Citadel Securities Says AI Chips Will Drive a $500 Billion Debt Wave

Citadel Securities expects AI chip purchases and data center construction to generate more than $500 billion in additional debt by 2028. A 36Kr item distributed through RSSHub highlighted the forecast on August 3, citing reporting from Jiemian.

The figure matters because the financing would not support conventional assets alone. A growing share would fund processors that age faster than the buildings and power systems surrounding them.

That mismatch creates the central conflict. Borrowers need enormous amounts of capital, but lenders must recover their money before newer chips weaken the value of their collateral.

Citadel Securities reportedly estimates that the new supply would exceed 5% of the Bloomberg US Corporate Bond Index by 2028. It also expects many chip-backed obligations to mature within three to five years.

The forecast has not appeared in a publicly accessible Citadel Securities research note. Its exact definitions and calculations therefore remain difficult to audit independently.

Still, the direction fits observable changes in the market. Hyperscalers once financed most AI infrastructure from their own cash flows. Developers, cloud providers, chip buyers, and special-purpose vehicles now rely increasingly on public bonds and private credit.

That shift moves the AI race beyond technology stocks. It places chip replacement cycles, data center leases, power availability, and model revenue inside the portfolios of bond funds, insurers, and private lenders.

The $500 Billion Forecast Changes the AI Financing Debate

Citadel Securities is describing a shift from corporate spending toward a distinct market for financing AI hardware.

The reported forecast covers debt used for chip procurement and related data center facilities through 2028. It includes financing from public bond markets and less transparent private credit channels.

That distinction is important. Debt raised by a large technology company carries the credit of the entire corporation. Project debt often depends on narrower assets, leases, guarantees, or customer contracts.

The $500 billion estimate also appears to describe additional issuance, not the total cost of the AI buildout. Estimates for overall infrastructure investment are much larger.

A Dallas Fed analysis places expected AI data center investment between $3 trillion and $5 trillion over the next three to five years. It estimates that hyperscalers internally funded roughly $500 billion to $600 billion of investment since 2023.

Hyperscalers are large cloud and technology companies that operate computing infrastructure at global scale. The group includes Amazon, Microsoft, Google, Meta, Oracle, and other major AI buyers.

These companies can still fund significant investments with operating cash. However, their plans increasingly extend beyond what even large balance sheets can absorb without tradeoffs.

They also work with developers, infrastructure funds, and cloud specialists. Those partners usually lack comparable cash reserves, making external financing essential.

Citadel Securities reportedly expects the new debt to equal more than 5% of a major US investment-grade bond benchmark. Even if the final amount differs, the comparison captures the potential scale.

AI financing would no longer sit at the edge of credit markets. It would become large enough to influence portfolio construction, corporate bond supply, borrowing spreads, and interest-rate risk.

That change is already visible. The Dallas Fed cites Wall Street estimates centered on $300 billion of AI-related investment-grade issuance during 2026.

It calculates that such issuance could create as much as $360 billion in ten-year-equivalent duration. Duration measures how strongly a bond’s value responds to changes in interest rates.

That supply would equal about one-eighth of the duration coming from US Treasury issuance during the year. Corporate AI spending could therefore influence markets far beyond technology credit.

The $500 billion forecast is not simply another estimate of capital expenditure. It signals that lenders are becoming a core part of the AI production chain.

Investors are no longer funding only companies with broad revenue streams. They are financing individual campuses, leased capacity, and processors whose economics depend on future AI demand.

This creates a sharper question than whether AI usage will grow. Lenders must decide whether specific contracts and assets can produce enough cash before their obligations mature.

Why AI Chip Debt Needs a Shorter Clock

The useful life of AI processors makes chip financing fundamentally different from financing a data center building.

A data center shell, electrical system, or cooling plant can operate across several hardware generations. The chips installed inside it lose economic value much faster.

Citadel Securities reportedly expects many new obligations to carry maturities of three to five years. That timeline reflects the speed of hardware improvement and replacement.

The Dallas Fed notes that major technology companies depreciate semiconductors over roughly five to six years. Buildings and physical infrastructure commonly carry accounting lives of ten to fifteen years or longer.

