I used to think return on investment was the cleanest concept in finance.
You put money in. You get more money out. You measure the difference. You pretend it was obvious the whole time.
Then I started looking at AI companies—and suddenly ROI felt like trying to measure fog with a ruler.
Because in the AI world, the real question isn’t just “what’s the return?”
It’s: what’s the return on compute?
And that’s where things get messy, expensive, and just a little bit absurd.
The New Arms Race Nobody Fully Understands
There’s a quiet competition happening right now, and it doesn’t look like anything we’ve seen before.
It’s not just about revenue growth. It’s not even about user growth.
It’s about compute.
Who has it. Who can afford it. Who can turn it into something that actually makes money before the electricity bill shows up like an uninvited guest.
Because AI isn’t like traditional software. You don’t just write code and scale it infinitely with minimal cost.
Every model, every query, every fancy generative output—it all burns compute.
And compute isn’t cheap.
So suddenly, the companies leading AI aren’t just tech companies. They’re infrastructure companies. Energy companies. Capital allocation machines.
And if they get the math wrong?
They don’t just lose market share. They burn billions.
What “Return on Compute” Actually Means
Let me strip this down without turning it into a textbook.
Return on compute is basically this:
How much value does a company generate for every dollar it spends on compute?
That includes:
- GPUs and specialized chips
- Data center infrastructure
- Energy consumption
- Model training costs
- Inference (serving the model to users)
If a company spends $1 billion on compute and generates $2 billion in value, great—you’ve got something resembling efficiency.
If it spends $1 billion and generates hype, press releases, and a slightly better chatbot?
Congratulations. You’ve built a very expensive demo.
The Illusion of Scale
Here’s the first trap: scale looks like success.
When an AI company announces it’s investing billions into infrastructure, people assume that’s a sign of strength. And sometimes it is.
But sometimes it’s just… spending.
Because scaling compute doesn’t automatically scale returns.
In fact, it can do the opposite.
The more compute you throw at a problem, the more you risk diminishing returns. You get marginal improvements that cost exponentially more.
It’s like paying ten times as much for a 2% performance boost and then convincing yourself it was worth it because the graph went up.
This is where return on compute becomes brutally important.
Because it forces you to ask:
“Is this actually efficient, or are we just flexing our budget?”
The Big Players (And Their Very Expensive Habits)
Let’s talk about the companies everyone watches.
They’re not just competing on products—they’re competing on how effectively they turn compute into outcomes.
Some are doing it well. Others are… experimenting aggressively.
NVIDIA: Selling the Picks and Shovels
NVIDIA isn’t just part of the AI story—it is the infrastructure.
Every major AI company depends on their GPUs. Which means NVIDIA has done something brilliant:
They’ve positioned themselves to benefit regardless of who wins.
From a return-on-compute perspective, NVIDIA doesn’t need to optimize usage—it just needs to sell more compute.
And they’re doing it at margins that would make most industries uncomfortable.
If AI is a gold rush, NVIDIA isn’t mining. It’s selling the equipment—and charging a premium for every shovel.
Microsoft: Monetizing at Scale (Carefully)
Microsoft has taken a different approach.
It’s not just investing in AI—it’s integrating it into products people already pay for.
Office, cloud services, enterprise tools—these are environments where AI can generate incremental value without needing to justify itself from scratch.
That’s key.
Because return on compute improves dramatically when:
- You already have distribution
- You already have paying customers
- You can layer AI on top instead of building from zero
Microsoft isn’t trying to win with the biggest model. It’s trying to win with the most monetizable one.
Alphabet: Efficiency vs Ambition
Alphabet is in a weird position.
On one hand, it has some of the most advanced AI research in the world.
On the other hand, it has to justify that research in a business model built on advertising.
Which creates tension.
Because better AI doesn’t always translate into better ad revenue. At least not immediately.
So Alphabet has to balance:
- Pushing the frontier
- Maintaining profitability
- Avoiding turning its core business into a cost center
Return on compute here isn’t just about efficiency—it’s about alignment.
Meta: Spending Like It Means It
Meta Platforms has made one thing very clear:
It is willing to spend.
Billions on AI infrastructure. Billions on open models. Billions on building capabilities that may or may not pay off in the near term.
From a return-on-compute standpoint, this is… bold.
Meta’s strategy seems to be:
“If we build enough capability, the returns will figure themselves out.”
Which is either visionary or terrifying, depending on your tolerance for risk.
The Dirty Secret: Inference Costs Matter More Than Training
Everyone loves to talk about training models. It sounds impressive. It involves massive clusters, cutting-edge chips, and numbers that make headlines.
But the real cost?
Inference.
That’s the cost of actually using the model—every time someone asks a question, generates an image, or interacts with AI.
And inference happens at scale.
Millions. Billions of times.
So even small inefficiencies get amplified into massive costs.
If your model is slightly more expensive to run per query, that difference compounds fast.
Which means return on compute isn’t just about building the model—it’s about running it efficiently over time.
The Monetization Problem Nobody Has Fully Solved
Here’s where things get uncomfortable.
AI is incredible at generating value for users.
But generating value for users is not the same as generating revenue for companies.
People love using AI tools. They use them constantly. They integrate them into their workflows, their creativity, their daily lives.
But how much are they willing to pay?
That’s the question.
Because if usage scales faster than monetization, your return on compute collapses.
You end up with:
- High engagement
- High costs
- Unclear revenue
Which is a polite way of saying:
“You built something amazing, and now it’s expensive to maintain.”
Capital Efficiency: The Part Investors Actually Care About
Let’s bring this back to something investors understand: capital efficiency.
Return on compute is really just a modern extension of that idea.
It asks:
- Are you generating more value per dollar than your competitors?
- Are you improving efficiency over time?
- Are your investments compounding—or just accumulating?
Because in the AI era, capital allocation decisions are magnified.
A bad decision isn’t a small mistake. It’s a multi-billion-dollar miscalculation.
The Companies That Will Win (Probably)
If I had to make a prediction—and I will, because that’s half the fun—I’d say the winners in AI won’t necessarily be the ones with the most compute.
They’ll be the ones with the best return on compute.
That means:
- Efficient models
- Smart deployment strategies
- Clear monetization paths
- Discipline in capital allocation
It’s not about who spends the most.
It’s about who spends the smartest.
The Part That Makes Me Slightly Uncomfortable
Here’s the thing I can’t shake:
We’re in a phase where spending is being rewarded more than efficiency.
Companies announce massive AI investments, and the market responds positively. It’s seen as a sign of ambition, leadership, dominance.
But eventually, the bill comes due.
And when it does, the narrative shifts from:
“Look how much they’re investing”
to:
“Look how much they’re earning from it”
That’s when return on compute becomes unavoidable.
Final Thought (Because Every Investment Thesis Needs One)
If I had to summarize everything in one sentence, it would be this:
AI isn’t just a technology race—it’s a capital efficiency race disguised as innovation.
And return on compute is the scoreboard.
Right now, everyone is still playing.
Some are winning. Some are spending like they already won.
But eventually, the numbers will make it clear.
Because in the end, compute isn’t just power.
It’s cost.
And the companies that learn how to turn that cost into consistent, scalable returns?
Those are the ones that won’t just lead AI.
They’ll survive it.
Comments
Post a Comment