AI is becoming cheaper and more efficient. But that may not reduce its need for infrastructure. It may simply move the bottleneck somewhere else.
In my previous article, I explored the physical bottlenecks facing AI infrastructure. Data centers need electricity, grids, equipment, materials and skilled people.
But technology does not stand still. AI becomes more efficient, chips improve and new infrastructure gets built.
So shouldn’t today’s bottlenecks eventually disappear?
Perhaps.
But there is another possibility:
The bottlenecks may not disappear. They may simply keep moving.
1. Cheaper AI May Mean More Infrastructure, Not Less
AI is becoming dramatically cheaper to use.
According to the Stanford AI Index, inference costs have collapsed. Running a model with roughly GPT-3.5-level performance cost about $20 per million tokens in November 2022.
By October 2024, it cost just $0.07.
That is a decline of more than 280 times in about 18 months. Meanwhile, Stanford estimates that machine-learning hardware becomes around 40% more energy efficient each year.
Stanford AI Index 2025 — Research and Development
At first sight, this should reduce pressure on infrastructure.
However, efficiency is only half of the equation. The other half is how much AI we use.
Google provides a striking example.
In May 2024, Google’s models were processing around 9.7 trillion tokens per month. By May 2026, that figure had risen to more than 3.2 quadrillion.
That is an increase of more than 300 times in two years.
Google I/O 2026 — Sundar Pichai keynote
These figures are not a perfect comparison. Google’s products, models and user base all changed during those two years.
Still, the direction is difficult to ignore:
The cost of AI is falling rapidly, while usage is exploding.
This is where the Jevons paradox becomes relevant.
In the 19th century, more efficient steam engines needed less coal for the same work. Yet total coal consumption increased.
Why? Cheaper steam power created more uses for it.
AI could follow a similar pattern. As intelligence becomes cheaper, we may not simply perform the same tasks for less money.
Instead, we may find entirely new ways to use it.
Better AI → Lower cost → More applications → More usage → More computing power
Of course, efficiency still matters.
The International Energy Agency estimates that better hardware, software and infrastructure could significantly reduce future electricity needs.
However, demand is growing at the same time. In its central case, the IEA expects global data-center electricity consumption to more than double to around 945 TWh by 2030.
AI is the main driver of that increase.
So the real question is simple:
Will efficiency improve faster than demand grows?
2. Solve One Bottleneck, and Another Appears
AI does not really have one bottleneck.
Instead, it depends on a chain of interconnected resources:
Chips → Data centers → Electricity → Grid → Equipment → Factories and people
Suppose tomorrow the chip shortage disappears.
More servers can be installed. But those servers need data centers.
Build the data centers, and they need power.
Find enough power, and it still needs to reach the site. Now the grid may become the constraint.
Reinforce the grid, and demand rises for transformers, cables and other equipment.
Increase production of those, and the constraint may move again. Factories, materials or skilled workers could become the next problem.
Nothing necessarily went wrong.
Each problem was addressed. Yet the bottleneck survived.
It simply moved.
We can already see this mechanism at work.
The IEA estimates that grid constraints could delay around 20% of global data-center capacity planned for construction by 2030.
Part of the problem is simply time.
In advanced economies, new transmission lines can take four to eight years to build. Meanwhile, lead times for transformers and cables have roughly doubled.
IEA — Energy and AI Executive Summary
Now compare that with AI.
Models can improve within months. New applications can spread rapidly.
Physical infrastructure cannot move that fast.
AI moves in months. Infrastructure moves in years.
This creates a difficult choice.
Build too little, and bottlenecks remain. Build too much, and expensive infrastructure may eventually be underused.
So the challenge is not simply to build more.
It is to build the right capacity, in the right place, at the right time.
Infrastructure industries have always faced this problem. But AI adds something new:
Demand can change extremely quickly.
As a result, we are seeing a race between two very different speeds:
How fast AI demand grows — and how fast the physical world can adapt.
3. But AI Can Also Help Solve the Problem It Creates
So far, AI looks like the source of the problem.
But AI may also be part of the solution.
If physical supply cannot respond quickly enough, higher productivity could help it move faster.
For example, AI can help engineers analyse information, prepare documents and compare options. It can also make knowledge easier to find.
In manufacturing, AI can improve planning, quality control and predictive maintenance.
Electricity networks can benefit too. Better forecasting and optimisation can help operators use existing infrastructure more efficiently.
This is already moving beyond theory.
Alphabet says it is using AI to help accelerate the grid connection of new power plants. This forms part of its wider effort to expand energy and data-center infrastructure.
Alphabet — Advancing U.S. Energy Innovation and Infrastructure
Over time, AI combined with robotics may also help address shortages of skilled labour.
However, we should be careful.
There is no solid evidence that AI will suddenly make a power plant or transmission line 30% faster to build. Physical projects will remain physical projects.
Still, there is evidence that AI can improve productivity.
Researchers from Stanford and MIT studied 5,179 customer-support workers using a generative-AI assistant.
Productivity increased by around 14% on average. Among novice and lower-skilled workers, the gain reached 34%.
The researchers also found something important. AI appeared to help less experienced workers adopt some practices used by stronger performers.
That could matter for infrastructure.
