Return on Intelligence, Part 6: Breaking Point

Part 6: Breaking Point

Part 5 argued that AI changes not only software, but the cost structure of the firm itself. This chapter turns from structure to timing: what actually causes the market story to break. It is the hinge in the series, and the natural bridge into the post-crash world of Part 7.

The market does not stop believing in AI. It starts doubting who owns it.

That is the mechanism behind every belief break in a real-paradigm bubble. Not a sudden rejection of the technology, but a quiet erosion of the story about ownership. And when enough of those stories erode at the same time, the valuation can no longer hold.

That is the point people often miss. Markets can live with overvaluation for a long time if the story still works. Investors can tolerate losses, burn rates, dilution, aggressive forecasts, thin margins and stretched multiples if they believe the current discomfort is the price of owning the future.

The break comes when the story stops explaining reality. Not all at once. Not usually through one single event. More often, several pressures build until the shared belief shifts.

In the dotcom bubble, the internet did not suddenly become less important. The belief that changed was more specific.

Investors stopped believing that every company with a domain name was a future monopoly.

They stopped believing that traffic alone was enough, that customer acquisition cost did not matter, that first mover advantage could overcome bad economics, and that capital could buy time forever.

The future was still real. The investable story broke. That is what I expect with AI.

The AI bubble will not burst because people stop using AI. It will burst when enough people realize that the current market story misidentifies who owns the future.

The market story today

The current AI story is a bundle of beliefs.

• It says frontier models remain scarce and valuable.
• It says the companies with the best models can capture intelligence rents through APIs and subscriptions.
• It says hyperscaler capex is a sign of future demand rather than a possible overbuild.
• It says the chip and data centre buildout will keep compounding because AI demand is effectively insatiable.
• It says SaaS companies can defend their workflow ownership by adding AI features.
• It says enterprises will move from pilots to deep adoption fast enough to justify the infrastructure.
• It says AI agents will increase usage, increase productivity and remain dependent on centralized cloud intelligence.
• It says the first visible AI winners are probably the long-term owners of AI value.

That is a lot of story.

Each piece can be partially true.

The bubble breaks when too many pieces become less true at the same time.

Pressure 1: The model scarcity premium declines

Today, frontier models feel scarce.

Only a small number of companies can train and serve the best systems. That scarcity supports the idea that model labs can own the intelligence layer.

But the scarcity premium is under attack.

Models get smaller. Open models improve. Specialized models become good enough for narrow tasks. Distillation moves capability downward. Inference gets cheaper. Hardware improves. On-device and enterprise-local models become more practical.

The frontier keeps moving, but commercial value does not always require the frontier.

Many tasks require reliable, cheap, contextual, private, low-latency intelligence rather than the smartest possible model.

As that becomes obvious, the market will ask how much pricing power belongs to the model API.

If the answer is less than expected, the story weakens.

Pressure 2: Token economics become visible

A lot of AI adoption is still happening under soft economics.

Flat-rate subscriptions. Promotional pricing. Cloud credits. Internal experimentation budgets. Strategic subsidies. Bundled features. Investor-funded usage. Enterprise trials. Internal mandates to test AI.

This creates activity, but not always durable economics.

The hard question is what happens when usage is priced honestly.

Agentic workflows can burn tokens. Coding agents can burn tokens. Research agents can burn tokens. Customer support automation can burn tokens. Multi-step planning, retrieval, retries, tool calls, evaluation and monitoring all cost money.

If the output is valuable enough, this is fine.

But if much of the usage is exploratory, low-quality, redundant or requires expensive human correction, then customers will become more disciplined.

The AI market is currently full of enthusiasm about usage.

The convergence point comes when usage has to justify itself as margin.

Pressure 3: Enterprise ROI disappoints before it matures

AI is useful in enterprises.

That does not mean enterprise transformation is easy.

Real companies have messy data, legacy systems, compliance requirements, permissions, audit needs, security constraints, customer promises, regulatory obligations, internal politics and edge cases.

