Your CEO wants you to spend more on AI.
So does your board. So does the peer CFO who posted about their transformation on LinkedIn last week. The pressure is coming from every direction, and none of it is wrong. AI is the biggest shift this seat will manage in a generation, and the CFO who sits it out loses.
But here is the trap almost nobody is naming.
You can believe in AI completely, commit real budget to it, move faster than your competitors, and still lose. Not because the technology fails. Because you pointed the money at the wrong layer.
That is the collapse CFOs are not prepared for.
It is not a story about AI disappointing anyone.
It is a story about where the profit actually lands, which layer keeps it, and which layer quietly pays for everyone else’s. Most finance leaders are committing three years of budget to the layer that pays.
I am not asking you to take my word for it.
In the last 90 days, four stories came out that change how you should think about your AI budget.
Peter Thiel sold his entire Nvidia stake.
Nvidia posted a profit margin almost no company ever reaches.
OpenAI lost 21 billion dollars in a single year.
And a Bank of America analyst told clients the SaaS software you already pay for is being repriced for good.
Read them in order.
You will look at your own AI budget differently.
Let’s dive in.
Peter Thiel sold every share of Nvidia
He sold all of it.
In the third quarter of 2025, Thiel’s fund sold every share of Nvidia it owned. All 537,742 of them were worth over $100 million. That position was roughly 40% of the entire fund.
The most successful technology investor of his generation looked at the defining company of the AI era and walked out the door.
Thiel is not anti-AI.
He funded the first outside check into Facebook. He runs Palantir. He believes the technology is real. He compared this moment to the internet in the late 1990s: more than a nothing burger, and less than the total transformation everyone promises.
Both halves of that sentence are the point.
The internet was exactly as real as everyone believed in 1999. That did not save the companies that bought into the wrong layer at the wrong price.
Thiel’s discipline is one question.
Follow the money. Where is the profit actually landing?
He answered it himself at the Aspen Ideas Festival in 2024. He estimated that 80 to 85% of the money being made in AI is made by one company, Nvidia. He called it “very strange,” because the profit sits at the hardware layer, a place he said Silicon Valley “doesn’t even know much about anymore.”
Nearly all the profit in AI is made at the layer almost no CFO is buying.
Before you commit a dollar to AI, ask Thiel’s question. Which layer captures the profit, and which layer pays for it?
Nvidia makes the profit. Everyone else pays the bill
In one quarter, Nvidia earned $58.3 billion in net income on $81.6 billion in revenue, at a gross margin of nearly 75%.
Data center revenue alone was $75.2 billion, up 92% from a year earlier. Strip out one-time gains and the company still keeps more than 55 cents of every dollar as profit.
That is not a good business.
That is the best business in the history of business.
Now look at who is buying.
The big cloud companies are on track to spend somewhere between $500 billion and $600 billion on AI infrastructure this year. One recent estimate has them recycling more than 100% of their cloud revenue straight back into capital spending. They are running to stand still, and the profit from all that motion pools in one place: the chip layer.
Sequoia’s David Cahn named the gap.
He called it “AI’s $600 billion question,” the distance between what the industry spends building AI and the revenue AI actually produces.
That gap has only widened since.
The AI economy is a shovel business, and one company sells the shovels.
The SaaS layer you already pay for is being repriced in public
Dario Amodei, the CEO of Anthropic, said it plainly at a financial services event. If a software company’s only moat is that its product is complex and hard to build, that moat is disappearing.
In 2026, Salesforce fell around 30%.
Workday fell around 33%.
The main software index dropped more than 20% into a bear market, and by some counts one to two trillion dollars of software value evaporated in about a year.
On May 18, 2026, a Bank of America analyst named Tal Liani cut Salesforce to Underperform with a $160 target, well below the Street.
The per-seat software model is being repriced in real time, in public, by the analyst community. And CFOs are still signing three-year, per-seat renewals into the middle of it.
Every SaaS renewal on your desk this year is a different decision than it was two years ago. The vendor needs you more than you need them, and the analyst community just handed you the leverage. Renegotiate. Shorten the term. Tie price to usage, not seats.
One honest caveat, because your credibility depends on it: the software sector bounced off its lows by mid-2026. But a repricing is exactly when contract terms get rewritten, and that window is open now.
OpenAI’s own CFO is living the fight you are about to have
Here is the fight in its purest form, at the largest scale on earth.
OpenAI lost $20.9 billion on $13.1 billion of revenue in 2025. Its research spending alone, $19.2 billion, was larger than its entire revenue. And the company has committed to roughly $1.4 trillion in compute over the next eight years, against a revenue run rate of around $20 to $25 billion.
That is a commitment-to-revenue gap of nearly 100 to 1.
When OpenAI’s CFO, Sarah Friar, floated the idea that the government might backstop AI financing, the backlash was immediate, and she walked it back within a day.
When an investor pressed Sam Altman on how the spending squares with the revenue, Altman’s answer was not a spreadsheet. It was, roughly, if you want to sell your shares, I will find you a buyer.
That exchange is the CEO-versus-CFO tension of this entire era, playing out in public.
You are about to have a smaller version of that exact conversation.
Your CEO sees AI as the future and wants to spend.
You see the contracts, the burn, and the missing return. The board watches both of you. The only thing that resolves it is your ability to answer three questions the whole industry is currently failing to answer.
What did we sign for?
What are we using.
What is it producing?
The Bottom Line
Thiel sold his Nvidia stake and told you the profit sits at a layer you are not buying. Nvidia’s 75% margin proved him right.
The SaaS layer you already pay for is being repriced in public. And OpenAI’s own CFO is fighting the exact fight you are about to have, at a 100-to-1 scale.
The disciplined CFO does three things with this.
First, follow the profit, not the story. For every AI dollar, ask which layer keeps the money and which layer feeds it. Treat the application on top as disposable and the capability underneath as the thing worth owning.
Second, build the measurement before the commitment. Do not sign the three-year deal and hope. Put a simple test on every AI workflow. What did we sign for, what are we using, what is it producing. If it cannot answer the third one in ninety days, it does not get renewed.
Third, stay liquid on tools, committed to capability. The tooling in 2026 is disposable and a single model release can obsolete it. The capability compounds. Do not lock three years of budget into a layer one launch can erase.
Do this, and the collapse is something you watch happen to other companies.
Fail to do it, and you are the case study in someone else’s newsletter next year.
The CFO who commits budget to the story is replaceable.
The CFO who follows the profit is not.
See you on Thursday.
Whenever you’re ready, there are 2 ways I can help you:
If you have an interesting AI use case, or you have already put AI into your finance function, I’d love to see it.
I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai
Find me on LinkedIn
Hi Wouter,
I have built Evidentia - A financial platform that analysis TB or companies annual report and translates into meaningful financial output with deterministic engine that only explainableAI. More information about this can at www.finoptic.ai.