Why I'm Not Buying AI Stocks
And how to make money when mania grips the world.
I found a way to access a Bloomberg-level terminal for 90% less. Here’s how →
This isn’t an idea, for a change.
It’s an analysis of the current AI boom, from the perspective of a business owner.
Deep-value isn’t really about buying stocks that are cheap.
It’s about assessing the downside risk above anything else, and buying hard when that risk is virtually zero.
It’s about rational capital allocation.
This is purest when using your own hard-earned money, and maybe even debt that has real-world consequences if you fail to pay it back, with interest.
Because, when you buy things this way, you think long and hard about the risk of things going wrong.
But first, a story…
A few weeks ago we took the kids to a canal museum.
Contrary to what you’re almost certainly thinking, it was actually quite interesting.
Most importantly, it was free, and my kids are not old enough to know any better.
Anyway, in the UK we have an extensive canal system.
It’s just always been ‘there’ and I never really questioned it.
But going around the museum taught me some interesting things.
For example, one of the first canals was the Bridgewater Canal in Manchester.
It was built by a guy called Francis Egerton to transport coal from his mines, and reduce costs.
The project was an incredible success.
Coal prices literally dropped 50% as a direct result.
The savings allowed them to maintain margins but massively expand volume, while customers got cheaper coal.
Many other famous industrialists of the time copied the model and leaned into the canal boom.
These included brands like Wedgewood (pottery) and Cadbury (chocolate).
For those guys, the investment was almost always a raging success.
This was simply down to the massive increase in efficiency.
For instance, on a road, a horse can pull around one ton of goods at a time.
When pulling a canal barge, that same horse could pull more like 50 tons of cargo at a time.
We don’t need my big-buttoned calculator to figure out what this means for the bottom-line.
All in all, the UK built 4,000 miles of canals, all over the country.
It was an integral part of the UK industrial revolution.
Canal Mania
Investors soon piled into this boom.
In 1790, parliament authorised one new canal to be built.
By 1793, it authorised over twenty new canals.
The appeal was obvious.
Build a canal and charge businesses to transport their goods along them, and receive fat dividends forever.
Simple.
Sadly, it didn’t quite work out for most of those investors, because other people figured out how to build railways.
Trains were just as stable and reliable, but much faster, and operated all year round.
Within 40 years, the canals were on the brink of death and railways started buying them up just to totally eliminate the competition.
It seems weird that there would be a ‘canal mania’ but that’s exactly what happened.
If I was around in 1795, watching all these businesses generate massive gains, I would most likely have become interested in canals.
Business people (generally) made lots of money, and then switched to railways, and made even more money.
The UK economy experienced a very similar benefit.
Almost everyone else, basically lost money, because they were trying to ride a wave that flattened out and died.
And that brings me back to AI.
The Scale Of Spending
The adoption of generative AI has been astonishing.
ChatGPT reached 100 million users in roughly two months and achieved a level of adoption that appears faster than almost any previous technology.
The demand is obviously real.
However, unlike many internet businesses that came before it, AI requires vast amounts of physical infrastructure.
Every query requires computing power.
Every image generated requires computing power.
Every video generated requires computing power.
The result is an enormous global construction programme centred around data centres.
Recent estimates suggest that building a single 1GW AI data centre costs approximately $38bn.
Around 55% of that cost is spent on servers alone.
This means the AI boom is not simply a software story.
It is a story involving land, steel, electricity, water, transformers, networking equipment, memory chips and specialised processors.
Unlike traditional software businesses where marginal costs are tiny, AI requires continuous investment simply to keep serving growing demand.
This makes future extrapolations tricky, to say the least.
Inference Is The Key
Much attention has been given to the cost of training large language models.
Training a new foundational model reportedly costs between $50m and $200m and can take several months.
These figures sound enormous, but in reality they are relatively small compared to the scale of the big tech companies.
The far bigger issue is inference.
Inference is what happens every time a user interacts with an AI model.
Every question.
Every response.
Every image.
Every video.
Unlike training, which happens once, inference-costs scale directly with usage.
As adoption increases, the infrastructure required to support that activity expands with it.
This is why capital expenditure plans continue to rise so aggressively.
The real spending is no longer about creating models.
It is about building the infrastructure required to run them at scale.
The Economics Look Difficult
The question that concerns me most is simple:
“How much revenue will all this infrastructure actually generate?”
The answer is surprisingly difficult to justify.
Capex at Microsoft, Alphabet, Amazon, Meta and Oracle is expected to exceed $600b annually by 2026 and remain around that level for several years.
To justify those investments using current economics requires an extraordinary increase in future revenues.
I saw one estimate suggesting that maintaining current returns on invested capital would require roughly $413bn of additional annual revenue across those companies.
That equates to approximately $50 of additional spending for every person on the planet.
Maybe that will happen.
Maybe it won’t.
What I find most interesting is how the investment community seems to discuss the spending itself rather than the economic returns it must eventually generate.
To me, or any business owner, this is the only question worth asking.
Forecast Optimism
One of the most interesting analyses I have seen examined projected capex and projected future-revenues through 2030.
The conclusion was a bit worrying.
