AI Infrastructure Is Compressing Decades of Investment Into a Few Years
The AI investment curve looks extreme.
Compared with railways, canals, the dotcom boom, and other major investment cycles, AI investment has accelerated much faster from its pre-boom starting point.
The obvious conclusion is that AI is simply the largest speculative cycle we have seen so far.
I am not sure that tells the full story.
My theory is that AI infrastructure may be following a pattern closer to several cellular investment generations compressed into one much shorter capital cycle.
3G, 4G, and 5G each required enormous infrastructure investment. Those cycles overlapped, but they still developed across successive technology generations and over many years.
According to GSMA, mobile operators invested approximately $1.6 trillion between 2015 and 2023 alone, a period covering major 4G expansion and the rollout of 5G.
AI is different in one important way.
The infrastructure behind it is being built across several major layers at the same time. Capital is flowing into advanced chips, compute capacity, data centres, networking, cooling, and power infrastructure within the same investment window.
In other words, AI may be experiencing the investment intensity of several infrastructure generations without the same separation between them.
That does not mean current AI investment is automatically justified, or that overcapacity is impossible.
But it may help explain why the curve is so steep.
Maybe AI is not one infrastructure cycle.
Maybe it is several generations of infrastructure investment compressed into one.
AI investment is accelerating faster than past innovation booms
Investment growth from the pre-boom trough, comparing AI with major historical infrastructure and technology cycles.
The unusual feature is not only the level of investment. It is the speed at which capital is being deployed.
-
01AI investment is accelerating faster than past innovation booms. BIS data shows AI-related investment rising much faster from its pre-boom trough than railway mania, the dotcom boom, the Roaring 20s, or canal investment.
-
02The steep curve may partly reflect investment compression. AI is attracting capital into semiconductors, compute, data centres, networking, cooling, and power infrastructure during the same investment window.
-
03Telecom infrastructure offers a useful comparison. Mobile operators invested approximately $1.6 trillion between 2015 and 2023, while major network infrastructure developed across successive 4G and 5G deployment cycles.
-
04AI may be several infrastructure cycles happening at once. The comparison with 3G, 4G, and 5G is not technological. It is about capital deployment patterns and the concentration of infrastructure investment into a much shorter period.
-
05Investment compression does not eliminate overinvestment risk. Infrastructure can be strategically important and still be overbuilt. The valuation question is whether capacity, utilisation, and future economics eventually justify the capital deployed.
Topics covered in this article +
- Why the AI investment curve looks so unusual
- What telecom infrastructure tells us about investment cycles
- Why AI may be compressing several infrastructure cycles into one
- How investment compression may explain the steeper AI curve
- What AI infrastructure investment means for valuations
- Why investment compression does not remove overinvestment risk
- Key takeaways and answers to common questions
Why the AI Investment Curve Looks So Unusual
The BIS chart is striking because AI investment separates from previous investment cycles very quickly.
Railway investment, canal construction, the Roaring 20s, and the dotcom boom all experienced periods of rapid capital deployment. Their investment curves, however, developed more gradually. Within roughly three years of the pre-boom trough, AI-related investment had already risen to more than four times its starting level.
There is an important distinction in how the chart should be read. It does not compare the absolute dollars invested in railways, canals, the internet, and AI. Each cycle is measured relative to its own pre-boom trough, which means the chart is primarily showing the speed at which investment accelerated from its starting point.
On that basis, AI stands out very clearly.
The simplest explanation is speculation. Companies are concerned about falling behind, investors are competing for exposure to the next major technology cycle, and hyperscalers are committing enormous amounts of capital before the long-term economics of AI infrastructure are fully understood.
That is probably part of the story.
But investment acceleration can also be affected by how much infrastructure needs to be funded within the same period.
This is where AI differs from many previous infrastructure cycles.
The current AI buildout is creating capital requirements across semiconductors, compute, data centres, networks, cooling systems, and power infrastructure simultaneously. These are not isolated investment categories. More compute increases the need for physical capacity. More data centre capacity raises networking and cooling requirements. Larger facilities then create additional pressure on electricity generation and grid infrastructure.
The result is a concentration of capital requirements.
