The (ai) Implosion

How Cloud 2.0 turned human intelligence into the hyperscalers' next addressable market — and why IT Dave may eventually expose the flaw

The cynical interpretation of the current (ai) boom starts with a much less glamorous proposition than the creation of machine intelligence.

It starts with hyperscaler revenue.

For twenty years, AWS, Microsoft Azure and Google Cloud built extraordinary businesses by persuading organisations to move computing workloads they previously owned into infrastructure they rented.

Cloud 1.0 was brilliant economics.

Move the servers.

Move the applications.

Move the databases.

Move the storage.

Then charge continuously for running them.

Once the workload had been moved, technical dependency, data gravity, skills, contracts, integration and switching costs made reversing that decision considerably harder than making it in the first place.

But eventually the easy migration opportunity had to mature.

Businesses began optimising cloud estates, embracing hybrid infrastructure and, in some cases, repatriating workloads where public-cloud economics no longer made sense.

Then generative (ai) arrived.

And suddenly cloud growth accelerated again.

That is not conjecture. Synergy Research says global cloud infrastructure spending reached $143 billion in Q2 2026, 43% higher than a year earlier, and explicitly describes generative (ai) as the primary driver of that acceleration.

So perhaps the important question isn't:

What came after cloud?

It is:

What if (ai) is the next cloud workload?

Except this time the workload isn't merely computing.

It is human cognitive work.

1. Cloud 2.0

The commercial proposition changes dramatically.

Cloud 1.0 said:

Give us your computing infrastructure and rent it back from us.

Cloud 2.0 can say:

Give us increasing portions of your organisational intelligence and rent access to it back from us.

Coding.

Writing.

Research.

Knowledge retrieval.

Customer support.

Analysis.

Planning.

Decision support.

Agents.

Organisational memory.

The potential addressable market is enormously larger than servers.

The ultimate hyperscaler workload is no longer the server used by IT Dave.

It is IT Dave.

If Dave costs a business £100,000 a year and a hyperscaler can persuade the board that £20,000 of (ai), agents and infrastructure can perform Dave's work, the sales proposition is spectacular:

£100,000 payroll expense

becomes:

£20,000 recurring hyperscaler revenue

while the customer believes it has saved £80,000.

That is considerably more attractive than migrating another virtual machine.

2. AGI supplies the urgency

But selling another software tool creates a problem.

Businesses evaluate normal technology according to ordinary questions:

What does it cost?

What does it save?

What is the return?

What happens if we don't buy it?

Those questions slow procurement down.

AGI changes the psychology completely.

If businesses are told that current (ai) is the beginning of an intelligence transition which could produce human-level or greater intelligence within a few years, then not adopting becomes dangerous.

The proposition becomes:

Your competitors may acquire machine intelligence before you do.

Now the CEO doesn't merely have an ROI decision.

The CEO has a FOMO problem.

And that makes the extraordinary AGI narrative economically valuable whether or not AGI actually arrives.

The hyperscaler does not have to prove human replacement today.

It merely needs the customer to believe that moving towards it today is strategically unavoidable.

Cloud migration already provided the sales playbook.

Enterprise account relationships already exist.

Procurement frameworks already exist.

Cloud migration specialists already understand how to sell transformation, urgency and future dependency.

The product has changed.

The commercial machinery hasn't.

3. Move Dave's salary into cloud revenue

The first great economic promise therefore becomes labour substitution.

Don't think of (ai) merely as another IT budget.

Think of it as technology attempting to access the payroll budget.

That is a gigantic expansion of the available market.

Knowledge workers cost trillions.

If even part of that expenditure can be converted into compute, tokens, agents, storage, orchestration and subscriptions, the hyperscalers have found their Cloud 2.0.

And the more human cognitive activity passes through hyperscaler infrastructure, the more valuable that infrastructure becomes.

This leads to the next stage.

4. Cognitive lock-in

Cloud 1.0 created data gravity.

Cloud 2.0 can create something considerably more powerful:

Cognitive gravity.

A company's traditional information can at least conceptually be exported.

Documents.

Files.

Databases.

Code.

Records.

But after several years of organisational interaction with (ai), what exactly constitutes the organisation's accumulated intelligence?

