AI Infrastructure Structured Finance, the Desk That Keeps Data Center Debt Off the Balance Sheet

Big Tech’s AI debt does not show up on the balance sheet

You cannot find Big Tech’s AI debt by reading its balance sheets. A Nikkei study added up the off-balance-sheet obligations of Alphabet, Microsoft, Amazon, Meta and Oracle and arrived at $1.65 trillion. That is larger than the $1.35 trillion the same five companies report on their books, and roughly eight times what the number was four years ago. Meta alone carries about $420 billion off the books against $140 billion on them. Oracle’s off-balance-sheet obligations have grown around thirtyfold in four years.

The money did not disappear. It moved into other legal entities. In October 2025 Meta formed a joint venture with Blue Owl Capital for the Hyperion data center in Richland Parish, Louisiana. Blue Owl holds 80% of the venture and Meta 20%, while Morgan Stanley arranged $27 billion of debt and $2.5 billion of equity through a special purpose vehicle. PIMCO anchored the debt, which matures in 2049, amortizes fully, carries an A+ rating from S&P, and priced at roughly 225 basis points over Treasuries. Meta is the sole tenant of the campus, yet because the venture bears the risk, the $27 billion never lands on Meta’s own balance sheet.

Everything in that paragraph is the job. Which entity holds the asset, what ownership split keeps it off the consolidated statements, whether the tenant agreement reads as a lease or a service contract, whether the collateral is the shell or the GPUs inside it. The market has scaled with the technique: outstanding data center securitization grew from $4 billion in 2020 to $61 billion by 2026, and 2025 issuance alone ran about $15 billion in ABS plus $11 billion in CMBS. A semiconductor capital markets banker raises equity for the issuer through IPOs and follow-ons. This desk does the opposite, keeping the issuer’s balance sheet as clean as possible and attaching capital to the asset instead. Securing the servers is a procurement engineer’s problem; designing how the money that buys them gets repaid is this one.

The buy side has organized around it. Private credit funds, insurance capital and infrastructure managers now compete for hyperscaler paper, and sell-side desks project $30 billion to $40 billion of gross data center securitization supply per year through 2027. Every one of those deals needs someone who can sit between an accounting policy team, a rating agency and a lender syndicate and hold all three conversations at once.

You design the repayment structure, not the valuation

What this seat asks for is not valuation talent. It is the ability to design a repayment structure that survives a rating committee.

• Consolidation accounting and lease classification. Whether an SPV consolidates into the sponsor, and whether a contract is a lease or a purchase of services, determines the entire point of the deal. Moody’s counted $662 billion of data center lease commitments signed but not yet commenced as of February 2026. Being able to explain when and in what form that number lands on someone’s books is what gets you into the room.
• Residual value on the collateral. Data center bonds run five to twenty years while GPU replacement cycles run far shorter. For xAI’s Colossus 2, an SPV anchored by Valor Equity Partners raised $20 billion to buy Nvidia chips and lease them back to xAI over five years: $7.5 billion of equity including up to $2 billion from Nvidia, with the balance as chip-collateralized debt led by Apollo, Valor and Diameter. When the collateral is silicon rather than a company, your view of what that silicon is worth in year five becomes the loan terms.
• Securitization execution. Tranching for ABS and CMBS, managing rating agency dialogue, negotiating spread, and judging which of the two markets a given asset belongs in. ABS used to take roughly 70% of volume; in 2025 the split ran close to even.
• Power, site and tenant diligence. Interconnection queues, long-term power purchase agreements, tenant credit ratings, remaining contract term. Cash flow on these bonds ultimately comes down to how many years one or two tenants have signed for. Bankers who cannot follow an electrical conversation lose half the diligence.
• Spotting circular flows. Nvidia invests in OpenAI, OpenAI signs cloud contracts with Oracle, Oracle buys Nvidia silicon, and the same dollars appear as an investment, as revenue and as backlog on three different statements. Oracle points to $523 billion of remaining performance obligations. AMD’s agreement giving OpenAI warrants for up to 10% of the company at one cent a share belongs to the same family. Separating outside money from recycled money is where diligence starts.

You come from project finance, not the M&A desk

The entry route is not an M&A or ECM coverage team. Most people arrive from a project finance, securitization or leveraged finance desk. Others cross over from an infrastructure fund’s investment team, a rating agency’s structured finance group, or a finance practice at a firm like Latham or Milbank. What they share is that they learned the lender’s vocabulary before the underwriter’s. If you are starting out, an accounting major and clean modeling will not carry you; lease standards and collateral enforcement mechanics have to be studied on purpose.

For pay, project finance banking is the closest published benchmark. In the US that runs roughly $140K to $180K all-in for analysts, $250K to $350K for associates, $450K to $650K for VPs, $500K to $750K for directors, and $800K to $1.2M for managing directors. Move to the developer side and analyst comp drops to $90K to $150K with materially better hours. Bonuses here do not swing with the deal cycle the way semiconductor ECM does, but pipelines are tied to multi-year construction schedules, so the flow rarely stops cold.

Higher up the ladder the paths fan out: deal teams at credit and infrastructure managers such as Blue Owl, Apollo, KKR and Brookfield, treasury and structuring roles inside hyperscaler finance organizations, or a sector head seat at a rating agency. Whichever chair you take, the same open question follows you around. What is a five-year-old GPU cluster worth as collateral in 2029? Nobody has settled that yet, which is why the assumptions differ from one rating report to the next. Pull a recent data center ABS presale report and read the tranche structure and the residual value assumptions first; that document teaches more about this job in an hour than any interview guide.