Broadcom, Apollo, and Blackstone have created a $35 billion financing structure for Anthropic’s next gigawatt of AI compute, which is a normal sentence only if the last two years have fully broken the reader’s sense of scale.
Broadcom announced Monday that it has formed the AI XPV Platform with Apollo and Blackstone’s credit and insurance business as anchor investors. The platform is designed to finance more than 20 gigawatts of compute capacity through 2028, using Broadcom’s custom XPUs and networking systems for frontier AI labs including Anthropic and OpenAI. The first tranche is a $35 billion transaction led by Apollo, with Blackstone and global banks participating, to support more than 1 gigawatt of Anthropic compute infrastructure expected to deploy at Fluidstack-based sites starting in mid-2026.
Broadcom has become one of the central suppliers of custom AI silicon, the non-Nvidia path for labs that want lower per-token costs and more control over their accelerator stack. If Nvidia sells the default GPU, Broadcom sells the custom answer for customers large enough to have a default of their own.
Anthropic, meanwhile, now buys compute from Amazon, Google, CoreWeave, SpaceX, Broadcom, Fluidstack, and whoever else can plausibly turn power and silicon into usable tokens before the next model cycle. It is no longer enough for a frontier lab to have capital. It needs delivery schedules, grid interconnects, substations, networking, packaging, and financing structures that can survive a buildout measured in gigawatts.
The finance layer is where this gets interesting. Apollo described the transaction as committed capital across a multi-year draw schedule. It also described AI compute as an emerging asset class, defined by contracted cash flows, mission-critical utility, and worsening supply-demand pressure. That is private credit language applied to model training and inference.
The AI buildout has been discussed mostly as a technology bottleneck: chips, packaging, power, cooling, and data center shells. Those bottlenecks are real. But the Broadcom-Apollo-Blackstone structure says the financial bottleneck is being redesigned too. Instead of a lab or hyperscaler simply writing checks for hardware and facilities, the capital stack is being assembled around long-term demand for compute, with private credit underwriting the physical substrate of model progress.
The phrase “AI XPV Platform” is not especially elegant. It does, however, describe the shape of the market. Broadcom supplies custom silicon and networking. Apollo and Blackstone supply large pools of credit capital. Anthropic supplies demand that is large enough, urgent enough, and presumably contracted enough to make the structure financeable.
Infrastructure markets follow a pattern: scarcity, then bespoke deals, then the bespoke deals become repeatable platforms. The strange part is the speed. AI compute went from scarce cloud capacity to a 20-gigawatt private-credit platform in roughly the time it used to take a normal company to migrate from one SaaS contract to another.
Broadcom says the structure will help frontier labs train and serve models at lower cost and lower power. That may be true. The immediate effect is simpler: the cost of staying at the frontier is now large enough that it requires Apollo, Blackstone, Broadcom, global banks, and a gigawatt of Fluidstack capacity to make the next tranche of Anthropic compute legible.
The model is still software. The balance sheet is starting to look like utilities.
–
By the Control Plane Editorial Team