The AI Bubble Is Not About AI

The AI Bubble Is Not About AI, it is about expectations, incentives, and timing

Everyone in tech seems to agree we are in an AI bubble.

What people do not agree on is what exactly is inflated, or what happens if it deflates.

Some believe the technology itself is overhyped. Others believe the infrastructure spending is excessive. Others think the real excess is in company valuations, not capability. And a growing number of developers quietly suspect something else entirely: that expectations have simply raced far ahead of reality.

The confusion is not accidental. It is a natural side effect of a technology that is genuinely powerful, unevenly useful, and arriving faster than organisations know how to absorb it.

Bubbles usually form around something real

The word bubble tends to imply fraud or fantasy. But most historical bubbles formed around technologies that actually mattered.

Railroads transformed economies.
The internet reshaped communication and commerce.
Cloud computing changed how software is built and deployed.

In each case, the mistake was not believing the technology was important. The mistake was assuming that importance translated directly into short-term returns for everyone involved.

AI feels similar.

There is a real kernel of truth here. The models work. They are improving. They unlock capabilities that did not exist before. But that truth has been wrapped in a level of certainty, urgency, and financial expectation that no complex technology has ever lived up to on the first pass.

Infrastructure optimism versus execution reality

One of the most striking features of the current AI moment is the scale of infrastructure spending.

Data centres, GPUs, power grids, and long-term compute commitments are being built on assumptions about future demand that are difficult to verify today. The justification is often that compute is the bottleneck, and that whoever builds fastest will win.

That may turn out to be true.

But there is a difference between possible and inevitable.

Execution risk is not just technical. It is organisational, economic, and human. Even if the models continue to improve, building sustainable businesses around them requires alignment across product design, cost structure, regulation, and actual user behaviour.

History suggests those things rarely move in lockstep.

AI coding as a microcosm of the bubble

Nowhere is the gap between expectation and reality clearer than in AI assisted coding.

On paper, it looks like the perfect use case. Code is structured language. It is expensive to produce. It has clear success criteria. And in isolated tasks, AI tools can be astonishingly effective.

In practice, many developers report a more complicated experience.

AI tools excel at:

• boilerplate
• tests
• quick prototypes
• explaining unfamiliar code
• fixing narrowly scoped bugs

They struggle with:

• large codebases
• long-term architectural consistency
• context awareness
• respecting project-specific conventions
• knowing when they are wrong

This does not make them useless. It makes them situational.

The problem arises when productivity gains measured in controlled tasks are projected onto entire engineering workflows. The result is disappointment, confusion, and sometimes a quiet sense that something does not add up.

Productivity is harder to measure than we admit

A recurring pattern in developer feedback is perceived speed versus actual output.

Many engineers feel faster with AI tools. They move more quickly through tasks. They generate more code. They spend less time staring at a blank file.

But when measured over weeks or months, the gains are often smaller than expected. Sometimes they even reverse.

Why?

Because software development is not just typing. It is:

• reasoning
• debugging
• reviewing
• maintaining
• aligning with other humans
• paying down technical debt

AI can accelerate some of these steps. It can also quietly increase the load on others.

More code still has to be understood. More changes still need review. And code that looks correct but subtly violates a system’s design can be more expensive than code that was written slowly and deliberately.

Technical debt does not disappear, it shifts

One of the quieter risks in widespread AI assisted coding is the accumulation of technical debt in less obvious forms.

The code may compile.
The tests may pass.
The bugs may be fewer.

But complexity can still grow.

Verbosity increases. Patterns drift. Inconsistencies multiply. And because the output often looks polished, issues can be harder to spot early.

This is not a flaw unique to AI. It is a familiar problem in software engineering. The difference is scale and speed. AI makes it easier to accumulate debt quickly, especially under pressure.

The cultural amplification effect

AI tools do not exist in a vacuum. They amplify the culture they are introduced into.

In teams with:

strong conventions

clear architectural principles

disciplined review processes

AI can be genuinely powerful.

In teams without those things, it often magnifies existing problems.

This is why adoption outcomes vary so widely. The tool is the same. The context is not.

Why this still does not mean the bubble is fake

None of this implies that AI is a fad.

It implies that the timeline has been misunderstood.

AI is not replacing software engineers overnight. It is changing what software engineering looks like, unevenly and gradually. Some workflows will be transformed quickly. Others will resist change for good reasons.

The danger is not believing in AI.
The danger is believing that belief alone guarantees returns.

A more grounded way to think about the AI moment

Instead of asking whether we are in an AI bubble, a better question might be:

Where are expectations being priced as certainty?

That framing allows for nuance.

It leaves room for:

real breakthroughs

real failures

uneven impact

long adjustment periods

And it reminds us that bubbles do not only burst. Sometimes they deflate slowly, leaving behind infrastructure, knowledge, and capabilities that matter long after the hype fades.

AI is powerful.
AI is imperfect.
AI is early.

The tension we are seeing is not between truth and illusion, but between possibility and impatience.

If there is a bubble, it is not one belief. It is many overlapping assumptions, moving at different speeds, held by people with different incentives.

Understanding that may be more useful than predicting exactly when anything pops.