Expanding compute capacity has been a central driver of AI progress, but sustaining recent growth rates requires exponentially more capital. Frontier labs are planning infrastructure deployments with costs in the tens to hundreds of billions of dollars, raising the question whether financing will constrain future compute scaling.
Anthropic’s recent infrastructure program provides a concrete case study. In November 2025 Anthropic announced plans to invest $50 billion in U.S. compute infrastructure while its annualized revenue was under $9 billion. Subsequent disclosures show that nearly $50 billion of debt financing was arranged for the buildout. Much of that financing was put together in early 2026, before Anthropic’s revenues spiked, making the deals a test of lenders’ willingness to extend capital against projected future payments from AI companies.
Anthropic needed both compute systems and places to run them. The company secured debt to finance leases for more than 1 GW of Google TPU systems and for five datacenters operated by Fluidstack. Roughly $35 billion of debt was raised to purchase the TPU systems (with Broadcom conditionally supporting much of that financing), and about $15.2 billion of loans were issued across five datacenter projects to build 1.43 GW of critical IT capacity (with Google providing conditional support if Fluidstack stops paying rent). In these structures institutional investors provide most of the capital up front, while Broadcom and Google make the future payment streams more dependable.
How vendor‑supported financing works
Banks, insurers, and private credit funds manage large pools of institutional capital seeking relatively predictable returns. A long‑term lease on computing hardware or datacenter capacity creates a predictable series of payments that can be used as the basis for debt, with the underlying racks or completed facilities serving as collateral. This permits outside investors to front the construction or equipment costs and be repaid over time from lease payments.
Investors must still assess whether the lessee will actually make those payments. Anthropic’s revenue grew very rapidly, but rapid growth does not equal a long track record of stable cash flow. Lenders therefore face uncertainty about how a lab would perform through technological change, competition, or regulatory shifts and would normally demand higher returns to compensate.
Broadcom and Google bridge part of that gap by using their credit and commercial interest in deployments to backstop losses. Both suppliers profit if more TPUs or datacenter capacity are deployed, and both have longer operating histories and more diversified cash flows than Anthropic. By agreeing to absorb some losses if Anthropic or Fluidstack stops paying, Broadcom and Google can make the debt cheaper and attractive to a wider investor base without supplying the full cash themselves. This is vendor‑supported financing: suppliers use their balance sheet and contractual commitments to help finance deployments from which they benefit.
Compute financing: an Anthropic lease with Broadcom support
The compute financing provides clear evidence that institutional capital is available at the scale frontier labs require. Led by Apollo‑managed funds, with participation from Blackstone and global banks, investors committed $34.5 billion of debt to finance more than 1 GW of Google TPU systems for Anthropic. Of that amount, $30 billion is supported by Broadcom and $4.5 billion is not. The unsupported portion shows investors were willing to take direct exposure to Anthropic, while Broadcom’s support allowed most of the financing to be raised at lower interest rates.
The structure begins with Anthropic committing to a five‑year lease of the TPU systems. A dedicated equipment company, AI XPV Platform (a special purpose vehicle, or SPV), borrows from investors, purchases the racks, and leases them to Anthropic. Anthropic’s lease payments are used to service the debt, while the racks serve as collateral if payments stop.
The SPV isolates the assets, liabilities, and payment streams for investors, and defines when Broadcom’s support kicks in and how large its maximum exposure is without Broadcom having to buy the systems outright. Funding is released as systems arrive: capital is expected to be deployed in roughly 16 stages over a little more than a year, and about $24 billion is expected to be paid out by summer 2027. This staging keeps funded debt roughly aligned with the available collateral following a default; Broadcom’s potential exposure rises and then falls as Anthropic makes payments.
The debt is split into three tranches with differing seniority. The $6 billion A1 and $24 billion A2 tranches are senior to the $4.5 billion B tranche. A1 pays 1 percentage point above Treasury yields, A2 pays 5.75%, and B pays 8.5%. Broadcom’s backstop protects A1 and A2 investors but not B investors. If Anthropic defaults, Broadcom can take over the lease or arrange sales of the racks; sale proceeds repay investors and Broadcom covers remaining A1 and A2 shortfalls under the agreement, subject to a reported maximum exposure of $29 billion. B investors are repaid only after A1 and A2 and thus rely more heavily on Anthropic’s payments and recovery value from the racks.
