NextFin News - Citadel Securities says the AI buildout is about to enter a new financing phase. The firm is forecasting more than $500 billion of public and private debt by 2028 to bankroll the chips inside artificial intelligence campuses, a sum it says would amount to more than 5% of the Bloomberg US high-grade index. The message is not that chip demand is fading. It is that the capital structure behind that demand is changing fast enough to matter on its own.
The number is striking because it shifts the frame from spending to funding. The first phase of the AI boom was a race to secure compute. The next phase is a race to secure the balance-sheet capacity needed to pay for that compute, refinance it, and replace it before the hardware and the economics age out. Once chips are financed like project assets, credit markets stop being a side story and become part of the operating model.
What Citadel Securities Is Really Signaling
Citadel Securities published its forecast on August 3, saying technology companies could tap “another $500 billion-plus of debt in the public and private markets by 2028” to bankroll chips used in AI campuses. Jeff Eason, the firm’s head investment-grade desk analyst, said that amount would be “equal to more than 5% of the Bloomberg US high-grade index by 2028.” He expects the borrowing to skew shorter dated, around three to five years, because chip life spans are short, and he said a portion could come as 144A private offerings.
The mechanics matter. Chips depreciate far faster than the factories, offices, and utility projects that usually anchor project finance. They are also tied to a technology curve that can turn a cutting-edge server rack into yesterday’s inventory before the debt is gone. That makes the financing question unusually sensitive to timing. A lender is not just underwriting the borrower’s credit quality. It is underwriting the speed at which the underlying hardware loses relevance.
This is why the forecast is best read as an industrial-credit signal rather than a spending forecast. The likely funding mix - public bonds, private placements, and structured deals - implies that AI capital expenditure is spreading beyond the traditional balance sheets of the largest cloud providers. Some borrowers will still fund out of cash flow. Others will need dedicated debt for chips, data-center equipment, and linked infrastructure. The result is a more segmented market in which the financing terms themselves become part of the competitive advantage.
That shift is already visible in the spending plans of the largest buyers of compute. Amazon said in its second-quarter earnings update that it expects 2026 capital expenditures to be about $200 billion. Meta reported second-quarter results on July 30 and said it expects 2026 capex, including principal payments on finance leases, to be in the range of $130 billion to $145 billion. Alphabet also raised its 2026 spending outlook this summer. The exact numbers differ by company, but the direction is the same: the AI buildout remains enormous, and the market is increasingly asking how much of it will be funded with debt rather than retained earnings.
That is the first-order effect. The second-order effect is more important. If chips become financeable collateral, then AI infrastructure begins to look less like a software cycle and more like a credit cycle with technology features. The obvious beneficiaries are lenders, arrangers, and chip vendors with long order books. The less obvious consequence is that the market may start to value AI capacity not only by demand growth, but by the durability of its refinancing path. In plain terms, the question shifts from “Can these companies buy enough chips?” to “Can they keep rolling the debt that bought them?”
There is also a portfolio effect. More issuance tied to AI infrastructure means more supply for the same broad pool of investment-grade, structured, and private-credit capital. That can pressure spreads in adjacent sectors, especially for borrowers that do not have the same revenue visibility or collateral quality. It can also pull more lenders into equipment-backed structures that look closer to project finance than to classic corporate lending. When that happens, the AI trade ceases to be only an equity-market story. It starts to reorganize parts of the credit market around shorter-duration hardware.
Why This Looks Cyclical Now, but Structural Over Time
The short-term driver is cyclical. The AI spending burst is being led by a handful of hyperscalers and infrastructure developers that are still in an expansion mode. Capital expenditure cycles tend to cluster, overshoot, and then slow when utilization, pricing, or financing conditions tighten. That pattern is already visible in cloud and data-center spending: guidance has moved up, suppliers have benefited, and the market has begun to debate whether demand growth will stay ahead of the asset base.
Three historical comparisons help keep the cycle in view. First, the 2020-21 infrastructure boom showed how quickly abundant liquidity can make every growth narrative financeable. Second, earlier telecom and fiber buildouts showed how demand can be real while the financing structure still runs ahead of cash flow. Third, prior semiconductor upcycles showed that capacity can be added faster than end-demand, leaving lenders and equity holders to absorb the mismatch. The common thread is not technology; it is capital formation outrunning the speed at which cash flow proves itself.
