Jensen Huang responds to CoreWeave debt, GPU depreciation period, and DeepSeek inference gross margin

On Wednesday, after the GTC Berlin conference, NVIDIA CEO Jensen Huang and Silicon Valley tech investor Gavin Baker were interviewed by Yahoo Finance. In the interview, Huang directly addressed three core issues that have recently troubled AI investors: the debt financing risks of new cloud computing companies, the actual useful life of GPU assets, and the astonishing inference gross margin data disclosed by DeepSeek.

Regarding debt financing risks and GPU useful life, Huang explicitly stated that NVIDIA GPUs can last seven to eight years, far exceeding the four years generally assumed by the market, a characteristic that makes them eligible as loan collateral. He also noted that NVIDIA has provided guarantees for the resale residual value of chips in some financing projects to help new cloud computing companies such as CoreWeave reduce financing costs.

In addition, Gavin Baker cited the latest data disclosed by DeepSeek, saying that the GPU inference monetization price of this open-source model is as high as $38 per hour, far above current market rental prices, which may be exactly why Blackwell chip rental prices have risen for 17 consecutive trading days. Huang responded that strong demand is the fundamental prerequisite for solving all current financing and supply problems: "With demand, all problems can be solved."

These remarks have direct significance for the market. New cloud computing companies (NeoCloud) such as CoreWeave have recently sparked investor concerns due to their highly leveraged financing structures, and their stock and bond prices have remained under pressure. Huang's response attempts to reconstruct the market's understanding of AI infrastructure debt risk from the perspective of asset quality, while DeepSeek's inference gross margin data provides new evidence for the sustainability of GPU demand.

GPU depreciation period reassessed: from four years to seven or eight years

In the interview, Huang emphasized the durability of NVIDIA GPU assets, characterizing it as one of the core solutions to the current AI financing dilemma.

He said the market had generally used four years as the GPU depreciation cycle, but that assumption is now outdated. "People are discovering that what used to be considered a four-year lifecycle may actually be eight years or even longer." He attributed this characteristic to NVIDIA's continuous upgrades to the CUDA software stack, which keep the hardware productive throughout its lifecycle. He also confirmed that the expected useful life of the next-generation chip Vera Rubin is likewise seven to eight years.

Huang further explained what this means for capital markets: the high residual value of GPUs, their broad applicability (capable of supporting all AI models and cloud service providers), and their software programmability together form the basis of collateral that lenders can underwrite. "Most GPUs today have almost no residual value, but people now recognize that NVIDIA's computing assets have considerable residual value, so they can be used to back the debt itself." He said that under this logic, NVIDIA's ability to bear AI infrastructure financing risk is actually superior to that of any other asset class.

DSX Pod: compressing the gap period and accelerating capital returns

Addressing the core financing pain point for new cloud computing companies such as CoreWeave—the months-long gap between purchasing equipment and generating revenue—Huang introduced NVIDIA's newly launched DSX Pod solution, a modular deployment format that prefabricates AI factories inside containers.

He explained that the construction cost of a gigawatt-scale data center is currently about $50 billion, and it usually takes four to six months from equipment leaving the factory to completed deployment and operation. During this period, capital sits idle, but interest costs have already begun to accrue. By moving part of the manufacturing process forward into the supply chain, DSX Pod is expected to compress this cycle to three to four months, or even shorter.

Gavin Baker agreed, and using $1 trillion in AI capital expenditure as a reference, noted that if the deployment cycle is six to nine months, it means hundreds of billions of dollars in capital remain unproductive for a long time, creating systemic pressure on financing structures. Huang said, "If we can shorten the time from system shipment to deployment and operation from months to weeks or even one month, that would be a complete game changer for all AI factories."

DeepSeek inference gross margin data draws attention

In the interview, Gavin Baker cited the latest operating data disclosed by DeepSeek, providing a new quantitative perspective on the strength of current GPU demand.

According to his account, DeepSeek publicly stated that the GPU inference monetization price of its open-source model reaches $38 per hour, while mainstream GPU rental prices in the current market are far below that level. He believes this data may explain why Blackwell chip rental prices have risen for 17 consecutive trading days.

Huang responded that strong demand is the fundamental prerequisite for solving all current financing and supply problems. He cited strong demand from multiple institutions including OpenAI, Anthropic, Meta, xAI, Google Gemini, DeepSeek and Moonshot AI, and said: "Without demand, no problem can be solved; with demand, all problems can be solved." He also characterized the current moment as a key inflection point at which AI is formally switching from "invention mode" to "practical mode," and said the continued rise in token demand has solid fundamental support.

In addition, Huang also said that China is the only country in the world capable of achieving large-scale output simultaneously across three dimensions: chip manufacturing, open-source model development, and AI researcher training. "No country can train AI researchers on the scale that China does." He also said that the United States' advantage lies in having the world's top frontier laboratories, chips and computing infrastructure, and said NVIDIA will continue to compete actively globally, including in emerging markets such as Africa, Central Asia, Brazil and Egypt.

On the issue of AI employment, Huang directly characterized the current mainstream narrative as "completely wrong." Citing employment growth data from multiple industries including energy, chips, infrastructure, construction and healthcare, he argued that AI is actually creating more jobs at all levels rather than eliminating them. "The narrative about employment is almost entirely backwards; it is completely wrong." He also noted that AI data center builders should communicate in advance and actively integrate when entering communities to improve public perception, and said the industry as a whole has already made some progress in this direction.Risk Disclosure and Disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are suitable for their particular circumstances. Investment based on this article is at the user's own risk.