The demo is always cleaner than the business.

A chatbot that summarizes a document is impressive. A governed AI workflow that safely acts across customer data, legal constraints, pricing rules, finance systems, support systems and audit requirements is much harder.

The value is there, but it requires integration. The likely pattern is disappointment first, maturity later.

Many pilots will fail to produce measurable returns. Some internal AI programmes will become expensive experiments. Some consultancies will sell transformation without operational depth. Some companies will discover that the model was not the hard part. The hard part was workflow, data, permissions, governance, evaluation and change management.

This does not kill AI. It slows the revenue curve.

If valuations assume rapid enterprise conversion, that matters.

Pressure 4: Hyperscaler capex becomes a question instead of proof

Right now, AI infrastructure spending is treated as proof of demand.

The logic is simple: Microsoft, Google, Amazon, Meta and others are spending enormous amounts on data centres, chips, networking, power and cooling. They must see demand. Therefore the buildout validates the market.

That is true until it is not. At some point, capex changes meaning.

If utilization disappoints, if customers resist pricing, if model efficiency improves faster than demand, if local inference absorbs common workloads, if financing costs rise, if power constraints bite, or if companies guide toward more discipline, the same capex can be reinterpreted as overbuild.

This is psychologically important.

The market currently sees spending as confidence. The bubble cracks when spending starts to look like sunk cost.

The language will be subtle at first:

• optimizing utilization;
• moderating future expansion;
• focusing on capital discipline;
• aligning capacity with demand;
• improving efficiency;
• prioritizing high-return workloads.

None of those phrases means AI is dead. They mean the growth story is changing.

Pressure 5: SaaS growth slows and pricing power weakens

SaaS is part of the AI bubble even when it is not labelled that way.

Many SaaS companies are priced as if AI will enhance their existing products, defend their workflows and create new revenue streams.

That may be true for the strongest platforms.

But AI also attacks SaaS.

• If AI reduces headcount, seat-based pricing weakens.
• If AI generates custom workflows, generic SaaS workflows lose scarcity.
• If AI features become standard, they stop being premium add-ons.
• If enterprises become more capable of building internal tools, the buy-versus-build boundary shifts.
• If AI-operated companies use thinner software stacks, traditional SaaS expansion weakens.

The pressure may appear first as slower expansion, lower net revenue retention, more pricing scrutiny, consolidation of tools, fewer seats, longer sales cycles or customer resistance to AI surcharges.

That does not look like a dramatic crash at first. It looks like growth becoming ordinary.

For high-multiple SaaS, ordinary growth is dangerous.

Pressure 6: Thin AI wrappers fail

Every paradigm shift creates wrappers.

A wrapper is a company that packages a new technology into a narrow product before the durable platform structure is clear.

Some wrappers become real companies. Many do not.

The AI wrapper problem is obvious. A startup takes a model API, adds a workflow, a prompt library, a UI, a narrow use case, maybe some integrations, then raises capital as if it has created a durable product.

Some of these will work.

Many will be crushed by model providers, SaaS incumbents, open source, platform features or customers building their own versions.

When the wrapper layer starts failing visibly, the market mood changes.

Again, this does not mean AI is fake.

It means that not every use case is a company.

Pressure 7: Financing stress appears below the surface

The AI buildout is capital intensive.

Data centres, chips, power, cooling, memory, networking and long-term capacity commitments require enormous amounts of money.

Some of this is funded directly by highly profitable companies. Some is funded through debt, leases, partnerships, private capital, infrastructure vehicles, vendor financing and complicated commitments.

In boom phases, this all looks rational. Everyone wants exposure to the buildout. Future demand is assumed. Assets look strategic. Financing looks safe because the tenants and counterparties are high quality.

But infrastructure bubbles can break through financing stress before the public narrative changes.

