The Financial Times reported on August 5 that a Morgan Stanley-led bank group is preparing roughly $15 billion of loans for Nexus Data Centers' Hubbard, Texas campus and plans to refinance them into bonds, partly possibly via leveraged loans, shortly after drawdown. The roughly 2,000-acre site is backed by Google and leased to Anthropic.
This is not the first financing report on the project. Reuters and Bloomberg reported on July 30 that banks were discussing about $15 billion in lending. The FT's new information is the refinancing path: moving the loans into public bonds so banks exit the balance sheet faster.
The campus includes its own roughly 1.6GW natural-gas plant (Reuters, July 30), making power self-supply part of the financing design and putting fuel and generation risk inside the same capital structure. All reports are media sources; the banks and the project company have not issued official statements.
Read alongside the previous Atlas column on the Broadcom-Apollo-Blackstone platform, AI infrastructure financing is forming two parallel paths: private credit packaging silicon, sites and long-term usage commitments into one structure, and public bonds refinancing a single large campus. Both show that credit layering has begun.
How $15B moves from loans to bonds
The flow typically has two stages: banks lend first to fund construction, then the borrower issues bonds (partly possibly leveraged loans) as soon as possible after drawdown to replace bank credit with public-market funding. Banks transfer credit risk to bond investors while keeping underwriting and arrangement fees.
As soon as possible is the operative phrase. The shorter the time loans sit on bank balance sheets, the less capital and regulatory room they consume, and the sooner banks can underwrite the next deal. For bond investors, it means buying an asset that is still under construction and far from steady operating cash flow.
The bonds are expected to be speculative grade, priced on an uncompleted asset with a single anchor tenant. A speculative-grade price compensates completion probability, lease performance and power availability rather than mature-asset cash flow.
- 01Banks lend
Morgan Stanley leads roughly $15 billion for construction.
- 02Drawdown
Funds draw against construction milestones; Google support is not yet in force.
- 03Bond refinancing
Bonds issue shortly after drawdown; part may go to leveraged loans.
- 04Guarantee activates
Google's support covers the project only after completion.
Who carries construction risk
The FT reports that Google's guarantee only takes effect after completion. Construction delays and cost overruns sit with investors during construction, consistent with the previous Atlas argument that residual value and utilization decide risk: the earlier capital moves to public markets, the more credit depends on completion probability and tenant performance.
The guarantee's timing splits the same debt into two phases: before completion, credit anchors on the developer's execution and lease commitments; after completion, Google support enters the credit structure. Which phase an investor buys determines what risk they carry.
The $15 billion maps to a roughly 2,000-acre campus and about 1.6GW of self-supplied power, larger in scale and power self-sufficiency than typical campus financing. The on-site plant turns fuel prices, emissions permits and generation availability into credit variables: the plant is not just power infrastructure, it is part of debt-service capacity.
No guarantee during construction
Google support starts only after completion.
$15B / ~2,000 acres
Single-campus financing scale.
~1.6GW on-site plant
Fuel and generation availability enter the credit structure.
Anthropic lease
Tenant concentration is a pricing factor.
Scaling $15B down to megawatts and the lease
Against 1.6GW, $15 billion converts to roughly $9.4 million per MW. The figure includes land, buildings, the on-site plant and distribution, making it a campus-level financing density rather than a per-server capital cost; it is useful for magnitude only and should not be compared with per-cabinet costs.
What actually determines debt service is the match between lease revenue and debt: the tenure, payment structure and availability terms of the Anthropic lease are undisclosed, so interest coverage cannot be calculated. Lease disclosure will directly set the credit spread when public bonds price.
Unlike a typical corporate bond, cash flow here depends on a single tenant in a project that must be completed, making this project-finance-style credit: value follows the completion schedule, lease cash flow and on-site plant reliability rather than the issuer's balance sheet.
| Metric | Disclosure | Basis and boundary |
|---|---|---|
| Financing size | ~$15B | FT and Reuters reporting; no official announcement. |
| Campus size | ~2,000 acres | Hubbard, Texas; backed by Google, leased to Anthropic. |
| On-site power | ~1.6GW gas | Reuters July 30 figure. |
| Converted density | ~$9.4M/MW | Atlas conversion; campus-wide, not per-cabinet. |
| Guarantee timing | Effective after completion | Construction-period risk sits with investors. |
What could push the refinancing off course
The first layer is market risk. Speculative-grade bonds need demand, and construction-stage assets have a narrower issuance window when rates or spreads move. If issuance is delayed or shrinks, banks hold the loans longer and the refinancing path loses its point.
The second layer is completion risk. Building a roughly 1.6GW plant and a 2,000-acre campus together means delays and overruns directly erode bond value; a guarantee that only activates after completion leaves the largest timing risk with bond investors.
The third layer is lease and operating risk. Anthropic is the sole anchor tenant; demand changes, availability disputes or early termination would hit single-source cash flow, while gas prices and emissions regulation set the plant's actual generation cost.
Counter-evidence is equally clear: if banks never announce the loan, issuance comes in below $15 billion, or guarantee terms differ from reporting, public-market pricing is weaker than expected; early completion and energization would prove construction risk lower than feared.
Issuance window
Weak demand for speculative bonds delays or shrinks the deal.
Completion risk
Delays and overruns directly erode bond value.
Single anchor tenant
Demand changes and availability disputes hit cash flow.
Gas and emissions
They set the plant's actual generation cost.
The follow-up to $35 billion becoming 1GW
The previous Atlas column examined Broadcom-Apollo-Blackstone packaging silicon, sites and long-term usage commitments into one structure. Nexus shows the other path: refinancing a single large campus with public bonds and transferring risk to capital markets. Together they show that credit layering in AI infrastructure has begun and funding channels are moving beyond private-credit-only.
For banks, loan-to-bond conversion frees capital for the next deal, but failed issuance stays on the balance sheet. For bond investors, the construction-and-ramp spread comes with completion and single-tenant risk. For developers, broader funding may lower the cost of capital without changing power, chip or tenant fundamentals.
For 2027 and beyond, data-center bond supply itself becomes a variable: if this model replicates across campuses, public markets directly participate in AI infrastructure risk pricing, and credit events carry a wider industry impact radius.
Watch for official bank announcements, bond size and ratings, disclosure of Google's guarantee terms, the tenure and payment structure of the Anthropic lease, and project completion and energization dates.
IDC ATLAS VIEWThe $15 billion refinancing plan pushes AI data-center capital structure into public markets: banks lend, bonds take over, and Google's guarantee activates only after completion. Construction risk therefore lands on investors, bonds price at speculative grade, and the roughly $9.4 million per MW conversion is magnitude only; lease disclosure is the real pricing variable. Official confirmation is still pending, but the direction is clear: financing layering and completion probability are becoming the core credit variables of AI infrastructure.
