Reference price: $56.22 (August 31, 2026). Multiples move with the price; figures come from official filings with the US regulator.
What this company actually does
It prepares the study notes for artificial intelligence. AI models learn from examples: millions of texts, images and conversations classified, corrected and labeled by hand. Someone has to do that work, and that someone is people. Innodata has 10,107 employees — mostly in India and the Philippines — doing exactly that for the big AI labs.
Be careful with the "technology" label on the stock screener. Run this number: $317 million in revenue with 10,107 employees = about $31,400 per person per year. A real software company generates $300,000–500,000 per employee.
This is not a software company. It is a labor company that works for software companies. And that changes everything.
The good
Revenue has multiplied 5.5x in five years. From $58M in 2020 to $170M in 2024 (+95%), $252M in 2025 (+48%) and $317M over the last twelve months. Last quarter: $92M against $58M a year earlier — +58.6%. This is real, it is in the official accounts, and it is not up for debate: demand for the service exists and it is enormous.
No debt and $240M in cash. Cash went from $117M to $240M in one quarter. Goodwill is $2M: it has acquired nothing — everything it owns, it built.
It earns a lot on very little capital. Return on invested capital ~29%, return on equity ~38%. And its published accounts are exact: announced results match regulatory filings to the cent in all three years that can be cross-checked.
The bad
Problem 1: revenue up 39%, earnings up 4%. Over the last twelve months revenue grew +39% and EPS +3.9%. Same story for the full year: 2025 sold 48% more than 2024 and each share earned three cents more ($0.89 → $0.92).
Problem 2: it pays itself in stock, and the bill has multiplied by six. From $4M a year in 2023–2024 to $11M in 2025 and a $26M annual run rate in 2026. Against roughly $58M of annualized profit, nearly half of what it earns is being paid out in new shares.
Problem 3: every dollar of revenue leaves less behind than a year ago. Gross margin has slid from 45.2% (winter 2024) to 38.3% (winter 2025). And this is not bad luck — it is the nature of the business. If labeling twice the data takes twice the people, doubling revenue costs you double the headcount. There is no economy of scale — the exact opposite of what makes a software company worth 40 years of earnings.
Problem 4: insiders have sold almost half of their stake. Insider ownership is down 48%, to 5.08% of the company — with the stock coming down from $125.
Problem 5: down 55% in three months. From a high of $125.14 to $56.22. It moves nearly three times as much as the market (6.6% on a normal day) and 14% of the float is betting it falls further.
Problem 6: only two profitable years in its recent history (2024 and 2025) — and between them, earnings grew 3.4%. From 2021 to 2023 not a single analyst covered it: the company was abandoned until the AI wave arrived.
Problem 7: customer concentration cannot be verified in the machine-readable part of the filings. For a company selling training data to AI labs, this is the decisive number: if one customer is 40% of revenue and builds an in-house team, revenue collapses from one quarter to the next. It is the first thing to check before putting a dollar here.
Problem 8: the cash it truly generates doesn't cover your cost of money. Against a $1.93B enterprise value, 2025's clean $36M of cash flow yields 1.87% — less than the cost of financing.
The question that decides everything
Only one of two things can happen: either it gets the machine to do the work 10,107 people do today — using AI to label AI data — and margins explode and this is worth the price and more; or it keeps needing one person per labeled data point, and then it is a services company growing fast but earning little, valued as if it were software.
Not both. And for now the numbers say the second: margins are not rising — they are falling. One detail that does not help: the people best placed to know the answer — management — just sold half of what they owned.
What we would do
Not buy it at this price. And we want to be fair: this is not vaporware. It genuinely sells, owes nothing, has acquired nothing, and publishes clean accounts. The problem is that growth is not making it more profitable, and at 42 times earnings the only thing justifying the price is a margin take-off. It is doing the opposite.
| Price | What you would be paying for |
|---|---|
| Around $34 | 25 years of this year's earnings. Where it traded a few months ago. |
| $34–50 | Reasonable if 40% growth holds. |
| $50–60 ← it is here ($56.22) | Paying for all of 2027 in advance. |
| Above $60 | Paying for margin improvement — which is not happening today. |
The three things we would watch
1. Gross margin back above 45. It is at 38 and falling. That number is the whole thesis: if it rises while revenue grows, the company has automated its own work and this becomes something else.
2. Stock compensation per quarter. Running at $26M a year, six times 2024. If it keeps climbing with revenue, your slice of the company shrinks exactly as fast as the company grows.
3. The customer breakdown in the written report. The one number that can break the company in a quarter. If a single customer crosses 25–30%, the risk changes category.
Next earnings: early November. No rush from any direction.
In one sentence
It prepares the study notes for artificial intelligence with ten thousand people, has quintupled revenue in five years, and each share earned three cents more — because every new data point it labels needs another person to label it, and the people who understand that detail best, management, just sold half of what they owned.