Return on Intelligence, Part 7: After the Crash

Part 7: After the Crash

Part 6 described the convergence that breaks belief. This chapter follows the story through the crash and into the mature AI economy, where the first narrative has failed but the technology keeps spreading. It also sets up the final ownership question answered in Part 8.

The dotcom crash did not end the internet.

It began the mature internet.

That is not how it felt at the time. At the time, it felt like collapse. Companies disappeared. Careers changed. Offices emptied. Investors lost money. Founders who had been treated like prophets became examples of excess. The story broke.

But the infrastructure did not vanish. The habits did not vanish. The technical direction did not vanish. The public did not go back to a pre-internet world.

Instead, the market sobered up.

Weak companies died. Useful assets were absorbed. Valuations reset. Business models became more disciplined. The fantasy layer burned away. The companies that survived had to become real businesses.

Then the internet became more important than ever. That is the part that matters for AI.

The AI crash, when it comes, will not be the end of AI. It will be the beginning of the real AI economy.

The crash is not the end of the paradigm

A crash feels like rejection. It is often maturation.

In a discovery bubble, the market is not rejecting the technology. It is rejecting the first overfunded story about the technology.

The internet story of the late 1990s had to fail before the mature internet economy could form.

The same may be true of AI.

The first AI story says frontier model labs, hyperscaler data centre buildouts, SaaS incumbents with AI features and thin AI wrappers are the natural owners of the future.

Some of those companies will survive. Some will become stronger. Some will be absorbed. Some will disappear. Some will remain important but less profitable than expected. Some will become suppliers rather than platforms.

The crash will not decide whether AI matters. It will decide which parts of the first story were wrong.

That is a different question.

What happened after dotcom

The internet after dotcom became both more useful and more centralized.

The early web was strange, open, personal, chaotic and independent. It was full of small sites, weird communities, hand-built pages, forums, experiments and businesses that looked nothing like the platform economy that followed.

After the crash came consolidation. Search became a dominant interface. Cloud became the new infrastructure layer. Social networks centralized identity, attention and distribution. Mobile operating systems created new app store gatekeepers.

Then the next layer hardened. Streaming platforms reshaped media. Online advertising became surveillance infrastructure. The consumer became the product. Walled gardens replaced much of the early web’s open weirdness.

This is the uncomfortable truth about paradigm shifts. The mature form is often more useful and more extractive than the early form.

The mature internet gave us convenience, scale, search, communication, ecommerce, software distribution, cloud computing, mobile apps, maps, video, payments and access to knowledge.

It also gave us attention markets, behavioural tracking, platform dependency, algorithmic manipulation, creator precarity, app store taxes, surveillance advertising and a handful of companies with extraordinary power.

The crash did not prevent that.

It helped create the conditions for it.

AI will also centralize after it decentralizes

AI has an interesting tension.

On one side, models will get smaller, cheaper, more local and more embedded. That pushes toward decentralization.

On the other side, the companies that control devices, operating systems, clouds, identity, enterprise systems, data, app stores and workflow platforms are already large. That pushes toward centralization.

Both things can be true.

AI capability may become more distributed while AI power becomes more concentrated.

This is what happened with the internet. Publishing became easier for everyone, but distribution centralized. Communication became easier for everyone, but attention centralized. Software delivery became easier for everyone, but platforms centralized. Commerce became easier for everyone, but marketplaces centralized.

AI may follow the same pattern.

More people and companies will be able to build, automate and create.

But the operating layers may centralize around a small number of powerful control points.

Those control points may include:

• chips;
• device ecosystems;
• operating systems;
• cloud infrastructure;
• app stores;
• identity systems;
• enterprise data platforms;
• workflow orchestration;
• model routing;
• governance and audit;
• payment and billing;
• distribution and discovery.

The mature AI economy may not be owned by the chatbot brands that made AI famous.

It may be owned by the systems that decide where AI runs, what it can see and what it is allowed to do.

Weak model labs disappear or are absorbed

After the crash, the model landscape will likely consolidate.

Training frontier models is expensive. Serving them is expensive. Talent is scarce. Infrastructure is scarce. Distribution is hard. Enterprise trust is hard. Consumer attention is hard. Regulatory compliance is hard. Safety and evaluation are hard.

There will not be unlimited room for every model lab to become a major platform.

• Some will be acquired.
• Some will become specialized providers.
• Some will become research teams inside larger companies.
• Some will pivot to enterprise tooling, safety, evaluation, vertical models, agent infrastructure or deployment platforms.
• Some will disappear.

The surviving labs will need more than model quality.

They will need distribution, trust, cost discipline, enterprise relationships, developer ecosystems, integration depth or control of some strategic layer.

In the mature AI economy, the model is not enough.

Infrastructure gets reused

One of the recurring features of paradigm-shift bubbles is that badly timed infrastructure can still become useful later.

The investor loses money. The asset survives.

The first story burns off. The physical substrate remains, waiting for the next owner and the next use.

Dotcom overbuild did not mean the world never needed internet infrastructure. It meant the early capital structure and timing were wrong.

AI data centres, chips, networking, power contracts, cooling capacity and tooling may follow a similar path.

• Some infrastructure may be overbuilt for the first story but valuable for the second.
• Capacity funded by one group may be bought cheaply by another.
• Projects justified by one demand curve may serve a different demand curve later.

This is why the crash should not be confused with technological reversal. The physical and software infrastructure of AI will not vanish.

It will be repriced.