Accounting depreciation does not perfectly measure market value. However, it illustrates the mismatch facing lenders.

A processor can remain functional after five years while becoming less competitive. New chips can deliver more computing work per unit of power, floor space, or operating expense.

That efficiency matters because electricity and infrastructure capacity are limited. A customer may prefer a newer cluster even when an older one can still run the same software.

Borrowers financing chips must therefore recover their investment quickly. They need high utilization, dependable customers, and contracts that survive changes in hardware economics.

A shorter loan term reduces the period during which lenders face technological obsolescence. It also creates more frequent refinancing events.

That tradeoff shifts risk rather than eliminating it. A borrower that cannot repay principal from operating cash must return to the market when its loan matures.

Refinancing conditions may be unfavorable at precisely the wrong moment. Demand could weaken, interest rates could rise, or newer hardware could reduce the collateral’s resale value.

The same problem applies to residual value. Lenders often assume that financed equipment retains some recoverable value after a default.

That assumption is harder to defend for specialized AI accelerators. Their resale prospects depend on software support, interconnect compatibility, export rules, and available power infrastructure.

A chip also produces little value by itself. It requires servers, networking, cooling, software, and a site with dependable electricity.

Removing equipment from one failed project does not guarantee that another operator can deploy it economically. Transaction costs can further reduce recovery values.

Short maturities can align debt with hardware life, but they raise the pressure on revenue. Customers must pay enough during a compressed window to cover interest, principal, and operating costs.

That can lead to higher rental rates or stronger minimum commitments. It can also encourage borrowers to use contracts that transfer more risk to AI developers and cloud customers.

The model works best when capacity is preleased to a creditworthy tenant. A long-term commitment turns an uncertain server fleet into something closer to contracted infrastructure.

However, the tenant still needs a viable business. A strong corporate guarantee matters only when the guarantor can support its commitments throughout the loan’s life.

The hardware cycle therefore becomes a financing cycle. Each new generation can trigger fresh borrowing while older loans remain outstanding.

If demand continues rising, that cycle can support rapid expansion. If demand pauses, lenders may discover that several generations of equipment compete for fewer workloads.

Private Credit and Rule 144A Move the Risk Out of View

AI infrastructure is growing through financing channels that offer speed and flexibility but provide less public visibility.

Citadel Securities reportedly expects some of the coming debt to reach investors through Rule 144A offerings. Rule 144A allows qualified institutional buyers to trade certain privately placed securities.

These securities are not registered like ordinary public offerings. Issuers can move quickly while reaching large institutions, including insurers, pension funds, and asset managers.

That structure suits data center projects. Each campus can use customized collateral, lease terms, guarantees, and repayment schedules.

A private placement review found that data center operators added more than $40 billion in Rule 144A placements after November 2025.

The same review described a $27.3 billion bond offering connected to Meta and Blue Owl Capital. The financing supports a two-gigawatt data center campus in Louisiana.

That transaction differs from the shorter chip debt in the Citadel Securities forecast. Its bonds reportedly mature in 2049 because the structure finances a long-lived campus and contractual cash flows.

The contrast shows why “AI debt” is not one asset class. A bond secured by a leased building does not carry the same risk as a loan against processors.

Private structures can combine both. A special-purpose vehicle, or SPV, is a separate legal entity created to own assets and assume project obligations.

An SPV can buy chips, lease them to a customer, and use the lease payments to service debt. The sponsor may keep those obligations outside its primary corporate balance sheet.

That isolation can protect the sponsor from direct leverage. It can also leave creditors dependent on a narrow set of contracts and assets.

The approach has already reached major AI companies. Apollo and Blackstone announced an infrastructure platform backed by an initial $35 billion loan.

According to compute financing details, the platform is intended to help Anthropic lease Google-designed chips through Fluidstack data centers. Broadcom also participates in the structure.

The platform targets more than 20 gigawatts of computing capacity through 2028. Its scale demonstrates how private capital can connect chip vendors, data center operators, AI developers, and lenders.

The arrangement also shows why conventional corporate debt statistics provide an incomplete picture. The hardware can sit in an SPV rather than on the AI company’s balance sheet.

