Nobody Gets a Bonus for Asking Whether the AI Actually Works

There is a wonderfully inconvenient question sitting in the middle of the artificial-intelligence investment boom.

It isn't:

Clearly we can.

It isn't:

Clearly it can.

And it isn't even:

Almost certainly.

The uncomfortable question is much simpler:

Can a probabilistic generative system reliably take an established body of complex work, change precisely what needs changing, propagate all necessary consequences and preserve everything that should remain unchanged?

Because that is what much of economically valuable human work actually consists of.

And if the answer is "not yet", there is another rather awkward question:

Why are we spending hundreds of billions of dollars making the generator bigger before establishing whether generation is actually the limiting problem?

There may be a perfectly rational answer.

Because almost nobody currently making money from the AI boom is financially rewarded for asking that question.

Welcome to the most magnificent shovel business in history

The five major AI hyperscalers are expected to spend roughly $750 billion on capital expenditure during 2026.

AI-related global debt issuance is also accelerating. The Bank of England says AI-focused companies crossed an important threshold in 2025 when investment requirements exceeded their ability to finance the build-out entirely from internal cash flows.

External capital is increasingly entering through bonds, private credit, structured finance, special-purpose vehicles and other arrangements.

This is no longer merely a software experiment.

We are building power stations, substations, data centres, transmission infrastructure, cooling systems, networks and semiconductor factories to support it.

And the commercial logic appears beautifully circular.

There is nothing inherently fraudulent about this.

There is, however, something economically magnificent about it. Because almost every turn of the wheel creates measurable economic activity before anybody has to establish the final productivity return.

Follow the incentives

Imagine being the chief executive of a hyperscaler. Your competitors announce another 5GW of AI capacity. What exactly are you supposed to do?

Announce that you have developed an interesting philosophical concern about semantic invariance?

"Thank you, everyone. Rather than spending another $80 billion on AI infrastructure this year, we have decided to investigate whether probabilistic generation can ever provide sufficiently controlled cumulative knowledge for autonomous engineering."

Your share price would probably discover semantic invariance on the journey towards the floor.

The rational response is obvious. Build.

Microsoft builds because Amazon builds. Amazon builds because Google builds. Google builds because Meta builds. Meta builds because Microsoft builds.

Nobody needs to believe every dollar will ultimately produce an adequate return. They merely need to believe that not building would be more dangerous than building.

That changes the investment decision completely. The question is no longer:

It becomes:

That is an extraordinarily powerful mechanism for capital expenditure.

The career risk is beautifully asymmetric

Now move down the organisation. Imagine you're responsible for AI strategy.

Which decision would you make?

This is not stupidity. It is incentive-compatible behaviour.

The personal downside of underinvesting in a successful revolution is enormous. The personal downside of overinvesting alongside the entire industry is distributed.

That is how perfectly intelligent people can collectively produce extraordinary investment cycles.

And Wall Street doesn't particularly want the question either

Investors have their own problem. They need a growth story.

AI provides perhaps the greatest growth narrative available in global markets. It promises new software markets, automation, cloud demand, semiconductor demand, robotics, autonomous agents, scientific discovery and eventually the replacement or augmentation of enormous quantities of expensive human cognitive labour.

That last part is particularly important. Human knowledge work represents an enormous pool of economic cost. If machines can perform much of it, the theoretical addressable market is almost unimaginably large.

But that assumption quietly moves between two propositions.

The distinction is financially inconvenient. So it is much easier to measure tokens, users, benchmark scores, inference volumes, cloud revenue and model capabilities. All of those numbers go up. Everyone feels better.

The metric that would spoil the party

Imagine hyperscalers reported a different number:

Now run that measurement over three years.

Suddenly things become less comfortable. Because AI could make the initial creation of something ten times faster while making its subsequent maintenance only 20% faster. Or not faster at all. It could even create additional downstream cost.

That number is considerably less exciting than tokens per second.

The really nasty question

There is an even more fundamental problem. Generative AI creates candidate states. Engineering organisations maintain established states. Those are different activities.

An engineering system doesn't spend its life repeatedly asking: "Please design me a powertrain."

It spends decades saying:

Each change operates upon everything previously established.

So the economically important capability isn't simply intelligence. It is controlled cumulative intelligence.

Can the machine make the necessary change, discover every legitimate consequence and preserve everything unaffected?

If it cannot, humans and deterministic engineering systems remain responsible for settling its proposals. At which point AI may still be enormously useful. But the economic model becomes different.

We haven't replaced the engineer. We've built the world's most extraordinary proposal generator.