These industries do not simply need more workers. They need more experienced workers.
AI may help people access knowledge and solve problems faster. In turn, that could shorten the learning curve.
Of course, customer support is not infrastructure construction. Productivity gains also vary greatly by task.
But the mechanism matters.
We now have two forces working at the same time:
Better AI → More demand → More infrastructure needed
and:
Better AI → Higher productivity → More infrastructure capacity
That leads to perhaps the most important question in this article:
Which grows faster: AI-driven demand or AI-driven productivity?
If demand wins, bottlenecks could persist.
If productivity catches up, some constraints may ease.
And if capacity eventually grows faster than demand, today’s shortages could become tomorrow’s excess capacity.
A bottleneck is not necessarily permanent.
4. In the End, the Economics Must Work
There is one final part of the story: Capital.
Expanding AI infrastructure requires enormous investment.
Companies need data centers and chips. Those facilities need electricity, which requires generation, grids and electrical equipment.
Behind them sits another layer of investment in factories and supply chains.
The scale is already extraordinary.
Alphabet expects $175–185 billion of capital expenditure in 2026. Meta expects another $130–145 billion.
Alphabet — Q4 2025 Earnings Call and 2026 CapEx Guidance
Yet Alphabet still expects to remain supply-constrained through 2026.
The reason is simple: parts of the supply chain take a long time to expand. Today’s capacity therefore depends partly on investments made years earlier.
This illustrates the problem perfectly:
Capital can finance more capacity. But it cannot make physical capacity appear instantly.
Capital also has its limits.
Eventually, these investments need to create economic value.
That value does not have to come only from AI subscriptions. It can appear throughout the economy.
An engineer may work faster. A factory may reduce downtime. A company may automate administrative work.
A grid operator may use existing assets more efficiently. New products and services may also emerge.
So the key question is broader than AI revenue:
Will AI create enough economic value, soon enough, to justify the infrastructure being built for it?
The words soon enough are important.
There are three different clocks:
AI demand can change in months.
Financial conditions can change in quarters.
Physical infrastructure takes years.
Capital has to bridge those timelines.
Higher interest rates can make that capital more expensive. Tighter financial conditions can also reduce the appetite for new investment.
And if returns disappoint, companies may question further spending.
Meanwhile, the infrastructure may still take years to build.
If AI creates enough value, the cycle can continue:
AI improves → Usage grows → Investment rises → Infrastructure expands → The bottleneck moves
But if returns disappoint, investment may slow.
Projects can be cancelled. Planned capacity may never be built. Eventually, some existing infrastructure could even become underused.
At that point, the limiting factor is no longer chips, electricity or transformers.
It is the return on capital.
So Where Does Scarcity Move Next?
Perhaps AI’s bottlenecks will eventually disappear.
For now, however, they appear to be moving.
As computing capacity expands, scarcity can move towards power.
As power expands, it can move towards grids and equipment.
Then, as industrial capacity responds, constraints can move further down the chain.
At the same time, AI may help the physical world respond faster.
Capital can finance that response — as long as the economic returns justify it.
So we should not think of AI infrastructure as a fixed list of bottlenecks.
It is a moving chain of scarcity.
For investors, that distinction matters.
As I discussed in How to Analyze a Commodity Like an Investor, strong demand alone does not create a great investment.
Supply needs to remain constrained long enough to create value.
But even scarcity is not enough.
A shortage of transformers does not make every transformer manufacturer attractive. Likewise, a constrained grid does not make every utility a good investment.
A company still needs to capture the economic value created by that scarcity.
Expectations matter too.
As explored in Why Markets Reward Surprises… Not the Best Companies, even a powerful structural trend can make a poor investment.
The market may already have priced it in.
So perhaps the most useful question is no longer:
What does AI need more of?
Instead, it is:
Where will scarcity move next — and who will capture the value when it gets there?
Ultimately, four forces will decide the answer: demand, productivity, physical capacity and capital.
Or, more simply:
It is whether productivity and infrastructure can catch up with AI demand — before capital loses patience.
The opportunity may not be in identifying today’s bottleneck.
It may be in understanding where it moves next.
Sources
AI costs and hardware efficiency
Stanford University — 2025 AI Index
Growth in AI usage
Google I/O 2026 — Sundar Pichai keynote
AI and electricity demand
International Energy Agency — Energy Demand from AI
Grid constraints
International Energy Agency — AI and Energy Security
Infrastructure lead times
International Energy Agency — Energy and AI Executive Summary
AI and productivity
Brynjolfsson, Li & Raymond — Generative AI at Work, NBER
AI applied to energy infrastructure
Alphabet — Advancing U.S. Energy Innovation and Infrastructure
AI infrastructure investment
Alphabet — Q4 2025 Earnings Call
Meta — Q2 2026 Results
Further Reading on Portefeuille Serein
How to Invest in the AI Boom: The Bottlenecks Markets May Still Be Underestimating
How to Analyze a Commodity Like an Investor
Are We Entering a New Commodity Supercycle?
Copper: The Strategic Commodity of the Decade?
Why Markets Reward Surprises… Not the Best Companies
Data and sources checked as of August 2026. This article is intended for general educational and analytical purposes. It does not constitute personalised investment advice.
Laisser un commentaire