Since the beginning of this year, driven by artificial intelligence (AI), U.S. stocks have rebounded sharply, and technology stocks have also "regained their glory" and formed the "Big Seven" including $Microsoft (MSFT.US)$ and $NVIDIA (NVDA.US)$. Everything seems to be looking very good, but it should be noted that currently only Nvidia is really making money from AI.

Whether it is the "leader" Microsoft or the "rising stars" Google, Meta, and Adobe, these companies are still in the stage of integrating AI into their products and have not yet truly "monetized" it, that is, making real money from AI. At present, many companies' AI services are free. Only Microsoft dares to increase the price of Copilot by 83%, but consumers may not buy it yet.

Although they have not yet truly achieved profitability, it has become a fact that technology companies are investing heavily in the AI ​​field and hoarding GPUs. According to Wall Street analysts, Nvidia's GPU sales may exceed $50 billion by the end of this year.

At this point, investors can't help but wonder, can technology companies recoup their capital by purchasing GPUs on such a large scale when their profit prospects are still unclear? In the end, will it be all in vain? If it can be repaid, when can it be repaid?

David Cahn, a partner at venture capital firm Sequoia, recently published a calculation. Cahn believes that every $1 of GPU spending corresponds to approximately $1 of data center energy costs. In other words, under conservative estimates, if NVIDIA can sell $50 billion of GPUs by the end of the year, data center spending will be as high as $100 billion.


Then, assuming a 50% profit margin, the AI ​​industry needs $200 billion in revenue to recoup the cost of the upfront investment. But Cahn pointed out that there is currently only $75 billion in annual revenue, leaving a shortfall of $125 billion.

Doubts come

Guido Appenzeller, special advisor to Silicon Valley venture capital giant A16Z and founder of AI startup 2X, refuted Cahn's views and overturned his arguments word for word.

Overall, Appenzeller's core argument revolves around the belief that artificial intelligence will become a ubiquitous component in almost any software-containing product. He asserted that large investments in GPU infrastructure, even as high as $50 billion, could easily be amortized against the massive $5 trillion in global IT spending.

He not only overturned Sequoia’s estimate of AI’s profitability, but also pointed out that Sequoia’s most fundamental problem was that it underestimated the impact of AI’s historic revolution.

Specifically, Appenzeller first pointed out that Cahn was a "clickbait" and tried to use a number like "$200 billion" to attract people's attention, but in fact his calculation process was completely wrong.


Appenzeller pointed out that Cahn added together the purchase cost (capital expenditure) of the GPU, the annual operating costs, the cumulative revenue during the GPU life cycle, and the annual revenue from AI applications, and came up with a seemingly exaggerated figure of $200 billion. But he believes a more appropriate calculation would be based on the annual rate of return GPU buyers receive on their investment costs.

Secondly, he also believes that the electricity cost of GPUs has also been exaggerated. According to Appenzeller, an H100 PCIe GPU costs about $30,000 and consumes about 350 watts of power. Taking into account servers and cooling, the total power consumption is likely to be around 1 kilowatt.


Calculated at an electricity price of US$0.1/kilowatt, this H100 GPU will require only US$0.15 in electricity for every US$1 spent on GPU hardware during its five-year life cycle, which is far lower than the US$1 estimated by Cahn.

But most importantly, Appenzeller believes, Cahn ignores the scale of the AI ​​revolution. He pointed out that AI models are an infrastructure component just like CPUs, databases and networks. Now, almost all AI software uses CPU, database and network, and this will be the case in the future.


So, can the AI ​​industry earn enough US$200 billion? Appenzeller gave an affirmative answer, and more than that, as network infrastructure, the revenue it creates will exist in different forms in each department.

Therefore, he concluded that AI will subvert all software, and Cahn's so-called "AI revenue gap" does not actually exist.