It potentially exists across:

conversation histories,

prompts,

agent memories,

RAG systems,

embeddings,

vector databases,

tool histories,

summaries,

workflows,

model-specific behaviours,

orchestration state,

human corrections,

decisions,

exceptions,

and countless interactions between them.

There is currently no generally accepted portable object equivalent to:

"Everything our organisation has cumulatively learned through this system."

You can move the source files.

That does not necessarily recreate the accumulated cognitive state.

And this is where the deepest problem in the entire (ai) proposition appears.

5. AI Cumulative Hiatus Delta

We have called that problem:

AI Cumulative Hiatus Delta — AI CHD

Each word matters.

AI

This problem concerns current generative (ai): systems primarily built to produce outputs probabilistically from learned representations and supplied context.

Cumulative

Human knowledge is not merely a sequence of fresh answers.

Engineering, organisations, science, law and individual intelligence depend upon carrying previous state forward.

Something changes.

Other things must remain true.

Dependencies must survive.

Decisions have consequences.

Errors must be corrected without destroying everything that remains valid.

Knowledge therefore has to accumulate coherently.

Hiatus

There is a break between the extraordinary ability of current (ai) to generate the next plausible state and its ability to establish an authoritative continuous state across repeated change.

It can create another answer.

That does not automatically mean it has preserved what must remain true from every previous answer.

That break is the hiatus.

Delta

The delta is the missing capability required to move from one side to the other:

probabilistic generation

to

dependable cumulative knowledge continuity.

So the concise definition is:

AI Cumulative Hiatus Delta is the missing capability between probabilistic generation and dependable cumulative knowledge continuity.

Or in ordinary English:

(ai) can create the next version. It cannot yet independently guarantee what must stay true while changing only what genuinely needs to change.

That matters enormously if you are proposing replacing Dave rather than merely helping him.

6. Kant saw the underlying problem 245 years ago

And this is where the story becomes rather extraordinary.

Immanuel Kant was wrestling with a structurally similar problem in the first edition of the Critique of Pure Reason in 1781, substantially revised in 1787.

Kant was not talking about computers.

He was asking what makes coherent experience and knowledge possible in the first place.

A stream of disconnected representations isn't sufficient.

Something must synthesise them into a unified experience.

They must belong to a continuing subject.

Objects must possess sufficient persistence and identity for experience to be coherent.

Causality, relation and other organising structures cannot simply be arbitrary associations.

Kant's transcendental unity of apperception addresses the requirement that our representations belong together within a unified consciousness. His Transcendental Deduction argues that mere empirical association isn't sufficient to establish the necessary objective unity of experience.

That's remarkably close to the structural problem current (ai) encounters.

A model can produce:

and each state can look extraordinarily convincing.

But that doesn't establish:

as the lawful continuous evolution of the same state.

Producing plausible succession is not the same as establishing continuity.

Kant was effectively distinguishing association of representations from the conditions required for their objective unity.

Our (ai) problem has an eerie resemblance.

Models have become astonishingly good at association.

The industry has still not demonstrated how scaling association alone crosses into the constitutive continuity required for dependable cumulative intelligence.

That is the Kant problem underneath AI CHD.

And after trillions of dollars of projected investment, there remains no established answer to:

What architectural development turns probabilistic succession into independently grounded cumulative intelligence?

More parameters don't answer it.

A larger context window doesn't answer it.

More memory doesn't answer it.

RAG doesn't answer it.

Agent orchestration doesn't answer it.

Those systems can help enormously.

But they move information around the model.

They do not by themselves establish the missing grounding condition.

Kant identified the philosophical requirement in 1781.

Current (ai) has yet to demonstrate how it crosses the delta.

7. Plausible Continuity Retention

And now something economically fascinating happens.

Because current (ai) does not intrinsically maintain dependable cumulative organisational intelligence, infrastructure has to compensate.

The system retains:

history,

embeddings,

vectors,

documents,

summaries,

agent states,

checkpoints,

tool interactions,

memory layers,

and increasing quantities of organisational context.

That gives the next model invocation enough historical material to produce something resembling continuity with the last invocation.

We should call the resulting infrastructure burden:

Plausible Continuity Retention

And this produces a remarkably cynical possibility.

Part of what is reported as increasing (ai) cloud consumption may not represent increasing productive intelligence at all.

It may represent the expanding infrastructure required to compensate for the absence of cumulative intelligence.

