Nvidia data center: from GPUs to $525 million of notes

On 3 June 2025, Cologix and Lambda announced the installation of NVIDIA HGX B200-accelerated clusters inside COL4, a data center in Columbus, Ohio. Supermicro systems and Blackwell GPUs would give Midwestern companies access to computing capacity for training, fine-tuning and running artificial-intelligence models.

Forty-eight days later, Cologix announced a second transaction involving the same building. This time, the subject was finance rather than computing performance: COL4 backed $525 million of secured notes, structured with a five-year fixed term.

The sequence offers a rare view of the financial machine Nvidia now wants to operate at a much larger scale. GPUs are installed in a building. Customers pay for computing power or for the infrastructure that makes it possible. The facility’s contracts, rents and property value then support a bond refinancing. Released capital can support more facilities that may later be refinanced in the same way.

This mechanism does not turn Lambda’s GPUs into bonds. The notes are backed by COL4 and a set of rights connected with the facility. No public document reviewed establishes that the chips installed by Lambda belong to the collateral pool. Distinguishing the compute, the building and the contracts is precisely what reveals where the risks sit.

Five key points

A data center normally passes through several financing markets before it reaches bond investors.
Structured refinancing usually comes after the facility has been built, leased and made capable of generating sufficiently predictable revenue.
Note proceeds can repay existing debt and make capital available for subsequent developments.
An anticipated repayment date of about five years can coexist with a 25- to 30-year legal maturity and substantial refinancing risk.
COL4 illustrates the mechanism Nvidia wants to scale, but nothing legally connects its 2025 issuance to the $500 billion programme announced in 2026.

Two markets inside one building

Several economic layers coexist inside COL4.

Cologix supplies the facility, power, cooling, physical security and connectivity. Lambda operates a cloud service specialised in AI. Supermicro provides the systems. Nvidia provides the HGX computing architecture and Blackwell GPUs.

The joint announcement of 3 June 2025 says Lambda customers can use the clusters on demand or through a consumption commitment providing access across generations of compute. Cologix then described four Columbus data centers totalling 500,000 square feet and 80 MW, connected by fibre and to more than 50 network and cloud providers.

Two markets therefore operate in the same building.

The first sells computing power. Its value depends on GPU performance, availability, software, compute pricing and Lambda’s ability to attract users.

The second sells operational infrastructure. Its value depends on occupancy, customer contracts, available electricity, operating expenses, facility reliability and the ability to re-lease capacity.

The Cologix issuance finances the second market. GPUs make the site more useful and may support demand. The published documents do not, however, allow the chips to be treated as collateral for the notes.

That distinction matters. A chip can lose value quickly when a new generation arrives. A sufficiently powered, connected and adaptable data center may continue hosting new equipment. Technology risk has not disappeared. Part of it has shifted to the building’s ability to accommodate future generations.

Building the bond

A data center does not normally move straight from a building permit into a bond fund. The Structured Finance Association describes a succession of funding sources suited to different phases: construction loans, project finance, corporate debt, private credit and specialist equipment finance, followed by ABS or CMBS once the asset is stabilised.

Construction lenders accept delay, cost-overrun, grid-connection and equipment-availability risks. Bond investors normally arrive later. They finance an operating asset with contracts and income that can be documented, modelled and rated.

The boundary remains imperfect. A facility may still be expanding, some capacity may remain vacant and construction work may continue. The logic is to isolate assets mature enough for operating income to support debt service.

The COL4 documentary record

Cologix’s announcement is not the only public document.

The Form ABS-15G filed with the SEC on 24 June 2025 identifies the issuer as Scalelogix ABS US Issuer, LLC, Series 2025-1. The KPMG agreed-upon procedures report names Cologix and Deutsche Bank Securities, identified as the structuring agent. It covers six data center agreements intended for the collateral pool. KPMG compared attributes including tenant name, leased critical load, rent, contract dates and renewal options with documents supplied by Cologix.

The report is neither an audit of those contracts nor an opinion on collateral value or the notes’ capacity to repay. KPMG says so explicitly. The public version also discloses neither the tenants’ identities nor the values tested.

On 26 June, S&P Global Ratings assigned preliminary ratings to three classes totalling $540 million: a $466 million class A-2 rated A-, a $49 million class B rated BBB-, and a $25 million class C rated BB-. The agency described collateral consisting primarily of real-property interests in one data center, personal property and fixtures, tenant leases, reserves, certain accounts and equity interests in the property-owning entities.

Cologix ultimately announced a $525 million closing. The public documents reviewed do not explain the $15 million reduction from the preliminary structure or disclose the final class allocation. Nor do they establish whether Lambda’s contract was among the six agreements examined.

The asset actually being financed

In direct issuance structures described to the market, a dedicated entity owns or controls the data centers. It may own the land and buildings outright or hold leasehold interests. Collateral typically includes real-estate interests, accounts, customer contracts, insurance proceeds and certain rights associated with operating the facilities.

