I wonder if the historical analogy breaks in two interesting places.
First, we don’t yet have evidence of the platform effects that made railroads and the internet self-reinforcing networks. So far the stronger evidence is that AI makes particular forms of cognition cheaper.
Second, previous infrastructure booms left substantial durable assets after the capital-market excess was washed out. GPUs depreciate technologically in 3–5 years, models in months, and much of the data-centre investment is specialised. Perhaps the most durable asset being created is energised land.
So if there is an AI bubble, I’m not sure the post-bubble economy inherits the same kind of productive infrastructure. The analogy may be describing the investment cycle while importing the mechanism that made previous infrastructure bubbles ultimately productive.
Good question to ask. I should probably have been more specific in the piece about what the 'leave behind' looks like.
In shipping logistics, the container was the innovation that allowed a standardised global logistics network to develop. The leave behind was the system it enabled, not the containers. Similarly, although the railroad boom left behind a lot of track, arguably it was the new transport system with all its stations, connections, logistics, etc that was the leave behind. Tracks last a long time, but every section gets swapped out, upgraded or improved over time.
In this wave of change, I don't see either the LLMs or GPUs as the leave behind. Most models currently are based on 6+ years of depreciation for GPUs, but the models are way more transient and disposable than that.
For me, the leave behind is re-tooled, re-imagined networks and socio-technical systems based on intelligence as a utility. I think that will be the platform that this era creates for the future. which includes everything from harnesses, context, knowledge systems, etc through to agentic orchestration, governance and management. Given the speed of innovation in open./ small LLMs and in more efficient computing, I think LLMs and GPUs will ultimately be commoditised, just like containers.
I can believe the durable “leave behind” is the socio-technical system you describe. But that raises the question of whether that system has the same economic leverage as the systems created by previous infrastructure waves.
A durable layer of harnesses, context, orchestration, governance and management could certainly make cognition cheaper and organisations better at using it. But that doesn’t necessarily produce a self-reinforcing platform effect, or a large aggregate productivity effect. And we’ve yet to see evidence of a platform effect.
AI may leave behind a durable new way of organising intelligence; and that may still only add a relatively modest amount to aggregate TFP/GDP. The historical analogy seems to assume the second follows from the first.
I agree we have not yet seen the platform effect, which is consistent with previous waves. Systems take time to develop, and then hopefully a gradual compounding of efficiency innovations on top (same with standardised shipping, etc). Let's see...
But I think the prize is bigger than just better / cheaper ways to use cognition in companies. I think it is a re-wiring of what we mean by organisational operating systems, and indeed what we mean by software as well. So it will take time, and it will not be an easy transition. It has the potential to significantly reduce the cost of action in business and other domains, but that is notoriously hard to measure using old methods like TFP/GDP.
Yes, and I think this gets to the distinction I was trying to make.
Unit drive, for example, eliminated a lot of porters. Replacing steam with electric traction could deliver an immediate fuel saving. Both are productivity improvements, and both could be part of a much larger transformation. But neither is the compounding mechanism.
The really interesting historical effect came when the new capability changed what other firms could do, which changed what their suppliers could do, which changed the economics of other activities, and so on. The value compounds across the boundaries between firms, rather than simply accumulating inside each firm.
So I agree that an AI-native organisational operating system could be a very significant 'leave behind'. The question I’m still stuck on is what makes it compounding: what is the mechanism by which making one organisation better at using intelligence makes the next organisation more valuable at using it?
That’s the platform effect I think the historical analogy is quietly assuming. And I don’t think we’ve seen it yet.
I wonder if the historical analogy breaks in two interesting places.
First, we don’t yet have evidence of the platform effects that made railroads and the internet self-reinforcing networks. So far the stronger evidence is that AI makes particular forms of cognition cheaper.
Second, previous infrastructure booms left substantial durable assets after the capital-market excess was washed out. GPUs depreciate technologically in 3–5 years, models in months, and much of the data-centre investment is specialised. Perhaps the most durable asset being created is energised land.
So if there is an AI bubble, I’m not sure the post-bubble economy inherits the same kind of productive infrastructure. The analogy may be describing the investment cycle while importing the mechanism that made previous infrastructure bubbles ultimately productive.
Good question to ask. I should probably have been more specific in the piece about what the 'leave behind' looks like.
In shipping logistics, the container was the innovation that allowed a standardised global logistics network to develop. The leave behind was the system it enabled, not the containers. Similarly, although the railroad boom left behind a lot of track, arguably it was the new transport system with all its stations, connections, logistics, etc that was the leave behind. Tracks last a long time, but every section gets swapped out, upgraded or improved over time.
In this wave of change, I don't see either the LLMs or GPUs as the leave behind. Most models currently are based on 6+ years of depreciation for GPUs, but the models are way more transient and disposable than that.
For me, the leave behind is re-tooled, re-imagined networks and socio-technical systems based on intelligence as a utility. I think that will be the platform that this era creates for the future. which includes everything from harnesses, context, knowledge systems, etc through to agentic orchestration, governance and management. Given the speed of innovation in open./ small LLMs and in more efficient computing, I think LLMs and GPUs will ultimately be commoditised, just like containers.
I can believe the durable “leave behind” is the socio-technical system you describe. But that raises the question of whether that system has the same economic leverage as the systems created by previous infrastructure waves.
A durable layer of harnesses, context, orchestration, governance and management could certainly make cognition cheaper and organisations better at using it. But that doesn’t necessarily produce a self-reinforcing platform effect, or a large aggregate productivity effect. And we’ve yet to see evidence of a platform effect.
AI may leave behind a durable new way of organising intelligence; and that may still only add a relatively modest amount to aggregate TFP/GDP. The historical analogy seems to assume the second follows from the first.
I agree we have not yet seen the platform effect, which is consistent with previous waves. Systems take time to develop, and then hopefully a gradual compounding of efficiency innovations on top (same with standardised shipping, etc). Let's see...
But I think the prize is bigger than just better / cheaper ways to use cognition in companies. I think it is a re-wiring of what we mean by organisational operating systems, and indeed what we mean by software as well. So it will take time, and it will not be an easy transition. It has the potential to significantly reduce the cost of action in business and other domains, but that is notoriously hard to measure using old methods like TFP/GDP.
Yes, and I think this gets to the distinction I was trying to make.
Unit drive, for example, eliminated a lot of porters. Replacing steam with electric traction could deliver an immediate fuel saving. Both are productivity improvements, and both could be part of a much larger transformation. But neither is the compounding mechanism.
The really interesting historical effect came when the new capability changed what other firms could do, which changed what their suppliers could do, which changed the economics of other activities, and so on. The value compounds across the boundaries between firms, rather than simply accumulating inside each firm.
So I agree that an AI-native organisational operating system could be a very significant 'leave behind'. The question I’m still stuck on is what makes it compounding: what is the mechanism by which making one organisation better at using intelligence makes the next organisation more valuable at using it?
That’s the platform effect I think the historical analogy is quietly assuming. And I don’t think we’ve seen it yet.