I find that AI defies almost every investment technique I have ever developed. When DeepSeek was released last year, I thought we would see some correction in pricing in Western AI relative to Chinese AI. That is pricing for Nvidia/TSMC chips would fall back towards SMIC levels (SMIC makes chips for DeepSeek). Contrary to my expectations, the pricing differential has continued to widen.

I then thought the financing troubles of Oracle might also signal signs of a rethink on AI spend. Previous spikes in Oracle CDS in 2022 and 2008, were years of weak markets. The market has happily ignored Oracle weakness.

But these potentially negative signals have not worked. If anything, the AI boom seems to be broadening out. We have now seen with the recent surge in DRAM and Nand prices. Or in other words, the pricing of commodity semiconductors are starting to close the gap with the cutting edge Nvidia ones.

The broadening out of the AI boom has also including the share prices of electricity generating equipment producers. Doosan Enerbility is one such company - and the share price has returned back to the peak of the China boom.

2008 was the time to get rid of your Doosan shares - as this is a VERY cyclical industry. And back in 2008, China was the real driver of electricity generating growth. China did not stop adding generating capacity, but it did grow at a slower rate, from 2008 onwards, and Doosan fell 99% from peak to trough. When I look at Oracle, Nand pricing and Doosan, I do get flashbacks of 2008.

If that is the bear case for the AI boom, then this is the bull case. The “idea” of a modern economy has been set in stone since the 1960s or 1970s. Access to constant electricity, a working road network, and various other mod-cons. The US was the first to achieve this modern society, and basically the world has been striving to catch up. Mexico, which is a neighbour and shares many features with the US, was dirt poor in the 1960s, and now inching closer to developed status, we can see its development in its use of electricity. I have used log scale for ease.

What you can see is the Mexican growth in energy consumption skyrocketed in the 1960s and 1970s as it developed, and then slowed from 1980s, to stagnation since 2010. The US growth has really seen stagnating growth from 2000. What I am saying, is that in essence, we had reached a lifestyle that did not require more energy (in fact on a per capita basis, energy consumption was falling). Whenever you hear that the US has not built a new refinery, or new power station for 40 years - this is not market failure. This is the market saying that the outlook for power demand did not justify the investment. Obviously this is starting to change with AI. EIA predicts growing demand and supply going forward.

The trillion dollar question then, is how many data centres do we need? And how much energy would this consume? If we look at data centres, plainly the US is in the lead. But what we are seeing is that US demand is now constrained by its ability to produce electricity. China has less data centres, but obviously not constrained by energy capacity.

Ask ChatGPT how many data centres does the US need and you get an estimate ranging from 100GW to 500GW - so large multiples from what we have today. One of the big problems with this is that the US power grid is constrained by connection capacity. That is, data centre supply is constrained. When supply is constrained, companies tend to over pay and over order. The reason for this is by making them themselves the biggest and most important customer, they are hoping to jump up the queue in getting supplied. This leads to a negative doom loop when supply catches up, as double orders get cancelled. This is why these industries are considered cyclical. That being said, power generation equipment supplier, GE Vernova has powered to new highs even with oil price doubling.

I have started to think about semiconductors in a different way. If the post war growth model was built around cars and fossil fuel usage, then this growth was only really constrained in the 1970s, when energy prices surged. Using a Bloomberg analysis, energy costs stayed below 4%of GDP until the 1970s, and then surged to 12%, and this ushered in the era of Washington consensus, and age of energy efficiency. That is total power generation stayed constant even as GDP grew.

Lets say we are in an AI age, where the limiting resource is now semiconductors. Can we make a graph like above? Well I am going to try. This uses data from Semiconductor Industry Association. Using a few estimates, I get to a 1% spend of world GDP on semiconductors in 2026, from a 0.4% in 2000s. So the build out of personal computer through the 1980s and 1990s drove the first bull market in semis, which then stabilised, and now we are in a new era of AI. The big bullish argument for semiconductors is that AI is going to permanently increase semiconductor demand going forward, as more and more of our activities require AI and data centres. Could we be spending 3% of GDP on semiconductors going forward? 5% to 6% on energy is normal, and 3% of defence is considered the going rate. Are semiconductors now the life blood of the modern world? Probably, would be my guess. The question is when does supply catch up.

Backward looking, mean reversion believers would tell you that we are back at dot com peaks, and we are going to see a crash. But having seen the use of AI in the battlefield in both Ukraine and now Iran, and the various AI related lay offs been conducted by tech companies, AI seems the real deal to me. As also pointed out before for companies like Microsoft, AI is now a life and death issue, so they will keep spending, and this will likely keep everyone else spending too. If I look at Nvidia/TSMC chips, NAND and DRAM, they all seem to reflect this reality. But with the semiconductor stack, I find areas that are not pricing in this future yet. I quite like the wafer business, where pricing has been weak due to slow smartphone and PC volumes, but an increasing share of wafers are now going to AI products. This reminds me of the DRAM set up - but before prices skyrocketed. That is pricing is discouraging capacity expansion in wafers, even though pricing is encouraging capacity expansion everywhere else in the tech stack. I like that risk/reward.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.russell-clark.com/subscribe