The More AI Succeeds, the More Dangerous It Is to Own GPUs

AI might really change the world, but the first people to buy GPUs may be the first ones stuck paying the industry's bill.

This is not an argument that AI doesn't work.

Quite the opposite — what I'm actually worried about is this: AI truly succeeds, demand truly explodes, the whole industry truly gets bigger.

And the earliest people who paid for GPUs, built data centers, and signed power contracts may still not make money.

That is the most counterintuitive risk in AI capex.

If AI eventually fails because demand never shows up, the picture is easy to analyze: not enough demand, too much capex, share prices fall.

The harder case is the other one:

The real trouble isn't AI failing — it's AI succeeding, and the payoff never reaching the people who bought the GPUs.

That sounds contradictory, but it's not the first time it's happened. A lot of infrastructure works this way: the direction call was right, society benefited, and the people who paid for it first still didn't get paid.

Fiber, shale gas, solar — different industries, the same trap.

Fiber won, shareholders lost

Around 2000, the whole world believed the internet would change everything. That belief turned out to be entirely correct. Video, cloud computing, streaming, mobile internet — nearly all of it was built on cheap bandwidth.

But the people laying fiber back then did not get paid for being right.

Global Crossing was one of the most representative companies of that era. The market loved it, and the story was simple: internet traffic would explode, and the world would need more network capacity.

The story was right. The math was wrong.

Global Crossing was founded in 1997 and announced its first transatlantic submarine cable, AC-1, that same year; it entered service in 1999. From there the company acquired, issued debt, and built out its network across the globe.

It spent well over ten billion dollars building telecom networks, and filed for bankruptcy protection in 2002. By 2003, Singapore Technologies Telemedia picked up its core assets for pennies on the dollar — just $250 million.

Its problem was not that nobody used the internet. It was that telecom networks have a brutal economic structure: the fixed cost of building the network is enormous, while the marginal cost of selling one more unit of bandwidth is very low. Once competition starts, prices get pushed to levels many companies cannot survive.

So the lesson of the fiber bubble is not that the internet didn't work.

Quite the opposite:

The internet was real and the payoff was real. The payoff just did not go to the people who laid the first fiber.

Society got the infrastructure.
Later founders got cheap bandwidth.
Consumers got lower prices.
The early shareholders and creditors of the builders paid the bill.

Fiber's problem was that prices got crushed. Shale gas is more painful: it's not that demand never comes, it's that it comes too slowly.

Demand will come, but it may come too late

U.S. shale gas is a similar story. After natural gas production ramped up quickly, prices stayed cheap for years.

Cheap gas did eventually create new demand:

Coal plants switched to gas, chemical capacity came back onshore, LNG export terminals got built. By 2025, EIA data shows natural gas generating roughly 40% of U.S. electricity.

But none of that demand appeared overnight.

Supply can arrive in three years. Demand may take five or ten to catch up. Power plants, liquefaction terminals, and chemical plants are not built in a few quarters.

That gap in between is where shareholders get buried.

The representative name is Chesapeake Energy. Once a star of the American shale revolution, it snapped up land an order of magnitude faster than its peers and became one of the largest natural gas producers in the country. The industry's direction wasn't wrong — gas demand did eventually show up. But Chesapeake blew out supply and crushed prices itself, filed for bankruptcy protection in 2020, and never lived to see demand catch up with supply.

So the question is not only whether demand is elastic.

The sharper question is:

Can demand show up faster than supply expands?

Solar just moved the trap to a new industry. The cost of solar power has fallen roughly 90% since 2010, and global installations have exploded.

That is a victory for the energy transition. But module manufacturers have faced a permanent price war, and many of the early stars never converted industry growth into shareholder returns.

Look at Q-Cells, Suntech, and SolarWorld. Q-Cells was once a global leader in solar cells and filed for insolvency in 2012. Suntech was briefly the largest module maker in the world and was pushed into restructuring by its creditor banks in 2013. SolarWorld, a flagship of German solar manufacturing, entered insolvency proceedings in 2017.

None of this is coincidence. A lot of infrastructure works this way: the construction phase is when capital is tightest and price wars are easiest to start; the ones who actually capture the payoff are usually the people who later build on top of cheap infrastructure.

That is exactly what makes me cautious about AI right now.

What GPUs fear isn't going unused — it's going obsolete too fast

Many people explain AI through the Jevons paradox: as efficiency improves, unit cost falls and total usage rises instead.

When steam engines got more efficient, coal consumption went up, not down. Computing costs have fallen for decades, and global demand for compute has only grown.

I think that direction is probably right.

