Despite its lukewarm adoption to this day (except for non-GenAI AI applications, such as search or social media algorithms), AI’s role in the economy has grown unprecedentedly in recent months by masking the miseries of what’s otherwise a clearly recessionary global economy.

With Germany on a two-year recessionary run and poised for a third, and France’s politicians unwilling to make the necessary reforms to save its economy from an inevitable debt explosion and subsequent bailout, Europe appears to be a lost cause. Moreover, China is still seeing massive deflation and stagnant internal consumption.

However, the US isn’t in a sweet spot either. According to estimates, without AI investment, GDP growth in the last quarter would have been flat. And just in the last few days:

Therefore, whether you believe the AI buildout is worth it or not for incumbents is one story; the fact that it is “saving” the economy is another (at least the metrics).

Hence, many investors, including yours truly, believe that the market’s capacity to withhold the apparent signs of overextension of valuations (with the Buffett indicator, his preferred way to evaluate how pricey a market is, a simple percentage measurement of the value of the S&P500 vs the US economy, being at historic highs of 217%), hinges in whether Hyperscalers can continue to keep the party going via brute AI spending.

But will they?

The short answer is that, for now, Hyperscalers (Meta, Google, Amazon, Microsoft, and Oracle) are willing to keep the lights on for a little bit longer and are expected to not only continue investing but grow their AI Capex to half a trillion in 2026, as per a Goldman Sachs report:

All but Oracle (which is financing most of this via debt—for a deep dive into Oracle’s big bet, read here) are expected to surpass $100 billion in capital expenditures for 2026.

Just to put these numbers into perspective:

But how is no one panicking? Two example data points strengthen the narrative: Oracle’s RPO for the following years (the amount of booked revenue they have) is almost $500 billion, and Microsoft has another $400 billion; in total, nearly a trillion dollars is booked between the two.

Consequently, what makes the “AI bubble” different is that, this time, it’s being financed by the wealthiest corporations on the planet’s free cash flow (and debt). Except for Oracle, it’s not as if they are borrowing money to do this (although they are increasingly doing so, especially Meta); they actually have the cash.

However, here’s the thing: willingness and feasibility are distinct concepts. And this difference matters a lot in AI.

When people discuss these companies and their Chinese counterparts “investing in AI”, what values are we discussing, and what does it really mean?

Here’s a quick rundown of the deals announced just in the last six months in the US, the Middle East, and South Korea:

Which leads us to the question: What does it mean to ‘invest in AI’?

While I would need three additional newsletters to explain the AI supply chain in detail, and I would still fall short, here’s a quick rundown of the different phases:

So when discussing ‘AI CapEx’ beyond the money, this is the physical world of ‘AI investing’, and here’s where the real constraints arise.

Therefore, we will examine the physical constraints from two perspectives: energy and manufacturing. The first one is what everyone, including incumbents, tirelessly points to, but the latter may be even more bottlenecked.

The main energy issue is that we aren’t adding nearly enough power as we need right now, compared to the demand projections we are seeing based on the announcements.

Data centers these days are measured in gigawatts. That means a single site requires more than 1 billion watts per hour to work. But how much power is that in terms of AI throughput? Let’s answer this by computing how much one of these data centers would take to train GPT-4.

Most are being built around NVIDIA’s GB300 NVL 72 servers, a rack-scale AI server that houses 72 GPUs (or 144 GPU chiplets because each Blackwell GPU has two GPU dies) and requires between 120 and 142 kW of power to work, depending on what OEM (the infrastructure companies that build the server around NVIDIA’s chips) you ask.

Thus, a 1 gigawatt data center can house around 7,600 of these servers, each with 720 PFLOPs of compute throughput at FP8 precision. That means the total compute power of this ‘AI factory’ is 5,472,000 PFLOPs, or 5.5 sextillion operations per second, or 5.5×10²¹ operations.

To put that number into perspective, assuming no issues (reality is a little bit more complex), this cluster would take one hour to train GPT-4 (2×10²⁵ total training FLOPs divided by 5.5×10²¹ FLOP/s = 3636 seconds).

Yes, one hour to train the state-of-the-art model back in 2023 (and it was SOTA for more than a year afterward), and required more than three months of training.

But the biggest issue here is not the AI racks themselves; it’s all that comes before, which really puts into question whether this is actually feasible.