As OpenAI’s ChatGPT went mainstream in 2022, the AI trade began in earnest. Nvidia was at the center of all the action. Nvidia’s GPUs (Game Processing Units) were tailor-made for LLM (Large Language Model) training. The proverbial gold rush was on as more players joined the LLM race. However, as is initially the case, the biggest beneficiaries were Nvidia and its ecosystem partners - the shovel and wheelbarrow makers - racks and server makers like DELL, SMCI, ASUS, GigaByte, etc., their suppliers like Vertiv, Micron, etc., and obviously TSMC. The list is illustrative, but you get the drift.

As the AI race went global, by 2023, the Biden Administration became concerned about China gaining an upper hand. Sanctions and export controls were imposed on high-end Nvidia chips. Interestingly, by the end of 2023, Nvidia’s shipments to Singapore skyrocketed. By early 2024, rumblings — among a small group of market watchers — about Nvidia bypassing sanctions by using Singapore as a jump-off point for moving GPUs to China.

In late 2024 and early 2025, I was informed — couldn’t verify — that Nvidia’s GPUs had been supplied to the largest companies in China, and the military. I was also told that US customs officials were on the ground in Shenzen working alongside Chinese customs to stem the flow of Nvidia GPUs into China. Suddenly, Nvidia GPUs hit marketplaces like eBay, Facebook, and made-in-china[dot]com, available at pennies on the dollar, ready to be shipped at a moment’s notice from suppliers in Shenzen and Guangdong. It was apparent that shenanigans were going on.

SMCI was caught up in investigations after Hindenburg Research released a report on SMCI in August 2024 — I had released a similar report in May 2024 — but I figured Nvidia had become too big to investigate and prosecute. Well, long story short, Nvidia’s sales to China are now completely gone; zero, zip, nada. However, rumblings persist over whether Nvidia is now using other Southeast Asian countries as a jump-off point to supply China. That said, China has raced ahead. Whether China reverse-engineered or innovated is a moot point; it now has its own chips.

It was ho-hom in the AI space in early 2025, relatively speaking, and then one LLM use case took off. Code generation. As OpenAI and Anthropic vied for the pole position, software engineers. consulting companies, IT departments, and tech companies started using LLMs for code generation. Demand for tokens went ballistic. Suddenly, there is not enough compute available.

To enable AI, one needs data centers. That means huge buildings, humongous amounts of electricity and water, and, goes without saying, the whole semiconductor supply chain.

The physical requirements, such as land, buildings, water, and electricity, require permits and local approvals. The biggest bottleneck for data center buildout now is the lack of electricity availability. The grid is just not ready to handle the electricity requirements of a data center, as a single data center requires electricity similar to that used by cities like San Diego or Miami. Those on the ground and in the know knew that the electricity grid capacity was a bottleneck about to hit the data center space and were quite vociferous about it. I have covered this on this blog and have spoken about it on social media since 2025, and recently wrote about it while covering GE Vernova.

That said, the projected need for electricity led to the hunt for trades in that space in the second half of 2025. First, gas turbine makers like GE Vernova (GEV) were chased as they got sold out well into 2028. The other beneficiaries were Caterpillar (CAT) - yes, we need to move a lot of earth, and they also have a power division - and infrastructure solutions providers like Quanta Services (PWR). Then, as people realized the grid capacity would take time to build out, the hunt and chase began for on-site power generation equipment and solutions providers like Bloom Energy (BE).

As the electricity bottleneck hit, it led to more than half of the data centers scheduled to be built in 2026 being delayed or canceled. That had a few second-order effects. Prices for existing compute rose sharply, benefiting the likes of CoreWeave (CRWV). Also, some of the already procured Nvidia GPUs — a depreciating asset — are now sitting in warehouses or empty buildings gathering dust.

As the realities of physical constraints were playing out, people started looking into other potential bottlenecks. It showed up everywhere in the semiconductor supply chain. Memory, interconnect, optics, storage, packaging, you name it. Stocks in these spaces, like Lumentum (LITE), Micron (MU), and Sandisk (SNDK), to name a few, went parabolic, generating enormous wealth for shareholders within six to seven months.

