Key Takeaways
The $600 Billion AI Secret: Why Big Tech is Quietly Quitting NVIDIA
- Nvidia guarantees to lease back unused GPUs from small cloud providers if they can't find customers
- The strategy helps Nvidia reduce dependence on hyperscalers who are building rival AI chips
- Small providers get both GPU financing and data center funding through Nvidia's guarantees
Nvidia is offering small cloud providers a deal they can't refuse: buy our GPUs, and if you can't sell the compute, we'll lease them back ourselves. The arrangement, reported by The Information, lets Nvidia diversify its customer base away from Amazon, Microsoft, and Google while bankrolling a new generation of AI infrastructure companies.
How the financing works
The chip company is acting less like a hardware vendor and more like a financial backer. Nvidia promises to cover unused GPU capacity if young cloud providers can't find AI developers to rent it. In exchange, Nvidia takes a cut of the provider's cloud revenue.
The guarantee solves two financing problems at once. Without it, startups struggle to secure loans for both data center leases and the GPUs themselves. A single H100 chip costs upward of $40,000, and training clusters require thousands of them.
“Nvidia kills two birds with one stone. If Nvidia only guaranteed the building leases, then you're still stuck with, 'How do you finance the GPUs?' But when Nvidia guarantees it will pay for unused compute capacity, the GPUs get financed and the data center gets financed.”
— Unnamed data center executive, via The Information
Why Nvidia wants out of hyperscaler dependence
Amazon, Microsoft, Google, and Meta buy the majority of Nvidia's chips. That's a concentration risk. Each of those companies is also building custom AI silicon: Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia. Every chip they design in-house is one they won't buy from Nvidia.
By propping up smaller cloud providers, Nvidia creates alternative distribution channels. If CoreWeave, Lambda, or other startups grow into viable compute platforms, Nvidia has customers who lack the resources or incentive to build their own chips.
Jensen Huang, Nvidia's CEO, has said publicly that "the more different types of clouds, the better for us." This financing strategy is the execution of that principle.
The scale of Nvidia's market position
Nvidia controls an estimated 70 to 95 percent of the AI training chip market. Its data center revenue hit $26 billion in a single quarter in fiscal 2024. The company's market cap exceeds $3 trillion. That dominance gives Nvidia the balance sheet to act as a lender of last resort for GPU-hungry startups.
Nvidia's venture portfolio already exceeds $50 billion in value. The company has invested in dozens of AI startups over the years. But this financing model goes further. Instead of taking equity stakes, Nvidia is backstopping operating risk, essentially betting that GPU demand will remain high enough that the company never has to make good on its buyback promises.
What this means for startups shopping for compute
For AI teams comparing cloud options, the arrangement suggests smaller providers may offer competitive pricing and availability. If Nvidia is guaranteeing their inventory, those providers have less pressure to mark up GPU time to cover financing risk.
The trade-off: these providers remain locked into Nvidia hardware. Teams betting on alternative chips, whether AMD's MI300 series or custom accelerators, won't find them at Nvidia-backed clouds.
Logicity's Take
This is vendor financing dressed up as startup support. Nvidia isn't being charitable; it's buying market share insurance. For AI builders, the practical upside is real: more cloud options with GPU availability. But the deeper implication is that Nvidia is cementing its position by making the entire ecosystem financially dependent on its success. If you're evaluating compute providers, ask whether their Nvidia-backed financing affects their flexibility to offer alternative hardware down the line.
The risk Nvidia is taking
The buyback guarantees only work if GPU demand stays high. If the AI training boom slows, or if inference workloads shift to cheaper hardware, Nvidia could end up leasing back chips it can't sell. That's a lot of capital tied up in depreciating inventory.
Still, Nvidia's bet is straightforward: AI infrastructure demand will keep growing, and being the lender makes Nvidia indispensable to the next generation of cloud providers. If that bet pays off, the company owns both the chips and a chunk of the revenue from running them.
Frequently Asked Questions
What is Nvidia's GPU buyback guarantee?
Nvidia promises small cloud providers that if they can't find customers to rent GPU capacity, Nvidia will lease back the unused chips itself. This reduces the financial risk for providers buying expensive AI hardware.
Why is Nvidia funding AI startups instead of just selling chips?
Nvidia wants to reduce its dependence on hyperscalers like Amazon, Microsoft, and Google, which are building their own AI chips. Backing smaller providers creates alternative customers less likely to compete with Nvidia.
How much does an Nvidia H100 GPU cost?
An H100 GPU costs upward of $40,000, and AI training clusters require thousands of them, making financing a significant barrier for smaller cloud providers.
Does Nvidia take equity in the startups it backs?
This financing model differs from traditional venture investment. Instead of equity stakes, Nvidia takes a cut of the cloud provider's revenue in exchange for the GPU buyback guarantee.
The growing demand for AI agents is one driver behind the GPU compute crunch Nvidia's financing addresses
Need Help Implementing This?
If you're evaluating cloud providers for AI workloads or need help understanding GPU pricing dynamics, reach out to the Logicity team. We track infrastructure options and can help you find the right fit for your compute needs.
Source: The Decoder / Maximilian Schreiner
Huma Shazia
Senior AI & Tech Writer
Produced with AI assistance and reviewed by the Logicity editorial team. Learn more in our Editorial Policy.