Even under highly optimistic assumptions, where future revenues flow almost entirely through to profits, most of the major AI infrastructure investors still struggle to generate attractive returns on the capital being deployed.
That analysis does not prove the investments are wrong.
But it does highlight how dependent the entire thesis is on future outcomes that remain highly uncertain.
The current spending is clearly not justified by near-term revenue expectations.
The only credible justification is that these assets will generate substantial profits many years into the future.
Investors are effectively underwriting a future that does not yet exist, and has shown very little signs of existing so far.
Infrastructure Monopolies
My little trip to the canal museum provides a useful framework here.
During Britain’s canal mania, investors poured huge amounts of money into canal construction.
The underlying economic logic appeared sound.
The demand for transporting goods was real.
The infrastructure was useful.
The total addressable market was enormous.
Yet many investors still earned disappointing returns.
Some canals proved valuable.
Many did not.
Then railways arrived and changed everything.
Useful infrastructure does not guarantee monopoly profits.
I see obvious parallels with AI infrastructure today.
Investors are assuming that massive spending on data centres will create enormous economic value.
It probably will.
But history suggests that building infrastructure and earning monopoly profits are two very different things.
Asset Lifetime
Another concern (for me) is asset longevity.
Much of the cost of AI data centres is concentrated in servers and computing equipment.
I saw some other analysis that suggests these assets may only have useful lives of three to five years.
That feels like an impossibly short period in which to earn back the original investment.
Land and buildings may endure.
Grid connections may endure.
But the most expensive components apparently depreciate rapidly.
This creates a very different economic profile from traditional infrastructure.
Investors are effectively betting that today’s spending can be recovered before the next generation of technology makes current hardware obsolete.
This is another tricky assumption, given the sheer pace of the advancement.
AI Is Still Being Subsidised
Another concern I have is that many AI services remain heavily subsidised.
The objective appears to be acquiring users and market share rather than generating profits.
OpenAI’s video platform Sora reportedly generated lifetime revenue of around $2.1m while costing approximately $15m per day in inference expenses.
The economics were so unfavourable that the service was recently shut down.
It’s a reminder that technological capability and economic viability are not the same thing.
Many AI applications are impressive.
The harder question (which no one seems to be worrying about) is whether they can ever become sufficiently profitable.
Cheaper Alternatives May Emerge
The bullish case for AI infrastructure rests on the idea that demand for intelligence is effectively unlimited.
Maybe that is true.
But even if demand is unlimited, that does not mean demand for expensive infrastructure is unlimited.
Many tasks can potentially be solved using smaller, specialised models that require only a fraction of the computing power.
Another piece of analysis I read involved podcast transcription for a company called Overcast.
Using a large commercial AI model would reportedly have cost them approximately $30m per year.
Using a specialised smaller model apparently reduced the cost to roughly $10,000.
The productivity gain was achieved either way.
The expensive infrastructure was not necessary.
This is another thing that would keep me awake at night (if I was betting my money on AI infrastructure).
The Financing Structure
The financing arrangements behind the AI boom are becoming increasingly complex.
We’ve all seen the ‘circle-jerk’ memes on social media.
Large technology companies are funding construction from operating cash flow.
However, substantial amounts of debt and private credit are also entering the system.
Special purpose vehicles are becoming common.
Suppliers are financing customers.
Capital is moving through increasingly complicated structures.
Leverage, opaque financing structures and circular funding arrangements have appeared repeatedly throughout financial history during periods of speculative excess.
When I see risk being shifted around the financial system in creative ways, I become more cautious rather than less.
Useful Isn’t Necessarily Profitable
This is really my ‘bottom-line’ concern.
I have no doubt that AI will be useful.
I have no doubt that automation will continue expanding.
I have no doubt that important new businesses will emerge.
But usefulness does not automatically translate into attractive shareholder returns.
The current market narrative often assumes that because AI is transformative, the infrastructure being built today must therefore be enormously valuable.
History suggests that conclusion is far from certain.
Canals were useful.
Railways were useful.
The internet was useful.
Yet many investors who financed those booms earned disappointing returns.
The winners often emerged later and in unexpected places.
The Missing Downside Protection
The core issue is not whether AI changes the world.
The core issue is whether today’s infrastructure spending can generate returns that justify today’s valuations.
At present, I struggle to answer that question confidently.
What I see is:
• Massive capital expenditure.
• Extremely optimistic assumptions about future demand.
• Short-lived hardware assets.
• Subsidised economics.
• Complex financing structures.
• Returns on invested capital that already look difficult to justify.
That does not mean AI is a fraud.
It does not mean the technology will fail.
It simply means that investors may be confusing technological success with investment success.
History repeatedly shows that the two are not the same thing.
My concern is that the infrastructure boom currently underway may eventually look more like the canal mania than investors realise:
Useful, transformative and economically important, but incapable of generating the monopoly profits required to justify the capital invested.
For a deep-value, business-minded investor like me, there is just far too much uncertainty and not enough downside protection to get involved at this point.
If, like me, you prefer investments with full downside protection and a massive margin of safety, you might like my newsletter.
An annual subscription also gets you access to exclusive bonus content that you might also enjoy.
I found a way to access a Bloomberg-level terminal for 90% less. Here’s how →