This hunt went global; obviously, the semiconductor supply chain is global. South Korean stocks like Samsung and SK Hynix joined the party. And now that Interactive Brokers (IBKR) has enabled access to South Korea, retail investors have piled on. KOSPI is up over 5% as I write this (05/04/2026); Samsung and SK Hynix are over 50% of the index. Large social media accounts have started promoting small and microcap semiconductor-related stocks in South Korea and Japan.

How can memes be far behind when all of this is going on? Toto, a high-end toilet maker in Japan, went parabolic, gaining over 10% each day for days. The shitco references were obvious on the surface; however, to be fair, 50% of Toto’s operating profit comes from electrostatic chucks (ESCs), which are essential components in the production of 3D NAND flash memory chips. Not to forget, Toto is the second largest manufacturer of ESCs globally.

I only hope investors are aware of the risks of getting into illiquid small-caps and microcaps in emerging markets. Once the tide turns, there would be no exit liquidity.

The craziness, if one were to call it such, or euphoria, has given us a blast from the past of the dot-com era. On April 15, Allbirds, a struggling footwear and apparel provider, sold off its core business and brands for $50 million and announced it was pivoting to the AI data center business. BIRD was trading at 2.50. It spiked to a high of 24.31 and has now settled at 5.88.

And how can I forget the old dog Intel (INTC)? Intel has finally managed to cross its all-time high made in 2000. The primary driver was the need for more CPUs as AI inference goes mainstream. Simplistically, AI-training’s fundamental computation engine is matrix multiplication for which GPUs are best suited. Meanwhile, inference requires more orchestration for which CPUs are best suited. For training, the CPU-to-GPU ratio was 1:8, whereas for inference, the CPU-to-GPU ratio is currently expected to be 1:1, if not higher. What beats me is that it took so long for the markets to figure this out. I guess nobody at the big shops bothered to talk to an engineer.

Another beneficiary of inference is memory. Much more will be required as time goes by. If that were not enough, due to overwhelming demand, memory players like Micron and SK Hynix have gained long-term pricing power. Memory orders were generally quarter-to-quarter or one-year contracts. As shortages hit, spot prices have more than doubled and are expected to move higher. The bigger impact on PnL for Micron and SK Hynix is the pricing leverage in long-term contracts, which now include stiff penalties for breaking contracts.

Now Sandisk (SNDK) has joined the party. Sandisk’s CEO said the company has signed five long-term contracts, three of them worth $42 billion, and "if they walk away from a contract, I get that money."

I continue to believe, as I explained in April, the semiconductor industry will remain cyclical, and I am not sold on the hype. The market has gotten way ahead of itself to the point that the semiconductor space is in a bubble. However, if you disagree with my opinion, you may look at the individual names I have highlighted above or play the theme via the following ETFs:

It would be prudent to take into account the following quote from the CEO of Micron:

I am waiting for the “software eats the world” moment to arrive. A great opportunity is opening up for software entrepreneurs in the coming years. Let the incumbents cannibalize each other. Let "SaaS Apocalypse" play out. Better to be a Google than Alta Vista, Yahoo, or Netscape.

That is a humongous amount of CAPEX — almost 2% of GDP — coming from just five names. It doesn’t include other hyperscalers or data center build-outs by other players in other parts of the world. The direct beneficiaries of this CAPEX are the bottleneck names discussed above.

Morgan Stanley, in a recent report, estimates $851 billion (revised up from the actual values in the graph above) and $1,116 billion in CAPEX from these five names for 2026 and 2027, respectively. That is $1.97 trillion over two years. Mind-blowing to say the least. However, there is a funding gap of ~1.5 trillion that will have to be financed by someone else. Most likely via debt.

For years, the success of these companies rested on a simple premise: capital-light, almost infinitely scalable businesses capable of generating massive free cash flow while still funding growth, new projects, and, eventually, shareholder returns.

That premise is now being challenged.

If the core profile of these businesses is shifting from asset-light platforms to something closer to industrial infrastructure companies, the market should not value them using historical multiples. The real risk is not that AI fails, but that these companies end up with structurally lower free cash flow conversion, shorter asset lives, and a permanently heavier capital base. This, as we shall see later below, has started happening.

The best thing that can happen to these hyperscalers is that algorithmic improvements may eventually outpace the need for ever-growing computational power. I strongly believe that is going to happen. People have the right to scoff at me; however, scoff at Sergey Brin at your own peril.

Any slowdown in CAPEX due to lower compute demands would bludgeon the ongoing rally in the bottleneck trades mentioned above.

