at which the AI ecosystem is turning to it simultaneously.
Visible corporate borrowing also understates the financial commitments. AI infrastructure is increasingly financed through leases, project finance, construction lending, equipment-backed debt, private credit, infrastructure funds and special-purpose vehicles. A data center does not have to appear as debt on a hyperscaler’s balance sheet for someone to have borrowed money to build it.
The economic exposure consequently spreads among technology companies, banks, insurers, pension funds, alternative asset managers, infrastructure investors and bondholders.
That does not automatically make the boom unsustainable. Similar structures routinely finance airports, telecommunications networks and power infrastructure. But they make AI’s true leverage more difficult to measure and raise an increasingly important question: Who bears the risk if the economic value of the infrastructure turns out to be lower than expected?
From Technology Risk to Capital Risk
During 2023 and 2024, investors mostly asked technological questions. Would generative AI work? Would Nvidia retain its semiconductor advantage? Could Google defend search? Which model developers would win?
By summer 2026, another question had become equally important: What return will all this capital earn?
AI can produce astonishing technological progress and still generate disappointing financial returns if the price of delivering intelligence falls faster than companies can monetize it. Fierce competition makes that possibility more plausible. Better, cheaper models benefit customers while potentially making it harder for providers to earn sufficient margins on enormous infrastructure expenditures.
Cheap intelligence is good for customers. Expensive infrastructure needs profitable intelligence.
The tension becomes more consequential once outside capital enters the equation. Equity can absorb enormous losses without creating contractual defaults. Debt creates scheduled obligations. Interest must be paid, and principal must eventually be repaid or refinanced.
The bond market is already reacting to the extraordinary supply. Reuters reported Aug. 21 that as much as $220 billion of 2026 issuance from major technology companies had been directed toward AI infrastructure under the methodology it cited. Amazon’s recent $25 billion transaction reportedly priced around 120 basis points over Treasuries, approximately twice the comparable spread a year earlier.
Amazon had not suddenly become a questionable borrower. Its debt had become less scarce.
That distinction is fundamental. For the strongest hyperscalers, the immediate concern is less default risk than market risk: spreads can widen and existing bonds can lose value simply because investors are being asked to absorb enormous quantities of similar securities.
Goldman estimated AI-related companies represented approximately 18% of investment-grade issuance during 2026 but roughly 40% of issuance with maturities of 15 years or longer. Amazon had risen to the highest duration weight in an investment-grade index after ranking 20th the previous year, while Google’s position had climbed dramatically.
This matters because investors can acquire more AI exposure without consciously deciding to do so. As technology companies issue more debt, their representation in debt-weighted indexes increases. Passive and benchmark-aware bond funds consequently own more of their securities. A client already exposed to Amazon, Alphabet, Microsoft, Meta and Nvidia through an equity index can therefore accumulate additional exposure to the same AI investment cycle through a corporate-bond fund.
Wall Street Searches for More Capital
When traditional investment-grade investors reach concentration limits, financing does not necessarily stop. It moves. AI borrowers are tapping overseas bond markets, private credit, infrastructure funds and special-purpose financing vehicles. Alphabet, for example, raised A$5.5 billion, or approximately $3.9 billion, in its first Australian-dollar bond transaction on Aug. 19.
Private markets offer an even larger potential pool. Goldman estimates approximately $4.5 trillion of dry powder exists across private credit, infrastructure, real estate and private-equity strategies. One particularly striking example appeared Aug. 20, when Reuters reported that Broadcom was seeking more than $60 billion—and potentially as much as $100 billion—of financing for AI chip and computing projects connected with companies including Anthropic and OpenAI. The contemplated structure reportedly involved a special-purpose vehicle and major alternative-asset managers.
The result is no longer simply “technology company issues bond and builds data center.” Chip suppliers, AI laboratories, hyperscalers, infrastructure investors and private lenders can participate in the same capital structure. That complexity feeds another concern: circularity.
