We know exactly what AI costs, but we are unable to measure what it produces. And that is a serious problem

We know very well the cost of developing AI: mammoth data centershe electricity consumption skyrocketedhe tech capex through the roof… The problem is that it seems that all this is not having a return, or not enough to justify tremendous investment. The fear of the bubble is justified, but maybe we were wrong and the problem is another: that our measuring tape is broken.

The hidden production. In an extensive and in-depth analysis in the newsletter Semianalisysuse the term ‘dark output’ in reference to the economic value that AI is generating, but which current measurement systems do not see well and therefore does not have an impact on GDP. This hidden production has two aspects:

• Hidden production by substitution: These are jobs that used to be done by a human for a price and that can now be done by AI for a fraction of that cost. There is a very graphic example with the writing of wills, a job that historically cost $400, which had dropped to $150, and in a single year AI has plummeted to $0.50. The work is done, but the economic transaction disappears from the data.
• New production that remains hidden: On the other side are the jobs that were not done because they were too expensive, but that AI has made so cheap that they can now be done. The example that Semianalysis provides are the bibliographic reviews whose price was up to $2,000 and that made them a very exclusive service. Now with AI you can do one of these reviews on all types of projects. The problem is that the economic trace is non-existent, except for the use of tokens or payment of subscriptions.

Why it is important. The thesis of the analysis is that we are not facing a bubble, but that we are not measuring well the return that AI is producing and that is a problem that goes far beyond a simple statistical debate. Macroeconomic data is the metric by which investors detect real growth, central banks adjust interest rates, and companies decide whether to hire or automate. Making decisions of this caliber based on inaccurate data can have serious consequences.

The difficulty of measuring it. Services and intellectual labor are much more complicated to measure than physical goods. It is very easy for a furniture factory to measure whether new machinery allows it to manufacture more chairs in less time. AI is helping to do tasks such as programming, writing documents, summarizing them or creating briefings and the way we measure it is the tokens consumed. The problem is that consuming more tokens can result in enormous benefits for the company, but they can also produce bad code and bad summaries. The value is in the production, in the output, not in what we spend to get to it.

Precedents. Something similar happened during the computer boom in the 80s and 90s. At this time, macroeconomic data were not capable of detecting what the computer revolution was bringing. The solution did not arrive until 2013, when R&D and investment in intellectual property were included in GDP accounting. The result was that 3.6 trillion dollars were added retroactively, showing that in the year 2000 alone it represented 30% of the GDP.

The other precedent is the so-called care economy, in reference to all the domestic and care work carried out mainly by women without receiving remuneration. The International Labor Organization estimated in 2018 that 16.4 billion hours of unpaid care work were performed, which would be equivalent to 11 trillion dollars or 9% of global GDP.

Yes, but. That it is necessary to update our measuring tape does not detract from the fact that investment in AI infrastructure is truly dizzying. In 2025, big tech companies will invest $410 billion in AI and in 2026 the plan is to exceed the 650 billion dollars. The chief economist of Golman Sachs said that the contribution of all this crazy investment to US GDP was “basically zero.” In this sense, it is as risky to say that we are facing a bubble about to burst due to excess spending, as it is to assume that there is immense invisible wealth justifying every dollar invested.

Image | Xataka with Gemini

In Xataka | “The biggest mistake of all time”: Bill Gates let slip 400 billion when Microsoft didn’t buy Android