Multiple funding techniques — also common in other industries — have been used to fund the AI CAPEX. The following are the four common ways it is being done:

Optically, the circular nature of a lot of this looks bad. However, as I said earlier, these funding methods are nothing new. Yes, if OpenAI, whose future revenues are used to secure a lot of this CAPEX, goes kaput, there might be short-term issues. However, these companies have other customers and internal revenue-generating projects to fund this CAPEX.

For instance, Anthropic reported that it has reached $44 billion in ARR. That in and of itself is amazing. The question remains whether it is sustainable, as these companies have no moat, can be disrupted within a few months by a competitor with a better model, and there are no switching costs. These facts raise the question of whether these commodity companies, such as Anthropic, OpenAI, etc., will have pricing power and turn profitable.

The reason I am calling them commodity companies is that they are trained on the same underlying data scraped from the internet and from scanned books. Eventually, the models will converge when neither one is any better. Proprietary data is the moat, and that differentiation is what the market has yet to make among companies supposedly getting disrupted by LLMs.

I wanted to take a moment to explain the basics of how CAPEX is accounted for in the three financial statements — income statement, balance sheet, and cash flow statement — and the impact on earnings over time.

Say I have to buy some equipment, which costs $120. Here is what happens in the books when I buy the equipment:

So, in the case of hyperscalers, even though hundreds of billions of cash is being spent now, the expense is spread out over six years, which is fine under GAAP. The issue starts happening when the CAPEX is regular and on short-lived assets. Depreciation and amortization expenses start climbing, and if debt is used, interest expense starts rising; thus, put together, it starts impacting margins and net income unless revenue is able to outpace the growth in these expenses.

As I said earlier, as this whole AI CAPEX thing plays out as it stands today, the hyperscalers will start resembling industrial companies and will have to be valued at lower multiples, i.e., stock prices have to fall. This has already happened with Oracle, and further derating is expected.

Now, there are two ways to keep the expense down, to keep earnings elevated, to keep the markets happy:

Last week, apart from Oracle, the four other large hyperscalers reported great earnings. Google’s earnings were celebrated, with the stock rising 10% as it reported a record $62.6 billion in net income. However, there was a catch. Half of the net income came from gains in non-marketable equities that Alphabet holds. These are stakes in Anthropic, SpaceX, and other ventures. There is nothing wrong with this, as it is reported as required under GAAP.

Of late, Other Income has been inflating hyperscalers’ Net Income, thus EPS, and it should be noted that most of this income is vaporware.

One of the quickest ways to judge the quality of earnings is to look at the Cash Flow Statement. If Cash from Operations is less than Net Income, the earnings are generally of poor quality, and in more egregious cases, could be a sign of accounting fraud.

Let’s look at how the five hyperscalers stack up on this front and FCF in Q1 2026.

There is a lot more that goes into analyzing a company, and the exercise here is not to do that for each one. The point here is to show the deterioration in FCF at some of the companies and the possibility of these companies going extremely FCF negative in future years if investments in AI don’t pan out. Also, in Q1 2026, the combined FCF of these five companies has fallen 32% Year-over-Year.

The big tell will be how analysts start valuing these companies. Historical metrics can’t be used. If they start using EV/EBITDA multiples, then we know the time has come to assign lower multiples to these companies. I have a feeling the timing of that might be sooner than what most people have penciled in right now.

I don’t like using EBITDA for various reasons, but I have to play along, as many or most are using it. My personal reasons aside, this is how Charlie Munger defined EBITDA.

Even though I love catching shennanighans at companies, I am bullish by nature; most certainly at the index (S&P500) level. I stated the same last month. Even though I don’t let macro events dictate my long-term holdings, there are times when I do raise cash levels. Now is one of those times.

The Strait of Hormuz “situation” has lasted longer than I thought it would. Global stocks of crude, gasoline, diesel, and jet fuel will hit critically low levels by the end of May, at which point prices will go parabolic. First, the spike; then crude will crash along with all asset classes as a recession hits. The following chart, even though from the 1970s, does make one wonder how the current situation will play out. The variables might be different today, but history sometimes just rhymes.

Two stocks I have written negatively about are reporting this week.

As is customary, I would request you not to gamble with short-term options. I can only request, at the end of the day, what you do is your choice.

Until next time, have fun!!!