Building AI infrastructure takes a lot of money, but the companies buying computing capacity aren't always the ones paying to build it. Behind many of today's biggest AI data centre projects, there's a whole network of developers, investors, lenders and customers, all taking on different financial responsibilities.
On 4 November 2026, TeraWulf announced that Google would provide financial support for certain obligations under its $19 billion lease arrangement with Anthropic. It's a good example of how complicated AI infrastructure financing can get when several companies are involved in getting one project off the ground.
But having the money to build a data centre and actually delivering the computing capacity customers are paying for are two different things. For enterprises planning long-term AI investments, the question isn't just whether a provider has enough funding.
It's also who's responsible for delivering the infrastructure, what they're relying on to make that happen and what happens if something goes wrong.
Who Actually Pays for AI Infrastructure?
The answer depends on which part of the infrastructure we're talking about, because building a data centre, owning it and selling the computing services it provides aren't necessarily the same job. One company might buy the land, build the facility and arrange its power supply. Another might operate the data centre or provide the GPU computing services that customers use.
Sometimes the same company handles several of these jobs, but that's not always the case. Then there are the lenders and investors putting up the money to make everything possible. They expect something in return, whether that's loan repayments, a share of the business or another form of financial return.
Customers can also help fund development by signing long-term contracts or paying for some of their capacity upfront. All these arrangements help spread the cost of building AI infrastructure across different organisations. The catch is that they don't necessarily spread the financial risks in the same way.
Take the November 2026 TeraWulf credit support agreement, for example. TeraWulf's subsidiary, Raylan Data, has a 20-year lease agreement with Anthropic for around 401 MW of critical IT capacity at a data centre campus in Kentucky. TeraWulf is developing the facilities, while Anthropic has agreed to lease the capacity as it becomes available.
Google has now agreed to provide financial support for certain payments Anthropic is responsible for under those leases. So, TeraWulf is building the infrastructure, Anthropic is agreeing to pay to use it, and Google is helping protect against specific financial risks if Anthropic can't meet its obligations.
That doesn't mean Google is paying for the entire project or promising that construction will finish on time. Each company has its own responsibilities, and understanding where one company's obligations end and another's begin is an important part of understanding who's carrying the financial risk.
How AI Infrastructure Financing Shifts Risk
Companies have several ways to raise the money they need for AI infrastructure. They can borrow against equipment, use future customer payments to support loans, bring in investors or arrange guarantees that protect against certain financial problems.
According to S&P Global Ratings' February 2026 analysis, around $237 billion was spent globally on data centre buildings, power and cooling infrastructure in 2025, with another $283 billion expected in 2026. That's a lot of money to find before a project can start earning revenue.
And while these financing arrangements can help companies get projects built, each comes with different obligations and risks.
Customer contracts and secured financing
One way AI infrastructure providers raise money is by using the payments they expect to receive from customers to help secure loans. Think of it this way. A lender may be more willing to finance a new data centre or GPU deployment if the provider already has customers who've agreed to pay for the computing capacity once it's available.
Those contracts give the lender a clearer idea of where the money to repay the loan will come from. In October 2026, Lambda announced a $1 billion secured financing arrangement to help fund GPU infrastructure for three customer deployments. The loan is backed by the computing equipment, related infrastructure and payments expected from those customer contracts.
Lambda can also draw down the funding in stages as the infrastructure reaches certain commissioning milestones, rather than taking the full amount upfront. Another option is getting customers to contribute directly to development costs. For example, IREN's $9.7 billion agreement with Microsoft, announced in November 2025, includes a 20 per cent prepayment.
That gives IREN money towards the infrastructure it's agreed to provide. Both approaches can help providers cover the upfront costs of building AI infrastructure. But they also mean the financing depends, at least partly, on customer commitments and the provider's ability to deliver what it's promised.
Credit backstops and financial guarantees
Sometimes a company needs more than a customer contract to secure financing. It may also need another organisation to provide some financial protection if things don't go according to plan. That's where arrangements like credit backstops come in. A credit backstop is essentially a promise to provide financial support if certain payment obligations aren't met.
However, exactly what it covers depends on the agreement. Under TeraWulf's arrangement, Google's support applies to specific obligations under Anthropic's leases. If Anthropic stops making payments or becomes insolvent, Google has several options under the contract, including taking over a lease or paying a termination fee in certain circumstances.
There are also conditions controlling when Google's support starts and ends. This gives TeraWulf some protection against the risk of Anthropic being unable to pay, which can also make the project easier to finance. What it doesn't do is guarantee that the buildings will be finished on time, the equipment will arrive or the completed infrastructure will perform as expected.
Joint ventures and asset value risk
Another approach is to bring in investors who own part of the infrastructure while another company agrees to use it over the long term. In October 2025, Meta announced its Hyperion data centre joint venture with funds managed by Blue Owl Capital. Blue Owl-managed funds own 80 per cent of the venture, while Meta owns the remaining 20 per cent.
Meta will lease the facilities and has also provided what's known as a residual value guarantee for certain circumstances during the first 16 years of operations. Put simply, this means Meta has agreed to provide some financial protection if the facilities are worth less than a specified amount under certain conditions.
Any payment would be subject to the agreement's limits. It's quite different from Google's promise to support certain Anthropic lease payments. And that's the important distinction across these financing models. Different companies can take responsibility for different risks, even when they're all involved in funding the same infrastructure.
Why Financial Backing Doesn't Guarantee Compute Capacity
Getting the funding together is a big step towards building AI infrastructure. But it doesn't mean the computing capacity is ready for customers to use. A provider can have customer contracts signed, financing arranged and equipment ordered while its data centre is still being built.