The customer sees:

Our (ai) knows more about our business.

The underlying architecture may actually be doing:

store more → retrieve more → reconstruct more → feed more context back into the generator.

That creates a remarkable revenue property.

Productive use can remain flat.

The accumulated continuity estate can continue growing.

Therefore:

Useful work: flat.

Stored cognitive context: rising.

Cloud consumption: rising.

Hyperscaler revenue: rising.

That means cloud growth and economic productivity are not necessarily equivalent.

Some cloud growth could represent the cost of overcoming the limitations of (ai) rather than evidence that those limitations have been solved.

That is a radically different interpretation of Cloud 2.0 economics.

8. Then the ROI problem appears

Eventually businesses have to measure what actually happened.

Dave didn't disappear.

Perhaps fewer Daves were required.

Perhaps Dave became dramatically more productive.

Perhaps some jobs genuinely vanished.

But businesses also acquired:

(ai) subscriptions,

compute,

integration,

cybersecurity,

agent infrastructure,

governance,

verification,

Plausible Continuity Retention,

cognitive clean-up,

and human checking.

The beautiful spreadsheet assumption:

Dave → (ai)

starts becoming:

Dave + (ai) + infrastructure + oversight.

That can still be enormously useful technology.

But it isn't the economic proposition implied by human-replacement intelligence.

And that distinction is lethal to FOMO economics.

Because useful software has to demonstrate ROI.

Emerging superhuman intelligence only has to create fear of being left behind.

9. The financing machine is already ahead of the evidence

Meanwhile hyperscalers have committed themselves to an unprecedented infrastructure programme.

The Bank of England reported in July 2026 that (ai)-focused companies reached an important inflection point in 2025: required investment exceeded their ability to finance it internally.

External financing accelerated sharply during the first half of 2026.

The Bank records growing use of public debt, private credit, leveraged finance, structured finance and off-balance-sheet vehicles.

The private-credit share of financing (ai) investment increased from 9% in 2024 to 34% in 2025.

And hyperscaler 2028 capex expectations rose from below $600 billion in December 2025 to over $1 trillion by July 2026.

The Bank explicitly warns that uncertainty remains about:

future compute demand,

adoption,

monetisation,

productivity gains,

asset depreciation,

and the financial consequences if assumptions behind the investment prove wrong.

So the money has started moving ahead of demonstrated economic outcomes.

And that sets up GatesGate.

10. GatesGate

In August 2026 Bill Gates dramatically escalated his assessment of (ai)'s employment consequences.

He now talks about far fewer jobs, argues that governments may need to reserve certain work for humans, and has discussed taxing (ai) and robotics to alter the incentive to replace people.

His change of position was sufficiently dramatic that he described the situation as "crazy" and "insane."

Viewed benevolently, Gates is warning society about a profound technological transition.

Viewed through our cynical hyperscaler thesis, something else happens simultaneously:

GatesGate turbocharges FOMO.

The message reaching every board becomes:

The technology is progressing so quickly that one of computing's most famous figures thinks governments may have to protect humans from it.

What CEO now wants to tell shareholders:

We decided not to participate?

The fear of being left behind becomes stronger precisely when the infrastructure programme requires ever greater adoption.

Whether Gates intended that effect is irrelevant.

The economic effect is what matters.

11. And then Dave discovers his pension is involved

This may become the politically explosive part.

Dave isn't merely being told that (ai) might eliminate his job.

Dave is saving for retirement through a pension.

The UK government is simultaneously encouraging workplace pension capital into private markets.

Under the Mansion House Accord, the target is at least 10% of relevant defined-contribution pension assets in private markets by 2030.

Government modelling suggests that could mean approximately £74 billion invested in private markets rather than roughly £26 billion under the previous baseline.

At the same time, the Bank of England says private markets, private credit and structured finance are becoming increasingly important sources of capital for (ai) infrastructure.

We cannot trace Dave's individual pension contribution directly into a particular Microsoft or Oracle data centre.

But the structural loop is real:

Dave works.

Part of Dave's wages goes into his pension.

Pension capital increasingly enters private-market investment.

Private markets increasingly finance (ai) infrastructure.

The infrastructure is sold to Dave's employer partly on the proposition that it can reduce human labour.

Dave is the labour.

And suddenly:

Dave is helping finance the capital substitution intended to make Dave unnecessary.