Semiconductor companies sell hardware to AI companies. AI companies need capital to buy that hardware. Technology companies and financial institutions help finance the infrastructure. AI laboratories purchase computing services from hyperscalers. Hyperscalers purchase more chips. Investors value participants partly on revenues generated by one another.
Such arrangements are not inherently irrational. Vendor financing and strategic investments are common throughout industrial history. The critical question is whether economically productive end-user demand ultimately supports the system.
AI has a “last payer” problem. Somebody eventually has to generate enough economic value to pay for the chips, data centers, electricity, interest expense and investment returns accumulated along the chain. A financial institution must generate enough productivity to justify its AI spending. A retailer must increase sales or lower expenses. A pharmaceutical company must accelerate research. A software developer must produce more software with fewer resources.
If those gains materialize broadly, AI infrastructure could eventually resemble electricity grids or telecommunications networks: extraordinarily capital intensive upfront but capable of supporting decades of economic activity. If they do not, the financing chain becomes more vulnerable.
AI Does Not Have to Fail for Investors to Lose
This is perhaps the most important distinction in the AI investment debate. Artificial intelligence can transform the economy and still produce disappointing investment returns. Railroads transformed commerce but repeatedly bankrupted investors. Telecommunications infrastructure became essential to the internet while producing enormous losses after the dot-com boom. Technological necessity and investment profitability are different concepts.
AI could follow a similar pattern. Models could improve dramatically while competition pushes prices lower. AI could become ubiquitous while customers resist paying enough to generate attractive margins. New chips could make existing infrastructure obsolete faster than expected. Enterprises could discover that some AI applications cost more to operate than the productivity they create.
The risk increases farther down the credit spectrum. Lending to a trillion-dollar diversified technology company that happens to build data centers is very different from lending to a single-purpose project or specialized provider whose economics depend almost entirely on continued growth in AI compute.
Technological obsolescence adds another complication. A 30-year bond can finance infrastructure containing equipment that changes every few years. The building may last decades; the chips inside it may not. Fixed-income investors therefore need to distinguish physical longevity from economic longevity.
What AI Debt Means for Portfolios
AI debt is gradually becoming an asset class inside multiple asset classes. There are hyperscaler bonds, data-center loans, infrastructure debt, utility bonds, equipment financing, private credit and potentially securitized products. All depend to varying degrees on the same underlying wager: AI demand will become sufficiently large and profitable to justify today’s infrastructure.
That creates opportunities. Higher issuance can provide attractive yields from strong borrowers, and supply-driven spread widening may allow active managers to buy high-quality credits at better valuations. But it also creates hidden concentration.
A client might own hyperscalers through an S&P 500 fund, lend to them through an investment-grade bond fund, finance their data centers through an infrastructure allocation, own utilities expanding generation for those facilities and participate in construction loans through private credit. Five investment wrappers can represent one underlying economic thesis.
For advisors, the appropriate response is not blanket avoidance of AI-related credit. It is greater awareness of how AI exposure is migrating across portfolios. Advisors should increasingly look through traditional asset-class and sector labels to understand how much client wealth ultimately depends upon continued AI capital spending and successful monetization.
The bond market may ultimately become one of the most useful referees of AI’s financial sustainability. Equity investors possess theoretically unlimited upside and can tolerate extraordinary uncertainty. Bondholders do not. Eventually, lenders ask who will repay the money.
Artificial intelligence has not broken the bond market, nor is there compelling evidence that AI borrowing alone explains elevated Treasury yields or broader fixed-income volatility. Something subtler is occurring. AI is changing who borrows, how much they borrow, how long they borrow for, what bond indexes contain, where private capital flows and which risks connect supposedly different asset classes.
The financial system is adapting to AI. Now AI has to justify the financial system being built around it. For wealth managers, that creates a new portfolio question. Clients have spent the past several years asking how much artificial intelligence they should own.
The more important question may soon be: How much artificial intelligence have they already financed?
Researched by DWN Staff