That does not make the economic obligation disappear. The project still depends on lease payments, contractual commitments, and the customer’s ability to generate revenue.

Private credit is attractive because lenders can negotiate protections unavailable in standard bonds. Those protections can include collateral controls, minimum payments, reserve accounts, and restrictions on additional borrowing.

Borrowers gain speed and customization. They also avoid relying entirely on the capacity of public investment-grade markets.

Yet opacity complicates risk assessment. Investors cannot easily determine how many structures rely on the same customers, vendors, or revenue assumptions.

One AI developer might support commitments across several providers. One chip supplier could appear throughout multiple collateral pools.

The projects may look diversified by legal entity while remaining economically concentrated. A slowdown at one major customer could affect numerous lenders.

Private-market valuations also adjust less frequently than public bond prices. That can delay visible recognition of weakening collateral or declining demand.

The result is a financing system that expands capacity quickly but reveals stress slowly. Credit markets may learn about shared exposures only when refinancing or contract disputes force disclosure.

The Real Contest Is Contracted Cash Flow Versus Fast Obsolescence

The central test is whether binding customer revenue can outrun the declining economic value of AI hardware.

Supporters of data center debt point to real assets and long leases. They argue that preleased facilities resemble established infrastructure more than speculative technology ventures.

That argument is strongest for buildings supported by highly rated tenants. Contracted rent can continue even when short-term demand fluctuates.

Ratings agencies have accepted that logic in several transactions. Some securitizations combine multiple facilities, customers, and leases to reduce dependence on one site.

The logic weakens when financing moves from buildings toward chips. Hardware generates cash only when customers actively use or reserve its computing capacity.

A three-year customer commitment can support a three-year loan. A vague expectation of future model demand cannot provide the same protection.

The issue is not whether AI has useful applications. It is whether usage produces revenue at the time and scale assumed by financing documents.

Model developers face significant competition. Falling inference prices can expand adoption while reducing revenue per unit of computing work.

Inference is the process of running a trained model to generate an answer or prediction. It represents a growing share of day-to-day AI computing demand.

Efficiency improvements create another tension. Better software can reduce the amount of hardware needed for each task.

That is positive for customers. It can be negative for projects whose repayment models assume steadily rising consumption of computing capacity.

New processors can also lower the cost of each output. Older clusters may need discounts to remain competitive, reducing their projected cash flow.

Credit quality therefore depends on more than occupancy. Lenders must understand the customer, workload, contract, hardware generation, and replacement plan.

A data center can be fully leased while still carrying hidden concentration. Several leases may ultimately depend on one AI laboratory or one cloud platform.

Guarantees can reduce that risk, but only when they are enforceable and supported by a strong balance sheet. Not every project receives a full corporate guarantee.

Some structures depend on limited guarantees, vendor support, or minimum revenue commitments. The distinctions can determine whether lenders recover principal after a problem.

Wall Street is already paying closer attention. Bloomberg reported that 34% of surveyed global fund managers viewed hyperscaler capital spending as the likeliest source of a future systemic credit event.

That share had reportedly doubled from the previous month. US private credit remained the larger concern, selected by 42% of respondents.

These responses do not establish that an AI credit crisis is approaching. They show that investors increasingly recognize concentration and financing risk.

The strongest deals still have meaningful defenses. They include long contracts, diversified tenants, committed power, established sponsors, and conservative loan-to-value ratios.

Riskier deals rely on optimistic utilization, uncertain power delivery, concentrated customers, or aggressive assumptions about hardware resale values.

The market may price both groups differently as disclosure improves. That process would be healthy if investors can identify the differences before stress arrives.

The Citadel Securities forecast raises a separate supply problem. Even sound borrowers may pay more when hundreds of billions in new bonds compete for investor demand.

Corporate bond buyers have limited risk budgets. More technology issuance can crowd out other companies or require higher yields to attract capital.

The Dallas Fed argues that AI financing can also affect the Treasury yield curve through duration supply and interest-rate swaps. The effects can spread beyond the original borrowers.

This means successful AI projects can still increase financing costs across the economy. Credit pressure does not require widespread defaults.

A flood of issuance can widen spreads, increase term premiums, and change which borrowers receive capital. Smaller issuers would probably feel that pressure first.

