Nebius' acquisition of Inferize lets its GPUs spend less time sitting idle, so they earn back their cost faster. That makes Nebius a better borrower, but it does nothing for the chips its secured lenders hold as collateral. Nebius has bought itself a better business, but its lenders are still holding the same chips.
When I last looked at Nebius, it had just raised $775mn in its first secured debt deal, backed by GPUs and the contracted cash flows they generate. My concern was that the chips face the risk of becoming technologically obsolete before the loan matures in 2030. Nebius’ recent acquisition of Israeli startup Inferize may not eliminate that risk but could materially change its economics.
The paradox in GPU-backed lending
For cloud providers like Nebius, chips are a major component of . Therein lies the credit risk: the company spends heavily upfront on infrastructure whose economic value depends on how much productive work it can generate before the next generation of hardware makes it uncompetitive. This exposes a paradox in GPU-backed lending: the technological progress that lenders are betting on can make the borrower stronger whilst making the obsolete faster. The question going forward for lenders is who captures the gains when becomes more productive, is it the company, the shareholders, or the investors holding chips as security?
What Inferize actually does
Founded in January 2026, Inferize was acquired by Nebius within its first year. The deal was reportedly valued at a range between $100-$150mn, according to Calcalist. It tackles a problem known as the ‘GPU idle tax.’ Demand for fluctuates and customers require near-instantaneous responses; unless a model is already loaded, bringing a GPU online, known as a ‘,’ can take a significant amount of time. So, providers keep additional GPUs available and running even when they are not actively processing requests. Inferize’s technology reduces these cold-start times from “tens of minutes to seconds,” according to Inferize CEO Guy Bortnikov. This means that GPUs can be brought online much closer to when demand arrives, reducing the cost of computing capacity sitting idle.
A better borrower, not better collateral
For Nebius, this means more , and more revenue from every GPU it owns. For the lenders behind the $775mn, it means very little. Their loan is repaid from a contract with an investment-grade customer, which pays whether GPUs are busy or not. The value Inferize creates sits outside their collateral pool, making Nebius a better borrower in the long-term, not by making its collateral any better, which is a much more important distinction.
If Nvidia launches a new generation of chips that makes Nebius’ existing GPUs uncompetitive, Inferize cannot stop that. What it can do is help Nebius earn back its capex much faster. That benefit goes to the company and trickles down to its shareholders and , not to the lenders holding the chips.
The technological progress that lenders are betting on can make the borrower stronger whilst making the collateral obsolete faster.
Faster payback, lower risk
A shorter technological life becomes less problematic when the company can generate sufficient cash flow during that period to recover its initial investment. So, whilst Inferize does not extend the of the GPU, it improves the economics of the life they already have.
That matters because the risk I originally identified was not simply that GPUs depreciate, but that they could become obsolete before Nebius has recovered the capital used to buy them. Faster payback still matters for Nebius’ creditors as it shortens the period in which the company depends on chips that may be eroding in value. This has three tangible implications:
- The higher should support cash generation, meaning that Nebius has more capacity to service its debt.
- If the company needs fewer additional GPUs to support the same level of growth, its future capex could fall.
- Faster payback reduces the risk that the chips securing the debt lose significant economic value before the debt can mature.
The acquisition has not removed the technology risk but made that risk cheaper to carry.
The same paradox, elsewhere
The same paradox is already appearing elsewhere. In July, inference provider General Compute secured a facility of up to $400mn through from SambaNova as collateral, which use far less power than comparable GPUs. An early sign that some providers are focusing on the importance of cheaper, more efficient chips over newer GPUs.
What lenders are really lending against
Even as bond yields rise globally, investors continue to finance the AI build-out at scale, such as the recently announced $60bn debt package supporting Anthropic’s lease of Google’s , the largest chip financing deal to date. With more advancements in inference-specific chips, the next phase of AI infrastructure finance may be less about lending against more compute, and more about lending against more productive compute.
Nebius and General Compute point to the same conclusion: the value in AI infrastructure is moving away from the chip itself and towards how well it is used. Inferize makes Nebius’ GPUs earn more; SambaNova’s chips make General Compute cheaper to run. In both cases, the borrower gets stronger. But the lender cannot repossess efficiency. What it holds is hardware, and that hardware is still racing towards obsolescence.
For lenders, the question is no longer simply whether the chips will hold their value, but whether this is the right infrastructure to lend against. And that may be the more important shift for AI infrastructure credit: until lenders can find a way to secure the productivity and not just the hardware, every gain in AI efficiency will strengthen the borrower faster than it strengthens the collateral. So, Nebius has bought itself a better business, but its lenders are still holding the same chips.
- https://www.calcalistech.com/ctechnews/article/byzggtiqgx
- https://nebius.com/newsroom/nebius-acquires-inferize-to-strengthen-nebius-token-factorys-production-inference-stack
- https://26806167.fs1.hubspotusercontent-eu1.net/hubfs/26806167/gated/nebius-economics-of-ai-clusters-whitepaper.pdf
- https://www.onlogic.com/uk/blog/what-is-a-gpu-a-beginners-guide/
- https://thebasispoint.finance/notes/lending-against-obsolescence
- https://www.ft.com/content/5b9c8ce3-d07c-46f9-8cea-cbc4e7f8ccca?syn-25a6b1a6=1
- https://techcrunch.com/2026/07/17/why-the-first-gpu-financiers-are-turning-to-inference-chips-in-a-400-million-deal/

