NVIDIA, major Wall Street investors team up on “AI factory” financing meant to unlock $500 billion in third-party capital

NVIDIA is moving to recast how artificial-intelligence data centers are funded, pitching “AI factories” as investable infrastructure and announcing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build independent financing platforms for third-party capital.

In a blog post published Tuesday, NVIDIA said the effort is aimed at mobilizing more than $500 billion of aggregate third-party capital over time to support the buildout of AI infrastructure. NVIDIA emphasized that the capital would be assessed and underwritten independently by the financial partners, while NVIDIA would provide the underlying AI factory platform.

The company framed the shift as a move away from project-by-project chip buying and one-off data center construction toward financing structures that resemble other long-lived infrastructure categories. In NVIDIA’s depiction, an AI factory is designed to produce revenue for operators and off-takers, with a value profile that can improve over time due to software upgrades and ongoing demand across multiple AI workloads.

NVIDIA said it is not just selling GPUs as discrete hardware, but a complete AI factory platform. It described NVIDIA DSX AI factories as combining accelerated computing, networking, systems software, AI frameworks and a developer ecosystem, and said one factory can serve multiple customers and workloads. NVIDIA also linked redeployability to a globally adopted architecture used across major clouds and system makers, arguing that capacity can be shifted if customer needs or operators change.

The company said the software layer is central to the “productive asset” argument. It highlighted CUDA, its programming platform for accelerating AI workloads, saying each software generation improves performance and efficiency for already-installed infrastructure and can extend an AI factory’s useful economic life. NVIDIA cited the durability of its Ampere-based A100 chip, introduced in 2020, saying it remains in active commercial use for training, fine-tuning, inference and high-performance computing, with customers continuing multi-year commitments that it said can extend A100’s economic life toward a decade.

NVIDIA also pointed to market pricing as evidence of sustained compute economics. It said one-year H100 rental pricing rose from about $1.70 per GPU-hour in October 2025 to about $2.35 per GPU-hour in March 2026, and that cross-provider on-demand median pricing rose from roughly $2.00 per GPU-hour in October 2025 to about $2.70 in June 2026. For Blackwell capacity, NVIDIA said reported B200 cloud rates have spanned approximately $5.30 to $7.05 per GPU-hour.

To address access-to-capital constraints, NVIDIA said each financing platform is designed for qualified AI labs, enterprises and AI clouds to access AI-factory infrastructure at scale. NVIDIA described an underwriting model in which investors independently assess demand, utilization, cash flow and residual value for each opportunity, while NVIDIA supplies the platform and assumes a limited role in risk support.

NVIDIA added that in some cases it may provide a residual-value support mechanism for up to 25% of an opportunity, assessed project by project. The company said this support is residual-value based and intended to complement, not replace, independent underwriting. It argued that the level of support can be lower than other compute-financing arrangements because NVIDIA compute is designed to be fungible, software-upgradable and redeployable across a broad ecosystem of customers.

Still, important details are not disclosed in the announcement. NVIDIA did not specify the structure of any particular financing vehicles, the expected terms such as duration, interest rates or covenants, or how residual value would be measured in practice for each hardware generation. It also did not name which facilities or customer projects the platforms are already funding, or whether NVIDIA’s limited residual support would apply consistently across partnerships.

For investors and the broader AI buildout, the next question will be whether these platforms can turn compute demand into repeatable, bankable cash flows at scale, and how quickly capacity is financed once projects are underwritten on utilization and residual value assumptions. The test case will likely be the interaction between software-driven performance improvements, shifting AI workloads, and the ability of financing partners to place long-term capital without overreaching on demand forecasts.