Debt service comes from net operating income after expenses required to run the site. Investors are not simply purchasing a stream of invoices. They are financing an operating business ring-fenced inside a dedicated legal perimeter.

The manager must keep the facility running: find or replace customers, negotiate contracts, maintain infrastructure, manage suppliers, procure electricity, control expenses, invoice and collect. Cash flow depends both on existing contracts and on this continuing operational work.

The Latham & Watkins request submitted to the SEC says these transactions generally have loan-to-value ratios of no more than 70%. Roughly 30% of appraised value would therefore remain as equity in the property-owning entity. The same letter describes an anticipated repayment date of approximately five years and a final maturity of 25 to 30 years. These are general market characteristics, not disclosed parameters for COL4.

Why the transaction releases capital

The word financing can describe two different events.

The first brings new money into a development. The second replaces earlier funding that has become too expensive, too short or too concentrated. Data center securitisation often performs the second function.

Once a facility has been completed and leased, its risk profile improves. Bond investors may accept a lower return than lenders who bore construction risk. New notes can then repay construction debt or another existing facility. The sponsor recovers borrowing capacity and can redeploy some of its capital into another project.

Cologix says the proceeds would support strategic investment across its Scalelogix and Digital Edge portfolios to meet demand from inference AI and cloud workloads. The release does not allocate the proceeds among debt repayment, investment and any distributions.

Switch provides a more explicit example. In April 2026, the operator raised $768 million. Net proceeds were to repay existing debt and fund general corporate purposes. The issuance added a Reno facility of about 1.4 million square feet and more than 52 MW to the collateral pool. After the transaction, the vehicle contained 11 data centers across five markets and served more than 550 customers.

Switch had already raised approximately $4.2 billion through five issuances since 2024. Securitisation thus becomes a financial production line: stabilise, refinance, add an asset, issue another series, then repeat.

Five years hidden inside thirty

The central risk in many transactions sits at the anticipated repayment date, or ARD, rather than the final legal maturity.

A common structure combines:

an ARD around five years after issuance;
little or no scheduled principal amortisation before that date;
a final legal maturity of as much as 25 to 30 years;
a higher cost of debt after the ARD;
a cash sweep allocating excess cash to repayment of the notes.

The expected outcome is a refinancing or repayment at the ARD. The final maturity gives investors a long legal runway if that repayment does not occur. It does not mean the principal is designed to amortise gently over thirty years.

KBRA’s methodology explicitly addresses ARDs, amortisation triggers, the allocation of excess cash and the possible use of property-sale proceeds. The Structured Finance Association explains that these mechanisms encourage refinancing or deleveraging when notes remain outstanding beyond the expected repayment date.

The economic risk is therefore a five-year refinancing wall. If rates rise, a facility loses value, a major customer leaves or investors reject a new issue, the issuer may be forced to carry more expensive debt and devote a larger share of cash flow to repayment.

COL4’s published five-year fixed term is compatible with this pattern, but does not prove it. The available public documents disclose neither its precise ARD, its final legal maturity nor its post-five-year mechanisms. Missing terms cannot be filled with parameters borrowed from another transaction.

Diversify the customer base or secure one giant tenant

The credit quality of a data center note depends heavily on the facility’s commercial model.

A hyperscale data center may be leased for 15 years to one highly rated technology company. Revenue appears predictable, but the departure or renegotiation of that customer can affect almost the entire transaction.

A colocation portfolio instead serves hundreds or thousands of customers. Each contract carries little weight, but the operator must continually renew, replace and expand that customer base.

DataBank 2026-1 illustrates the second model. The pool contained 36 data centers, 1,757 customers and 257.6 MW of critical load. The largest customer represented only 3.9% of annualised monthly recurring revenue. Yet the weighted average remaining contract term was just two years.

Diversification reduces exposure to a single default. Short contracts increase renewal work and expose revenue to market conditions. The two models distribute risk differently.

Bond investors therefore need answers to five questions: who pays, for how long, what share of revenue depends on the largest customers, how much capacity must still be leased, and who bears electricity, maintenance and upgrade costs?

The building is not enough

A data center is often described as real estate enhanced by contracts. That description understates its dependence on energy, technology and an operational ecosystem.

Facility value depends on electricity that is actually available, infrastructure redundancy and cooling, connectivity, investment required for new servers, electricity prices and the ability to re-lease the building if the original customer leaves.

A facility built for a density that later becomes inadequate may require expensive upgrades. A dedicated substation may become underused if demand moves elsewhere. A very long contract can protect revenue while delaying adaptation. Concentration among a few strong customers may reassure rating agencies while creating a large renewal cliff.

Credit analysis must therefore model more than the probability that a tenant pays next month. It must examine defaults, renewals, operating expenses, residual value, the time needed to find a replacement customer and the ability to refinance.

COL4: established facts and limits

The Columbus timeline establishes a material connection between the Nvidia ecosystem and the bond market.

