Goldman Sachs Hyperscaler Debt Bet Turns Cautious as AI Bond Supply Surges

Goldman Sachs has turned underweight on major AI borrowers despite their strong balance sheets, making hyperscaler debt supply the conflict investors can no longer ignore.

Lindsay Rosner, head of multi-sector investing at Goldman Sachs Asset Management, disclosed the position on September 24. According to her portfolio stance, the firm still believes in artificial intelligence growth. It is resisting the bonds financing that growth because too much issuance is reaching the market.

That distinction changes the AI investment debate. The immediate question is not whether Microsoft, Amazon, Alphabet, Meta, or Oracle can service their obligations. It is whether investors should accept today’s spreads before hundreds of billions in additional bonds compete for the same capital.

Goldman Sachs Research expects gross borrowing from five hyperscalers to rise sharply through 2027. Investors are responding by demanding larger concessions, favoring traditional corporate issuers, and limiting concentrated exposure to AI-related credit.

The market has therefore separated two ideas that once moved together. An investor can remain optimistic about AI demand while becoming less willing to own the debt funding its infrastructure.

Goldman Sachs Hyperscaler Debt Position Targets Supply Risk

Goldman Sachs Asset Management is treating heavy supply as an investment risk even when default risk remains low.

Rosner’s underweight position applies to the largest AI borrowers rather than the entire technology sector. An underweight allocation means the portfolio holds less exposure than its benchmark or normal allocation would imply.

The decision reflects the amount of paper entering the market. Corporate bonds are priced through both credit fundamentals and supply-demand conditions. A financially healthy company can still issue bonds that subsequently underperform if investors receive more comparable debt than they can comfortably absorb.

That mechanism matters because hyperscalers are not financing one isolated acquisition. They are funding a multi-year construction cycle involving data centers, computing hardware, networking equipment, land, and power infrastructure.

Goldman Sachs Research estimated that AI-related borrowers had issued nearly $500 billion of debt by early August 2026. Its own discussion of the AI debt market described the scale and duration of issuance as unusual for corporate credit.

The figure includes more than direct borrowing by the largest cloud platforms. It also covers financing connected to data centers, technology suppliers, utilities, and other parts of the infrastructure chain.

That broader perimeter can make exposure difficult to measure. A bond fund may own a hyperscaler directly, finance a data-center landlord, and hold utility debt supporting the same expansion. These positions can share one economic driver even when their industry labels differ.

Rosner’s stance does not amount to a prediction that hyperscalers will default. The major platforms still generate substantial operating cash flow, hold valuable assets, and possess wide access to capital markets.

Instead, the position reflects a narrower judgment about relative value. If a buyer expects repeated issuance, waiting can produce a better spread or a more favorable new-issue concession.

A new-issue concession is the extra yield offered to place fresh bonds compared with an issuer’s existing debt. Concessions often rise when supply becomes harder to distribute.

That pattern has already appeared in individual transactions. One investment manager reported that Amazon raised $25 billion across eight bond tranches in July. The deal followed a $37 billion transaction in March, bringing Amazon’s twelve-month total to $107 billion.

Demand remained substantial, but the terms became more favorable to buyers. According to the manager’s bond-market analysis, Amazon paid concessions of 12 to 22 basis points across the July tranches.

That exceeded the broader investment-grade average cited in the analysis. A basis point equals one-hundredth of a percentage point, so small changes can materially affect large institutional portfolios.

The change is easy to miss because every deal can still sell successfully. Yet successful placement does not guarantee strong secondary-market performance, especially when another large transaction may arrive within days.

Goldman Sachs hyperscaler debt caution therefore begins with market structure. The borrowers remain credible, but the balance between available bonds and available buyers is shifting.

AI Capital Spending Is Consuming More Corporate Cash

The pressure comes from a financing transition, as AI investment approaches the cash generated by the companies building it.

Hyperscalers initially funded much of the AI buildout from operating cash. Their profitable advertising, software, commerce, and cloud businesses provided an advantage unavailable to smaller infrastructure companies.

That advantage has not disappeared. The problem is that capital spending has grown closer to the scale of operating cash flow, leaving less room for dividends, repurchases, acquisitions, and financial flexibility.

