NVIDIA, Time to Call an End to the Party

According to the AWARE Original piece above, NVIDIA’s share price corrected by about 30% after some analysts, alongside former President Trump’s tariff threats that began in February, pointed out that parts of Microsoft’s new data center construction plans had been canceled, reigniting concerns about demand for AI accelerators (GPUs).

I quoted and endorsed Mizuho Securities analyst Jordan Klein’s comment that it “just looks to me like part of normal course operations,” emphasizing that concerns over AI accelerator demand were still premature at this stage.

Going back to the piece I wrote in July 2024, “AI’s 700 Trillion-Won Problem”, I addressed the revenue and profitability issues around AI raised by David Cahn, a partner at Sequoia Capital, one of the largest VCs in the US. I argued that his claim “failed to account for the time value of money,” and that even the “productivity gains” that do not show up as revenue are sufficient grounds to invest in AI. This has been borne out by numerous companies that have dramatically boosted the productivity of their existing workforce with AI.

Subsequently, in the most recent Q2 (calendar year) earnings releases from Meta and Microsoft, it was confirmed that AI infrastructure investments are indeed helping to drive tangible revenue and productivity gains. As a result, NVIDIA, the flagship name in AI accelerators, achieved the historic feat of becoming the first company in the world to reach a market capitalization of $4.4 trillion.

Coincidentally, that $4.4 trillion market cap is very close to the $4.5 trillion valuation I had assessed as reasonable earlier this year in the AWARE Private Community channel. We do not currently hold NVIDIA in the Model Portfolio, but we did at the start of the year, so there was intense interest among AWARE members in what constituted fair value for NVIDIA.

The Model Portfolio first added NVIDIA on September 25, 2023, buying at $42.33 per share (split-adjusted). Most people probably don’t remember, but even back then, when the price was less than a quarter of where it is now, many argued that NVIDIA was overvalued relative to its earnings, pointing to its high P/E ratio (a debate that became moot as earnings exploded shortly thereafter). There were also persistent concerns that the surging demand for major LLM (large language model) services led by ChatGPT was only temporary and would soon plateau.

Nevertheless, over the past two years or so, US tech giants commonly grouped under “Big Tech” — Meta, Microsoft, Google, and others — have accelerated their CAPEX spending on AI accelerators and data center expansion, at a pace that has already exceeded their originally ambitious plans. Naturally, NVIDIA, which dominates the AI accelerator market and represents the single largest line item in AI data center CAPEX, has seen its earnings grow in lockstep with this rising spend. TSMC’s most advanced 2-nanometer process capacity is reportedly fully pre-booked, leaving no room for new customers.

An analyst at Exante Data estimated the impact of AI spending on US GDP based on NVIDIA’s data center revenue. AI spending, which accounted for about 0.2% of GDP in 2022, has risen to 2.0% in 2025.

While it is difficult to take these estimates at face value (as the author himself stresses), they are very helpful in identifying the trend. The share has grown tenfold in just three years. The contribution of AI spending to GDP growth was also estimated, and if we take the 0.7% contribution in 2025 at its nominal value, we arrive at the conclusion that the US economy would have contracted this year were it not for AI.

CAPEX spending on AI data centers now accounts for 1.2% of US GDP, surpassing the telecom bubble of 2020 driven by 5G commercialization. It is also estimated to be on par with the ratio seen during the dot-com bubble of the early 2000s, when internet cables were being laid nationwide. The only historical episode where even more was spent was the railroad construction boom of the 1880s, when the discovery of large gold deposits in California triggered the Gold Rush and opened the era of full-scale westward expansion, spurring rapid construction of transcontinental railroads. Most of the rail network that exists in the US today was built between 1880 and 1900, which is also why US railroads are notorious for being slow and outdated. The key point is that rail infrastructure is still in use more than 100 years after it was built. By contrast, the cutting-edge AI data centers being built today are highly likely to have no practical use 100 years from now. In accounting terms, their depreciation period is short — in other words, the payback period on the investment is short.

In DCF (discounted cash flow) models used to value companies, the perpetual growth rate cannot exceed GDP growth; it is physically impossible. The perpetual growth rate assumes (1) the company exists forever and (2) generates cash flows over that entire period. If a company were to grow faster than GDP indefinitely, it would eventually become larger than the economy itself. Since GDP is calculated as the sum of all economic activity, a single company cannot exceed the total.

What this thought experiment is meant to convey is that NVIDIA, which accounts for roughly 25–35% of AI data center investment and related spending, has become extraordinarily large. The fact that NVIDIA (NVDA) is the world’s largest listed company by market capitalization makes this intuitively obvious. Based on fiscal 2022, the company’s revenue was $26.9 billion, and by fiscal 2025 it had grown to $130.4 billion, roughly a 3.8-fold increase. If we assume that the share of AI data center spending in the economy doubles from here to 4.0% by 2028, and apply the same ratio to NVIDIA’s growth, the company’s revenue three years from now would be $185.1 billion, up 42%. Those are strong numbers, but the growth rate slows sharply compared with the past three years: on a compound annual growth rate (CAGR) basis, the last three years come to 68.7% per year, while the next three years are 12.4% per year.

One could slice scenarios extremely finely—assuming, for example, that AI’s share of the economy grows to 20%, or tweaking Nvidia’s share of AI data center spending up and down—to build a very detailed forecasting model, but I don’t think that’s what really matters. What is clear is that, because of physical constraints, the probability that Nvidia can keep its current revenue growth rate falls exponentially the longer that growth is sustained.

A party is a gathering held to celebrate something good or to mark an event. It’s formed under a positive momentum that breaks up the continuity of everyday life, and you can think of the moment when everyone raises a glass for a toast as the climax of that momentum. But the fact that it’s time to go home when the party ends doesn’t mean that the positive momentum that brought us there simply disappears.

When acceleration is constant, velocity converges toward infinity over time. The problem is that nothing can travel faster than light, which means acceleration cannot actually remain constant as in the graph. Physicists resolved this through the theory of relativity, but since this piece is not meant to be a physics lecture, I won’t go down that rabbit hole.

In our economic reality, there comes a moment for every company when its acceleration converges to zero. Because acceleration is the derivative of velocity, once acceleration hits zero, velocity no longer changes. The theoretical reason we invest is to earn risk-adjusted returns in excess of the risk-free rate. Let’s borrow a physics analogy and define the pursuit of excess returns as the search for assets with positive acceleration. How would you assess Nvidia’s acceleration?

As mentioned earlier, the end of the party doesn’t mean the momentum is gone. It simply means the plane has finished taking off and has reached the phase of cruising at high altitude.