Published on Jun 26, 2026 • Updated on Sep 24, 2026
This article was written by Adem Tumerkan
Content Editor, Dunham | 2025 ThinkAdvisor Luminary Award Winner | 2026 Wealthies Finalist — Thought Leader of the Year | Macroeconomics, markets, geopolitics & global trends
Subscribe to get the latest insights delivered straight to your inbox.
We care about your data in our privacy policy
By now, everyone is aware of the AI bubble boom.
AI data centers. Trillion-dollar capex plans. NVIDIA chips sold out before they even release. Power companies stressed. Dead firms slapping AI in their name and soaring >100%. And every CEO on earth is suddenly claiming AI is central to the business.
But the sheer scale of this buildout is still hard to wrap your head around. . .
As I wrote recently, the AI capex boom has already outspent every megaproject in modern history. At the current pace, the AI data center buildout has consumed close to $1 trillion in just six years1.
To put that in perspective, adjusted for inflation, the U.S. railroad boom lasted more than 70 years and cost roughly $550 billion. The Interstate Highway System cost about $620 billion and took 36 years to build 46,000 miles. The Apollo missions cost around $260 billion from start to finish.
AI blew past all of them in a fraction of the time.
And it’s not slowing down.
So, either we’re in the midst of a multi-decade long productivity boom - or just the largest bubble in history.
Well, there’s certainly a case for both arguments right now.
And the best way to understand the tension is through two old economic ideas - Jevons Paradox and the Solow Paradox.
Put the two together, and you get the strange AI cycle we’re living through right now - a historic “layer cake” buildout in power, chips, data centers, copper, cooling, training models, etc. and surging demand - but still no broad productivity boom that clearly matches the size of the spending.
Let’s break it down.
Jevons Paradox - named in 1865 after English economist William Stanley Jevons2 - simply means that when something gets more efficient, the cost of using it falls. And when the cost falls, demand actually surges.
Why is that a paradox? Because we logically expect efficiency to save resources (which is supposed to be the point), yet in practice, it often triggers such a big expansion in use that total consumption goes up instead of down.
And for a true Jevons Paradox to occur, three conditions must hold - workers or systems must actually get more productive from the efficiency gain, that productivity must translate into lower prices, and demand for the cheaper resource must be elastic enough to explode in response.
The classic example Jevons looked at was coal in 19th‑century England.
As steam engines became more efficient, they used less coal per unit of work. But instead of using less coal, even more was being used. That’s because as steam power became cheaper and more useful, it spread everywhere – from factories and railroads to shipping and manufacturing.
That’s Jevons Paradox.
Now let’s apply that to AI.
Every few months, we hear that chips are getting faster, models are getting smarter, and inference costs are falling. At first glance, that sounds like it should reduce the pressure on the system.
But here’s the thing.
Token prices have fallen sharply as competition intensified. But instead of costs going down, total usage exploded.
And since the AI boom is a layer cake - this has ripple far from just AI tokens.
The International Energy Agency (IEA) expects global data center electricity use to more than double by 2030, reaching about 945 terawatt-hours3 (that’s slightly more electricity than Japan – the world’s fourth largest economy - uses today).
In the U.S., data centers are expected to drive a large share of electricity demand growth between now and 2030, with their share of total power use potentially doubling or more. This will further stress an already aging, creaky grid – causing consumer power prices to rise and potential for energy rationing or blackouts.
Put simply, cheaper AI costs may just widen the systemic problems.
Now here’s the other side.
In 1987, economist Robert Solow4 wrote one of the most famous lines in modern economics:
“You can see the computer age everywhere but in the productivity statistics.”
That became known as the Solow Paradox - the idea that you can pour money into new technology and see it everywhere, yet not see a clear payoff in the productivity numbers for many years.
Companies were buying computers. Offices were filling them with software. Workers were changing how they did their jobs. But measured productivity didn’t immediately surge.
Why? Because buying technology is not the same thing as reorganizing around technology.
