Concentration in the market and pricing
A few big tech companies now dominate the stock index.
The largest tech platforms account for unusually high shares of the S&P 500 and global indexes.
AI stories explain most of the stock market’s gains since the end of 2022.
A slight shock, such as a surprising competitor or regulatory move, can move trillions in market value in a single day.
The dipseq incident in early 2025, where a cheap Chinese model briefly wiped out large amounts of market cap, showed just how fragile sentiment is. When the narrative changes, it can move very quickly.
to spend which advances the current return
Capital spending on AI infrastructure has entered historic territory.
Big tech companies collectively spend hundreds of billions of dollars each year on data centers, GPUs, and power.
Some estimates put AI-related capex at over $500 billion annually for several years.
In contrast, direct AI service revenues are still very small, and in some segments are measured in the tens of billions rather than the hundreds.
Consulting and research reports stand at an odd point: most businesses experimenting with generative AI are yet to see a significant impact on their P&L.
Extensive studies show that the majority of AI initiatives so far show little or no measurable ROI.
Many projects improve individual productivity, but not overall margin or revenue growth.
AI is often still stuck in pilot mode, not deeply embedded in tasks.
You can justify the heavy initial investment for a while. You can’t do it forever if the profit story remains vague.
Circular And aggressive Financial assistance
Some AI contracts and investments are designed to keep the music playing.
Vendors pre-purchase large blocks of cloud capacity from each other.
AI labs commit to spending huge sums on specific infrastructure providers.
These commitments are then reflected as an increase in future revenue from the supplier, even if the buyer does not yet have a direct way to recoup that amount.
This is not cheating, but it creates a feedback loop in which the rosy assumptions on both sides reinforce each other. If a piece cracks, the loop can open quickly.
Physical barriers: Energy, Cooling, and Earth
AI is no longer just software. It is solid, copper and MW.
Modern AI data centers can use as much electricity as a large city.
Local grids, water supplies, and permitting processes are beginning to deteriorate.
Governments and regulators are asking whether building unlimited AI is compatible with climate goals and local infrastructure.
If power or cooling becomes a hard constraint in key regions, some existing capex projects will need to be rolled back. Such a hard stop is a classic trigger for asset restatements.