The productivity story makes more sense to me when assistance and delegation are kept separate. A tool can make someone faster without becoming a reliable coworker. The stronger claim needs evidence about handoff, review, ownership and where the work actually lands.
I think we should set aside the GPT assumption for a moment—blockchain passed many of the same early tests. Technological capability does not predict economic transformation, so we should focus on what we can actually observe.
On the other hand, if we separate the economic activity due to the speculative data centre build-out from realised productivity gains, the evidence for firm- or economy-wide transformation is much scarcer. Your data shows this tension nicely.
There is a lot of comparison with electrification, but the analogy needs care. Electrification’s breakthrough came from decoupling power consumption from generation. The key shift was not replacing steam engines with electric motors; it was changing the architecture of production by relocating power generation and distributing power through a network. Steam power generation benefits from the square-cube law: larger centralised generators are more efficient, and the more demand the network attracted, the more valuable the infrastructure became.
The benefits we see from LLMs today emerge from the tip of the tool: the interaction between a model, prompt, and context. We have not seen a network effect from AI. This is not to say such a network will not emerge, but we should be honest that we do not see one today.
So perhaps the more interesting question is not:
> “How much is AI contributing to GDP?”
but:
> “How much of the current AI GDP contribution survives after removing the construction boom required to make AI possible?”
The evidence that micro AI productivity will filter through macro is the same as it was for blockchain. If there is a difference then it’s yet to be found.
Critical notes, especially in the closing paragraphs. It also seems as if the firms that are expanding either capital intensity or capex are also the ones most likely to refine their workflows (probably already occurring with newer models as noted) or unlock human bottlenecks.
This reminds me of the adoption of ERP software. For the first two decades or so, there were few productivity gains. Now ERP is fundamental to a company and its absence would reduce productivity dramatically.
The reason was that it took time to redesign workflows, incentives, and complementary processes across the company, to use Tedeschi's words. Downstream bottlenecks needed to be removed for the micro-gains to become macro-gains. Technology doesn't improve productivity; use of the technology does.
There is no reason to believe that AI will be any different. The question is, what is the time scale? Will it be two decades (certainly not), ten years, five years, or two? My money is on the middle of the range.
This has two consequences: people and society will have more time to adapt (good luck!) and the markets will be much disappointed in how rapidly the massive capital investments will pay off (good luck!).
Still waiting to see if those meager gains persist once everyone has the same AI and competition levels up ... early productivity spikes may fade when the edge gets competed away.
Would it be correct to infer that the evidence covered in this piece broadly favours O-ring models of automation? The kind of economic analysis that people like Alex Imas have put forward? And with regard to redesigning workplaces and workflows, this is a process that could take years or decades, at least if we’re going off prior technologies, right?
The productivity story makes more sense to me when assistance and delegation are kept separate. A tool can make someone faster without becoming a reliable coworker. The stronger claim needs evidence about handoff, review, ownership and where the work actually lands.
I think we should set aside the GPT assumption for a moment—blockchain passed many of the same early tests. Technological capability does not predict economic transformation, so we should focus on what we can actually observe.
We have abundant evidence that LLMs are powerful tools. Google's latest research is a good example of this: https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/
On the other hand, if we separate the economic activity due to the speculative data centre build-out from realised productivity gains, the evidence for firm- or economy-wide transformation is much scarcer. Your data shows this tension nicely.
There is a lot of comparison with electrification, but the analogy needs care. Electrification’s breakthrough came from decoupling power consumption from generation. The key shift was not replacing steam engines with electric motors; it was changing the architecture of production by relocating power generation and distributing power through a network. Steam power generation benefits from the square-cube law: larger centralised generators are more efficient, and the more demand the network attracted, the more valuable the infrastructure became.
The benefits we see from LLMs today emerge from the tip of the tool: the interaction between a model, prompt, and context. We have not seen a network effect from AI. This is not to say such a network will not emerge, but we should be honest that we do not see one today.
So perhaps the more interesting question is not:
> “How much is AI contributing to GDP?”
but:
> “How much of the current AI GDP contribution survives after removing the construction boom required to make AI possible?”
Does seem like a big difference with Crypto: micro level productivity using AI will eventually feed through into economic growth
The evidence that micro AI productivity will filter through macro is the same as it was for blockchain. If there is a difference then it’s yet to be found.
Critical notes, especially in the closing paragraphs. It also seems as if the firms that are expanding either capital intensity or capex are also the ones most likely to refine their workflows (probably already occurring with newer models as noted) or unlock human bottlenecks.
This reminds me of the adoption of ERP software. For the first two decades or so, there were few productivity gains. Now ERP is fundamental to a company and its absence would reduce productivity dramatically.
The reason was that it took time to redesign workflows, incentives, and complementary processes across the company, to use Tedeschi's words. Downstream bottlenecks needed to be removed for the micro-gains to become macro-gains. Technology doesn't improve productivity; use of the technology does.
There is no reason to believe that AI will be any different. The question is, what is the time scale? Will it be two decades (certainly not), ten years, five years, or two? My money is on the middle of the range.
This has two consequences: people and society will have more time to adapt (good luck!) and the markets will be much disappointed in how rapidly the massive capital investments will pay off (good luck!).
Excellent paper!
Still waiting to see if those meager gains persist once everyone has the same AI and competition levels up ... early productivity spikes may fade when the edge gets competed away.
This chart looks familiar: https://substack.com/home/post/p-194305598
Very interesting! Thank you for the write-up!
Would it be correct to infer that the evidence covered in this piece broadly favours O-ring models of automation? The kind of economic analysis that people like Alex Imas have put forward? And with regard to redesigning workplaces and workflows, this is a process that could take years or decades, at least if we’re going off prior technologies, right?
Excellent post, btw
You may find this useful too https://www.federalreserve.gov/econres/notes/feds-notes/the-ai-buildout-and-the-economy-publicly-available-data-to-assess-ais-impact-20260717.html