The capital required to build artificial-intelligence infrastructure is pushing technology companies deeper into project finance, with guarantees increasingly used to make data centres and accelerator deployments investable without requiring every dollar of development cost to appear as conventional corporate borrowing.

An analysis by the Financial Times estimates that technology companies are using guarantees to support as much as $300 billion of AI-related exposure across data centres and chips, widening a financing approach already visible in regulatory filings from Nvidia, Broadcom, and Meta.

The structures differ substantially and should not be confused with debt already due. Companies are providing varying forms of support if customers default, leases end, or infrastructure is worth less than agreed, allowing project vehicles and outside investors to provide capital against stronger credit protection.

The distinction is important because AI infrastructure increasingly combines hyperscalers, chip suppliers, model developers, data-centre operators, banks, and private capital inside one project, with different organisations carrying different parts of the economic risk.

Nvidia shows how the structure works

Nvidia disclosed guarantees capped at an aggregate $105 billion in connection with infrastructure for SB Energy’s PORTS-Pike development in Ohio, where OpenAI is expected to lease capacity using Nvidia compute.

The support covers an initial 4.25GW of IT load and phases in as facilities enter service. Specified tenant defaults can trigger the guarantees, while exposure falls over the relevant lease periods and can terminate under defined circumstances.

Techopia examined the Nvidia structure when it was disclosed in August. The broader financing pattern now shows that conditional support is appearing across several large AI infrastructure arrangements rather than remaining peculiar to one campus.

Broadcom has disclosed another structure involving AI equipment and customer leases, while Meta has reported residual-value guarantees associated with a data-centre venture.

Residual value becomes part of credit

These mechanisms protect capital providers against scenarios where a customer walks away or the infrastructure is worth less than assumed. That becomes especially relevant to AI because some of the physical estate may last for decades while accelerators, networking, cooling requirements, and rack densities can change much faster.

The financing model therefore looks increasingly like project finance rather than software investment. A technology company can support a facility through leases, purchase commitments, guarantees, or contracts without borrowing the full cost or owning the entire asset.

That flexibility allows pension funds, infrastructure investors, banks, and private-credit providers to carry more of the immediate development cost, provided they have sufficient confidence in both future cash flows and the companies standing behind them.

A guarantee improves that confidence but does not eliminate economic risk. If demand remains strong and tenants perform, the company providing the guarantee may never make a material payment. If projects fail or assets become obsolete more quickly than expected, the contingent promise can become a real obligation.

Hardware makes valuation harder

AI complicates infrastructure valuation because the useful life of the computing technology can be much shorter than the useful life of the building. A powered shell may remain attractive for years, but only if its electrical and cooling design can support future hardware densities.

Financiers therefore have to estimate not simply whether a tenant will pay but what the facility and equipment would be worth if the tenant disappeared. A project highly specialised around one customer or hardware generation may be more difficult to repurpose.

Those questions matter in Europe as well as the United States. European developers are competing for capital to finance data centres, AI factories, and sovereign-compute initiatives while purchasing much of the same accelerator hardware and operating against similar construction constraints.

The cost assigned to project risk will ultimately influence which developments attract capital and what customers pay for the resulting compute.

AI competition extends into capital structure

A model developer can sign a large compute agreement, but someone still has to finance the land, grid connection, cooling, equipment, and construction before the service produces cash. Companies with strong balance sheets can lower that financing cost by standing behind parts of the transaction.

The approach can expand the industry’s effective funding capacity without forcing the largest technology groups to own every facility. It also connects their financial position more directly to projects sitting outside their conventional balance sheets.

The maximum figures consequently require care: contingent guarantees are not equivalent to borrowing the same amount on day one. The economic exposure nevertheless becomes more important if the AI infrastructure cycle weakens.

When utilisation is high, those guarantees may remain largely contractual support in the background. If customers fail or assets lose value faster than expected, they provide a route through which some of the financial risk from the infrastructure boom can travel back towards the technology companies that helped make the projects financeable.