Compute Will Be Like Oil: A CEO's Take on AI Cost Strategy

Most companies are managing AI costs backwards. They treat token spend as a badge of progress — the bigger the bill, the more "AI-forward" the company. Doodle CEO Christian Fielitz argues the opposite: AI cost is an investment decision, not a status symbol, and the companies that win will be the ones that decide deliberately what to insource, what to outsource, and where to keep control. Compute itself will commoditize — "compute will be like oil" — but that does not mean your bill goes down. It means the strategic question moves from how much can we spend to what should we own.

This piece distills Fielitz's framework for thinking about AI costs, drawn from his June 2026 conversation on the Vlad Catcher Show.

The strawberry-field problem: consider who's giving the cost advice

Before taking anyone's advice on how much to spend on AI, Fielitz says, look at who benefits from the answer.

"It's a little bit like saying you have a strawberry field and you claim, 'Hey, we love strawberries, everybody should eat more strawberries.' Of course, if somebody eats only 10 strawberries, they're not really in love. You should have 20, 50, 500 strawberries."

The people loudly insisting that more tokens equals more progress are very often the people selling tokens. That doesn't make them wrong — but it makes their incentive worth naming. "Just buy tokens, otherwise you're doing it wrong" is not a strategy. It's a sales pitch wearing the costume of one.

The takeaway: separate the signal (AI creates real leverage) from the sales motion (you should consume as much of it as possible, as fast as possible).

Why compute will commoditize — and why that won't lower your bill

Fielitz is unambiguous about the long-run direction of raw compute cost:

"The marginal cost will go down. Compute will be like oil, or like energy that you pull out. For me, it's one of the things that will for sure happen."

So far that sounds like good news for buyers. Here's the catch. The cost of raw compute falling is not the same as the cost of your AI usage falling — because the economics one layer up are badly out of balance:

"We know these statistics and the analysis around the current investment into infrastructure are not being met by the investments in the application layer, or the value being created through the application layer."

In plain terms: enormous capital is going into AI infrastructure, but the value being generated at the application layer hasn't caught up. When that gap closes, it tends to close in the direction of higher prices for the people consuming the service. Which leads to the part of this every operator should internalize.

The cloud-cost history lesson: a five-year-old playbook, playing out again

We have seen this movie before, and recently.

"If you blindly just put everything into LLMs and outsource your knowledge and your creation to other systems, you will inadvertently come to a point where your provider will increase cost — which happened, by the way, in cloud infrastructure whatever five years ago. All of a sudden it's like, 'Oh my.' That led to multi-cloud environments, that led to, 'Hey, should we basically have our own data centers again?'"

The cloud parallel is the most useful mental model in the whole conversation. Cheap, abundant cloud compute encouraged companies to put everything in one provider's ecosystem. Then prices moved, lock-in became expensive, and the pendulum swung back toward multi-cloud and even repatriation to owned data centers.

AI is on the same arc, just earlier. The companies that "blindly put everything into LLMs" today are building the same dependency that cloud-native companies spent the back half of the 2010s unwinding. The lesson isn't avoid AI — it's avoid blind dependency.

What to insource vs. outsource: a control-point framework

So what do you actually do with this? Fielitz's answer reframes AI cost as a question of control, not consumption:

"Having a control of AI costs — knowing what you can outsource and what you should deliberately insource as a control point, as a competitive advantage — is a very important conversation. We have it continuously."

A practical way to apply the framework:

• Outsource the commodity. General-purpose model capability, where switching costs are low and no proprietary advantage is created, can sit with external providers. This is the "oil" — buy it on the market.

• Insource the control points. Anything that is a genuine competitive advantage, or that touches data you cannot afford to hand over, should be kept under tight control. For Doodle, security and privacy force this: "Some of the things, even if we wanted to outsource, we need to keep under tight control."

• Translate cost into investment. When you deliberately insource a capability, you are not just absorbing a cost — you are buying a control point. "You translate cost into an actual investment and say, okay, we're going to keep control of it."

The discipline is in the word deliberately. The failure mode isn't outsourcing or insourcing — it's doing either one by default instead of by decision.

When AI costs more than the humans it was supposed to free

There is also a near-term cost trap that's easy to miss in the enthusiasm:

"If you just blindly let people use LLMs — and not just AI tools — you see that costs are actually higher than humans, already."

Unmanaged, per-seat AI usage can quietly exceed the cost of the human work it was meant to augment. That's not an argument against AI; it's an argument for instrumenting it. If you can't see what AI is costing per workflow versus the value it's producing, you can't manage it — and "we're proud of our AI bill" becomes a very expensive sentence.

The paradox at the end of the road: who buys when no one's left?

Fielitz closes the cost conversation with the question that reframes the entire AI-spend debate:

"If everything is AI and there are no humans left in the companies — who's going to consume your product? Because agents buying from agents… so don't forget."

It's half a joke and half the most important question in enterprise AI strategy. The value of automating everything is bounded by the existence of someone to sell to. An economy of agents transacting with agents, with no human demand underneath it, isn't a market — it's a closed loop. The implication for cost strategy: optimize for durable human value, not for maximum automation. The point of freeing up time isn't to remove people; it's to let them do the work only people can do.

The bottom line

Fielitz is careful not to oversell his own certainty: "We're not saying we'll figure it out. I'm just saying we're having active discussions and trying to look at it." That humility is the point. AI cost strategy in 2026 isn't a solved equation — it's a discipline:

1. Discount advice from people who sell the input. (The strawberry field.)

2. Expect compute to commoditize — but expect your provider's prices to move anyway. (The infrastructure/application gap.)

3. Remember the cloud. Blind dependency gets repriced; plan for it now.

4. Decide your control points deliberately. Outsource the commodity, insource the advantage, translate cost into investment.

5. Instrument usage. Unmanaged AI can cost more than the humans it augments.

6. Keep humans in the loop — they're also your customers.

Treat AI spend the way you'd treat any serious investment: with a thesis, a control plan, and a clear view of the value on the other side.

This piece is drawn from Christian Fielitz's June 2026 interview on the Vlad Catcher Show.