Instead of moving through clearly separated infrastructure generations, AI is forcing several major infrastructure layers to expand during the same investment period.
The steep curve may therefore reflect more than investor enthusiasm.
It may also reflect several generations’ worth of infrastructure requirements arriving at once.
What Telecom Infrastructure Tells Us About Investment Cycles
Telecom provides a useful comparison because mobile networks also required enormous amounts of infrastructure capital.
The difference is how that capital was deployed over time.
3G expanded mobile internet access and created a new generation of data demand. 4G required another major infrastructure cycle to support higher speeds, greater capacity, and increasingly data-heavy applications. 5G then introduced further investment in network density, equipment, and higher-performance infrastructure.
These investment cycles were never perfectly separated. Operators continued upgrading existing networks while deploying newer standards, and some physical infrastructure supported multiple network generations.
Still, the broader pattern was successive.
Each new cellular generation created another period of infrastructure expansion, and operators had years to increase network capacity, extend coverage, upgrade equipment, and adapt to changing patterns of demand.
The scale of this investment was enormous. According to GSMA, mobile operators invested approximately $1.6 trillion between 2015 and 2023, largely to deploy 4G and 5G infrastructure.
That figure is useful because it shows how capital-intensive modern digital infrastructure can become.
But the more important comparison for AI may be the timeline.
Telecom moved through 3G, 4G, and 5G across successive technology generations. AI infrastructure is currently trying to scale compute, semiconductors, data centres, networks, cooling, and power capacity during the same period.
The analogy is not that chips are equivalent to 3G or that data centres are equivalent to 4G.
The comparison is about the concentration of capital deployment.
This is the basis of my theory.
AI may be experiencing something closer to the investment intensity of several cellular infrastructure generations, but without the same separation between investment cycles.
What telecom funded across successive generations, AI may be trying to fund within one compressed investment period.
That would not prove current AI capex is justified.
But it could explain why the investment curve rises so much faster.
Why AI May Be Compressing Several Infrastructure Cycles Into One
The telecom comparison becomes more interesting when we look at how AI infrastructure actually scales.
AI does not depend on one isolated infrastructure category. Each major layer creates demand for the next.
More AI workloads require greater compute capacity. Greater compute capacity increases demand for advanced semiconductors, accelerators, and servers. Those systems need physical data centre capacity, high-speed networking, storage, and increasingly complex cooling systems.
Then comes power.
Larger AI facilities require enormous and reliable electricity supply. As data centre capacity expands, the infrastructure discussion moves beyond technology companies and into generation capacity, transmission, grid connections, and physical energy infrastructure.
The important point is that these requirements are developing together.
AI is not waiting for the semiconductor cycle to finish before data centres expand. Data centre investment is not waiting for networking infrastructure to reach maturity. Power requirements are already becoming part of the same capital allocation discussion.
That creates a compounding infrastructure effect.
Investment in one layer can expose a capacity constraint elsewhere, which then triggers another round of capital deployment. More chips create demand for more facilities. More facilities expose power constraints. Power constraints create new infrastructure requirements.
This is why I think the comparison with several cellular generations is useful.
3G, 4G, and 5G each introduced new infrastructure requirements across successive technology cycles. AI may be facing a similar breadth of capital requirements, but with much less time separating them.
The current investment cycle is trying to solve compute, physical capacity, connectivity, cooling, and power constraints simultaneously.
That is what I mean by investment compression.
AI may be pulling several infrastructure investment cycles into the same capital allocation window.
More AI workloads
Higher model usage and larger workloads increase demand for compute capacity.
More compute infrastructure
Additional chips and servers require physical facilities, networking, and cooling.
More data centre capacity
Larger facilities create higher electricity demand and expose grid constraints.
More supporting infrastructure
Power and capacity constraints create another round of capital deployment.
How Investment Compression May Explain the Steeper AI Curve
This brings us back to the BIS chart.
The AI investment curve rises much faster than the historical cycles included in the comparison. It is easy to read that slope as evidence of a larger speculative boom.
That may be part of the explanation, but the speed of capital deployment also matters.
If major infrastructure requirements are spread across successive cycles, investment has more time to accumulate. Capital is deployed, capacity is built, demand develops, and the next major infrastructure cycle follows.
