On August 10, 2026, NVIDIA signed memoranda of understanding (MOUs) with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create independent AI infrastructure financing platforms. The goal is to attract over $500 billion in third-party capital to deploy "AI factories" and turn GPU computing into a new investable asset class.
What Happened
The agreements are structured as memoranda of understanding (MOUs), and final contracts have not yet been signed. Under the partnership, the financial institutions are creating separate platforms to finance the construction and operation of GPU-based data centers using NVIDIA GPUs. At the event, Jensen Huang presented the thesis that AI computing generates stable revenue, has versatility due to the DSX architecture, is resold between operators, and improves over time through CUDA updates. NVIDIA is ready to guarantee up to 25% of the residual value of projects on a deal-by-deal basis.
Context
An analogy with commercial real estate or roads: GPU clusters are being assessed for the first time in history as a bankable asset capable of servicing debt. Rating agency Moody's had previously warned about the uncontrolled growth of debt among tech companies due to capital investments in AI — this is precisely the problem motivating the search for alternative funding sources. The industry is facing a shortage of computing power: GPU rental prices are steadily rising, and major companies are competing for a limited supply of chips.
Why This Matters for the Industry
If the financing model works, the financial pressure from Moody's is relieved from major tech companies: the construction of AI infrastructure will shift to specialized financial operators. For the industry, this means the emergence of new independent GPU rental providers not tied to traditional cloud giants. At the same time, the dependence of the entire AI infrastructure on the NVIDIA ecosystem is strengthened — margin is shifted from direct hardware sales to financial engineering. The risk of rapid depreciation of GPU equipment with the release of a new generation remains the main challenge: traditional loan terms may not align with the chip lifecycle.
Why This Matters for Users
Current market GPU rental rates show a growing barrier to entry: H100 has increased from ~$1.70 per GPU-hour in October 2025 to ~$2.35 in March 2026, while Blackwell (B200) costs $5.30–$7.05 per hour at cloud providers. If the model is successfully implemented, third-party capital could expand supply and reduce inference costs by 20–40% over two years through competition and scale. For startups and mid-sized businesses, this potentially opens access to Blackwell-level computing power without the need for direct purchases. In the current quarter, the situation for engineering teams and research groups does not change — the infrastructure has not yet been built.
What Is Still Unknown / Limitations
All agreements are at the MOU stage — there are no final contracts or launches of financing platforms. The claim that the DSX architecture is a "de facto standard" relies exclusively on NVIDIA's marketing materials: no technical specifications, open-source components, or independent benchmarks for DSX have been presented. The 25% residual value guarantee is a financial mechanism, but the problem of real liquidity of GPU equipment with the emergence of new chip generations remains unresolved. There are currently no specific implementation timelines or public pricing plans from third-party operators.
Sources
- NVIDIA Blog — AI Factory Compute Is Becoming an Investable Asset Class
- NVIDIA News — Official Partnership Announcement
- CNBC — Nvidia, Wall Street asset managers partner on $500B AI push
- Jensen Huang — Twitter/X Announcement
Author
Look at AI, editorial team