Large hyperscalers retain strong cash flows and broad access to capital. Independent cloud providers and developers depend more heavily on receptive credit markets.

They also face greater customer concentration and refinancing exposure. A modest change in lender appetite can therefore alter their expansion plans.

The primary contest is not public debt against private debt. Both markets can finance productive assets, and both can misprice risk.

The real contest is contracted cash flow against obsolescence. Every structure must answer who pays, for how long, and under what conditions.

Three Signals Will Test the Forecast Before 2028

Issuance volume, contract quality, and refinancing performance will determine whether the projected debt wave remains financeable.

The first signal is actual AI-related bond and private credit issuance. Citadel Securities has offered a large forecast, but market data must confirm its pace.

Investors should separate corporate borrowing from project-level financing. They should also distinguish buildings, power equipment, networking systems, and chips.

Those assets carry different useful lives and recovery values. Combining them into one headline figure can hide the market’s underlying risk.

The Dallas Fed’s estimate of $300 billion in 2026 investment-grade issuance provides one benchmark. A sustained pace near that level would support the broader debt-wave thesis.

A sharp slowdown would tell a different story. It could reflect project delays, investor resistance, weaker demand, or renewed use of corporate cash.

The second signal is the quality of customer commitments. Announced capacity does not equal contracted revenue.

Investors should watch how much new capacity is preleased and whether tenants provide full guarantees. Minimum payments and termination clauses deserve similar scrutiny.

They should also track concentration across projects. A customer supporting several SPVs can create a larger exposure than any single financing reveals.

The distinction between committed and expected demand will matter as projects reach completion. Capacity without binding customers will face more refinancing pressure.

The third signal is the performance of early chip-backed loans. Delinquencies are only one measure, and they often appear late.

Utilization, rental pricing, contract renewals, collateral marks, and maturity extensions can reveal stress earlier. Loan amendments may show that initial projections were too aggressive.

Secondary-market spreads offer another useful indicator for tradeable securities. Wider spreads suggest investors want more compensation for credit or liquidity risk.

Refinancing will provide the clearest test. A project that replaces maturing debt without added support has demonstrated continuing lender confidence.

A project that requires sponsor cash, stronger guarantees, or expensive extensions has not necessarily failed. However, its original economics deserve reassessment.

Hardware performance must also be compared with debt schedules. A cluster can operate reliably while losing commercial relevance faster than lenders expected.

The market should therefore focus on revenue per unit of installed capital. Raw computing capacity says little about whether a project can repay its obligations.

One encouraging path remains available. AI usage can expand enough to keep both new and older systems busy at different price points.

New processors may serve demanding training and inference workloads. Older hardware can support smaller models, batch processing, and less time-sensitive applications.

That tiered market would extend useful lives and improve collateral recovery. It would resemble other technology markets where older equipment remains valuable after premium demand moves forward.

The negative path begins when efficiency gains outpace workload growth. Providers would compete by cutting rates while still carrying fixed financing obligations.

Short maturities would then expose weak projects quickly. Private structures might delay public recognition, but they would not prevent the economic loss.

A cautious interpretation is therefore more useful than a bubble verdict. The forecast describes a financing requirement, not a guaranteed issuance total or default outcome.

Its importance lies in the mechanism. AI expansion is converting expectations about future software revenue into debt secured by present-day physical assets.

That conversion can accelerate construction because lenders provide capital before all demand materializes. It also transfers part of the technology cycle into credit portfolios.

Developers and enterprise buyers should watch the consequences closely. Financing conditions can influence which cloud regions open, which chips become available, and how providers structure long-term contracts.

Knowledge workers will feel the effects indirectly through product access and usage limits. Higher infrastructure costs can shape subscriptions, model choices, and service reliability.

Bond investors face the most direct exposure. They need asset-level answers that broad claims about AI adoption cannot provide.

Citadel Securities has put a large number on the approaching market. The next task is to determine how much of that debt finances durable cash flow.

Watch the first refinancing cycles, not only the latest construction announcements. They will show whether AI chip debt is becoming dependable infrastructure credit or a recurring bet on tomorrow’s demand.