Goldman Sachs analysts estimated that hyperscalers would spend roughly $770 billion on capital expenditures in 2026. That amount was equivalent to about 100 percent of their projected operating cash flow, according to an account of the capital shift.

Capital expenditure includes more than chips. Companies must build structures, install networking systems, secure electricity, cool equipment, and connect facilities to customers across multiple regions.

Not every dollar in those budgets goes to generative AI. Cloud expansion supports conventional workloads, and infrastructure can serve several products over its useful life. Still, AI demand is driving much of the incremental capacity.

When capital expenditure absorbs most operating cash flow, management teams face several choices. They can issue debt, sell equity, reduce buybacks, slow construction, or combine those options.

Debt has obvious attractions for companies with investment-grade ratings. It avoids immediate shareholder dilution and spreads infrastructure costs across the years when the assets should generate revenue.

The drawback is a fixed obligation. Interest must be paid even if demand forecasts prove optimistic, pricing falls, or newer hardware reduces the value of existing facilities.

Debt also moves part of the AI investment decision from corporate boards into public markets. Each offering asks bond investors to validate spending plans through the price they accept.

That transfer changes the information investors need. Equity holders focus on growth, margins, market share, and the future value of AI services. Credit investors also ask how frequently the company will borrow and where its bonds will sit on future supply calendars.

Goldman Sachs expects the five largest hyperscalers to issue roughly $250 billion of bonds in 2026. Its estimates put 2027 issuance near $400 billion, while capital spending is expected to remain elevated.

Those estimates are forecasts rather than contractual commitments. Companies can revise budgets, delay facilities, use leases, or redirect financing between currencies and markets.

However, the direction is already visible. Reuters reported that expected gross hyperscaler issuance for 2027 had risen to $420 billion, a 60 percent increase over the 2026 estimate.

At the same time, total US corporate issuance through August reached $1.9 trillion, up 30 percent from the prior year. The combination means AI borrowers are competing within an already crowded market.

The pressure extends beyond technology issuers. Large volumes of similarly rated debt can pull demand away from industrial, financial, healthcare, and consumer companies.

Investors compare bonds by yield, maturity, liquidity, and expected volatility. A portfolio that absorbs more technology debt must either attract new inflows or reduce another holding.

This is why hyperscaler debt supply matters even to people who never buy a technology bond. The financing requirements can influence borrowing costs across investment-grade credit and affect capital available elsewhere.

AI Growth Confidence Now Conflicts With Bond Discipline

The central reversal is that belief in AI no longer requires enthusiasm for every security used to finance it.

For much of the AI boom, operational success and financing confidence reinforced each other. Rising demand encouraged greater investment, while strong company valuations made that investment appear affordable.

The bond market is now imposing a distinction. A successful technology strategy can coexist with mediocre bond returns if issuance repeatedly cheapens an issuer’s existing securities.

This is the main logic behind Goldman Sachs hyperscaler debt positioning. Rosner can accept the long-term growth case while declining to own as much of the most abundant paper.

That is not inconsistent with other Goldman Sachs views. Different divisions and strategies address different securities, horizons, and client objectives.

For example, Goldman Sachs equity research argued earlier in 2026 that investors should prefer hyperscaler shares over semiconductor stocks. That thesis concerned where AI economics might accrue within the equity market.

A fixed-income portfolio faces another calculation. Its upside is generally limited to contractual payments and spread tightening, while its downside includes spread widening, duration losses, and deteriorating credit metrics.

Duration measures how sensitive a bond’s price is to changes in interest rates. Long-dated technology bonds can lose value when Treasury yields rise, even if the issuer’s financial condition stays stable.

Supply adds another source of pressure. A new bond offered with a higher yield can make an older bond less attractive, forcing its market price lower.

The distinction also explains why default-focused analysis misses the present issue. Investors do not need to believe Microsoft or Alphabet will miss payments to prefer a bank, industrial company, or government bond at a better relative valuation.

Reuters found that highly rated corporate credit had split into two groups by late September. AI-related issuers faced greater caution, while traditional borrowers received stronger demand.

