Return on Intelligence, Part 1: Echoes | rebecca powell

Part 1: Echoes

The business cards in the drawer

I still have two packs of business cards from people I met at investor and founder events in San Francisco around 2000.

I kept them for posterity.

At the time, I did not know quite what I was keeping. I was working with a company inside the dotcom bubble. I spent time around people seeking investment, pitching ideas, trying to explain why the internet would change everything, and why their particular company would be one of the companies that survived and captured the upside.

There was a strange atmosphere around it all. It was not just greed, although there was plenty of greed. It was not just naivety, although there was plenty of that too. It was a feeling that something genuinely enormous was happening and that everyone in the room was close enough to touch it.

The future felt visible. Not clear. Not mature. Not understood. But visible.

Net Gain and the future hiding in plain sight

One of the books that seemed to follow people around the tech and startup scene in San Francisco at the time was Net Gain: Expanding Markets Through Virtual Communities by John Hagel and Arthur Armstrong. It was published in 1997, by Harvard Business School Press, and in the circles I was moving through it had something close to cult status. Not cult in the mass-market sense. Cult in the way certain books move through a scene because they give people language for something they can already feel but cannot yet fully explain. This book was giving the startup and tech communities a blueprint for the future. It described how e-commerce and communities were going to become central to the economy. This was subliminal and nobody had really drawn out such a clear map before.

What made Net Gain strange in hindsight is that it was, in effect, describing social networks before we had the phrase “social network” in the way we use it now. It was not describing Facebook, LinkedIn, Reddit, Twitter or the platform economy directly. Those things came later. But it understood something fundamental: that the internet was not only a publishing medium, a shopping channel or a cheaper distribution mechanism. It was a place where people would gather, identify with communities, exchange information, build trust, form preferences and create new markets around shared interest. Wired captured the mood at the time, noting that the book had been handed around at “digerati summits” and had travelled through the scene by word of mouth.

That is what I mean when I say the future felt visible but not clear. Net Gain was not right about every detail, and the businesses that tried to act on that idea in the late 1990s often had the wrong timing, wrong cost structure or wrong monetization model. But the basis of the argument was largely correct. Communities did become markets. Identity, trust, recommendation, reputation, belonging and network effects did become core economic infrastructure. The book saw the outline of the mature internet while the technology was still in its adolescence.

This is exactly the kind of thing that makes a paradigm-shift bubble so hard to reason about from inside it. The early language sounds overexcited because it is reaching for something that does not properly exist yet. Some of the companies built around that language will be ridiculous. Some will disappear. Some will burn absurd amounts of capital. But underneath the bad timing and the failed companies, there can still be a real prediction trying to get out.

You could feel the shape of something. You could feel that the internet was going to change commerce, communication, media, work, identity, finance, software and culture. The mistake was not believing that. The mistake was believing that this first generation of companies, with their first generation of business models, cost structures, burn rates and investor decks, already knew how that future would work.

I went there as a technical consultant for an M&A, but the acquirer were instead impressed with me enough to ask me to move to Silicon Valley and work for them. Sometimes I wonder what my life would have been like had I said yes, but that was not my story and I stayed in London. Leaving the UK would take another 4 years.

Well, I’m going to be bold and maybe a little arrogant and say, I want this to be my Net Gain moment for AI. I can feel the shape of something real. I can feel that AI is going to change commerce, communication, media, work, identity, finance, software and culture. The mistake would be to believe that this first generation of companies, with their first generation of business models, cost structures, burn rates and investor decks, already knows how that future will work. There are aren’t many benefits to getting old, but one of them is recognizing patterns over a long period of time. I have seen this movie before. I know the emotional texture. I know the pattern of belief and disillusionment. I know the way the future feels obvious in hindsight, but is still so hard to get right in the moment.

Most of the companies represented in those business cards no longer exist.

That is not a small detail for me. It is the physical memory of a market story that collapsed. Names, logos, job titles, phone numbers, clever domains, all printed with confidence on expensive card stock. Nearly all of them are gone.

But the internet did not disappear.

That is the part people forget when they talk about bubbles. They talk about the stupidity of the companies that failed, but they forget how much of the failed belief later came true in a different form.

The dotcom bubble did not prove that the internet was fake. It proved that the market had correctly identified a paradigm shift and then mispriced almost everything around it.

By paradigm shift, I mean more than a new product category or a fashionable new tool. I mean a change in the underlying logic of a system: what becomes cheap, what becomes scarce, where the power sits, and which kinds of companies eventually own the mature market.

A paradigm shift is a fundamental, transformative change in the basic concepts, practices, and underlying assumptions of a field, system, or society. It occurs when the usual, accepted way of thinking or doing something is replaced by a new, different model, often driven by new information or technological advancements.

That distinction matters, because I think we are now living through the same kind of moment with AI.

Not the same technology. Not the same market. Not the same infrastructure. Not the same companies.

But the pattern feels familiar:

• the same emotional texture;
• the same feeling that something real has arrived before anyone understands its final form;
• the same rush to price the mature version of a technology while it is still early, awkward and incomplete;
• the same confusion between believing in the paradigm and believing in every company claiming to own it;
• the same danger that the crash, when it comes, will be misread as proof that the whole thing was fake.

It will not be fake. That is the point. The AI bubble will burst because AI is real.

The internet was not wrong. The first market story was wrong.

It is easy to laugh at the dotcom bubble now.

Pets.com became the punchline. A sock puppet, a Super Bowl advert, and a business model that seemed ridiculous in hindsight. People still use it as shorthand for everything absurd about that period.

But step back from the company and look at the idea.

Was selling pet supplies online absurd? No, Chewy exists. Amazon sells pet supplies. Supermarkets deliver pet food. Subscription commerce exists. Heavy, bulky, recurring household purchases are now a perfectly normal part of online retail.

The absurdity was not the concept. The absurdity was the timing, the economics, the logistics, the customer acquisition cost, the infrastructure, the maturity of consumer behaviour, and the belief that a first-wave company could spend its way through all of those missing pieces quickly enough to justify the valuation.

Pets.com was not wrong in the simple way people like to think. It was directionally right and structurally early.

That is often more dangerous than being wrong.

A bad idea can die cleanly. A good idea arriving too early can absorb almost unlimited capital, because every piece of evidence against it can be reinterpreted as temporary friction.

The same was true of online clothes shopping.

In the late 1990s and early 2000s, the objection seemed obvious. You cannot try clothes on through a screen. Sizing is inconsistent. Colours are unreliable. Returns are annoying. People want to feel fabric. Fashion is personal. Shops are social.

All of that was true. It just was not permanently true.

Over time, the missing infrastructure formed around the idea. Broadband improved. Product photography improved. Mobile commerce arrived. Payment systems became trusted. Reviews became normal. Free and easy returns changed behaviour. Warehouses became more sophisticated. Delivery became faster. Consumers became comfortable buying things they had never touched.

Now Zalando exists. ASOS exists. Fashion ecommerce is normal.

The sceptics were not stupid. They were looking at an early, awkward version of a new market and asking whether it could already do the work of a mature one.

The market made the opposite mistake. It looked at the mature destination and priced it into companies that were still learning to walk.

This is the pattern I want to describe in this series.

Not every bubble is the same. Some bubbles are mostly fraud. Some are mostly leverage. Some are mostly speculation. Some are collective madness wrapped around a thin asset.

But some bubbles happen because the world has genuinely changed and the market cannot yet work out where the value will settle.

Those are the dangerous ones. Those are the ones that can be both stupid and right.

2008 felt different

The financial crisis of 2008 did not feel like dotcom to me. It had a different texture.

The dotcom bubble was not clean or innocent, but its emotional centre was technological possibility. The internet was obviously useful. The question was not whether it mattered. The question was how, when, and who would make money from it.

2008 was different. It was a balance sheet crisis, a leverage crisis, a fraud crisis, a ratings crisis, a broken incentives crisis. It was built on opacity and mispriced risk. It was not society discovering a new general-purpose technology. It was finance eating itself and then handing the bill to everyone else.

That distinction matters because people use the word bubble too loosely.

If everything overvalued is just a bubble, we lose the ability to distinguish between a fake story and an early story.

Dotcom was an early story. AI is an early story.

Crypto is more complicated. I do not think it should be dismissed entirely. It may still become a genuine paradigm shift if it eventually displaces fiat currency, settlement systems or trusted intermediaries at scale. But so far, much of its mainstream economic life has been speculation, regulatory arbitrage, niche settlement, corruption, fraud, and early utility in grey or black markets. That does not mean nothing important can come from it. It does mean that its final form is still unproven.

AI feels different. The utility is already obvious.

Developers use it. Writers use it. Students use it. Analysts use it. Designers use it. Customer support teams use it. Researchers use it. People who do not care about technology use it.

That does not mean the valuations are sane. It means the bubble is more dangerous.

The easiest bubbles to dismiss are the ones built on fantasy. The hardest bubbles to survive are the ones built on something true.

Society can feel the future before it understands the business model

Every rare technological shift has a strange early period, but the pattern is usually recognizable in hindsight.

It often moves through something like this:

1. Pioneer phase: the technology arrives looking like a toy, a curiosity, or a niche obsession.
2. Discovery phase: people start to feel the implications before they can explain the business model.
3. Speculation phase: capital, talent and language rush in ahead of settled economics.
4. Overbuild phase: the market starts pricing a mature economy before the infrastructure, habits and profit pools exist.
5. Belief break: the first story stops explaining reality well enough to support the valuations.
6. Consolidation and maturity: the technology keeps spreading, but ownership shifts and the real winners often emerge later.

It happened with railways. It happened with electricity. It happened with computing. It happened with the internet. It happened with mobile. It happened with cloud.

The first version of the story is almost always immature.

The first railway mania was not the final railway economy. The first electrical companies were not the final electrical grid. The first personal computer companies were not the final computing industry. The first web portals were not the final internet. The first mobile apps were not the final mobile economy. The first cloud startups were not the final cloud platform structure.

The first story is usually a cartoon version of the final story.

But cartoons matter.

They compress the future into something people can point at. They create the first wave of capital, talent, language, infrastructure and public imagination.

The problem is that markets do not merely imagine. They price.

And once the market starts pricing phase six while the world is still somewhere around phases two, three or four, the bubble begins.

AI is at exactly this stage.

People can feel the future. They can feel that software is changing. They can feel that work is changing. They can feel that knowledge work, creativity, analysis, coding, support, operations, education and business process design are all becoming fluid in a way that was not true before.

They are right to feel that.

But the market is already trying to decide who owns the future.

• Is it the model labs?
• Is it the hyperscalers?
• Is it the chip companies?
• Is it the SaaS incumbents adding AI features?
• Is it the device and operating system companies?
• Is it the companies with proprietary data?
• Is it whoever owns identity, permissions, workflow, governance and orchestration?
• Is it a new class of AI-operated companies that do not yet exist at scale?

The answer matters because capital markets do not reward vague transformation. They reward ownership, pricing power, margins and durable control.

That is where I think the current story begins to fracture.

The mistake is confusing the gateway with the destination

In the early internet, gateways mattered enormously.

For many people, AOL was the internet. It was access, identity, email, chat, content, community, billing and a sense of safety. It translated the chaos of the open web into something ordinary people could use.

That was historically important.

But being the gateway to a paradigm in its childhood does not guarantee ownership of the mature paradigm.

The mature internet was not AOL.

It became search, broadband, cloud, social networks, ecommerce, mobile operating systems, app stores, streaming, online advertising, creator platforms and walled gardens.

The value migrated.

Some of the early companies survived. Many did not. Some infrastructure was overbuilt and later reused. Some companies became footnotes. Some assets were absorbed. Some ideas reappeared years later with better timing.

The public story changed.

That is the lens through which I see ChatGPT.

ChatGPT made AI legible to normal people. That is an enormous achievement. For many people, ChatGPT is AI in the same way AOL was once the internet.

But that does not mean a chatbot interface, or a centralized model API, is the final economic structure of AI.

It may be the gateway.

The mature form may be something much more distributed, embedded, local, specialized and invisible.

This is where the current AI market story becomes fragile.

A great gateway can create a category. It does not automatically own the category forever.

A paradigm shift is not a company

One of the most important lessons of dotcom is that a paradigm shift and the companies funded around that paradigm are not the same thing.

• The internet was real. Many internet companies were not durable.
• The internet transformed commerce. Many ecommerce companies failed.
• The internet transformed media. Many media startups failed.
• The internet transformed software. Many early software-as-a-service companies failed or were absorbed.
• The internet transformed advertising. Many adtech companies were replaced, consolidated or turned into infrastructure for larger platforms.
• The internet transformed almost everything, but not in the way most pitch decks described at the time.

This is the warning for AI.

• AI can be a real paradigm shift and OpenAI/Anthropic can still be mispriced.
• AI can transform business and Anthropic-style model API economics can still be overestimated.
• AI can make companies more productive and SaaS incumbents can still lose pricing power.
• AI can create enormous demand for compute and some data centre investments can still turn out to be badly timed.
• AI can change software forever and thin AI wrappers can still vanish.
• AI can be the future and investors can still lose money funding the wrong layer.

That last sentence is the whole point.

A paradigm shift does not tell you where the profit pool settles.

It tells you that the map is being redrawn.

Why the current AI bubble feels familiar

The current AI moment has the same strange combination of truth and overreach that I remember from dotcom.

The truth is obvious.

AI is useful. It is not just a demo. It is not just a toy. It is not just autocomplete. It is not just a better search box.

Used well, it changes the cost and speed of producing text, code, analysis, images, summaries, plans, prototypes and decisions. It compresses the distance between intent and output.

That matters, but then comes the overreach.

• Because AI can write code, people assume software development as an economic activity is about to collapse into a few prompts.
• Because AI can answer questions, people assume enterprise knowledge management is solved.
• Because AI can produce plausible plans, people assume autonomous agents can run complex organizations without supervision.
• Because companies are spending heavily on AI infrastructure, people assume demand will compound smoothly enough to justify every data centre, every GPU order and every financing structure.
• Because a model is best today, people assume it has a moat tomorrow.
• Because a SaaS company adds AI features, people assume it has defended itself.
• Because a startup calls itself AI-native, people assume it belongs to the future rather than to the first wave of overfunded experiments.

That is how bubbles form around real technologies.

Reality provides enough evidence to make the fantasy plausible.

The market is pricing the mature AI economy before we know its shape

The mature AI economy will not look like the first AI economy.

That is the safest prediction I can make.

The first AI economy is visible now:

• model labs racing for frontier capability;
• hyperscalers building enormous data centre capacity;
• chip companies selling into the buildout;
• SaaS companies bolting AI features onto existing products;
• enterprises running pilots;
• startups wrapping APIs;
• consultants selling transformation;
• investors looking for the next platform shift.

Some of this will survive. Some of it will become foundational. Some of it will be absorbed. Some of it will become embarrassing.

The mature AI economy may look very different.

Models will likely become smaller, cheaper, more specialized and more local. The best frontier models may still matter enormously, but not every task will need to call a giant remote model. Much intelligence will move closer to the device, the browser, the operating system, the enterprise tenant, the application and the workflow.

Software itself will become more fluid. Today, companies buy SaaS because custom software is expensive, slow and risky. If AI makes custom software cheaper, faster and more maintainable, then large parts of SaaS begin to look like a historical workaround rather than a permanent final form.

Organizations will change too. The most radical AI businesses may not be companies that sell AI. They may be companies that are mostly run by AI. Tiny human teams, large automated surfaces, generated internal software, AI-driven support, AI-driven operations, AI-driven reporting, AI-driven research, AI-driven sales workflows and continuous adaptation.

This is not a distant philosophical idea. It is the direction of travel once software creation, workflow execution and business coordination all become cheaper at the same time.

That is why the capital markets story matters.

The market is not simply betting that AI is useful. It is betting, implicitly, on where the economic control points will be.

I think many of those bets are wrong.

The crash is a change in belief

Bubbles do not burst the moment something becomes overvalued. They can stay overvalued for a long time. They burst when belief changes.

That belief change is rarely neat. It does not require everyone to wake up on the same morning and decide the future is cancelled. More often, the story loses its ability to explain contradictory evidence.

In dotcom, the internet did not stop being important. What changed was belief in the particular companies, burn rates and business models that had been funded.

At some point, investors no longer believed that every company with a domain name was a future monopoly.

• They no longer believed that traffic alone was enough.
• They no longer believed that losses were automatically evidence of ambition.
• They no longer believed that first-mover advantage could overcome bad unit economics.
• They no longer believed that capital could buy time forever.
• The future was still coming, but the story had broken.

That is what I expect in AI.

The bubble will not burst because people stop using AI. It will burst when enough people realize that the current story about who owns AI is wrong.

The convergence point may be a combination of pressures:

• model capability improves, but not fast enough to justify every valuation;
• smaller, local and open models reduce the scarcity premium of centralized APIs;
• enterprises discover that AI value is real, but integration is slower and harder than the demos implied;
• token costs become visible once subsidized and flat-rate usage gives way to harder economics;
• hyperscaler capex becomes harder to justify at the same growth rate;
• SaaS companies see pricing pressure, seat pressure and growth pressure;
• thin AI wrappers fail;
• investors ask who really owns the future.

The technical story does not have to fail. The market story only has to become unbelievable.

After the crash, the real economy begins

The strangest thing about the dotcom crash is that, in hindsight, it looks less like the death of the internet and more like the beginning of its mature life.

The crash cleared out companies, reset valuations, consolidated assets, disciplined business models and created the conditions for the next phase.

The internet after the crash became more useful, more profitable and more central to everyday life.

It also became more centralized, more extractive and more controlled.

The early web was weird, open, independent and chaotic. The mature web became search monopolies, social platforms, cloud platforms, app stores, walled gardens, algorithmic feeds, advertising markets, surveillance capitalism and the consumer as the product.

The mature internet was not the utopia many people imagined. It was better in some ways, worse in others, and much more powerful than the bubble version.

AI may follow the same pattern. The crash will not be the end of AI. It may be the moment AI becomes serious.

After the crash, weak model labs may disappear or be absorbed. Infrastructure will be reused. Inference will become cheaper. AI will become embedded into operating systems, devices, browsers, enterprise platforms and workflows. SaaS will consolidate. Custom workflow software will become more viable. AI-operated companies will become real competitors. The power map will shift toward chips, devices, private data, identity, governance, orchestration and systems of record.

The social consequences will be enormous. Labour markets, education, creative work, software development, management, services and the structure of companies themselves will all be affected. That deserves its own essay.

This series is narrower. It is about business, capital markets and the mistake of pricing the wrong ownership layer.

The first story fails

When I look at those old business cards, I do not see proof that everyone was stupid.

I see proof that people can be close to a real future and still be wrong about almost every investable expression of it.

That is the lesson.

The internet was real. The first story failed. AI is real and I believe the first story will fail again.

• Not because AI is useless.
• Not because the demos are fake.
• Not because nothing will change.

But because everything will change, and the market has not yet understood where that change finally settles.

The AI bubble will burst because AI is real.

That is where this series begins. I hope you’ll enjoy this journey with me into my vision of the future.