I run a software company. We’ve been building learning management systems for more than twenty years, and right now we’re using AI to rebuild almost everything.
It’s extraordinary.
Work that once required months can now be accomplished in days. Architectural problems that would have consumed weeks can be explored in an afternoon. I’m using more artificial intelligence than I ever imagined I would.
And I expect to use considerably less of it tomorrow.
That should worry anyone lending hundreds of billions of dollars to build AI data centres.
The Great Technical Debt Liquidation
For decades, the software industry has accumulated technical debt.
Obsolete languages, aging architectures, fragile integrations, databases designed around constraints that disappeared fifteen years ago. Work that needed doing but couldn’t be economically justified.
AI has changed that calculation almost overnight.
Suddenly, rebuilding a twenty-year-old application isn’t an absurd proposition. It’s an opportunity.
And we’re all doing it.
The resulting demand for AI computation is enormous. But there’s a problem with projecting that demand into the future.
Technical debt is a stock, not a flow.
We’re attacking decades of accumulated work because the economics have finally changed. Once we’ve rebuilt the software, we don’t need to rebuild it again tomorrow.
Of course, software will continue to evolve. New requirements will emerge, and cheaper development will make new projects practical.
But that doesn’t mean today’s extraordinary consumption represents tomorrow’s normal demand.
We’re confusing the energy required to build the factory with the energy required to run it.
The AI Building My Software Isn’t the AI Running It
At Thinking Cap, we’re using sophisticated frontier models to redesign architecture, write code, and solve difficult problems.
But the resulting system is largely conventional software.
When a learner completes a course, I don’t need a large language model to determine whether the completion criteria have been satisfied.
When an administrator enrolls ten thousand people, I don’t need AI to perform ten thousand enrollments.
I need reliable software.
Where AI belongs, it will often handle relatively narrow tasks: interpreting questions, providing guidance, resolving ambiguity, or helping people understand information.
Much of that can be handled by inexpensive, smaller models.
Occasionally, I’ll need something considerably more capable.
And I’ll call it.
AI will be the consultant you never hire.
You don’t keep a brilliant consultant sitting in your office all day waiting for something difficult to happen. You call when you need expertise.
With AI, that consultation might last thirty seconds and cost pennies.
The advice could be worth thousands.
The AI provider might earn three cents.
That’s a profound difference between economic value created and revenue captured.
The Economics Will Find Their Equilibrium
I don’t believe AI development will stop when the rebuilding frenzy subsides.
As AI becomes cheaper, I’ll create more software. I’ll keep adding capabilities until the cost of the next one exceeds what I can reasonably expect to recover.
That’s how businesses operate. It’s how employment operates. You hire people when their work justifies their compensation.
AI will find a similar equilibrium.
Lower prices will encourage more consumption, but greater consumption doesn’t necessarily mean greater spending.
A thousandfold increase in useful AI activity doesn’t guarantee a thousandfold increase in AI revenue.
It might not even produce a tenfold increase.
And the cost of inference keeps falling. Models are becoming more efficient, smaller models are becoming more capable, and competition is increasing.
The intelligence is becoming a commodity.
The enduring value will increasingly lie in what people build with it.
OpenAI Made Money From Me. Until It Didn’t.
In 2025, OpenAI was getting my money.
In early 2026, Anthropic was getting it.
Today, Moonshot’s Kimi gets the business.
I’m not simply shopping for the cheapest model. I’m buying the intelligence that best serves my purposes at a price that makes sense.
My appetite for AI hasn’t diminished.
My loyalty to any particular provider is approximately zero.
And yet the financial commitments being made by the major AI companies seem to assume extraordinary growth years into the future.
OpenAI has announced compute commitments measured in hundreds of billions of dollars.
Oracle and other infrastructure providers are investing enormous sums to satisfy that anticipated demand.
But where is the money ultimately coming from?
OpenAI has substantial revenue and has raised extraordinary amounts of capital. What it hasn’t demonstrated is the ability to generate enough operating cash to fund commitments of this magnitude.
Its suppliers invest against promises of future payments. Those promises become contracted revenue and backlog, helping justify borrowing and construction.
And the whole arrangement depends on customers continuing to spend.
Customers like me.
Who have already moved twice and might move again next month.
The contracts can be legitimate. The accounting can be legitimate.
And the underlying economic assumptions can still be disastrously wrong.
A Temporary Bottleneck Isn’t a Permanent Market
Today’s compute shortage is real.
But a shortage today doesn’t establish a shortage five years from now, especially when the industry is simultaneously building enormous new capacity and finding more efficient ways to use what already exists.
