A new playbook for AI capacity, off the balance sheet
On August 10, Anthropic announced Theseus Infrastructure, a strategic partnership with Macquarie Asset Management and GIC, Singapore's sovereign wealth fund, to develop, own, and lease dedicated data center capacity across the United States. The structure is deliberate: funds managed by Macquarie and GIC will own the platform and fund the majority of equity for each project, while Anthropic signs on as the anchor tenant under long-term lease agreements. Anthropic gets guaranteed capacity without the capital intensity of building and owning the facilities itself.
We have watched every major AI lab wrestle with the same math problem this year: training and inference demand is growing faster than any single balance sheet can fund through debt and equity alone. Theseus solves that by importing a financing model straight from infrastructure investing, where pension funds and sovereign wealth vehicles chase long-duration, contracted cash flows. Macquarie Asset Management alone manages roughly $497.6 billion in assets and operates across 30 markets, and GIC has investments across more than 40 countries. This is patient capital, not venture money, underwriting Anthropic's compute roadmap.
Why the ownership structure matters more than the headline
The press language, purpose-built facilities with Anthropic as anchor tenant, sounds like every other cloud capacity deal announced this year. It is not the same thing. When a hyperscaler builds a data center, it owns the asset, controls the roadmap, and can repurpose the facility if a customer churns. Under Theseus, Macquarie and GIC own the real estate and power infrastructure, and Anthropic's exposure is contractual, a lease obligation rather than a depreciating asset. That distinction changes who absorbs the risk if AI demand cools, if power prices spike, or if a site underperforms.
For enterprise buyers, this matters because it previews how AI capacity contracts will be structured industry-wide. Expect more labs and cloud providers to route new capacity through asset-owning joint ventures rather than corporate balance sheets, which means the contractual terms enterprises sign with AI vendors will increasingly sit on top of a second layer of financing agreements the buyer never sees. Understanding that layered structure is now part of vendor due diligence, not an afterthought for the finance team.
The electricity commitment is the real story for CIOs
Buried in the announcement is a detail that deserves more attention than the financing structure: Anthropic has committed to cover electricity price increases passed on to consumers near these data center sites. That is an explicit acknowledgment that gigawatt-scale AI campuses raise local power prices, and that communities and regulators are watching closely enough that AI labs now need to underwrite the political risk directly. It also mirrors commitments hyperscalers have started making in states where data center growth has become a local political flashpoint, suggesting this kind of rate protection is becoming table stakes for any large AI infrastructure deal rather than a one-off concession specific to Anthropic.
We think this commitment previews a coming cost category for every enterprise buying AI capacity: power-adjacency risk. If your AI vendor's data centers strain local grids and draw regulatory scrutiny or litigation, that risk eventually shows up in pricing, in capacity delays, or in service interruptions during grid stress events. CIOs negotiating multi-year AI capacity agreements should start asking vendors directly how they are underwriting power costs and community impact, because the answer increasingly determines whether committed capacity actually shows up on schedule. This is not a compliance checkbox exercise, it directly affects whether the training runs and inference workloads you are budgeting for next year land on time and at the price you modeled.
What this means for build versus buy
Theseus is also a signal about where AI infrastructure ownership is heading generally. A year ago, the build-versus-buy debate for enterprise AI capacity was framed as cloud provider versus on-premises GPU cluster. Today the more relevant question is which layer of ownership sits behind whichever option you choose. Hyperscaler capacity increasingly comes with its own behind-the-meter power deals, and now AI-native labs are adopting asset-light structures backed by infrastructure funds. Neither path gives the enterprise buyer direct visibility into the underlying capital stack.
The practical takeaway is that vendor risk assessment for AI capacity now needs an infrastructure-finance lens alongside a service-level-agreement lens. Ask who owns the data center, who financed the power plant behind it, and what happens to your contracted capacity if that financing structure is stressed. Those questions used to be irrelevant to a software procurement conversation. They are central to it now, and procurement teams that keep treating AI capacity purchases like a standard SaaS renewal will miss the financing risk sitting one layer beneath the contract they are signing.
The competitive pressure this creates
Anthropic's move will not stay unique for long. OpenAI, Microsoft, and Google have all leaned on variations of vendor financing and joint-venture structures to fund AI infrastructure this year, and Theseus gives Anthropic a way to match hyperscaler-scale capacity commitments without hyperscaler-scale balance sheets. Expect competing labs to announce similar infrastructure funds partnerships within the next two to three quarters, likely with different sovereign wealth or pension capital partners chasing the same contracted yield.
For CIOs, the upside of this trend is more competitive capacity supply as multiple labs race to lock in gigawatts through parallel financing vehicles. The downside is that capacity commitments become harder to compare apples-to-apples across vendors, since each one now carries a different, largely opaque financing structure behind it. Building a standard set of questions for AI infrastructure due diligence, ownership, financing partner, power sourcing, and lease terms, is becoming as important as comparing price-per-token across vendors.
