PIMCO, Goldman Sachs and Barclays Discuss Who Will Pay for AI Capex

The AI market has reached a stage where the main question is no longer model architecture but balance-sheet structure. Capital expenditure on data centers, chips and power makes up one of the largest investment cycles in decades, and the free cash flow of the biggest tech companies is no longer always enough to cover it. At the Bloomberg Future of Fixed Income 2026 conference in New York, the topic got its own panel. Lotfi Karoui of PIMCO, Amanda Lynam of Goldman Sachs and Brad Rogoff of Barclays discussed with host Romaine Bostick where risks and opportunities will appear in credit markets.

An important caveat: Bloomberg has published only an announcement of the session. The piece contains no transcript and no specific estimates of volumes, spreads or timelines. So this article does not cover what the speakers said, but why the makeup of the panel itself is telling.

Why lenders are at the table

PIMCO is one of the largest bond portfolio managers in the world. Goldman Sachs and Barclays arrange bond issues and lend. These are three different points on the chain through which money reaches a data center construction site: the buyer of the debt, the analyst and the underwriting bank. When players like these discuss AI in public, debt financing is no longer a niche.

The context is familiar from 2025 deals. Oracle issued bonds worth roughly $18 billion. Meta raised about $27 billion through a structure with Blue Owl for its Hyperion campus, and PIMCO was among the large buyers of those securities. These figures come from last year's reporting, not from the panel, but they show the scale: AI infrastructure is already being financed through public and private credit, not only out of profits.

Where the risks hide

Judging by the description, the panel was built around two questions: where the money will come from and who will take on the risk. For a credit investor, these break down into several practical points:

• Issuer concentration. A few hyperscalers and their partners account for a noticeable share of new issuance, and index portfolios risk becoming skewed toward a single theme.
• Asset lifespan. Bonds run for decades, while accelerators become obsolete within a few years. The maturity mismatch is built into the structure.
• Off-balance-sheet structures. Project SPVs and leasing move debt outside the sponsor's reported accounts, which complicates credit analysis.
• Demand dependence. Debt payments rely on expected revenue from AI services, which is so far growing more slowly than costs.
• Power and grid constraints. A delay in grid connection turns a finished facility into an idle asset.

What investors and companies should do

For credit portfolio managers, the practical takeaway is simple: count AI exposure separately from industry categories. The same bond can be classified under technology, real estate and utilities, yet its risk is one and the same. It is worth looking at the tenants' contract base, the length of their commitments and who holds the residual value of the equipment.

For tech companies, the picture is the reverse. Access to cheap debt today belongs to those with a convincing credit rating and cash flow, while for everyone else the cost of borrowing will depend on market sentiment toward AI as a whole. Any cooling in demand will show up quickly in spreads.

The full picture will come from the session recording, once Bloomberg publishes it. For now, one thing is clear: the question "who pays for AI" is now being debated at bond market conferences, and the answers come from the people who buy those bonds.