There’s an argument that cheaper computation will simply generate more demand.
It might. Jevons paradox is real enough.
But Jevons paradox isn’t a business plan.
It doesn’t guarantee that demand will absorb every new data centre, that prices will remain high enough to service the debt, or that today’s providers will capture tomorrow’s revenue.
We could have an explosion in AI usage, a collapse in AI prices, and a financial disaster for infrastructure investors.
All three could happen simultaneously.
And Canada Wants In?
Canada is considering enormous investments in sovereign AI computing capacity.
I understand the motivation. We shouldn’t be entirely dependent on foreign governments and corporations for critical technology.
But I’m not convinced that owning vast quantities of rapidly depreciating computing hardware is the best way to achieve independence.
At Thinking Cap, we’ve approached the problem differently.
How much of the information we send to an AI actually needs to identify anyone?
Surprisingly little.
If I’m asking AI to reason about a learner’s progress, it often doesn’t need the learner’s name. If I’m asking it to interpret an administrative problem, it doesn’t necessarily need the identities of the people involved.
Our software can retrieve the relevant information, substitute temporary identifiers, and present the AI with the relationships and facts it needs.
This isn’t perfect anonymization. Context can still reveal identities, and some information requires stronger protections.
But why insist that the intelligence must reside wherever the sensitive data resides?
Keep the data under Canadian control. Keep the business logic under Canadian control. Use whichever intelligence offers the best capability, cost, and acceptable risk.
Build sovereign computing capacity where it’s genuinely necessary.
And invest in Canadian companies creating valuable applications of AI.
Because that’s where I suspect much of the enduring opportunity lies.
The Balance-Sheet Illusion
Capital expenditure has a peculiar ability to make optimism look like an asset.
Spend billions constructing a data centre and you have billions of dollars of infrastructure on your balance sheet.
That doesn’t mean you’ve created billions in economic value.
You’ve created an asset whose value depends on future customers paying enough to recover its cost.
Meanwhile, the spending generates revenue for suppliers, contracted backlog for infrastructure providers, and apparently reassuring evidence of demand throughout the industry.
Everyone can be acting rationally according to their own incentives while collectively building far more capacity than the market can support.
And when the assumptions fail, the buildings don’t disappear.
Neither does the debt.
I wonder how much of this infrastructure would be built if investors had to finance it from the operating profits of the companies ultimately consuming the computation.
I suspect considerably less.
The World Is Still Made of Things
There’s something surreal about the financial importance we’re assigning to artificial intelligence.
Trillion-dollar valuations. Infrastructure commitments rivaling the economies of entire countries.
And yet the world’s economy still requires ships to be built, steel to be forged, and beans to be picked.
AI can improve how we design ships, manufacture steel, and grow food.
Those improvements could be enormously valuable.
But the ships still need steel. The steel still needs energy. The beans still need soil, water, and sunlight.
Intelligence is an input into economic activity.
It isn’t a substitute for economic activity.
And there’s no law requiring the suppliers of intelligence to capture anything approaching the value their intelligence helps create.
A farmer might save a million dollars through better decisions informed by AI.
That doesn’t mean the AI provider earns a million dollars.
It might earn twenty.
Electricity transformed the world. That didn’t make every electricity producer infinitely valuable.
Telecommunications transformed the world. That didn’t protect every telecommunications investor from ruin.
The internet transformed the world. That didn’t make the infrastructure investments of the dot-com boom economically sound.
Society benefited enormously, even when investors financing the transformation lost fortunes.
AI could follow the same pattern.
The Wrong Question
I don’t doubt that artificial intelligence is transformative.
I’m betting my company’s future on it.
What I doubt is the assumption that the extraordinary amount of AI we’re consuming during this transformation represents the amount we’ll need afterward.
The industry is asking how much computing capacity it needs to satisfy today’s demand.
It should be asking how much of that demand comes from constructing tomorrow’s systems, and how much will remain when those systems are operating.
Those are profoundly different questions.
I’m not predicting that AI demand will collapse.
I’m suggesting something more troubling for investors.
AI could succeed beyond anyone’s expectations.
It could become ubiquitous, indispensable, and extraordinarily cheap.
It could transform almost every industry while generating far less revenue for its suppliers than today’s infrastructure commitments require.
The world could win.
And the people financing the data centres could lose.
AI is a good investment. I’m happy Canada is thinking seriously about it.
But given the choice, I’d rather we built a few new shipyards, ports, and high-speed rail lines.

