How $35 Billion Becomes 1GW—and Where the Risk Sits

Broadcom, Apollo and Blackstone launched the AI XPV Platform in June 2026. The initial $35 billion capital solution supports more than 1GW for Anthropic, expected to deploy at Fluidstack-based sites beginning in mid-2026. The platform is designed to enable more than 20GW of Broadcom XPU and networking capacity through 2028.

The announcement calls it a repeatable framework. It widens the funding base to insurance capital and private credit, but it does not disclose complete asset ownership, customer minimums, residual-value support, termination rights or unit cost by gigawatt.

Apollo president Jim Zelter later described the initial financing as five-year amortizing paper and emphasized downside and residual-risk control. That is a public management description rather than a published contract. This column analyzes the mechanism without inventing undisclosed guarantees.

$35 billion cannot be divided cleanly by 1GW

The initial solution spans custom Broadcom XPUs, networking, site deployment and financing. Dividing $35 billion by 'more than 1GW' overstates unit build cost because the envelope may include equipment, construction financing, interest reserves, expansion options and several sites.

The 20GW figure is a platform objective through 2028, not contracted, energized or billed capacity. An individual project becomes deliverable only when equipment orders, sites, customer commitments and closed financing are present together.

Credit travels from model demand into data-center assets

Four layers have to work together. Anthropic supplies long-term demand; Broadcom supplies a custom XPU and network roadmap; Fluidstack deploys the sites; Apollo and Blackstone translate contracts and equipment into debt cash flow.

Anthropic reduces the need to own every asset up front, while Broadcom gains funded orders. Lenders receive scheduled amortization and accept that the technology turns much faster than conventional infrastructure.

1. 01
   Demand commitments support revenue

   Minimum use, payment duration and counterparty credit determine amortization; the detailed terms are not public.

2. 02
   The roadmap sets residual value

   Custom XPUs and networks have a narrower secondary market than general servers. Early obsolescence lowers recovery.

3. 03
   Site delivery controls timing

   Power, cooling, networking and data-hall delays leave purchased equipment waiting while interest and depreciation accrue.

4. 04
   Private credit absorbs the mismatch

   Five-year amortization shortens exposure, but customer payments, asset disposition and refinancing still have to align.

Unused capacity and generational obsolescence are hardest to price

A long-term compute contract can lock in nominal demand without guaranteeing hourly utilization. Model efficiency, lower inference prices or workload migration may reduce renewal value even when the customer keeps paying. Flexible termination terms would return risk to the project company and lenders earlier.

Residual value is the second variable. Data halls, power and cooling can operate for more than a decade; custom accelerators have shorter economic lives. The project needs substantial early cash recovery or an upgrade path that reuses the facility.

Concentration is the third. The first project depends on Anthropic, Broadcom and Fluidstack acting together. Public information does not establish substitution rights or loss allocation if one layer changes.

Scaling from 1GW to 20GW requires project-level evidence

The next financing headline is not enough. Each project needs a record of site, energization, equipment, customer, contract duration, financing close and commissioned capacity. Platform scale matters only after several projects repeat those milestones.

If five-year amortization recovers most principal inside the hardware's economic life and older data halls accept the next generation, the structure can become a major channel for AI capex. Delays, weak utilization or poor residuals will first appear in private-credit pricing.