Comparing A2 and B highlights the economic effect of the backstop: investors provided $4.5 billion without Broadcom support at an 8.5% yield, versus 5.75% for A2 with Broadcom support. That 2.75 percentage‑point difference reflects the combination of seniority and vendor support and is an upper bound on the value of Broadcom’s backstop to investors.
Together the tranches show institutional investors were prepared to take direct exposure to Anthropic and the TPU racks’ value, while Broadcom’s support helped mobilize a much larger pool of capital on better terms. Anthropic did not need to hold $35 billion in cash: future lease payments attracted the financing, and an interested supplier helped make most of it cheaper.
Datacenter financing: Fluidstack rent with Google support
Investors likewise provided capital to build datacenters under a similar structure. Across five sites, project companies issued about $15.2 billion of debt to construct 1.43 GW of critical IT capacity that Fluidstack will lease for Anthropic’s deployment. This debt was raised before construction began: investors supplied the upfront construction capital backed by the expected future rent once facilities are delivered.
Each site follows the same basic project‑finance pattern. A developer supplies the site, power hookups, permits, and construction capability. A project company borrows from institutional investors, builds and owns the datacenter, and leases the completed capacity to Fluidstack. Fluidstack pays rent as capacity is delivered, which normally services the investors’ interest and principal; the completed facility serves as collateral if payments stop.
Google makes the future rent more dependable. The exact arrangements vary by project, but Google protects investors against some losses if Fluidstack ceases payments. At Lake Mariner, for example, Google can pay missed rent and assume the lease, or fund a termination payment applied to project debt. That role mirrors Broadcom’s in the compute financing: Google places its credit behind a deployment from which it benefits commercially, while institutional investors provide most of the cash. This support increases the capital available and lowers the returns investors demand.
Using several developers lets the buildout proceed faster and leverages scarce inputs (power connections, interconnection work, permits, equipment, and construction pipelines) already controlled by different companies. By working across developers and power markets, Anthropic and Fluidstack can assemble capacity in parallel rather than waiting for a single developer to source and build every site.
Lake Mariner illustrates how the pieces fit together. TeraWulf is the developer there and committed to fund completion of data center buildings. As each building is delivered, Fluidstack begins paying rent under an initial ten‑year lease with two five‑year extension options. That rent is the usual repayment source for investors, while Anthropic pays Fluidstack for the capacity and related deployment and operating services. Google’s lease‑related backstop becomes effective with the corresponding lease, and Google received rights to acquire TeraWulf shares in exchange for providing support. Investors therefore depend on TeraWulf to complete facilities, Fluidstack to pay rent after delivery, Google to provide support if Fluidstack defaults, and the facilities themselves as collateral.
Lake Mariner lacks an otherwise comparable tranche without Google support, so the exact interest‑rate reduction attributable to Google can’t be isolated. The $3.2 billion of debt there nonetheless pays about 7.75%, indicating investors still price material risks—likely construction delays and conditions or limits on Google’s support. TeraWulf’s completion commitment and Google’s backstop address different transaction risks: one helps ensure delivery of facilities, the other improves rent dependability once leases begin.
Implications for frontier compute growth
In the near term, financing is unlikely to be the binding constraint on frontier compute expansion. Nearly $50 billion of debt has been raised for an infrastructure program announced when Anthropic had under $9 billion of annualized revenue. Anthropic did not need to raise the full buildout cost itself, and other AI companies are unlikely to need to either.
The market supporting these transactions is expanding quickly. Anthropic’s annualized revenue rose from roughly $9 billion at the end of 2025 to more than $47 billion by May 2026, and OpenAI had passed $25 billion by February 2026. That growth increases labs’ ability to commit to future compute purchases and gives investors more confidence to finance them. A financing structure assembled around a much smaller Anthropic therefore supports the idea that labs could finance substantially larger buildouts as the frontier labs and market grow.
Participants are already trying to scale the approach into a larger market. Broadcom, Apollo, and Blackstone described the $35 billion compute financing as an initial transaction in a platform intended to support more than 20 GW of frontier‑lab deployments, including Anthropic and OpenAI, through 2028. Early transactions establish contracts, market prices, and an investor base; as systems deploy and payments arrive they create a performance record that should make subsequent deals easier to evaluate. If these financings perform as expected, more institutions may become comfortable with the asset class, enabling larger and cheaper future deployments.
In sum: vendor‑supported debt and project financing converted future demand for compute into tens of billions of dollars of upfront capital. That combination of institutional lenders, developers, long‑term leases, and supplier backstops suggests financing itself is unlikely to be an immediate bottleneck for frontier compute growth.