Lambda did deploy NVIDIA HGX B200 clusters inside COL4.
Cologix did close $525 million of notes backed by COL4.
The financing is intended to support infrastructure growth that includes AI demand.

It does not establish that:

Nvidia, Lambda or Supermicro guaranteed the notes;
the GPUs belong to the collateral pool;
Lambda’s revenue alone services the debt;
the $525 million financed acquisition of the clusters;
the issuance belongs to the $500 billion-plus programme Nvidia announced on 10 August 2026.

COL4 demonstrates a mechanism. It is not a transaction executed by the six new platforms.

The shorthand “bonds backed by Nvidia GPUs” would therefore be attractive but unsupported. The public documents tell a different story: chips can create demand for a facility whose contracts, property rights and cash flows become financeable. Bond debt can then release capital for the owner. Nvidia benefits indirectly from a system capable of financing the infrastructure its customers require.

The machine behind the $500 billion

The first article in this investigation showed that Nvidia’s $500 billion announcement described future financing capacity combining equity, debt, co-investors, guarantees and platforms that partly existed already. The cycle described here shows how the same unit of capital can be used repeatedly.

A fund supplies project equity. A bank or private lender finances construction. Once the data center is stabilised, a bond issue can repay that debt. The sponsor can then invest the released capital in a second project. That facility may later follow the same path.

One dollar of initial capital can therefore support several buildings sequentially without becoming several simultaneous dollars. The recycling is economically real. It also makes announcements about “mobilised” capital harder to audit. Depending on the definition, the same capital can be counted when a fund is raised, when a construction loan is made, when the project is refinanced and when the proceeds are redeployed.

The Structured Finance Association, citing Barclays estimates, puts outstanding data center securitisations at $61 billion in 2026, up from $4 billion in 2020. Average data center ABS issuance is estimated at about $600 million. At $525 million, COL4 is close to that scale.

The same report cites a Morgan Stanley estimate that global data center investment through 2028 will require $2.9 trillion, with roughly $1.5 trillion financed outside the cash flows of large hyperscalers. These forecasts are not commitments. They explain why Nvidia, asset managers and banks want to turn stabilised data centers into repeatable financial products.

Public questions for Cologix, Lambda, Nvidia and the arrangers

The public documents available as of 11 August 2026 do not answer the following questions. l0g is putting them publicly to the companies concerned and will incorporate any response supported by verifiable evidence.

To Cologix and Deutsche Bank Securities: what is the final allocation of the $525 million among the note classes, and what happened to the $15 million difference between the structure preliminarily rated by S&P and the amount announced at closing?
To Cologix and Deutsche Bank Securities: which assets and rights make up the final collateral, including the categories of personal property and fixtures, contracts, accounts and equity interests in property-owning entities?
To Cologix and Lambda: was Lambda’s contract among the six agreements examined by KPMG, and what share of rents or leased critical load did it represent?
To Cologix, Lambda, Supermicro and Nvidia: who owns the HGX B200 systems installed in COL4, and who bears their replacement, upgrade and residual-value risk?
To Cologix: what were the appraised value, loan-to-value ratio, DSCR, anticipated repayment date, legal maturity and post-five-year mechanisms at closing?
To Cologix: how were the proceeds allocated among debt repayment, investment in new facilities and any distributions to shareholders?
To Cologix and Deutsche Bank Securities: which contractual protections cover a major tenant departure, insufficient power, technical obsolescence of the facility and a failed refinancing?
To Nvidia: can COL4 or the refinancing of comparable assets be counted toward the $500 billion target, and what method will prevent the same capital from being counted repeatedly as it is redeployed?

The next paradox

The market calls these instruments data center ABS. Cologix and Switch use that term. Rating agencies analyse them with structured-finance tools: payment waterfalls, security packages, reserves and anticipated repayment dates.

Yet on 29 July 2026, SEC staff agreed, for the specific facts presented, that certain direct data center issuances are not asset-backed securities under the Exchange Act. The response is a non-binding staff view, not a general Commission rule.

The contradiction is only apparent. The market is describing a financial technique. US law is applying a particular statutory definition. The third article will examine that boundary, the rules that may cease to apply and the transactions that will test it.

Main sources

Method and limitations

Evidence cut-off: 11 August 2026.
S&P’s $540 million structure was preliminary. Cologix announced $525 million at closing. No public document identified provides the complete final class allocation.
The SEC filing establishes the existence of six agreements intended for collateral, but publishes neither their values nor the tenants’ identities.
No public document identified proves that Lambda’s GPUs belong to the collateral pool or that its revenue alone services the debt.
The $61 billion, $600 million, $2.9 trillion and $1.5 trillion market estimates come from a professional report citing Barclays and Morgan Stanley. They describe a market or projected funding needs, not firm commitments.
This article describes a financing mechanism and its risks. It is neither a complete credit analysis of the notes nor investment advice.