Until the facility is finished, the systems are installed and everything has been tested, that capacity might not actually be available. Financing can also come with conditions of its own. Some funding depends on construction milestones being reached, customers continuing to make payments or other parts of the project going according to plan.
In its July 2026 digital infrastructure credit outlook, S&P Global Ratings reported that close to 90 per cent of its new project-finance and corporate data centre ratings that year were speculative-grade. That doesn't mean 90 per cent of data centres are financially unstable.
The figure only covers the transactions S&P rated during that period, and a speculative-grade rating doesn't mean a project is expected to fail. It does show why it's worth looking at the financial arrangements behind a project alongside the provider's plans for actually delivering it.
When capacity commitments depend on future delivery
There's a difference between capacity a provider has agreed to supply and capacity that's already up and running. For example, CoreWeave's second-quarter 2026 results reported around 1.5 GW of active power compared with 3.7 GW of contracted power. Those aren't two ways of describing the same thing.
Active power is already supporting operations, while contracted power includes commitments for infrastructure that may still need to be developed or brought online. The same applies to GPU computing capacity. A provider can have customers lined up for future services without having every facility completed or every system installed.
TeraWulf, for example, expects to deliver its Anthropic capacity in stages between late 2027 and the second quarter of 2028. So, while a long-term contract can give customers confidence that capacity is being planned for them, it doesn't mean they can start using it immediately.
What happens when a financial or delivery obligation fails?
This depends on what's gone wrong and what the contract says should happen. If a customer stops paying, a financial guarantee might help protect the company providing the infrastructure. But if a data centre is delayed, that guarantee may not help the customer waiting for computing capacity.
The Nebius agreement with Microsoft is one example of how contracts can deal with delivery problems separately. Its disclosed terms include provisions covering late delivery, service availability and circumstances where individual capacity commitments can be terminated. These protections can give customers financial remedies if something goes wrong.
But getting compensation isn't necessarily the same as getting the computing capacity they needed. If an enterprise is depending on that infrastructure to launch an AI service or complete a project, a delayed deployment can still cause problems, even when the contract provides some financial protection.
That's why enterprises need to look at both the financial backing behind a provider and its ability to deliver the services it's selling.
What Enterprises Should Check Before Committing to AI Capacity
Enterprises don't need to understand every detail of how their infrastructure providers raise money. But they should have a clear picture of anything that could affect the computing capacity they're planning to rely on.
This is particularly important when they're signing long-term contracts for infrastructure that hasn't been built yet. Before committing, there are three areas worth looking at more closely.
Is the capacity available or still being built?
Start by finding out whether the computing capacity already exists or whether the provider still needs to develop it. If the infrastructure is operational, enterprises can look at its availability, performance and the service commitments the provider is willing to make.
If it's still being built, they need a clearer understanding of the delivery schedule and what's left to be done.
- Has the provider secured the funding it needs?
- Has the equipment been ordered?
- Are there construction or commissioning milestones that still need to be completed?
A signed contract doesn't necessarily mean all those things have happened. It's also worth asking how the provider will report on progress and whether the dates in the contract are firm commitments or estimates that could change.
Who is responsible if something goes wrong?
The next question is who the enterprise is actually relying on to deliver the service. A computing provider might lease its data centre from another company, depend on external investors or use a separate operator to manage part of the infrastructure.
Those arrangements can affect how the service is delivered, but they don't automatically change the customer's rights under its own contract. Enterprises should be clear about what their provider has promised and what happens if it can't deliver.
That means looking at service-level agreements, delivery conditions, compensation, termination rights and any arrangements for keeping services running if there's a problem. They should also check who benefits from any financial guarantees connected to the project.
A guarantee that protects a lender or data centre owner doesn't necessarily give the enterprise buying computing services any extra protection.
How much does delivery depend on other parties?
Finally, it's worth looking at how much the provider depends on other organisations to keep its commitments. Some providers rely on a small number of major customers to support their financing. Others need lenders to keep funding development or depend on outside companies to finish building the facilities they'll use.
None of those arrangements is automatically a problem. In many cases, they're what makes large AI infrastructure projects possible in the first place. But they can create complications if a major customer changes its plans, a funding arrangement falls through or a development partner runs into delays.
Enterprises should try to understand which relationships could affect the capacity they're buying, rather than looking only at the company named on their service agreement. For publicly listed providers, financial disclosures can offer some useful information. Other details may need to be discussed directly with the supplier during procurement.
And when the computing capacity will support an important workload, it's worth considering what alternatives are available if the provider can't deliver as planned. The aim isn't to avoid providers with complicated financing arrangements. It's to understand those arrangements well enough to make an informed decision about whether the service and its protections are suitable.
Final Thoughts: Reliable AI Capacity Depends on More Than Financing
AI infrastructure financing is becoming more complicated as companies find different ways to pay for the data centres and computing systems needed to support growing demand. Long-term contracts, loans, joint ventures and guarantees can all help get these projects built.
But they don't all protect against the same problems, and none can remove every risk involved in delivering computing capacity. For enterprises, knowing that a provider has financial backing is only part of the picture. They also need to understand who's responsible for delivering the infrastructure, what those commitments depend on and what happens if something goes wrong.
As more companies become involved in financing and delivering AI infrastructure, those relationships will become harder to ignore when making long-term technology decisions. And with the way AI infrastructure is changing, keeping up with those commercial and operational developments is becoming part of the job.
That's where EM360Tech's ongoing enterprise infrastructure coverage can help, giving technology leaders a clearer picture of the changes shaping the services their businesses rely on.
Comments ( 0 )