That changes Save IT Dave completely.

It is no longer merely a joke about protecting experienced IT workers.

It becomes an economic question ordinary people can understand immediately:

Why are my deferred wages financing infrastructure whose economic justification includes eliminating my future wages?

And if the human-replacement thesis subsequently proves exaggerated, Dave potentially gets hit twice.

His employer has increased technology expenditure.

His employment has been disrupted.

And his retirement capital may have participated in financing infrastructure whose valuation depended on stronger automation economics than ultimately materialised.

That's the moment Save IT Dave potentially becomes political.

12. The Lightbulb Moment

Eventually the entire story comes down to one tiny question.

Dave is asked:

Change the light bulb.

Current (ai) can become phenomenally good at that task.

It can identify the fitting.

Order the bulb.

Schedule the technician.

Analyse historical failures.

Optimise electricity consumption.

Generate the safety documentation.

Perhaps eventually a robot changes it.

And then Dave looks at the room and asks:

"Does it have to be a light bulb?"

Perhaps the roof should be glazed.

Perhaps the room shouldn't exist.

Perhaps nobody needs to work there.

Perhaps the underlying requirement wasn't illumination at all.

That is not simply doing the supplied task better.

It is questioning the frame that made the task appear necessary.

And that is where the difference between automation and intelligence becomes economically visible.

Current (ai) can generate suggestions such as "maybe don't use a light bulb."

That's not the point.

It can statistically generate the sentence.

The deeper issue is whether it possesses the independent grounding, persistent worldview, embodied consequences, cumulative continuity and authoritative relationship to reality from which that reframing actually becomes its own justified conclusion.

That's where Kant returns.

What makes all these representations belong to one coherent understanding?

What makes change legitimate?

What preserves identity?

What makes causality more than association?

What makes the next state a genuine continuation rather than another convincing representation?

After almost 250 years, those remain deep questions about intelligence.

And current (ai) has not demonstrated how scaling probabilistic generation crosses them.

The Implosion

And that's why the (ai) implosion does not require (ai) to fail.

ChatGPT still works.

Copilot still works.

Agents still work.

Models continue improving.

Robots improve.

Coding accelerates.

Research accelerates.

Businesses continue buying the technology.

The implosion is conceptual.

It is the moment enough people simultaneously realise:

We haven't created replacement human intelligence.

We've created an extraordinarily capable new technology.

And that tiny change in description produces an enormous change in economics.

If (ai) is approaching autonomous replacement intelligence:

buy immediately or die.

If (ai) is another extraordinarily useful technology:

show me the ROI.

And once the CFO starts asking that question, the entire hyperscaler proposition is exposed to normal economics.

How much labour actually disappeared?

How much productivity actually increased?

How much additional infrastructure was required?

How much was spent preserving plausible continuity?

How many humans remain necessary?

How portable is the accumulated cognitive workload?

How much does leaving the hyperscaler cost?

And ultimately:

Did we move Dave's salary to the hyperscaler — or did we keep Dave and add another enormous technology bill?

That is the Great (ai) Reclassification.

Not:

Artificial intelligence failed.

But:

Artificial intelligence was misclassified.

It was priced, financed and sold as the beginning of replacement intelligence.

It may ultimately prove to be something simultaneously less magical and extraordinarily valuable:

the most powerful human productivity technology yet created.

And when that becomes the accepted description, the technology doesn't implode.

The human-replacement narrative does.

The hyperscalers then have a rather uncomfortable problem.

They have built Cloud 2.0 around the proposition that human cognition represents the next enormous workload.

They have accumulated infrastructure partly financed by expectations of unprecedented future demand.

They have encouraged businesses to move cognitive workloads into systems that are difficult to make genuinely portable.

And because those systems do not possess dependable cumulative continuity, the customer may even be paying recurring cloud charges to compensate for the very missing capability that prevents the system from actually replacing Dave.

Meanwhile Dave is still at his desk.

His pension may have helped finance the data centre.

His company is paying the cloud bill.

And Dave is still the one asking:

"Does it have to be a light bulb?"

That is Save IT Dave.

That is AI Cumulative Hiatus Delta.

That is the problem Kant identified in 1781.

And, as of September 2026, the (ai) industry still hasn't demonstrated how it crosses that necessary delta.