[你懂的]  AI’s Next Battle May Not Be GPUs: Meet the Company Selling the Data and “Tests” Behind AI

When investors talk about AI, the names usually come quickly:

NVIDIA for compute, Broadcom for networking and custom chips, and Microsoft, Amazon and Google for cloud infrastructure and foundation models.

But there is another question that is becoming increasingly important:

Even if you have all the GPUs in the world, what will AI actually learn from without high-quality data?

And as AI moves beyond simple chatbots toward AI agents, reasoning models and enterprise applications, another question emerges:

How do we know whether an AI system is actually good?

Is its answer correct?

Does it hallucinate?

Can it complete a complex task as instructed?

And in high-stakes industries such as finance, healthcare, law and government, what happens when an AI system gets something seriously wrong?

That brings me to a relatively under-the-radar company:

$Innodata(INOD)$

INOD is not a GPU maker. It is not a foundation-model company either.

Instead, it operates behind the scenes, helping AI companies prepare data, train and optimize models, and evaluate how well those models actually perform.

And what makes the company particularly interesting to me is this:

INOD is no longer just selling an AI story. It is starting to turn AI demand into real revenue, profits and cash flow.

---

1. First, What Does INOD Actually Do?

Innodata has been around for more than 36 years.

That matters.

There are plenty of small-cap companies that suddenly added “AI” to their story over the past few years.

INOD isn't one of them.

The company started as a data-focused business and gradually shifted toward AI data engineering, model training, model evaluation and AI deployment support as machine learning and generative AI developed.

The company was already building AI language models and related data-production capabilities back in 2016–2017.

Today, its business can broadly be divided into four areas.

First: Training Data

Helping AI models obtain high-quality data.

Second: Post-Training

Once a model has been trained, it still needs to be refined and aligned with human expectations.

Third: AI Evaluation

Testing whether a model actually performs well.

Fourth: AI Enablement

Helping enterprises deploy AI into real-world workflows, including AI agents and systems capable of using external tools.

According to its 2025 annual report, INOD serves leading AI innovators, technology companies and large enterprises, including five companies within the so-called “Magnificent Seven.”

So here's a simple way to think about it:

GPUs provide the computing power.

The AI model is the brain.

INOD provides the textbooks, practice exercises, grading and exams.

And that's what makes this company interesting.

---

2. Why Could Data Become the Next AI Bottleneck?

For the past two years, the AI investment thesis has been relatively straightforward:

More AI → more GPUs → more data centers → NVIDIA benefits.

But as models become more advanced, simply throwing more computing power at the problem doesn't solve everything.

Imagine giving an AI system one million articles to learn from.

That doesn't automatically make the model smarter.

Those articles could contain:

Duplicate information.

Incorrect information.

Low-quality content.

Copyright issues.

Private or sensitive information.

Inconsistent formats.

Even contradictory answers.

So the more important questions become:

«Which data should AI learn from?»

And:

«How should that data be structured and prepared?»

That's where data engineering comes in.

And the problem becomes even more complicated in the age of AI agents.

A traditional chatbot might simply answer a question.

An AI agent could:

Search for information → understand the task → call external tools → perform calculations → modify files → check the result → try again.

If something goes wrong at any step, the final result can be wrong.

That means enterprises need increasingly sophisticated AI evaluation.

In other words:

Yesterday's AI needed a teacher.

Tomorrow's AI may also need an examiner.

And INOD is trying to become that examiner.

---

3. What Really Caught My Attention Isn't the Story — It's the Financials

If INOD were simply saying:

«“AI is going to be huge, therefore we have a huge opportunity.”»

I wouldn't be particularly interested.

There are hundreds of companies saying exactly that.

What made me take a closer look is that the AI demand is already showing up in the company's financial results.

Full-year 2025

INOD generated approximately:

$251.7 million in revenue

up:

48% year over year.

Adjusted EBITDA reached:

$57.9 million

up:

68%.

Net income was approximately:

$32.2 million.

That's no longer just an “AI concept.”

That's a real company generating real profits.

And growth accelerated further in 2026.

---

4. The Second Quarter of 2026 Was Even More Interesting

INOD's latest second-quarter results are particularly worth watching.

Quarterly revenue reached:

$92.1 million

up:

58% year over year.

Adjusted gross profit:

$45.4 million

Adjusted gross margin:

49%.

Adjusted EBITDA:

$25.4 million

up:

92%.

Net income:

$14.4 million

versus:

$7.2 million

a year earlier.

In other words, net income nearly doubled.

That leads to a much more important question:

Is INOD becoming more profitable as it grows?

So far, the trend looks encouraging.

Revenue is growing, but margins are improving as well.

Adjusted gross margin reached 49% in Q2 2026, compared with 43% a year earlier.

Why does that matter?

