The $3 Trillion AI ROI Question, Read From Colombo

The AI ROI question is back, and the number attached to it this time is genuinely hard to picture: $3 trillion. That is roughly how much revenue the AI industry needs to earn to justify what it is spending on data centers, chips, and power. I read TechCrunch's framing of it in Can AI answer the $3 trillion question?, and my first thought was not about OpenAI or Anthropic. It was about the freelancer in Kandy paying for an API key.

Because when the biggest companies in the world are worried about paying back their AI bets, the person most exposed to the fallout is not them. It is the small builder who wired their whole product to a model they do not control.

---

📊 Where the $3 trillion number comes from

The figure traces back to David Cahn, a partner at Sequoia Capital. His math is simple even if the total is not. He pegs AI infrastructure spending at roughly $1.5 trillion for 2026, and calls that "probably an underestimate" because memory costs and specialty chips keep climbing. To earn a return on that kind of capital expenditure, the industry needs to generate multiples of it in revenue. That is where the trillions come from.

Now hold that against what the leading labs actually earn today:

Company	Reported ARR	As of
Anthropic	~$60 billion	current
OpenAI	~$20 billion	November 2025

Key takeaway: Even the two most successful AI companies combined are earning tens of billions against a spending gap measured in trillions. The revenue is real and growing fast. The gap is still enormous.

That is the whole tension. Nobody serious is saying AI produces nothing. The question is whether it produces enough, fast enough, to pay for the concrete and silicon already going into the ground.

---

⚠️ Why this is not just a Silicon Valley problem

You might read all this and think it is a rich-country boardroom fight. It is not, and here is the person who made me take it seriously.

Torsten Slok, chief economist at Apollo Global Management, put the risk in blunt terms. If the hyperscalers — think Microsoft, Amazon, Google, Meta — miss their 2028 cash-flow projections, the damage does not stay contained:

"A slower payoff wouldn't just be a sector problem, it would risk tipping the economy into recession."

Read that carefully. He is not warning about one company's stock. He is warning about the broader economy, and a US or global slowdown reaches Sri Lanka fast through remittances, freelance demand, and tourism. If AI spending has quietly become a load-bearing column of global growth, then a wobble in AI ROI is a macro event, not a tech-news headline.

For a small team here, the practical translation is this: do not assume today's cheap or free AI pricing is permanent. Some of it is being subsidised by companies racing for market share on borrowed time.

---

⚡ The one number that actually helps small builders

Here is the part of the story I found genuinely useful, and it is easy to miss under the trillion-dollar noise. The article notes that OpenAI's latest model is 54% more token efficient on coding tasks.

That single line matters more to a bootstrapped builder than the entire CapEx debate. Efficiency gains are the mechanism by which any of this reaches us. If a model does the same job for roughly half the tokens, your bill roughly halves for that work, whatever the labs are spending upstream.

• Fewer tokens per task means lower cost per feature you ship.
• Better efficiency partly offsets any future price increases the ROI reckoning might force.
• The trend, not the headline number, is what you should track month to month.

If you want to feel this in your own budget rather than take my word for it, run your real usage through our AI API cost calculator. Plug in the number of requests you actually make and see what a 50% efficiency shift does to a monthly invoice. That is the difference between an interesting news story and a line item you control.

---

🛠️ How to build so the reckoning can't sink you

I am not going to pretend to know whether this is a bubble. Cahn is measuring a real gap; the labs are growing real revenue. Both can be true. What I can do is build so I survive either outcome. A few rules I follow:

1. Never hard-wire a single provider. Keep your prompts and calls behind a thin layer you can re-point. If prices jump or a lab folds, you swap, not rewrite.
2. Start on free tiers, then measure. Prove the product works before you commit real money. Compare what each provider gives away with our AI free-tier comparison.
3. Track cost per user, not cost per call. A feature that is cheap per request can still bleed you at scale.
4. Prefer smaller models where they suffice. The efficiency trend is on your side; ride it instead of defaulting to the biggest, most expensive model.
5. Keep a no-AI fallback for anything critical. If a call fails or gets rate-limited, your app should degrade, not die.

Bottom line: The trillion-dollar question is theirs to answer. Your job is to make sure the answer, whatever it turns out to be, does not decide whether your product lives.

---

💡 What this means for you

If you are a student, a freelancer, or a two-person team in Sri Lanka building on top of these models, the $3 trillion question is not something you can solve, and you should not lose sleep over predicting it. What you can do is refuse to be a hostage to it.

The labs are spending like the payoff is guaranteed. Slok is reminding everyone it is not. Somewhere between those two positions is where you build: use the tools while they are cheap, keep your architecture portable, watch the efficiency gains flow into your costs, and never bet your business on a price that only exists because a giant is burning cash to win a race. Do that, and it genuinely does not matter to you whether AI answers its trillion-dollar question this year or in 2028.