The ARR Illusion in the Age of AI

— The moment GMV is labeled Annual Recurring Revenue, the business is built on sand

In recent feeds and headlines covering AI startups, a certain phrase repeats with exhausting frequency: "Reached $X million ARR in just n months."

At first, it is impressive. The second time, it invites skepticism. By the third, one inevitably arrives at the question:

"Is that actually ARR?"

This is not merely a matter of accounting tricks or semantic hair-splitting. It is a fundamental question about the structure of AI businesses, their sustainability, and how we ought to interpret the numbers presented to us.

What ARR Actually Is — A Concept Strict in its Simplicity

ARR stands for Annual Recurring Revenue. The two operative words here carry weight.

Annual: A full year must effectively elapse or be contractually guaranteed.
Recurring: The revenue must be structurally proven to repeat, not merely happen by chance.

Therefore, ARR is not:

"Last month's revenue × 12"
"This month's usage annualized"

These are run-rates. ARR is not a prediction; it is a result verified by time.

Fundamentally, For a company less than a year old to claim ARR is, conceptually, a contradiction.

The year has not passed. The recurrence has not been proven.

Yet Everyone Speaks in ARR. Why?

The reason is simple.

ARR is the lingua franca of SaaS. And the language of SaaS invariably connotes:

Stability
Predictability
High Valuation Multiples

Thus, many AI companies, even those far from reaching that stage, borrow the term ARR to describe their finances.

The problem begins exactly there.

The Real Numbers of AI Startups — Closer to GMV than ARR

Peer slightly beneath the surface of many AI startup revenue structures, and a pattern emerges.

A significant portion of what the customer pays
Flows immediately out the door.

For instance:

LLM API costs
GPU and compute expenses
Costs for external human-in-the-loop contractors

In this structure, the portion the company actually retains is but a fraction of the total payment.

Here, the correct concept is GMV (Gross Merchandise Value).

GMV represents the total value of merchandise sold through a platform. The critical distinction is this:

GMV is the "flow of money."
It is not the money the company earned.

If a platform mediates a transaction and takes a commission, GMV ≠ Revenue.

The Core of the Problem — When GMV Wears the Mask of ARR

The illusion specific to the AI industry arises here.

Contracts are signed monthly or annually.
Billing is recurring based on usage.
Dashboards display "how much was spent this month."

On the surface, it mimics SaaS perfectly.

But in reality:

Cost control lies with the model provider.
Pricing power resides externally.
The company takes only a thin slice from the middle.

In this structure, calling GMV "ARR" is, An act of consuming an unproven future as a present achievement.

To put it more bluntly, it borders on a bluff designed to induce a financing illusion.

On the Counterargument: "Won't Compute Costs eventually fall?"

A common rebuttal exists.

"Model costs are dropping, so even if margins are thin now, they will improve."

This statement is only half true.

Technology costs do decrease. However, demand always migrates to the newest, most powerful models.

People do not:

Choose GPT-3.5 because it is cheap.
Select a lower-tier Claude model on purpose.

Of course, low-cost models have their place—simple tasks, non-competitive domains.

But the question remains: Is the product you are building truly that kind of product?

We are cognitively greedy. Given the same time and money, we desire the smarter brain.

Therefore,

The assumption that "costs will fall"
Relies on the premise that "lower performance models will suffice."
A premise that rarely holds in reality.

This is Not a Numbers Problem; It is an Identity Problem

Ultimately, the discussion boils down to one thing.

Is this a Software Company?
Or is it a low-margin Reseller/Broker?

Neither is inherently right or wrong. But how they are valued must be completely different.

For SaaS, every dollar of revenue can be a unit of value.
For Reseller/Broker models, one must judge by:
The actual take rate (net revenue).
Contribution margin.
Control over costs.

A large GMV does not equate to a solid foundation.

Closing — Growing Numbers vs. Growing a Business

The spectacle of GMV-first growth often reminds me of a familiar scene.

The investor with $10,000 of capital, who trades dozens of times a day, boasting, "I moved millions in transaction volume this month."

The trades are many. The numbers are large. But what actually remains?

Mostly transaction fees paid to the exchange. The volume swells, but the asset base barely moves.

AI startups obsessed with GMV are not so different.

Transactions abound.
Money moves fast.
The numbers on the dashboard keep growing.

But the vast majority of that flow passes through to model providers and infrastructure giants, leaking straight out of the building.

To discuss GMV as if it were ARR in this context is not a demonstration of capital control, but rather an attempt to prove competence via the sheer volume of money that passed through one's hands.

From a consulting perspective, the questions are simple.

Is this company controlling its recurring revenue?
Or is it merely brokering atop recurring costs?
Does the margin structure improve over time?
Or does the fee burden simply grow with volume?

ARR is not a marketing slogan. It is a report card proving that a financial structure has withstood the test of time.

GMV can be a signal of growth. But it is, at best, a process metric. It cannot substitute for the foundation.

In the AI era, numbers have become larger and faster. That is precisely why we must be more honest with our terminology.

The moment you slap the label of ARR onto GMV, the business begins to look solid, but in reality, it is being built on sand.

Time will eventually reveal the truth. Whether you were a company that made many trades, or a company that accumulated value.

Time always distinguishes the two.