A data centre project delayed by power constraints. A financing structure that depends on aggressive utilization. A private infrastructure vehicle that cannot roll over debt on expected terms. A supplier with customer concentration. A chip order that gets pushed out. A capacity commitment that looks too large.

These are not necessarily headline events at first.

They are stress signals. The AI bubble may not crack first in a chatbot app.

It may crack in the financing of the physical world underneath it.

Pressure 8: The belief shifts from future ownership to future exposure

This is the psychological turn.

During the boom, exposure to AI is good.

• A company spends on AI infrastructure: bullish.
• A SaaS company adds AI features: bullish.
• A startup uses AI in the pitch: bullish.
• A cloud provider signs a huge capacity deal: bullish.
• A company talks about agents: bullish.

At the convergence point, the meaning changes.

AI exposure becomes a question:

• Are you capturing value or just spending money?
• Are you selling durable capability or subsidized usage?
• Are you defending a moat or training customers to expect cheaper software?
• Are you building infrastructure for future demand or overbuilding ahead of it?
• Are you using AI to lower cost or adding another expensive layer?
• Are your AI features differentiated or table stakes?
• Are you the owner of the workflow or a supplier to someone else’s workflow?

That is when the story breaks.

The market stops rewarding exposure and starts demanding proof of capture.

The convergence

• No single pressure has to be fatal.
• The model labs can survive smaller models.
• Hyperscalers can survive capex discipline.
• SaaS companies can survive slower growth.
• Enterprises can survive failed pilots.
• AI wrappers can fail without damaging the whole market.
• Infrastructure financing can absorb some stress.

The danger is convergence.

Several pressures arriving close enough together that the market narrative cannot absorb them:

• model scarcity declines;
• enterprise ROI slows;
• token economics harden;
• capex growth moderates;
• SaaS expansion weakens;
• wrappers fail;
• financing stress appears;
• investors ask who owns the future.

At that point, the market does not need proof that AI failed.

It only needs doubt that the current winners own enough of the future to justify their valuations.

The chart below is not a financial model. It is a conceptual picture of market psychology. The dark line is narrative belief. The accent line is contradictory evidence. The shaded gap between them is the belief buffer, the market’s remaining capacity to explain weak signals away as temporary, strategic, or worth enduring.

As long as the belief buffer exists, the story survives. Weak ROI can be called early learning. High token burn can be called transition cost. Giant capex can be called confidence. Slower SaaS expansion can be called temporary noise. The break comes when those explanations stop feeling persuasive enough to carry the valuation.

That is how a real-technology bubble bursts.

The sunk cost problem

Human belief is not rational at the end of a bubble. It is emotional.

People have reputations invested. Careers invested. Funds invested. Strategies invested. Public statements invested. Corporate roadmaps invested. Data centre plans invested. Hiring plans invested. Acquisition strategies invested. Product narratives invested.

The more capital is committed, the harder it becomes to admit the story has changed.

This is why bubbles often overshoot and sunk cost becomes identity.

The company cannot easily say the strategy was wrong. The investor cannot easily say the thesis was wrong. The executive cannot easily say the capex was excessive. The founder cannot easily say the moat was temporary.

So the story stretches and every contradiction gets explained away.

Until it cannot be. That is the breaking point.

The moment when maintaining belief becomes more painful than abandoning it.

The crash will be misread

When the AI bubble bursts, many people will say AI was hype.

That will be the wrong lesson.

The correct lesson will be that the first capital markets story about AI was wrong. The internet after dotcom became more important, not less.

AI after the crash may do the same.

The crash will destroy weak companies, punish bad capital structures, compress valuations, expose fake moats and force discipline, but the technology will continue moving.

Models will improve. Software will become more fluid. AI will become embedded. Companies will automate more. Some workflows will disappear. New business forms will emerge.

The bubble bursts because the story fails. The paradigm continues because the technology is real.

That is the convergence point.

In Part 7, I follow that break forward into the mature AI economy and the assets, power centers and business models that survive it.