When more and more people in the market begin to discuss whether AI will cause software to become fully commoditized—further compressing technology companies’ valuations and profit margins—NVIDIA CEO Jensen Huang’s answer is very direct:
What truly cannot be easily commoditized isn’t just the software itself, but the entire process of converting electronics into tokens.
In his latest interview, Huang Huang provides a complete account of how he understands this AI race, from NVIDIA’s supply chain, the CUDA ecosystem, and AI computing architectures, to hyperscale cloud customers, the China market, and U.S. export controls.
(Jensen Huang uses the “five-layer cake” metaphor to explain the evolution history of artificial intelligence)
His core argument can be condensed into a single sentence: AI is not a competition between single models, nor a competition between single chips. It is a “five-layer cake” battle spanning energy, chips, networks, software, the ecosystem, and the application layer—and what NVIDIA wants to do is the segment that is the hardest, but also the least likely to be replaced.
Jensen Huang: NVIDIA’s job is to turn electrons into more valuable tokens
Faced with outside skepticism, since many software companies’ valuations are under pressure because of AI, and NVIDIA, at its core, designs with TSMC for manufacturing, receives memory from SK Hynix and Samsung, and has assembly handled by Taiwan ODMs—could NVIDIA also be dragged under by the AI commoditization wave? Huang’s answer is: not that simple.
He believes NVIDIA’s role was never to do everything by itself. Instead, it is responsible for the most critical and most difficult part in the entire chain of conversion from electrons to tokens. As he puts it, NVIDIA’s input is electrons, its output is tokens, and the extremely complex conversion capability in between is what gives this company its reason for existence.
Huang Huang emphasizes that this conversion is not simply turning electricity into computational results. It must continuously increase the value of tokens—so that the same amount of computing can produce tokens with greater economic value and higher efficiency. This involves architecture design, packaging, memory, interconnects, algorithms, libraries, the software stack, and ecosystem collaboration—a highly engineering-driven and science-driven process that is still evolving rapidly. He believes it is unlikely to be fully commoditized.
He also further describes NVIDIA’s corporate philosophy: “do the most that is necessary, and do the least that is not necessary.” In other words, for parts that they don’t have to do themselves, they should hand them off to partners and the ecosystem to complete; but for parts that must be done and are extremely difficult, NVIDIA has to step in personally—and do the best.
These “tool-oriented software companies” may actually grow explosively because of AI
When it comes to the market’s concern that AI will squeeze the space for software companies, Huang Huang actually holds almost the opposite view. He points out that many software companies today are essentially tool makers—such as Excel, PowerPoint, or EDA companies like Cadence and Synopsys. The reason these companies haven’t yet seen a bigger explosion is not because tools will be eliminated, but because today’s agents are not yet good at using tools.
In his view, the number of agents in the future will grow exponentially, as will the number of tool users. This, in turn, will drive up the number of times tools are called and the licensing demand for the tools themselves. Take chip design as an example: today’s use of design tools is still limited by the number of engineers; but in the future, behind each engineer there may be multiple agents collaborating, and the density and frequency of design space exploration will be far beyond what it is today.
By then, the actual usage of tools like Synopsys Design Compiler, floor planner, layout tools, and design rule checker may instead surge dramatically.
In other words, Huang Huang does not believe AI will simply wipe out tool-oriented software companies. Instead, it is more likely to push them onto a new growth curve.
NVIDIA’s real moat is the upstream-to-downstream supply chain
When discussing NVIDIA’s large procurement commitments to the upstream supply chain in recent years—so much so that outsiders estimate it could accumulate to a scale of several hundred billion dollars in the coming years—Huang Huang did not deny that this is one of NVIDIA’s important advantages.
He said NVIDIA has indeed made a lot of upstream commitments, both explicit and implicit. The former are procurement commitments that are visible to the outside world; the latter are made by persuading leaders in the supply chain to be willing to invest in capacity expansion first. These investments happen not only because NVIDIA is willing to buy, but because suppliers believe NVIDIA has the ability to absorb that capacity and, through the huge downstream demand, sell it out smoothly.
That’s also why he views GTC not just as a product launch event, but as a “360-degree panoramic gathering” of the entire AI universe. In his eyes, one of the value of GTC is to let the upstream see the downstream, let the downstream understand the upstream, and let the entire industry chain jointly confirm that AI demand really will come—and at an enormous scale. Huang Huang even frankly admits that his keynote, to some extent, has a strong “educational” function, because he has to make the entire supply chain understand: why AI is coming, when it is coming, how big it will be, and how to prepare in advance.
This is also the reason NVIDIA has been able to continue expanding the flow of its supply chain in recent years. Huang Huang emphasizes that the supply chain doesn’t only look at cash flow; it also looks at turnover rates and demand visibility. If a company’s architecture and product turnover speed isn’t fast enough, the supply chain won’t be willing to build factories or expand production lines for it in advance. The reason NVIDIA can do these things is that its downstream demand is big enough and certain enough, and the entire supply chain can clearly see it.
Huang Huang isn’t afraid of bottlenecks; most bottlenecks are basically a two- to three-year issue
When asked whether the upstream can truly keep up with AI computing demand—especially as AI has already absorbed a large amount of TSMC’s advanced process and packaging capacity, and in the future, how could it possibly keep doubling year after year—Huang Huang’s stance is very clear: almost all manufacturing bottlenecks, in essence, are just a two- to three-year problem.
He gave an example: in the past, people often talked about CoWoS packaging bottlenecks, but nowadays hardly anyone talks about them—because the entire industry concentrated its efforts to resolve them within two years. TSMC has also treated advanced packaging and HBM as part of mainstream computing technology, not as something with special requirements. In other words, as long as the demand signals are clear enough, the industry chain will proactively rush in to fill the bottlenecks.
To Huang Huang, AI doesn’t bring jobs disappearing—it brings industry restructuring and a redistribution of talent demand. What really needs to be worried about isn’t whether certain occupations will completely disappear, but whether society is misallocating the supply of talent due to excessive fear.
He also said directly that issues like logic process, packaging, and HBM can all be solved within two to three years; what is truly slower and more troublesome is energy policy. Because whether it’s AI factories, chip manufacturing, advanced packaging, electric vehicles, robots, or reindustrialization, none of it can do without energy. If energy becomes the bottleneck, the expansion speed of the entire industry will be constrained.
This article Huang Jensen Huang’s latest interview: Can NVIDIA’s moat continue to hold? (Part 1) was first published on Lianxin ABMedia.
