Anthropic has created Theseus Infrastructure and teamed with Macquarie Asset Management and GIC to fund AI-focused data centres. The deal shoves billions of dollars of private-credit borrowing into a separate vehicle, keeping the debt off Anthropic’s balance sheet and pointing to a financing model that could push AI-service costs higher for developers.
Why the structure matters
Anthropic won’t own the facilities. Theseus will raise equity from its owners, borrow billions from private-credit funds, build the data centres, and then lease the sites back to Anthropic as the anchor tenant. The long-term lease binds Anthropic to regular payments while the borrowing stays on the books of the special-purpose vehicle (SPV), not Anthropic itself.
An SPV is a legally distinct company created solely to own an asset and raise debt against it. Because the SPV holds the debt, Anthropic’s financial statements show no liability, even though the company must make lease payments that reflect the cost of that borrowing.
The context behind the move
Gigawatt-scale compute powers large language models, but the upfront capital needed to build and run such facilities exceeds what most AI labs can comfortably fund. By off-loading financing to an SPV, firms like Anthropic, Meta and Oracle secure the compute they need without inflating their own balance sheets. Meta’s “Hyperion” project and Oracle’s sites for OpenAI follow the same pattern: a third-party vehicle raises capital, constructs the infrastructure, and then leases it back to the AI developer.
What developers stand to lose
- Capacity becomes a financing question – Rate limits and model availability on the Claude API depend on whether the underlying lease agreements stay in force. If the SPV’s debt service strains, the provider may curtail capacity.
- Pricing ties to debt service – Lease payments must cover private-credit interest and principal. If AI revenue falls short, the provider will likely raise API prices or impose tighter usage caps to meet its obligations.
- Vendor risk shifts to the debt market – Large private-credit funds now hold a hidden stake in the AI infrastructure that powers your applications. A downturn in credit markets could ripple through to the services you depend on, even though the risk doesn’t appear on the AI lab’s balance sheet.
What to watch next
- Lease terms and renewal clauses – Length and pricing mechanisms in the leases will dictate how quickly cost changes can be passed on to API users.
- Private-credit market conditions – Rising rates or tighter credit could increase borrowing costs for SPVs, which would flow through to lease payments.
- Industry adoption – If more AI firms adopt the off-balance-sheet approach, the pool of private-credit exposure to AI infrastructure will grow, amplifying systemic risk.
Practical steps for engineers and product teams
- Design for provider agility – Put model calls behind a vendor-agnostic layer so you can switch from Claude to another provider if pricing or capacity turns unfavorable.
- Stress-test cost models – Simulate scenarios where API prices change to see how your margins hold up.
- Track financing news – Treat data-center financing announcements with the same urgency as model releases; they often foreshadow upcoming cost shifts.
The bottom line: the price of AI is no longer just about the code you write. It’s increasingly about who holds the debt that powers the hardware behind the service. Developers and engineering leaders who ignore the financing side risk being caught off-guard when lease-driven cost pressures surface.