In Budget 2026, the federal government pledged to help Canadian enterprises and AI startups access 500 MW of sovereign AI compute through new financial tools, such as backstops and offtake agreements. If designed well, this could ease one of the hardest demand-side risks in AI infrastructure: finding offtakers. That makes it worthwhile to examine a topic that receives far less attention than power or GPUs: AI data center financing and the role customer commitments play in making projects bankable.
Why Hyperscalers Matter for AI Data Center Financing
Before an AI data centre reaches financing, it needs a customer willing to pay. Investors are most comfortable when that customer is a hyperscaler such as AWS, Microsoft or Google. The advantages are concrete:
- Scale. They commit to hundreds of megawatts. This concentrates risk in one tenant, but a strong tenant lowers overall risk and cuts management overhead, billing, support and credit monitoring.
- Long terms. Leases often run ten years or more, giving predictable cash flows. Lenders size term loans against those cash flows, so an investment-grade tenant lets a developer borrow a large share of project cost at lower rates. In effect, the lease underwrites the debt.
- Tenant credit. Lenders price the project closer to the tenant’s credit than the developer’s. A startup tenant pushes pricing toward venture-style risk, in contrast with an investment-grade hyperscaler. A recent example comes from the Stargate campuses, where Oracle signs long-term leases with developers and OpenAI purchases capacity from Oracle. Concerns about Oracle’s leverage and its dependence on OpenAI slowed syndication and made lenders demand firmer lease terms, and S&P later cut Oracle to BBB-. The episode illustrates a broader reality: financing terms are heavily influenced by tenant creditworthiness and long-term demand visibility.
- Take-or-pay. Large tenants often pay for reserved capacity whether or not they use it, which makes revenue predictable. Small companies usually want flexibility instead.
- Residual value. If a tenant leaves, the operator must re-lease the space. Lenders are comfortable when the site can be re-let to another large tenant, and less so when the building is customized or the tenant base is fragmented. AI campuses optimized for specific customers, architectures, or cooling technologies may have lower residual value than more generic wholesale facilities.
What Happens When the Customer Is a Startup?
Beyond those points, small companies create other problems for developers and operators:
- Scale mismatch. Sites are designed for tenants taking tens or hundreds of megawatts. A customer requiring a few megawatts may not justify the sales effort, contract complexity or a dedicated hall.
- Acquisition cost. Negotiating, diligencing and onboarding a small tenant costs nearly as much as a large one but earns far less.
- Credit evaluation. Startups often have limited operating history, uneven revenue and funding that depends on the next raise. Their position can change faster than a contract term.
- Funding-cycle mismatch. A startup may have 12 to 24 months of runway, while the operator wants a multi-year commitment.
- Collateral. Operators may require letters of credit, guarantees or deposits, which tie up cash many startups don’t have.
- Technical readiness. Dense GPU deployments need specific power density, cooling and networking. Some small teams lack the engineering capacity to specify and commission this, which adds risk for operators.
For a hyperscaler, the lease is the collateral. For a small company, the lease is the risk. That distinction sits at the heart of AI data center financing and explains why operators and lenders place such importance on customer quality and contract duration.
Can Neoclouds Bridge the Gap?
Small companies have several routes to GPU capacity. Hyperscaler clouds offer on-demand, reserved or committed-use instances. This is flexible, but costly per GPU-hour, and the newest GPUs can be scarce. Buying GPUs outright is capital-intensive and requires the expertise to commission and test clusters at scale.
A middle route is leasing from neoclouds and GPU specialists such as CoreWeave, Nebius, and Lambda, which often offer better pricing and more flexible terms. But the model is now under financial scrutiny. As of the date of this article, CoreWeave, one of the largest neoclouds, has seen its shares fall by over 60% from their 52-week high of 138.25 in May to under $50, and its bonds due in roughly five years yield over 15%, about 1,100 basis points above comparable US Treasuries. Fitch rates the company BB-, which is speculative grade. One risk it highlights is lease-term mismatch: long-dated data centre leases of roughly 15 years against customer contracts that generally run 3-5 years. While infrastructure obligations extend well into the future, revenue visibility often does not.
What Problem Is Ottawa Really Trying to Solve?
Canada already has an example of this dynamic. Bell’s 300 MW AI Fabric campus in Sherwood, Saskatchewan, a roughly $1.7 billion project, has its capacity pre-sold to two US neoclouds, CoreWeave and Cerebras. A further 900 MW is planned, contingent on customer commitments. The anchor tenants that make the project financeable are foreign neoclouds. Their credit quality, contract structures, and long-term viability therefore become part of the project’s risk profile. CoreWeave itself serves as an intermediary, aggregating demand from enterprises. If reports are correct that CoreWeave accounts for roughly 140 MW of the initial 300 MW campus, CoreWeave’s credit quality is part of the project’s risk profile.
The Canadian enterprises and AI startups that Budget 2026 seeks to support typically operate at a much smaller scale than the anchor tenants that make large AI campuses financeable today. Helping them access compute is the obvious goal, but there may be another angle. Ottawa may be attempting to create the kind of demand certainty that hyperscalers naturally provide. Government-backed offtake arrangements could allow infrastructure to be financed against aggregated demand from many smaller Canadian AI companies rather than a single large tenant. This may go beyond what the budget explicitly states, but it is a plausible path, and one that would make these projects far easier to finance.
Concluding Thoughts
Programs like the 500 MW pledge are needed to reduce friction in accessing compute, but design will decide their value. Who qualifies? Who hosts these smaller enterprises, and who receives the government guarantees? How will Ottawa vet both applicants and the operators providing sovereign compute? The pledge is a good start, but these questions remain open.
A program like this cannot eliminate demand risk, only reallocate it. The key question is whether the federal government ultimately assumes part of that risk through guarantees and offtake commitments, or merely facilitates transactions between buyers and providers. The answer will determine whether the initiative functions primarily as an access program or as a financing tool.
I have addressed supply side challenges, which include power, in separate posts. See for instance here and here.