For context, computers were spreading rapidly through offices in the late 1970s and 1980s, but U.S. productivity growth stayed anemic (if not worse than the previous ~25 years without computers).
It wasn’t until the mid‑1990s that productivity growth finally broke higher as firms reorganized around the new technology.
And we’re seeing the same thing with AI.
For example, a Fortune 500 call center that rolled out a generative AI assistant saw productivity jump about ~14%5, with the biggest gains going to junior staff. And developers in GitHub’s Copilot tests finished coding tasks 55% faster6 with AI than without it.
Impressive at the micro level.
But when researchers asked roughly 6,000 executives across four major economies what AI had done to their business so far, almost 90% said it had a marginal impact at best on productivity or employment7. Most use AI for maybe an hour or two a week.
Meanwhile, a 2025 MIT study8 found that about ~95% of enterprise generative‑AI pilots fail to generate any measurable returns on investment, despite tens of billions of dollars in spending.
Yet Goldman estimates AI capex could run into the trillions over this decade.
That’s a huge spending cycle against a payoff that is still hard to measure.
Even if AI makes certain workers and companies more productive, that doesn’t automatically mean the whole economy gets a productivity boom (besides the inflationary surge from aggressive spending).
Who captures the gains? Who loses income? How much spending disappears if firms use AI mainly to cut labor costs?
For instance, if AI boosts output inside a company but also reduces jobs, wages, or household demand, the macro data can look much messier than the company-level case studies suggest.
Now, of course, over time, AI will likely make a far bigger impact on the global economy.
But as Solow pointed out, that payoff – both good or bad – can take a long time.
And in the meantime, the real risk is that AI is increasingly funded by a ton of debt.
Thus, if expectations fall, market hype turns sour, or debt compounds faster than productivity gains can, that’s essentially how a bubble pops.
This is where Jevons and Solow meet the real world.
Put those together and you get the shape of the current cycle - a massive, physical buildout on the bottom with booming demand, but a slow, uneven productivity picture.
That’s why I think of AI as a layer cake.
At the bottom, you have the physical inputs.
In the middle, you have the compute backbone.
At the top, you have the intelligence and end-user productivity.
Each layer has its own opportunities – and its own bottlenecks.
For example:
The bottom layer gets pulled in first because the infrastructure has to exist before the software can scale.
That means potential booms in uranium, copper, natural gas, etc - but those are bottlenecked by permits that take years, mines that take a decade to open, and finite resources that are expensive and slow to pull out of the ground.
The middle layer gets paid as capacity expands – like data centers, chips, memory, servers, networking. But this layer has its own chokepoints.
For instance, Nvidia’s most advanced GPUs are still constrained by TSMC’s leading-edge fab capacity. And while data center construction is racing ahead (now surpassing office building)9, transformers and switchgear to actually connect them to the grid have lagged with multi-year backlogs. That means you can break ground on a data center today and still be waiting ~20 months for the substation equipment to power it.
The top layer – like the models, software, agents, and productivity tools we all use – is the last in line. But once the other layers are up, users and enterprises can then decide what they’re actually willing to pay for. This is where valuation lives or dies – because a model can be extraordinary and still get commoditized in 18 months if a cheaper competitor arrives. Enterprise software built on top of AI still has to prove it saves more than it costs.
And critically, these layers don’t move in sync.
For example - right now - demand for data centers is aggressive while the commodity supply chain is still catching up. Copper mines discovered today won’t produce at scale for ~20 years10.
Models are getting more powerful but power-hungry - and in some areas, there simply isn’t enough grid capacity to run the next generation of training clusters.
Meanwhile, the top layer is already being valued through software and agent potential before the bottom two layers have finished being built.
That mismatch is not a bug. It’s how the capital cycle works (I’ve written more about this before – read here). It just means the whole cake can be stressed, overbuilt, and underbuilt at the same time - just in different layers (each in its own feedback loops).
And that “later” is where Solow’s words should ring in our heads.
Goldman Sachs11 now expects the large technology companies leading this buildout to spend roughly $5.3 trillion from 2025 through 2030.