A compressed cycle behaves differently.
When compute, semiconductors, data centres, networking, cooling, and power capacity all require expansion during the same period, capital commitments can rise much faster from the initial investment trough.
The slope becomes steeper because the investment window is shorter.
This is the main reason I am cautious about interpreting the BIS chart as a simple comparison of speculative intensity. The graph clearly shows that AI investment is accelerating unusually quickly, but it does not tell us why that acceleration is happening.
My view is that part of the difference may be explained by infrastructure compression.
AI may be concentrating capital requirements that, in another technology cycle, could have appeared across several successive investment phases.
The telecom analogy helps illustrate the point. Imagine the infrastructure demands associated with several cellular generations arriving within a much narrower period. The resulting investment curve would probably look very different from a market that had years to move from one generation to the next.
That does not make every AI investment rational.
It also does not mean current capital expenditure will generate sufficient returns.
It simply changes how I interpret the steepness of the curve.
A faster investment cycle is not automatically evidence of a larger bubble. It may also be evidence that more infrastructure is being funded in less time.
Compute, chips, data centres, networks, cooling, and power
Several capital requirements arriving within the same investment window
Faster capital deployment from the pre-boom trough
What Investment Compression Means for AI Valuations
If AI infrastructure is being built through a compressed investment cycle, the valuation implications are not evenly distributed across the market.
Companies positioned around scarce infrastructure can benefit directly from the current buildout. Semiconductors, compute capacity, data centre infrastructure, networking, and power-related technologies may all attract premium valuations when investors believe supply constraints are difficult to replicate or strategically important.
That does not mean every infrastructure company deserves a premium.
The critical question is whether the company controls a durable bottleneck or is simply benefiting from temporary scarcity.
Those are very different valuation cases.
A company with proprietary technology, constrained capacity, strong switching costs, or a difficult-to-replicate position in the infrastructure stack may retain pricing power even as the market expands. A company whose advantage depends mainly on short-term capacity shortages may face a different outcome once supply catches up.
Applied AI companies sit on the other side of the cycle.
Many of them are infrastructure consumers. Their economics can improve as compute becomes cheaper, model access broadens, and AI infrastructure becomes more widely available.
That creates an important valuation split.
The companies benefiting from scarcity today may face pressure when infrastructure capacity expands. The companies consuming that infrastructure may see margins improve as the same constraints ease.
This is why I do not think “AI company” is a useful valuation category on its own.
Investors need to understand where the company sits in the infrastructure cycle and how its economics change as the market matures.
Three questions matter in particular.
Does the company control a scarce infrastructure layer or consume it?
Is current scarcity a durable competitive advantage or a temporary bottleneck?
Do the company’s margins and growth economics improve or weaken as AI infrastructure becomes more widely available?
The answers can lead to very different valuation conclusions, even for companies operating within the same AI market.
Investment compression may help explain the unusually steep AI capex curve.
But for valuation, the more important question is who benefits from that compression today, and who benefits when it eventually starts to unwind.
- 1 AI investment is accelerating unusually quickly from its pre-boom trough. The BIS comparison shows a much steeper early investment curve for AI than for railway mania, canal construction, the Roaring 20s, or the dotcom boom.
- 2 The steep curve may partly reflect investment compression. AI is forcing semiconductors, compute capacity, data centres, networking, cooling, and power infrastructure to scale within the same capital investment window.
- 3 The cellular infrastructure cycle provides a useful comparison. 3G, 4G, and 5G required successive rounds of infrastructure investment, while AI may be experiencing the investment intensity of several infrastructure generations within a much shorter period.
- 4 Mobile operators invested approximately $1.6 trillion between 2015 and 2023. The GSMA figure demonstrates the capital intensity of modern digital infrastructure, but telecom had more time to deploy capital across successive network generations.
- 5 A steeper investment curve does not automatically prove a larger speculative bubble. Faster capital deployment may also result from more infrastructure requirements being funded within a shorter period, although overcapacity and poor capital allocation remain possible.
- 6 Investment compression creates different valuation outcomes across the AI market. Companies controlling durable infrastructure bottlenecks may benefit from scarcity, while applied AI companies may benefit as compute and infrastructure capacity become cheaper and more widely available.