Inference becomes cheaper and more invisible

After the crash, inference costs will matter more.

The fantasy phase celebrates capability. The mature phase optimizes cost, latency, reliability, privacy and integration.

Models will be routed by task. Cheap local models will handle simple work. Specialized models will handle domain-specific work. Frontier models will handle hard cases. Enterprises will build policies that decide what data can go where. Devices will perform more inference locally. Operating systems will bundle AI into default experiences.

The user will see less of the model. AI will become a feature of the environment. That is when it becomes truly powerful.

The most important technologies often disappear into the background.

Electricity did. Networking did. Databases did. Cloud did. Machine learning already did in many areas. AI will too. The bubble version of AI is visible.

The mature version will be ambient.

SaaS consolidates and hollows out

The SaaS market after the AI crash will not simply die.

It will split.

Strong systems of record will remain important. Finance, HR, identity, ERP, compliance, regulated data, core operational records and deep enterprise platforms will still matter.

But workflow SaaS will be under pressure.

Generic workflows become less defensible when custom workflows are easier to generate. Seat-based pricing weakens when AI reduces human participation. AI features become standard. Customers become more willing to ask whether they need another SaaS product or whether they can generate the workflow themselves around existing systems of record.

• Some SaaS companies will consolidate.
• Some will be acquired by larger platforms.
• Some will become data and system-of-record providers.
• Some will become orchestration and governance layers.
• Some will become feature sets inside broader platforms.
• Some will be replaced by generated internal software.

This is not a collapse of software. It is a collapse of the generic workflow premium.

Software becomes more important. Packaged workflow software becomes less automatically valuable.

The rise of governed internal software

After the crash, companies will become more sober about AI.

• They will not stop using it.
• They will use it with more discipline.

That means more governance, not less.

The next phase will be about safe execution:

• which models can be used;
• which data they can access;
• which actions they can take;
• which workflows they can modify;
• which outputs require human approval;
• how performance is measured;
• how failures are audited;
• how generated code is tested;
• how AI tools are deployed and retired;
• how business rules are enforced.

This creates demand for enterprise AI control planes.

Not just chatbots, but control planes.

The companies that can govern AI-generated software, AI-operated workflows and AI agents inside real businesses will be strategically important.

That layer may become one of the mature AI economy’s durable profit pools.

AI-operated companies become real competitors

The crash will also make the AI-operated company more plausible.

During the bubble, many people will sell AI as a feature. After the crash, the best operators will use AI as a cost structure.

That is different.

A company with a small human team and a large automated operating surface can compete differently. It can personalize more, iterate faster, serve smaller markets, operate continuously and survive on lower margins.

This will begin in digital services, research, monitoring, content, software, operations, support, data enrichment and narrow B2B services.

Then it will spread.

The interesting thing is that these companies may not advertise themselves as AI companies.

They may simply be cheaper, faster and more adaptive.

That is when the transformation becomes hard to see from the outside but impossible to ignore in the numbers.

The consumer becomes more than the product

The internet taught us that free or cheap services can hide a deeper exchange.

The consumer became the product through advertising, attention, data and behavioural targeting.

AI may deepen that pattern.

The user is not only a target for ads. The user is a source of training signal, behavioural feedback, preference data, workflow data, voice, style, context, intent, correction and evaluation.

• Every interaction can improve the system.
• Every workflow can teach the system.
• Every correction can become part of the model or the product layer around the model.

In the mature AI economy, exploitation may not only mean showing better ads.

It may mean capturing the patterns of how people think, work, decide, write, negotiate, design, code, manage and communicate.

That is a deeper form of extraction. This is where the social consequences begin to overwhelm the capital markets story.

I do not want this series to become primarily about social impact. That deserves its own treatment, but it has to be acknowledged.

The mature AI economy may be useful, profitable and deeply extractive. The dotcom precedent suggests those things can coexist.

The crash creates the mature winners

Before the crash, everyone claims to be building the future. After the crash, the future becomes selective.

Capital becomes disciplined. Customers become demanding. Infrastructure gets repriced. Weak companies lose access to funding. Strong companies buy assets cheaply. Business models have to work. The fantasy layer fades.

That is when the mature winners emerge. They may not look like the bubble winners.

They may be:

• chip and memory suppliers;
• device and operating system owners;
• enterprise identity providers;
• systems of record;
• AI governance platforms;
• workflow orchestration systems;
• companies with proprietary data;
• companies that use AI to operate with radically lower cost;
• platforms that own distribution and execution rights.

Some current winners will remain. Others will be repriced. New ones will appear.

The crash clarifies the map.

The mature AI economy will be boring and enormous

The most important phase of a technology often begins when it becomes boring.

The internet became more powerful when it stopped being a special place you went to and became the substrate of everyday life. Cloud became more powerful when it stopped being a novelty and became the default way software was delivered.

Mobile became more powerful when it stopped being a category and became the normal interface to the world. AI will become more powerful when people stop talking about AI as a separate thing.

It will be inside the software. Inside the device. Inside the workflow. Inside the company. Inside the product. Inside the support process. Inside the operating model.

That is the mature phase. The crash does not prevent it. The crash accelerates it by killing the first story.

That is why I keep returning to the same point. The AI bubble will burst because AI is real.

And after it bursts, the real AI economy begins.

The crash is not where the argument ends. It is where the underlying map becomes easier to see.

In Part 8, I pull the whole series together and map the likely winners, pressured groups and control points in that mature economy.