Portfolio managers cited volume and unpredictable issuance schedules more often than solvency. The resulting buyer selectivity was visible through concessions and portfolio concentration decisions.

Predictability matters because bond funds manage cash and benchmark exposure across scheduled maturities. A borrower that communicates a clear annual funding plan gives investors time to reserve capacity.

AI infrastructure plans can change faster. A new model, customer agreement, chip generation, or power commitment can trigger additional spending and another bond sale.

Foreign-currency issuance has provided one release valve. Hyperscalers can reach buyers in euros or other currencies instead of relying entirely on the dollar investment-grade market.

Non-dollar bonds represented 33 percent of hyperscaler issuance in 2026, according to TwentyFour Asset Management. That was up from 14 percent in 2025.

Private credit, infrastructure funds, project finance, and asset-backed structures provide other channels. These arrangements can shift debt away from the parent company and connect repayment more directly to particular facilities or contracts.

They do not eliminate economic risk. A project remains dependent on construction, power availability, tenant commitments, technology demand, and refinancing conditions.

The financing structure determines who carries those risks and when losses appear. Public bondholders, private lenders, landlords, utilities, and shareholders can each absorb a different portion.

That distribution is becoming more important as AI infrastructure leaves the cash-rich core of major technology companies. Data-center developers and specialized cloud providers usually lack the same balance-sheet protection.

For investors, the practical conclusion is selectivity rather than blanket avoidance. A bond may offer attractive compensation after its spread widens, even when an earlier issue from the same company looked expensive.

Rosner’s underweight allocation captures that timing judgment. Goldman Sachs Asset Management appears willing to wait for supply to produce more favorable entry points.

Strong Balance Sheets Do Not Settle the AI Debt Debate

The bearish case can be overstated because the largest borrowers still possess exceptional cash generation and multiple financing options.

Supply pressure is real, but it does not automatically create a credit crisis. Bond markets have continued to finance large transactions, and investors still value liquid securities from recognizable investment-grade issuers.

Goldman Sachs analysts have argued that reasonable increases in leverage would remain compatible with investment-grade ratings for major hyperscalers. Leverage compares debt with earnings, cash flow, or other measures of repayment capacity.

The companies also retain choices that weaker borrowers lack. They can slow repurchases, reduce acquisitions, sell securities in several currencies, or phase construction over longer periods.

Demand for high-quality bonds can expand as yields become more attractive. Pension funds, insurers, foreign institutions, and income-oriented portfolios may step in when concessions rise.

This creates a self-correcting mechanism. Heavy issuance widens spreads, wider spreads attract buyers, and stronger demand helps the next deal clear.

Goldman Sachs research also found that credit markets were already taking on a greater role without an obvious breakdown. Spreads remained historically tight during parts of the issuance surge, despite the scale of borrowing.

BlackRock portfolio manager Amanda Brownback offered a similar interpretation in Reuters reporting. She attributed widening AI spreads mainly to supply-demand conditions, rather than a broad deterioration in credit quality.

That view challenges the most alarmist reading of Goldman Sachs hyperscaler debt caution. An underweight position can express patience and relative-value discipline without forecasting financial distress.

The harder uncertainty concerns returns on investment. Data centers have long operating lives, while chips and AI systems can improve within much shorter cycles.

A facility can remain useful after its original accelerators become dated. Operators can replace hardware, redirect capacity, or use the site for established cloud workloads.

However, retrofits require more capital. Power density, cooling, networking, and software requirements can also change, making some facilities less adaptable than anticipated.

Revenue visibility differs by company. A mature cloud platform can sell capacity across thousands of customers, while a single-purpose project may depend on one tenant or one contract.

Credit investors therefore need to look beyond an “AI-related” label. The underlying borrower, guarantee, collateral, tenant, maturity, and covenant structure matter more than the theme.

Concentration creates another concern. Index investors can accumulate exposure because large issuers receive larger benchmark weights after selling more debt.

That process can reward borrowing with additional passive demand. Active managers may respond by setting internal limits or holding less than the benchmark, which is precisely what an underweight position accomplishes.