As inference gets cheaper, many tasks that never penciled out will start to. Customer support, coding, design, data analysis, process automation — things that never used AI before may end up using it by default.

So I do not think the biggest risk to AI is that nobody uses it.

For someone buying GPUs, the genuinely dangerous combination is this:

AI usage keeps exploding, but the compute needed per task falls even faster.

Both can happen at the same time.

A cloud provider buys GPUs today, builds a data center, signs power contracts, and expects to earn it back over the next several years. Two years later, model architectures have improved, distillation has matured, inference costs have dropped, and the next generation of chips is dramatically faster.

The AI industry may still be growing.
Users may still be increasing.
Query volume may still be hitting records.

And that older batch of GPUs may already be heading toward obsolescence, with the rental income on them already crushed.

This is the biggest difference between GPUs and fiber.

Fiber can hold up over time. The glass sits in the ground and can still be put to use ten years later. Early shareholders lost, but the asset itself retained long-term value.

GPUs cannot wait. They are not twenty-year assets. They get replaced by new chips, diluted by algorithmic progress, and repriced downward as inference cost per task falls.

For whoever owns the cards, the faster AI advances, the shorter the economic life of the older ones.

So when I look at AI capex, I do not just ask whether future demand will be large.

That question is too coarse.

The real question is:

When demand finally shows up, who is holding the scarce resource, and who just bought a pile of machines that depreciate fast?

In the end, it comes down to who holds the supply constraint

Back inside an AI data center, GPUs are not the only asset. Some things behave more like fiber: land, power interconnection, the building, cooling, grid connection. These last longer and are much harder to scale quickly.

The IEA projects global data center electricity use rising from roughly 460 TWh in 2024 to over 1,000 TWh by 2030 — close to the entire electricity consumption of Japan today.

Power, transformers, transmission lines, and site locations are what look like real bottlenecks.

GPUs are different. They are the part of the capital stack that competition consumes fastest.

So in the AI value chain, what matters is not who talks loudest or who buys the most cards, but who actually holds something others can't get their hands on anytime soon.

If the constraint is GPUs, the GPU buyers make money.
If the constraint is power, whoever sells electricity and holds grid interconnection makes money.
If the constraint is models and products, whoever turns AI into high-retention software makes money.
If there is no constraint at all, the biggest winner is probably the user.

Which is why I say:

The more AI succeeds, the more dangerous it is to own GPUs.

Not because AI will fail, but because it may be great for society and terrible for the capital that went in early.

The genuinely exposed layer is whoever carries GPUs on the balance sheet and expects old compute to pay for itself over a long horizon.

An industry getting bigger only proves society needs it.
Shareholders making money requires something else: a supply constraint.

Without one, no amount of demand prevents prices from being competed down, handing the surplus to consumers and to the next layer of founders.

The investment implication is direct.

Investing isn't a bet on the industry getting bigger

So I no longer look only at AI revenue growth or model capability. I want to watch five things:

First, when capex growth starts to slow.
If every company is still forced to accelerate spending, competition has not ended and the pressure to buy cards is still rising.

Second, the true economic life of a GPU.
Accounting depreciation is one thing. What the market will actually pay for an older card is another.

Third, whether free cash flow is improving.
Revenue growth looks great, but if cash flow keeps getting eaten by capex, shareholders receive very little.

Fourth, whether credit markets start to turn.
As AI infrastructure leans more on debt, bond investors feel the pain earlier than equity investors do.

Fifth, whether hard constraints like power and data center capacity actually convert into profit.
A bottleneck is not the same as a profit. Utilities are regulated, data centers are cyclical, and power contracts carry costs. It still comes down to whether a specific company can keep the scarcity for itself.

So this piece is not an argument that AI is about to collapse.

What I actually mean is this:

Whether AI has a future, and whether this cohort of GPU buyers makes money, are two separate questions.

The internet had a future; the fiber builders went bankrupt.
Gas demand grew; many shale shareholders died first.
Solar reshaped the energy mix; module makers live in a permanent price war.
Air travel demand is highly elastic; the industry still struggles to earn excess returns.

AI may follow a similar path.

It may genuinely change the world.
It may genuinely create enormous demand.
It may genuinely raise productivity.

But if efficiency improves too fast, supply expands too aggressively, and old assets depreciate too quickly, the earliest GPU buyers may never see the day they break even.

That is the most counterintuitive risk in AI capex:

The cause of death is not AI failing — it's AI iterating too fast: the old assets never paid for themselves before they became obsolete.

Investing isn't a bet on the industry getting bigger — it's a bet on which layer keeps the profit.