The AI data-center boom has entered bond portfolios. A Bloomberg-syndicated EnergyNow piece puts the buildout above $3 trillion and says AI-related companies and projects tapped debt markets for at least $200 billion last year, likely an undercount because many deals are private. [1]

That is the balance-sheet sequel to Tuesday's paper, which put Anthropic's IPO option beside its compute bill and kept Nvidia's Vera Rubin claim tied to delivery, power and rack facts. Equity gets the story first. Credit often keeps the story longer.

The debt channels are not one market. The EnergyNow article names blue-chip bonds, junk debt, private credit and asset-backed pools. It describes special purpose vehicles built around long-term data-center leases, and the risk that the debt lives outside a hyperscaler's balance sheet while still depending on that customer's lease. [1]

That structure is the important part. A lease can make a project look stable because the tenant is large and the contract is long. It can also move risk into a vehicle whose creditors are betting on the tenant's continued appetite for capacity, the site's ability to deliver power, and the market value of a building designed around a fast-changing computing load. The debt may be outside the hyperscaler's balance sheet, but the economic dependency has not vanished. [1]

The quote that should worry sleepy bondholders comes from JPMorgan credit strategist Tarek Hamid: bond portfolios, historically tied more to rates and bank performance, are becoming correlated with technology companies' performance. [1]

That is a portfolio-construction story, not just a Silicon Valley story. A bond investor who thought she was buying duration, coupon and familiar industrial credit may now be buying exposure to AI demand forecasts, chip cycles, power constraints and cloud-customer concentration. The risk is not that every data-center bond is bad. The risk is that the asset class can inherit equity-style technology volatility while still paying debt-style upside. [1]

ETF Trends gives the June hook. Its June 3 AI market stack says hyperscalers are on track to spend roughly $725 billion this year, mostly on AI, and describes SoftBank, IBM, Snowflake, Robinhood and Nvidia as pieces of a broader capex and inference system. [2]

That capex number explains why credit has arrived. Equity markets can cheer the scale of the buildout, but someone has to finance land, substations, transmission upgrades, backup power, cooling, racks, shells and leases before a model answers a customer query. A $725 billion hyperscaler spending year is not just a demand signal for Nvidia. It is a funding requirement that pulls banks, private lenders and bond buyers into the same AI story. [2]

Anthropic's own Series H release shows why capital is racing the power curve. The company says it raised $65 billion at a $965 billion post-money valuation, crossed $47 billion in run-rate revenue, and signed compute agreements for up to five gigawatts with Amazon and five gigawatts of Google/Broadcom TPU capacity, plus SpaceX Colossus GPU access. [3]

Those figures are astonishing enough to require a sober verb. They do not prove that every data-center loan will perform. They do show why lenders believe the demand story has institutional buyers behind it. When a leading model company talks in gigawatts and names Amazon, Google, Broadcom and SpaceX in its compute stack, the infrastructure appetite is no longer a startup anecdote. It is a utility-scale capital call. [3]

Nvidia's Vera Rubin release adds the factory side: 350-plus factories across 30 countries ramping systems, five-rack POD architecture, photonics networking and fall shipments. [4]

The factory side and the credit side meet at the site boundary. A five-rack POD is not useful without enough power, cooling and network fabric to run it. A data center lease is not safe if the technology inside the building ages faster than the debt amortizes. That is the hidden duration mismatch in AI infrastructure: capital wants long contracts, while the compute layer keeps changing the definition of adequate capacity. [1][4]

The divergence is that stock coverage has a winner's grammar. It says Nvidia, Anthropic, SoftBank, Snowflake. Credit coverage asks who owns the lease, when the debt matures, whether power arrives, whether tenants renew and whether a data center built today can become obsolete before its financing is paid off. [1]

That grammar also changes who can be hurt. A venture investor knowingly buys volatility. A retail saver in a bond fund may not know how much AI-infrastructure exposure sits behind a familiar income label. Pension trustees, insurance portfolios and credit ETFs can end up underwriting the same buildout without participating in the stock-market euphoria that made it famous. [1]

That does not make the buildout fake. It makes it financed. Once the AI trade moves into fixed income, a pension account can own the same risk a venture investor cheers, only with less upside and a longer document.

-- THEO KAPLAN, San Francisco