TL;DR
· SemiAnalysis projects outstanding AI debt could reach $7.1 trillion by 2029, with AI infrastructure increasingly relying on credit markets.
· NVIDIA has launched revenue-sharing and credit support models for AI clouds, with Sharon AI and Firmus as the first partners.
· Backstops can reduce bank concerns, but declining rents, data center shortages, and contingent commitments remain key risks.

The report released by SemiAnalysis on July 6 brings the scale of AI infrastructure financing to the forefront: from 2024 to 2029, global AI capital expenditures could cumulatively reach approximately $11.1 trillion, with outstanding AI-related debt potentially exceeding $7 trillion by 2029, around $7.1 trillion.

This is not a simple GPU sales forecast. The core change discussed in the report is that AI construction is shifting from "tech giants buying GPUs with cash flow" to "banks and bond markets financing GPU clusters." If this projection materializes, AI-related debt could become a massive asset-backed financing category second only to the U.S. mortgage-backed securities market.

NVIDIA's role is also evolving. An official NVIDIA blog post on July 1 confirmed that the company has launched a "revenue-sharing and credit support model" for AI clouds, combining capital partners, cloud service providers, and data center projects to drive AI computing power construction. Sharon AI and Firmus are the first partners.

SemiAnalysis further assesses that NVIDIA may help Neocloud package GPUs, customer orders, and data center capacity into financeable assets through structures like GPU revenue backstops and revenue sharing. For lenders, the key is not how hot future AI demand will be, but whether the project can still generate cash flow to service debt in the worst-case scenario.

Global AI IT and Data Center Capital Expenditure and Debt Forecast: Cumulative CapEx of approximately $11.1 trillion from 2024 to 2029, with outstanding debt of about $7.1 trillion in 2029.

AI construction is getting more expensive, and banks first need to see who will pay the rent

In the past few years, AI infrastructure has been primarily borne by hyperscale cloud providers like Google, Amazon, Meta, Microsoft, and Oracle. These companies have cash flow, balance sheets, and internal AI demand, making financing relatively easier.

But as demand for AI training and inference continues to grow, relying solely on the capital expenditures of a few giants becomes insufficient to cover the entire computing power gap. SemiAnalysis expects annual AI CapEx to far exceed $2 trillion by 2028. GPUs, networking, storage, supporting CPUs, and data center construction will consume massive amounts of capital, and credit markets will become one of the funding sources.

This is where the financing difficulties for Neocloud lie.

Such new cloud service providers typically need to secure three things simultaneously: acquire GPUs, obtain data center capacity, and sign future customers. The hardest thing for banks to assess is whether GPU rents over the next few years can cover debt service. AI computing power rental prices change rapidly, customer lease terms vary, and GPU residual value and utilization are harder to estimate than traditional infrastructure.

NVIDIA's credit support model aims to give lenders a clearer cash flow floor. Officially described as revenue-sharing and credit support, SemiAnalysis characterizes the typical structure as a GPU revenue backstop.

6-Year Example Backstop Average Price of $2.36, Short-Term Lease Scenario IRR Could Reach 25%

The example structure provided by SemiAnalysis involves NVIDIA offering minimum revenue support for a specific GPU cluster over a 6-year period, with the price curve declining year by year, averaging about $2.36 per hour per GPU over the 6 years. If actual project rents exceed the backstop level, Neocloud and NVIDIA share revenue at approximately 40% to 60%.

This is not an official accounting term disclosed by NVIDIA but an indicative estimate in SemiAnalysis's model. Its appeal to lenders lies in partially converting a highly uncertain GPU leasing project into an asset with a minimum cash flow commitment.

Banks do not necessarily have to fully believe that AI rental prices will remain high. As long as the project still meets debt service coverage requirements under the backstop-triggered scenario, it may qualify for a loan. According to SemiAnalysis, for clusters backed by NVIDIA's AA/Aa2 rating, lenders require a debt service coverage ratio of at least about 1.3x, corresponding to a 70% to 80% loan-to-value ratio. Initial financing spreads may be higher than deals supported by hyperscale cloud providers but lower than CoreWeave's unsecured bond yield of around 10%.

Indicative Terms of NVIDIA Backstop: 6-year average of about $2.36/hour/GPU, with 40%-60% revenue sharing.

For Neocloud, a backstop is not just insurance but a critical condition for obtaining debt financing.

In the GB300 short-term one-year lease example, if the first-year rent is $6.75/hour and NVIDIA's share is 40%, Neocloud's 6-year project IRR is about 25.4%, with NVIDIA's average take rate around 18%. If market demand is insufficient and the project falls entirely under the backstop lease, Neocloud's IRR could be near zero or slightly negative.