There is also a forecasting risk in the supply estimates. The reported $420 billion projection for 2027 assumes companies follow through on current investment plans and use the expected financing mix.

Slower AI demand would reduce the need for capital, but it would introduce a different problem. Investors would then question whether existing spending can earn sufficient returns.

Stronger demand could improve revenues while encouraging even larger construction programs. In that case, better operating results might arrive alongside continued pressure on bond valuations.

Both scenarios weaken simple conclusions. The important question is not whether AI succeeds or fails, but whether revenue, cash flow, and financing capacity expand at compatible rates.

That relationship remains unresolved. Investors should not treat issuance forecasts as proof of a bubble, and they should not treat strong ratings as protection from mark-to-market losses.

Bond Supply Is Becoming a Price on AI Ambition

The cost of debt is emerging as a public scorecard for how confidently markets fund the next stage of AI infrastructure.

Three signals will determine whether Goldman Sachs hyperscaler debt caution proves timely. The first is the pricing of upcoming bond sales.

Investors should watch new-issue concessions, order-book coverage, and secondary trading after each large deal. Those measures reveal whether demand is keeping pace with supply.

A bond that prices with a large concession but rallies afterward suggests the issuer found a clearing level. A deal that weakens immediately can indicate that buyers remain saturated.

The second signal is the relationship between capital spending and operating cash flow in upcoming earnings reports. Investors need to see whether AI services generate cash quickly enough to reduce dependence on external financing.

Revenue growth alone will not settle the issue. Free cash flow, lease commitments, depreciation, and updated capital budgets will show how much financial flexibility remains.

Management guidance will also reveal whether companies can stage projects. Flexible construction schedules reduce financing risk, while rigid purchase or power commitments make spending harder to adjust.

The third signal is Goldman Sachs’ projected 2027 issuance pipeline. The reported jump toward $420 billion will strengthen the supply thesis if borrowers confirm funding plans and continue issuing across currencies.

It will weaken the thesis if spending slows, cash flow rises faster than expected, or alternative financing meaningfully reduces public bond supply. Investors should distinguish genuine risk transfer from cosmetic changes in structure.

Project finance can remove debt from a hyperscaler’s balance sheet while leaving the project economically dependent on that company. Lease obligations can also create recurring commitments without appearing as conventional bond issuance.

Broader market conditions will shape all three signals. Treasury yields affect the total income available from corporate debt, while fund inflows determine how much new supply buyers can absorb.

Goldman Sachs has noted that rising Treasury and AI-related issuance can place upward pressure on long-term yields. Goldman economists nevertheless estimated that the broader crowding-out effect remained smaller than some commentary suggested.

Their analysis put 2026 AI investment near $600 billion, or about 2 percent of US gross domestic product. It also represented about 10 percent of business fixed investment and 15 percent of equipment investment.

Those estimates show why financing has become a macroeconomic issue. The AI buildout now competes for capital, labor, power equipment, and construction capacity beyond the technology industry.

Yet the bond market is offering a more immediate test than any long-range forecast. It assigns a price each time an issuer asks investors to finance another facility.

That price will help determine whether companies continue at the planned pace. Higher borrowing costs can delay marginal projects, favor companies with stronger internal cash generation, and raise barriers for smaller competitors.

Developers and enterprise technology buyers should care because financing conditions affect supply. A more expensive capital base can influence cloud contracts, capacity commitments, and the geographic timing of new infrastructure.

Knowledge workers and AI product users will experience the effects indirectly. Infrastructure costs shape how providers price models, allocate compute, and decide which services deserve continued investment.

The next stage of the AI market will therefore depend on more than benchmark scores or product launches. It will also depend on whether investors believe the promised cash flows justify the financing required today.

Goldman Sachs Asset Management has chosen to wait for better compensation from the largest borrowers. That decision does not reject AI growth. It recognizes that a convincing technology story and an attractive bond are not the same investment.

The question for the coming quarter is concrete: will new hyperscaler deals reward patient buyers, or will resilient demand absorb the supply without lasting repricing? Watch concessions, cash conversion, and confirmed borrowing plans. Together, they will show whether today’s underweight position is an early warning or simply disciplined negotiation with exceptionally strong issuers.