My thoughts are this. I don’t know what the costs of building out AI are but they are extraordinarily huge. I don’t know what the technical and energy constraints are. I definitely don’t know what the financial returns will be, given the competition. Obviously it’s a useful technology but from an investment perspective it’s way too hard for me. Intuitively it all seems like a lot of resources compared to the human brain which is capable of much higher levels of intelligence fuelled by a cup of coffee a croissant and eight hours sleep. I keep coming back to the distinction between an efficiency tool versus actual intelligence. It’s definitely a game changing efficiency too and it’s definitely expensive.

The massive hyperscaler capex numbers (creeping toward $800B this year) are incredible. There is plenty of debate on the accounting depreciation tricks, like tech firms extending server life from 4 to 6 years on paper.
The incentives are obvious. Management has every reason to manipulate these accounting assumptions. Extending the useful life of a server lowers immediate depreciation expenses, which artificially pumps up reported operating income. Those inflated numbers keep the stock price high, allows them to raise cheaper capital to fund even more capex, and ensures insiders get maximum value when cashing out their Stock-Based Compensation. History shows that whenever there is subjective room in accounting rules combined with massive financial incentives, corporate behavior is completely predictable.

I imagine these physical assets face rapid economic degradation long before their paper accounting life wraps up.

The Law of Diminishing Returns. Pre-training scaling laws are slamming into a wall. We are burning exponentially more power and capital to get marginal, flattening gains in benchmark intelligence. If it takes a gigawatt cluster to move the needle 2%, the cost per useful output becomes prohibitively expensive.

Accelerated Obsolescence. AI servers lose value because of the opportunity cost of power. Since grid capacity is a hard ceiling, keeping older, power-hungry chips plugged in is economically irrational when new silicon offers 4x the compute per watt. Furthermore, modern AI densities are forcing a massive, expensive rip and replace transition from legacy air cooling to direct-to-chip liquid cooling infrastructure.
Ultimately, we are spending billions on AI factories that will be obsolete in 24 to 36 months, regardless of what the audited footnotes say. Unless these models transform into low-energy autonomous systems that actually generate cash returns, this capital cycle could turn out to be very disappointing.

Maybe someone a Berkshire understands it all but I also wonder if Buffett has made the Alphabet investment as a hedge against AI destroying many of Berkshire’s other businesses, as opposed to it being a high conviction investment that will give a great return. Buffett didn’t sound overly confident in the investment. I wonder if the Occidental investment is a similar hedge in that case against rampant inflation.

I know nothing about all of this but these are my idiotic thoughts regardless. It’s an interesting and important topic and I hope I learn more about it.