Because if revenue continues to grow, the company isn't necessarily just doing:

«Revenue increases → costs increase → little additional profit.»

Instead, we're beginning to see:

«Revenue growth → improving margins → faster EBITDA growth.»

That's operating leverage.

And operating leverage is exactly what investors want to see in a high-growth company.

---

5. Why Could Its Margin Profile Become Increasingly Important?

There's an important issue here that investors may overlook.

Historically, a large part of INOD's business could be described as:

People + data processing + customer projects.

That naturally makes the company look somewhat like a labor-intensive services business.

And that raises an obvious question:

If INOD relies heavily on humans to process and label data, could AI eventually replace INOD itself?

That's a completely legitimate concern.

But the company is not simply trying to add more people.

It is increasingly combining:

Automation + proprietary platforms + human experts + AI systems.

One particularly interesting development is its push into off-the-shelf datasets.

The difference is important.

The old model:

«Customer A needs a dataset → INOD builds it → INOD gets paid.»

The more attractive model would be:

«INOD creates valuable proprietary data assets → those assets can potentially be monetized across multiple customers.»

Those are two very different businesses.

The first looks more like:

A services company.

The second starts to look more like:

A data-asset company.

INOD said in its Q2 2026 results that off-the-shelf datasets and high-value pre-training projects contributed to margin improvement, while some of its data assets may retain intellectual property rights and potentially be monetized across multiple customers.

If this model scales successfully, the company's profit structure could become increasingly attractive.

---

6. Then There's the Bigger Story: AI Agents

This may ultimately be one of the most important parts of the INOD thesis.

AI is moving from:

Chatbots

toward:

AI agents.

A chatbot might answer:

«“Analyze this company for me.”»

An agent could actually:

Open the files → search the web → collect information → build a model → run calculations → write a report → check its own work.

That makes AI systems dramatically more complex.

And it creates a new problem:

How do you test an AI agent?

Simply asking:

«“Was the final answer correct?”»

may no longer be enough.

You also need to know:

Did it use the right tool?

Did it execute the correct steps?

Did it leak sensitive information?

Did it hallucinate?

Did it follow the company's rules?

Did it behave safely under unusual circumstances?

This is where AI Evaluation + Safety + Assurance becomes increasingly important.

And this is an area INOD is actively pursuing.

The company's annual report specifically discusses AI agents, tool-using systems, and the evaluation of AI capability, alignment and safety.

That's why I don't think the most interesting part of INOD is simply:

“AI data labeling.”

The bigger question is whether INOD can evolve from:

«Data Provider»

into:

«AI Lifecycle Partner»

In other words, can it participate throughout the AI lifecycle — from data preparation and training to evaluation, deployment and ongoing monitoring?

If it can, the potential value of the business would be significantly greater than that of a traditional data-processing company.

---

7. Its Customers Are Both an Advantage and a Risk

This is where we have to look at both sides of the story.

One of INOD's biggest advantages is that it has already entered the ecosystems of some very difficult-to-win customers.

The company says it serves five of the Magnificent Seven companies, along with leading AI research organizations.

But there's a catch.

The more important a large customer becomes, the greater the concentration risk.

In 2025, one DDS customer accounted for approximately:

58% of total company revenue.

That figure was about:

48% in 2024.

That's extremely high.

So if you simply see:

«“INOD has major technology customers.”»

and conclude:

«“Then the revenue must be safe.”»

That's the wrong conclusion.

Large customers also have significant bargaining power.

If a customer reduces projects, delays spending, builds more capabilities internally or switches suppliers, INOD could feel the impact quickly.

And many of its customer agreements are not permanent.

The annual report says most customer agreements can generally be terminated with 30 to 90 days' notice.

So:

High-quality customers ≠ risk-free revenue.

In fact, customer concentration is one of the most important things I would monitor going forward.

---

8. But INOD Is Also Trying to Diversify Its Customer Base

This is another part of the story worth watching.

If INOD can move from:

One dominant customer

toward:

5, 10 or 20 major AI customers

the quality of its revenue base could change dramatically.

The company has been expanding its customer base and moving into more AI applications.

Beyond large technology companies, it also works with industries including banking, insurance, financial services, retail and media.

The company also established a dedicated Federal Practice in 2025, targeting AI data engineering and deployment opportunities across areas such as defense, intelligence and regulatory agencies.

If government AI projects move from experimentation toward long-term deployment, that could become another potential source of revenue.

---

9. The Balance Sheet Is Worth Watching Too

At the end of 2025, INOD had approximately:

$82.2 million in cash and short-term investments

and no debt drawn.

By the end of June 2026, cash, cash equivalents and short-term investments had risen to:

$250.4 million.

But there's an important detail here.

The company said part of this balance was related to customer advances.

Excluding those customer advances, cash was approximately:

$134 million.