That is an enormous bet. Maybe it pays off. But the path of how it gets there matters big time.
The point is, infrastructure spending is upfront. Productivity gains are delayed. And somewhere between here and there is where volatility lives.
In the late 1990s, telecom companies laid more than 80 million miles of fiber optic cable across the U.S., raising close to $2 trillion in equity and $600 billion in debt to fund it12. The idea was that the internet was being adopted faster and cheaper, thus bandwidth demand would explode, and whoever built the network first would win the arms race.
And while they were right about the technology - what they got wrong was the timing.
By 2002, only about ~3% of that fiber was actually being used. The rest sat dark and idle.
Bandwidth prices collapsed by 90%13. WorldCom filed the largest bankruptcy in U.S. history (at the time). Global Crossing followed. And the sector lost more than $2 trillion in market value – which caused the NASDAQ to take nearly fifteen years to recover.
The irony is that the thesis wasn’t wrong – because the internet did eventually need every mile of that fiber (and more). But the debt compounded faster than the demand arrived and took years to recover (and a lot of pain).
Like all things in macro and markets, there are more variables than anyone can fully grasp.
There are things we can measure. Things we can’t. And things nobody sees coming until they’re two inches from our face.
But the framework is pretty clear.
All of those can be true at once.
That’s the Jevons and Solow problem in one sentence - the infrastructure bill arrives early, and the productivity curve arrives late – if it arrives at all.
As always, time will tell.
Is the AI boom a bubble or a real productivity revolution? Right now, both sides have solid evidence. AI spending has already blown past every past megaproject, and usage keeps climbing. But real productivity gains stay thin — McKinsey found 80% of workers report personal gains, yet only 37% of organizations see company-wide impact. The honest answer is that it's too soon to call this one either way.
Why hasn't AI shown up in productivity numbers yet? Businesses buy AI faster than they rebuild around it. Economist Robert Solow noticed the same lag with computers back in 1987 — new tech can spread everywhere and still take over a decade to move national productivity stats. BDO Canada found only 18% of companies have actually built AI into their daily workflows so far.
Does cheaper AI mean lower costs for businesses? Not really. This is Jevons Paradox at work — when AI gets cheaper per task, businesses just use it more, so total spending climbs instead of falling. That pattern needs three things: real productivity gains, lower prices from those gains, and demand that explodes in response. Right now, AI is nailing that third condition fast.
How much money is going into AI infrastructure right now? A lot. Goldman Sachs projects roughly $7.6 trillion in global AI infrastructure spending between 2026 and 2031, with yearly capex hitting about $1.6 trillion by 2031. CreditSights separately estimates the five biggest hyperscalers will spend around $602 billion in 2026 alone, a 36% jump from the year before.
Can AI make individual workers faster but hurt overall company output? Yes, and it's happening already. A 2026 ADB-led study found AI helped 62% of workers produce better work, but the other 38% turned in weaker final output because speed cut their motivation to double-check it. Researchers call this the "mediocrity trap" — faster individual work doesn't guarantee a more productive company.
This communication is general in nature and provided for educational and informational purposes only. It should not be considered or relied upon as legal, tax or investment advice or an investment recommendation, or as a substitute for legal or tax counsel. Any investment products or services named herein are for illustrative purposes only and should not be considered an offer to buy or sell, or an investment recommendation for, any specific security, strategy or investment product or service. Always consult a qualified professional or your own independent financial professional for personalized advice or investment recommendations tailored to your specific goals, individual situation, and risk tolerance. All examples are hypothetical and are for illustrative purposes only.
Information contained in the materials included is believed to be from reliable sources, but no representations or guarantees are made as to the accuracy or completeness of information. This document is provided for information purposes only and should not be considered as investment advice.
Dunham & Associates Investment Counsel, Inc. is a Registered Investment Adviser and Broker/Dealer. Member FINRA/SIPC. Advisory services and securities offered through Dunham & Associates Investment Counsel, Inc.