This is not favorable for equity returns but is critical for financing: project returns may be compressed, but debt service may still be covered. In other words, the backstop transforms a GPU cluster that "could make a lot of money" into a financeable asset that "could still service debt under stress scenarios."

GB300 Neocloud Return Comparison: Under 40% short-term lease sharing, 6-year IRR of about 25.4%; when fully triggering the backstop, IRR near zero or slightly negative.

Sharon AI and Firmus First to Implement, Asia-Pacific Projects Become Testing Grounds

NVIDIA has officially confirmed that Sharon AI and Firmus are the first partners for this revenue-sharing and credit support model.

A Sharon AI announcement on June 12 stated that the company has reached a 6-year strategic computing partnership with NVIDIA, with a 72MW AI factory in Australia deploying up to 40k Grace Blackwell GB300 units. Sharon AI's overall AI factory capacity is planned to reach 132MW, of which 102MW has been contracted, with over 55k NVIDIA GPUs expected to be deployed by mid-2027.

Firmus's Batam project in Indonesia is larger in scale. According to the official NVIDIA blog, the Firmus Batam project can scale to 360MW, deploying up to 170k NVIDIA GPUs. SemiAnalysis includes this project in its discussion, noting it primarily targets AI-native enterprises and inference service providers, potentially offering diversified lease terms.

These cases show that NVIDIA's credit support model is no longer just a financial model assumption but has entered the early project implementation stage. However, currently disclosed cases are mainly in the Asia-Pacific region, while the U.S. market still faces constraints such as data center capacity, power, and interconnection speed.

Data centers remain the hardest bottleneck. GPUs can be procured, customer demand can be contracted, but power, land, racks, cooling, and interconnection timelines are difficult to replicate quickly. The SemiAnalysis model also mentions that NVIDIA may need to directly lease data center capacity to help Neocloud bridge the supply-demand gap. The specific capacity and scale involved in this part are still estimates from the report and should not be equated with official NVIDIA disclosures.

NVIDIA Can Earn a Share but Will Also Take on Larger Long-Term Commitments

For NVIDIA, supporting GPU financing has two layers of benefits.

First, it can expand the scale of GPU sales and deployment. More Neocloud entities obtaining financing means more players can purchase and operate large-scale GPU clusters, reducing the AI computing power market's reliance on a few hyperscalers.

Second, it can potentially generate additional revenue sharing. According to SemiAnalysis's model estimates, if such structures continue to expand, NVIDIA's incremental income from backstops and revenue sharing could become significant, with high profit margins.

The costs are also clear. NVIDIA's off-balance-sheet or related disclosures of long-term commitments could grow rapidly. In the paid report section, SemiAnalysis estimates that NVIDIA's cloud service agreements or contingent guarantees could rise to hundreds of billions of dollars in the coming years. Since these numbers have not been confirmed item by item by NVIDIA, they are more suitable as model stress tests rather than actual liabilities.

NVIDIA Cloud Service Agreement Growth Forecast: According to the SemiAnalysis model, related long-term commitments may continue to accumulate with every 100MW of computing power support.

This is not traditional direct debt. But if the GPU rental market weakens and customer demand falls short, the probability of backstops being triggered increases, and NVIDIA would need to shoulder more minimum revenue support. Ultimately, the market will not only look at how much NVIDIA can earn from the share but also whether these commitments affect its own capital allocation and cash flow priorities.

The Biggest Test: Can Rents and Data Centers Hold Up?

The most striking aspect of this report is placing AI computing power construction within the credit market. As capital expenditures swell to trillions of dollars, GPU clusters are no longer just tech products; they become financing assets jointly evaluated by banks, bond investors, and cloud service providers.

However, the $7.1 trillion AI debt is still a forward-looking model projection, not a fact already realized. It depends on several premises: sustained expansion of AI demand, relatively high GPU utilization, rental prices declining at a controllable pace, data center construction keeping up, and lenders willing to accept cash flow models backed by NVIDIA's credit support.

The most likely issues are prices and implementation speed. If GPU rents decline faster than expected, Neocloud's returns under high revenue sharing and high financing costs will be compressed. If backstops are heavily triggered, projects may still service debt, but NVIDIA's commitments become heavier. If data centers, power, and interconnection are delayed, the GPU deployment timeline in financing models will also be disrupted.

NVIDIA's "backstop for GPU financing" story points to the next phase of funding sources for AI infrastructure. It can enable more computing power projects to obtain loans and may push NVIDIA to a more central position in the AI credit market. But whether this market can grow to $7 trillion ultimately depends on rents, utilization, and data center delivery.

Click to learn about job openings at BlockBeats

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia