I’ve long talked about AI’s limitations on a technical basis; it’s not the magical technology some make it to be. Huge potential, still mostly unmet.

But it’s hard to deny it’s a transformational technology. The question is more like: when? When will it change the world?

We’ll go back in history to answer that question. And you might not like the answer I have for you. And at the end, I’ll give you my two cents on what I think will happen.

You might not like that either.

Let’s dive in.

We, humans, love to ignore history. But as Mark Twain (allegedly) once said: “History doesn’t repeat, but it rhymes.” So, what does history tell us about AI’s chances to change the world?

And the short answer is that it will, but it will take time, maybe too much time.

Nobel Laureate Robert Solow had a great quote back in 1987: “You can see the computer age everywhere but in the statistics.”

At that time, it had been almost two decades since a group of engineers at Intel created the microprocessor in 1971, a development that eventually led to the personal computer.

But in 1990, three years after saying those famous words, just 20 million personal computers were sold. But PCs are hardly the only example; electricity is another great one. While the first commercial power station was built in 1882, it was not until 1920, four decades later, that electricity surpassed steam as the dominant form of horsepower in the US economy.

We can go even further back to another great example: steam engines. In their case, the diffusion took even longer. James Watt patented his steam engine in 1769.

By 1830, its penetration into the economy, measured by productivity, was still trivial, and we didn’t see its real impact until the third quarter of the nineteenth century, almost a century after Watt’s patent.

Circling back to PCs, it wasn’t until the mid 1990s that the PC revolution really took off, and it took an outrageous 3 decades to reach 50% PC adoption.

It was particularly slow in its contribution to the most important metric of them all: productivity. For the first two decades since its conception, the Internet revolution had very little impact on macro numbers, if any.

And it might be the case that it happens to AI, too.

What if a technology is just too good at its job? Several months ago (maybe more than a year, actually), I argued that AI could actually decrease the size of the economy.

The rationale was based on two ideas:

With enough sophistication (not the case today), AI lowers barriers to competition in its areas of impact, commoditizing those industries and leading to price wars in which prices fall faster than demand grows.

In other words, while price reductions do increase demand (Jevons’ Paradox), because GDP (Gross Domestic Product) is measured as price times quantity, if prices fall faster than quantities rise, the overall market shrinks.

Agriculture is always a great example. Despite producing more food than at any time in history, agriculture’s impact on major economies has nosedived (except in some African countries, where agriculture accounts for a minimal share of national GDP, with the US’s share falling below 1%).

Nonetheless, while in 1840 agriculture accounted for more than 60% of total jobs in the US, most developed economies have since transitioned to services as the main driver of employment.

The answer as to why is quite simple: huge increases in productivity (like the one we see in Sweden below) decreased prices severely, way more than demand rose, overall decreasing agricultural output.

Adding to the fact that many transformational technologies can shrink output, we have to add Baumol’s cost disease to the picture.

Baumol’s cost disease is the idea that some sectors become more expensive not because they are getting worse, but because they cannot raise productivity as fast as the rest of the economy.

Productive sectors have higher wages, attracting more talent. To retain talent, less productive sectors raise wages, creating a downward spiral of declining productivity.

A beautiful example is a string quartet.

In the 18th century, you needed four musicians and a certain amount of time to play a Mozart quartet. Today, you still need four musicians and roughly the same time. Productivity has barely improved.

But those musicians live in an economy where other sectors, like manufacturing or software, have become far more productive, so wages across the economy rise. To keep musicians from leaving for better-paid jobs elsewhere, orchestras must also raise wages, even though each performance is not much more “productive” than before.

This has a big effect on the macro picture, as a nation’s GDP becomes dominated by essential, low-productive sectors. Healthcare, education, and industries where productivity doesn’t change as fast as others but are essential, concentrate the majority of the output because quantity is not negotiable (they are essential), and prices continue to rise over time, driven by non-productive wage hikes.

The lesson? Fascinatingly, if AI’s impact is not pervasive across all industries, it may be extremely hard to see, which poses a very real danger to the industry: AI’s invisible output.

One thing most tech bros in San Francisco miserably fail to understand is that being useful does not automatically mean being more productive. At least not in the way we measure productivity.

Everyone and their grandma can see AI is useful, but justifying the alleged massive rises in productivity is proving way harder than we thought.

If we are spending several basis points of global GDP on AI, we should expect in return a technology that massively transforms productivity, meaning the world starts to do a lot more with a lot less.

But are we? We’ll talk about costs later, but the other big issue we face is what some are calling ‘invisible output’. In the sea of tokens being generated worldwide, many productivity gains “benefit no one.” Let me explain.