So it would be misleading to say:

«“INOD suddenly generated $168 million of free cash.”»

That's not what happened.

Still, even after accounting for those customer advances, the company's balance sheet appears relatively healthy.

For a small, fast-growing company, that's important.

---

10. So What's the Core Investment Thesis?

If I had to summarize the entire INOD thesis in one sentence:

«INOD isn't betting on which AI model wins. It's betting that whichever models win, advanced AI will need better data, evaluation and continuous optimization.»

That's what makes the company interesting.

Suppose the future belongs to:

OpenAI.

Google.

Meta.

Anthropic.

Amazon.

Or even a completely different AI architecture that doesn't exist today.

INOD could potentially serve them all.

Because it doesn't need to sell the model itself.

It sells the infrastructure needed to make those models:

better, more reliable and more useful in real-world applications.

It's somewhat like:

«Selling the picks and shovels — except the shovel isn't for the gold miner. It's the textbook, exam and quality-control system for the AI model.»

---

11. But I Wouldn't Call INOD “The Next 10-Bagger”

This is extremely important.

The biggest danger with a great company is often that investors become willing to pay too much for it.

The question isn't simply:

«Can INOD grow?»

The more important question is:

«How much future growth is already priced into the stock?»

If the market is already valuing INOD on the assumption that it will grow:

40%–50% annually for several years,

then even if the company grows 30%, the stock may still struggle because the valuation multiple could compress.

On the other hand, if INOD can maintain:

40%+ revenue growth

while continuing to expand margins,

add more customers,

reduce customer concentration,

and turn AI Evaluation, Agent and Federal AI opportunities into meaningful revenue,

then a higher valuation could eventually be supported by fundamentals.

That's why I wouldn't focus too much on the daily stock price.

I'd watch these five things instead:

① Revenue Growth

Can it sustain 30%–40%+ growth?

② Gross Margin

Can the roughly 49% margin continue to expand?

③ Customer Concentration

Can the dominant customer's share of revenue fall from 58%?

④ Adjusted EBITDA

Can EBITDA continue growing faster than revenue?

⑤ AI Evaluation / Agent Business

Can these become meaningful sources of actual revenue rather than just an interesting story?

If all five improve simultaneously, the INOD investment thesis becomes much stronger.

---

12. So Why Is INOD Worth Watching?

Because the market has become accustomed to thinking about AI in a very specific way:

GPU → NVDA

Networking → AVGO / MRVL

Power → VRT / CEG

But there is another layer of the AI ecosystem:

Data → Training → Evaluation → Deployment

And INOD sits directly within that chain.

More importantly, it has already demonstrated something many AI-themed small caps haven't:

This isn't just a PowerPoint story.

Revenue grew 48% in 2025.

Q2 2026 revenue grew another 58%.

Net income reached $14.4 million.

Adjusted EBITDA grew 92%.

And adjusted gross margin expanded from 43% to 49%.

Of course, INOD remains a small, high-growth and potentially high-volatility company with meaningful customer concentration risk.

I would not put it in the same risk category as broad-market ETFs such as VOO or QQQ.

But if you ask me:

«“Is there a relatively under-the-radar company that can help us understand how the next phase of AI could actually make money?”»

INOD is absolutely worth researching.

Because as AI moves deeper into commercialization, the competition may no longer be just about:

Who has the most GPUs?

It could increasingly be about:

Who has the best data?

Whose models are the most reliable?

Whose AI agents make the fewest mistakes?

And that may be exactly where INOD wants to position itself.

---

The Bottom Line

I wouldn't say “INOD is going to rise” simply because the AI story sounds attractive.

But when you look at the company from:

Business model → AI positioning → revenue growth → margins → profitability → cash → new growth opportunities → risks,

it looks more interesting than many small-cap stocks whose only connection to AI is a catchy presentation.

The Bull Case

AI agents, model evaluation and data engineering continue to expand rapidly.

INOD diversifies its customer base.

Margins continue to improve.

And the company evolves from a data-services provider into a broader AI lifecycle infrastructure partner.

The Bear Case

Customer concentration remains extremely high.

Project-based revenue proves volatile.

Large AI companies increasingly build data capabilities internally.

Or, perhaps most importantly, the stock valuation already prices in years of high growth.

That's why I wouldn't describe INOD as a “close-your-eyes-and-buy” stock.

But I do think it belongs on the watchlist of investors looking for less obvious AI picks-and-shovels opportunities.

And for me, the next earnings report matters more than the next AI headline.

I want to see:

Can revenue growth stay near 50%?

Can gross margin remain around 50% or move higher?

Can customer concentration finally start coming down?

Those numbers will tell us far more than another headline about how big AI could become.

Because maybe the first chapter of AI was about compute.

The second could be about data.

And the third may be about proving that AI actually works.

INOD is betting on the latter two.[思考]