Betting on AI is not an investment strategy
Why the economics of AI don't work the way most investors think
A few weeks ago I was sitting across from an investor. Smart person, well-connected, and an important part of the Norwegian startup ecosystem. At some point he told me that the most common sentiment among investors right now is that they won’t touch anything that isn’t about AI. It has become the default position for the ecosystem.
I wanted to argue. But I nodded.
Because there’s a question hiding inside that sentiment that I’m not sure most of the people making these bets have fully answered:
Do you understand what kind of business you’re actually investing in?
The SaaS Playbook Is Quietly Breaking
For the past two decades, software investing has followed a reliable script. You find a painful workflow, you digitize it, you charge per seat, you scale. Gross margins sit somewhere between 50 and 85 percent. Net revenue retention compounds over time. The model is elegant, predictable, and very well understood by investors.
AI breaks almost every assumption in that script.
To be fair, the shift away from per-seat pricing is something people are beginning to grasp. Usage-based models — where customers buy tokens or pay per interaction rather than per user — are increasingly the norm in AI products. Most sophisticated buyers understand this is where things are heading. But understanding the pricing alternatives is not the same as understanding the underlying economics.
Here is what is less discussed: when an AI agent does the work, your revenue is directly coupled to your compute costs in a way that traditional SaaS never was. Every query, every task, every resolved ticket has an infrastructure cost attached to it. And those costs are not stable. We are in the middle of a global energy crisis. Data center demand is exploding. The long-term price of compute is genuinely uncertain. No one knows where and when it settles.
This means the margin profile of an AI-native business is fundamentally harder to predict than a SaaS business. It’s not just that pricing has shifted from seats to usage. It’s that the cost side of the equation is now volatile in ways that traditional software investors have never had to model. A business that looks like it has 70 percent gross margins today might look very different in two years depending on energy prices, GPU availability, and how the hyperscalers choose to price their infrastructure.
If you’re writing checks into AI companies while mentally underwriting them as SaaS businesses, you may be building your thesis on a foundation that no longer exists.
The Tool That Won’t Become a Company
There’s a second problem, and it’s more uncomfortable.
Across industries, a new kind of operator is emerging: someone with domain expertise and just enough technical fluency to assemble powerful AI tools from the inside. They are not developers. They are not buying software in the traditional sense. They are building exactly what they need, when they need it, for close to nothing. A few days, a few tools, a working application. No engineering team. No procurement process. No vendor.
This matters enormously for startups whose entire premise is selling software to companies with complex legacy workflows.
The assumption has been: these organizations can’t build it themselves, so they’ll buy it from us. That assumption is eroding faster than most founders - and most investors - have registered.
The Customer Is Building Your Product
Nowhere is this more relevant than in Norway’s industrial heartland.
Right now there is significant energy - and significant capital- going into startups that promise to digitize heavy, manual workflows in sectors like maritime, energy, logistics, and construction. It’s a reasonable bet. These industries are genuinely behind. The pain is real. The willingness to pay has historically been there.
But here’s what I think is being underestimated: the most sophisticated operators inside these legacy companies are not waiting for a vendor to save them. They’re assembling tools internally. They’re automating from the inside out, quietly and without fanfare, because the technology now lets them do exactly that.
The startup that just raised a seed round to digitize offshore inspection workflows should be asking a harder question: not just “who else could build this?” but “could my customer build this themselves in six months?”
In 2024, the answer was usually no. In 2026, the answer is increasingly yes.
What Good AI Investing Actually Looks Like
None of this means AI is a bad bet. It means the bet requires more precision than “we only invest in AI.”
The questions worth asking are harder than they used to be.
What are the actual margins when you factor in compute costs - and what happens to those margins if energy prices keep climbing?
Is this a product or a workflow that a sufficiently motivated internal operator could replicate?
What happens to pricing power as foundation models get cheaper and more capable? Where exactly is the moat — in the model, the data, the distribution, or something else entirely?
These are not exotic questions. They are the basic questions of business model analysis, applied to a new context. But they require letting go of the SaaS mental model that has served the industry so well for so long.
AI is a capability. A genuinely transformative one. But capability is not a business model. And in the rush to not miss the wave, I worry that too many investors in the Norwegian ecosystem are funding the label rather than the logic.
The companies that will matter in five years are not the ones that added AI to a pitch deck. They are the ones that figured out a durable way to create and capture value in a world where the economics of software have fundamentally changed.
That is a much harder question. And it starts with asking it.
You can reach me by replying to this email or at contact@anjali.no.
Check out My website