← Back to feed

Broadcom launched the AI XPV Platform on June 9, 2026, alongside Apollo Global Management and Blackstone’s Credit and Insurance business. The vehicle is designed to fund more than 20 gigawatts of compute capacity using Broadcom’s custom XPUs and networking infrastructure by 2028.

Structure

The platform opens with a $35 billion initial tranche, led by Apollo in partnership with Blackstone. This is a private credit deal — not equity, not public financing. Apollo and Blackstone are the anchor lenders; Broadcom provides the silicon and networking technology; the AI labs consume the capacity.

The first deployment: Anthropic’s previously announced expansion to more than 1 gigawatt of compute at Fluidstack-based sites, expected to begin in mid-2026. OpenAI is named in the announcement as a target customer for future capacity on the platform.

Why This Structure

XPUs are Broadcom’s custom application-specific processors, built to run large model inference and training workloads for specific lab architectures. Google’s TPU is the most prominent precedent. Anthropic and OpenAI have both been deepening their Broadcom relationships as they move compute off NVIDIA’s general-purpose H100/H200/B200 stack toward custom silicon with better cost and performance per workload.

The private credit wrapper solves the same problem that Apollo and Blackstone solved for Google TPU buildouts: multi-gigawatt AI infrastructure requires 10-figure financing at timescales that exceed what most labs can self-fund, and the collateral — long-term compute contracts from creditworthy counterparties — lends itself to structured debt rather than venture equity.

Scale Context

20 gigawatts is the stated total target by 2028. For reference:

• The Stargate project (OpenAI + SoftBank + Oracle) targets 500MW in its first phase, scaling toward 5GW.
• Microsoft’s 2026 capex program is targeting roughly 10GW globally.
• The entire US datacenter power consumption was approximately 25-30GW total as of early 2026.

The AI XPV Platform is not an exclusive arrangement. Broadcom positioned it as an infrastructure platform that can serve multiple frontier labs, with Anthropic as the anchor tenant.

The Compute as Asset Class Pattern

This is the second major private credit vehicle for AI infrastructure announced in 2026 after Apollo and Blackstone’s earlier Google TPU deal. The structure is consistent: institutional capital, investment-grade-adjacent collateral from lab compute contracts, managed by the largest alternative asset managers. CoreWeave, IREN, and Fluidstack are all on similar paths.

The private credit market’s move into AI infrastructure represents a structural shift: compute capacity has become financeable infrastructure, not a venture bet. The question is whether 20GW of Broadcom XPU supply actually clears by 2028 on the demand trajectory the platform is betting on.