The structural piece is more durable. Once debt markets accept chips as a financeable asset class, the architecture of AI expansion changes. That is not a temporary sentiment shift. It is a regime change in how the buildout is funded. Public investment-grade issuance, 144A placements, and private credit structures each allow different investors to own different parts of the same economic story. The market is building a repeatable financing channel for a rapidly depreciating asset, and that channel is likely to survive even if the pace of borrowing slows.
That distinction matters because investors often confuse the two horizons. A cyclical wave can reverse quickly if demand slows or financing tightens. A structural change in funding architecture usually does not. If the AI economy is now being financed in shorter-dated debt, then the refinancing date becomes a key operating deadline. The new bottleneck is not only power or chips. It is the ability to refinance a hardware stack that ages faster than traditional industrial assets.
“We expect most of that issuance will likely be shorter-dated — around three to five year — to match the life-span of the chips.”
That line is the core of the story. It reveals the transmission mechanism. The market is not merely funding long-lived infrastructure. It is funding hardware with a short useful life and then relying on the next financing window to keep the buildout going. That creates a reflexive loop: more debt buys more compute, more compute requires more capital, and the next round of debt has to arrive before the first round looks old.
The strongest counter-thesis is that this is exactly what healthy demand-led expansion looks like. In that view, the largest AI spenders have strong enough cash flow to service the borrowing, and the debt is simply an efficient way to match long-lived competitive investments with longer-dated funding. Supporters of that argument would say the market is not overfinancing a bubble; it is building the industrial base for a new general-purpose technology.
That counter-thesis deserves respect because the demand side is real. Hyperscalers are still increasing spending, cloud revenue growth remains strong, and chip suppliers continue to benefit from dense order books. But the financing structure still changes the risk profile. If borrowing rises faster than monetization, the debt market begins to price not just growth, but the possibility that growth arrives too slowly for the liabilities attached to it. The warning signal is not whether AI demand exists. It is whether the assets financed by that demand can be refinanced on terms that remain attractive when the chips age.
The falsifying signal is specific. If hyperscaler capex growth slows materially while AI-linked credit spreads remain contained and refinance conditions stay benign, then the debt-binge thesis is overstated. If spreads on shorter-dated AI infrastructure paper widen, maturities shorten further, or lenders demand more collateral for the same assets, then the market is already treating chip financing as a separate risk class.
What It Means for Credit, Chips, and the Rest of the Market
The near-term beneficiaries are the obvious ones: chip vendors, underwriters, lenders, and structured-finance desks. They get more paper to place, more collateral to analyze, and more fees tied to the AI buildout. The medium-term beneficiaries are the firms that can turn debt-funded capex into high utilization and recurring revenue. The long-term exposed players are the borrowers whose revenue curves do not outlast the depreciation cycle on the hardware they financed.
That time-horizon split is the key to reading the story correctly. In the short term, more debt can accelerate the buildout and support supplier revenue. In the medium term, the market will test whether that spending translates into durable earnings power. In the long term, the question becomes whether the financing architecture itself has become a permanent layer of the AI economy or merely a stopgap that will need to be rolled repeatedly.
The base case is that financing remains available because the largest AI spenders still have enough scale and credibility to keep issuing. The upside case is that chip financing becomes a normal, standardized asset class, with repeat issuance, deeper investor participation, and clearer pricing for short-dated hardware risk. The downside case is that monetization lags, spreads widen, and the market starts to separate infrastructure winners from borrowers that need perpetual refinancing to keep up with depreciation.
The next catalysts are straightforward. Investors will watch the next round of capex guidance from hyperscalers, the pace of debt issuance tied to AI infrastructure, and any sign that lenders are asking for shorter tenors or tighter collateral terms. The most important sign to watch is not headline spending. It is whether the market begins to demand a higher return for financing hardware that can become obsolete before the debt does.
Citadel Securities’ forecast does not say the AI boom is ending. It says the boom is becoming a credit story. Once the market starts financing chips on three- to five-year money, the real question is no longer how fast AI can scale. It is how long the balance sheet can keep outrunning depreciation.
Explore more exclusive insights at nextfin.ai.