What two years ago would have required consulting my lawyer for any trivial issue can now, at least for some of them, be handled by AI.

As a self-employed person with two companies in Spain, I receive a lot of attention from ‘Hacienda’ (our IRS), receiving a decent amount of letters. These are famously cryptic, but I don’t need my lawyers anymore for such things; I just ask ChatGPT.

This means that my lawyer is not getting paid less. Of course, one could take the other side of the argument and say, “Well, but your lawyer can now attend to more people.”

Maybe, but that’s the thing; there’s a certain threshold of capabilities AIs can take away from what before would have been transactional interactions feeding into our GDP output metrics.

Don’t get me wrong, you would be unprecedentally foolish to think you don’t need a human lawyer anymore; it’s just that, in some way, AIs have taken the role of “consigliere for all menial stuff” that increases users’ productivity without appearing elsewhere.

Or, if somewhere, it’s on the AI Lab’s top line, if not for the fact that they have massively commoditized their businesses or, in the cases of ChatGPT or Gemini Search, offer many of such tokens for free.

But even if that ChatGPT supporting my business does appear in OpenAI’s top line (I am a paying subscriber), it’s on an entirely different sector and mixed with a lot of other stuff, making it almost impossible to measure accurately.

It must be mentioned that one can also be critical of the measurements themselves. For example, the way we measure the productivity of the public sector is, by definition, flawed.

Since you don’t pay for many of those services, the only way we can measure them is by considering production costs and using volume as the output.

A public school in Spain is “free”, so to measure its productivity, we just look at costs (wages and others) and measure output by the “volume of educational services delivered”, using metrics such as student-hours and even academic attainment. But as there’s no price, how productive they really are, how much value it’s delivering, well, it’s hard.

Nevertheless, the fact that the number of new graduates is growing doesn’t say anything about the quality of education. In fact, in Spain, we call this “titulitis”, a phenomenon in which we have a huge number of “highly educated” unemployed workers while construction workers, carpenters, and plumbers have all the work they want, and more.

Is Spain’s public university sector really that productive if we’re sending most of these kids directly into unemployment?

To be clear, I’m not saying public education is not a great success of society (I myself benefited from it during my undergraduate years, and I’m eternally thankful for it); I’m just questioning the quality of the measures of productivity.

But leaving this quite hard-to-deal-with problem aside for a moment, let’s go back to history to answer: why did it take so long for other technological disruptions to transform the economy?

In most cases, it was a mixture of three things: a solution looking for a problem to solve, unsophisticated approaches, and, of course, costs.

On the latter, steam engines and electricity were simply very early to the party. In the steam engine’s case, its very slow diffusion was held back by fuel inefficiency and very slow price declines.

This is the easiest example answer; it just wasn’t worth it until it was, and the technology had to develop and become cheaper to be fairly adopted.

Sounds familiar?

Electricity’s case was a little bit more nuanced, and quite frankly, much more interesting.

Although the “War of the Currents” between Edison and Tesla didn’t help, as Edison created a lot of fear around Tesla’s alternating current despite being much safer and, in hindsight, the only viable solution for long-distance transmission (we need very high voltages to transmit enough power without current, and thereby losses, being small enough, and also we only knew how to step down voltage using AC transformers), in this case the technology was quite ready.

In particular, not until we reframed our factories to electricity did electricity become adoptable. Surprisingly, electricity was only widely adopted in the 1920s, almost three decades after the first transformer was built.

What changed? Unit drives.

Not until factories were retrofitted to handle several individual unit drives (motors with their own, individual control units), forcing a complete redefining of the factory’s layout, did electricity finally make sense.

Put another way, part of the delay in exploiting the potential industrial productivity gains offered by electricity was simply due to the durability of old manufacturing plants that embodied technology adapted to the regime of mechanical power derived from water and steam.

Sounds familiar?

With PCs, it was actually mostly about being a solution looking for a problem to solve. A hilarious example of this was Apple’s marketing at the time, where they couldn’t even describe the use case and “begged” users to tell them what it was.

Sounds familiar?

It had some price dynamics, too: not until Intel was forced to drop prices due to competition and computers became affordable (especially with MOS Technologies selling its 6502 for $25, $150 in today’s dollars), was Steve Wozniak “allowed” to tinker with the technology, leading to the first Apple computer prototype.

Now that we know what determines diffusion, clear use cases, system readiness, and cost, it’s time to see how AI is faring in each one. And, ladies and gentlemen, it’s not good.

Behind the paywall, we analyze AI’s situationship across all levers of diffusion, while also using history again, a last-century, lesser-known revolution that offers incredible insight into how most AI investments are going to turn out.

Because AI will succeed. Most investors, however…