The $750 billion question

This creates a wonderfully cynical possibility. What if making models larger continues producing better proposals, but does not solve cumulative control?

Then the industry could spend another trillion dollars improving something that was never the ultimate bottleneck.

Not because the models are useless. Precisely the opposite. They will become spectacular.

The demonstrations will improve. Benchmarks will improve. Coding will improve. Video will improve. Reasoning will improve. Agents will improve. Every improvement will provide another justification for additional infrastructure.

Yet the difficult final mile — turning probabilistic proposals into dependable cumulative work — could remain stubbornly expensive.

That would be the most dangerous possible outcome financially.

Not AI failure. AI success without sufficient economic conversion.

Fortunately, we can borrow the money

The AI infrastructure boom began with some of the richest companies in history financing investment from enormous operating cash flows. That is changing.

The Bank of England says the sector reached an inflection point in 2025 as investment requirements exceeded internal cash-flow capacity. Debt financing is accelerating. Private credit is participating. Structured finance is participating. Off-balance-sheet arrangements are increasing.

The Bank notes that the five major hyperscalers accounted for only around 3% of outstanding US investment-grade corporate debt at the end of 2025, but more than 15% of issuance by early May 2026. AI companies have also become significant issuers in high-yield markets.

This doesn't mean catastrophe. These are some of the world's strongest companies. But something important has changed.

The industry is increasingly spending future cash flows to construct the infrastructure expected to generate those future cash flows. That makes the eventual productivity question considerably less academic.

Never ask a question whose answer might destroy your addressable market

Suppose somebody conclusively demonstrated tomorrow that probabilistic generative models, regardless of scale, cannot independently maintain controlled cumulative semantic state.

Would AI disappear? Absolutely not. Would GPUs become useless? Of course not. Would generative AI stop transforming industries? Almost certainly not.

But expectations might change. Perhaps AI replaces 20% of an engineering workflow rather than 80%. Perhaps every autonomous system requires a substantial deterministic settlement architecture around it. Perhaps human accountability remains necessary much longer than anticipated. Perhaps inference demand grows at 30% annually instead of whatever heroic number was embedded in a data-centre financing model.

Small changes in those assumptions can produce enormous changes in the present value of infrastructure.

So why rush to investigate the question? There are another 200,000 GPUs arriving on Tuesday.

Nobody needs to lie

This is the important part. There doesn't need to be a conspiracy.

Jensen Huang can genuinely believe accelerated computing will transform the world. Microsoft can genuinely see enormous demand. Amazon can genuinely need more capacity. Google can genuinely observe rapidly increasing inference workloads. Enterprises can genuinely report productivity improvements. Researchers can genuinely produce more capable models. Investors can genuinely expect enormous future profits. Banks can genuinely believe the borrowers are creditworthy. Pension funds can genuinely want exposure to long-duration infrastructure.

Every individual decision can be completely rational.

And the aggregate system can still be wrong about one assumption.

That is how interesting financial events tend to happen. Not because everyone was lying. Because everybody's perfectly reasonable model contained the same variable.

The variable nobody wants to discover is wrong

The assumption may be:

Increasing generative intelligence eventually produces increasingly autonomous economically valuable work.

Maybe it does. I hope it does.

Engineers may invent the missing architecture tomorrow. Formal systems, digital twins, deterministic settlement, executable requirements, semantic dependency systems and technologies we haven't conceived may eventually transform probabilistic generation into genuinely cumulative machine intelligence.

If that happens, the infrastructure may prove extraordinarily valuable. But notice the distinction. That would be another invention. It would not demonstrate that enough GPUs inevitably turn probabilistic generation into controlled cumulative intelligence.

And until that transition is demonstrated, the AI investment boom contains an extraordinarily inconvenient technical dependency.

So don't ask whether the AI is clever

That question is financially useless. Of course it is clever. The next one will be cleverer. And the one after that will make today's systems look primitive.

Ask instead:

Can it take an established complex system and make the thousandth consecutive change without silently corrupting what the previous 999 changes established?

Then ask:

Can it demonstrate what changed, why it changed, what consequences propagated and what remained invariant?

Then ask:

Can it do that more cheaply and reliably than the human engineering system it is supposed to replace?

And finally ask the question capital markets eventually ask of everything:

Did it generate more external cash than the system consumed producing it?

That is a considerably less exciting conversation than AGI. It may also be the one that matters.

Because the most cynical explanation for why hyperscalers aren't loudly asking the difficult question isn't that they know the answer.

It is that right now, uncertainty about the answer is economically much more convenient than finding out.