- Published
- October 25, 2026
- Reviewed
- October 25, 2026
- Author
- Gurbaksh Chahal
- Category
- Founder Insights
- Reading
- 16 min
By Gurbaksh Chahal
I was sixteen when I started ClickAgents. The internet was growing faster than the machinery needed to turn attention into a business. Publishers had audiences. Advertisers had budgets. Connecting the two was an opportunity.
That was 1998.
I went on to build BlueLithium and RadiumOne, and the same industry taught me three different lessons. First, distribution mattered. Then the intelligence behind that distribution mattered more. Eventually, the advantage moved toward systems that could use signals, make better decisions and prove the result.
I have watched that shift happen before. And today, watching hundreds of companies race to supply experts, datasets, agent traces and training environments to a small number of AI labs, I recognize the early stages of another one.
Gokul Rajaram recently compared the AI data business to the ad networks of the 2000s. He is right about the resemblance. But this isn't a historical analogy I learned from reading about ad tech. I was building in the middle of it.
What interests me is not how similar these two industries look today. It's what that similarity tells us about who will still matter tomorrow.
The question has always been the same: When everyone can sell access, what do customers still need you for?
ClickAgents: Access was the opportunity. Performance was the test.
When I launched ClickAgents, online advertising was still becoming a real market. Many advertisers didn't have an efficient way to reach independent publishers, and many publishers didn't have a reliable way to turn their traffic into revenue. Networks helped make that market work.
That was valuable. It was also the easiest part for competitors to understand.
If ten companies could offer access to similar audiences, buyers eventually compared what they were paying and what they received. Reach alone stopped being a convincing answer. The discussion moved toward cost per click, customer acquisition and actual performance.
The lesson I carried out of ClickAgents was simple: Aggregation creates an opening. Accountability earns the next order.
That distinction is getting lost in the excitement around AI data.
An AI data company can recruit thousands of experts, source private datasets and organize impressive delivery operations. Those are genuine achievements. But if the same experts work across several platforms and the lab specifies exactly what they should produce, the supplier has to answer a difficult question.
What did we build here that will get better with the next delivery?
A larger roster isn't automatically a stronger business. Neither is a larger pile of completed tasks.
There is nothing wrong with running a first-rate services company. I have enormous respect for the discipline it takes. But a services business and a compounding technology platform should not be valued or managed as though they are the same thing.
The first wave of AI data vendors is solving a real access problem. The next wave will be judged on what happens after access becomes easier.
BlueLithium: The impression wasn't the asset. Understanding behavior was.
By the time I started BlueLithium in 2004, the market had changed. There was more advertising inventory, more competition and more pressure to make each impression useful.
We focused on behavioral targeting.
That meant looking beyond the fact that an ad could be delivered. The more important question was whether the available signals helped us understand what a person might actually care about and make a better advertising decision.
BlueLithium eventually reached more than 120 million monthly users and was acquired by Yahoo in a transaction announced at approximately $300 million in 2007.
The value wasn't simply that we had a large network. It was the technology and understanding we built around the signal.
That distinction is almost identical to the choice facing AI data companies today.
A million rows of operating history are not the same as a million valuable training examples. Ten thousand coding-agent sessions are not ten thousand usable RL tasks. A recorded interaction is not automatically evidence that a task was completed correctly.
The hard work is in the context.
What was the person trying to do? What information was available at that moment? What decision did they make? What failed? What was corrected? What did the final state look like? How do we know the result was right?
A dataset becomes much more useful when those questions can be answered. Sometimes the original archive is the scarce asset. Sometimes it is the demonstrated expert judgment inside it. Sometimes it is the ability to turn a complicated sequence of events into a realistic task that a model can actually practice.
That is where I expect the premium to move: away from the volume of data a vendor can deliver and toward the intelligence and reliability it can preserve.
You can buy more data. You cannot buy your way around not understanding what it means.
RadiumOne: The signal gets more valuable when you can act on it.
With RadiumOne, we looked at a different part of the internet. People were sharing links, content and information across the open web. Those relationships contained signals of interest that traditional audience categories often missed.
Our work in social-graph and sharing technology was an attempt to make sense of that activity and turn it into better targeting and programmatic decisions.
What stayed with me was not the idea that more signals are always better. It was that disconnected activity becomes useful when the system understands the relationships, the timing and the outcome.
There is an important difference between collecting events and making those events actionable.
That is how I think about the next stage of AI data.
If an agent fails while updating a repository, the failed answer alone may tell you very little. The useful record might include the starting code, tools called, tests attempted, wrong assumptions, human corrections and the final verified result.
If an enterprise workflow fails, the story may be spread across an inbox, a CRM, a finance system and an approval log. Isolated records tell only pieces of it.
The valuable supplier doesn't just hand the buyer those pieces. It helps reconstruct the work, identify the failure worth teaching and create a way to measure whether the model has learned to handle it.
That is a system, not a transaction.
And it can improve, but only if the supplier gets meaningful feedback and is disciplined enough to use it. The next task should be better designed, easier to verify or less expensive to produce. Otherwise, activity is growing without a real advantage growing alongside it.
Revenue can grow faster than the value of the company.
I remember how fast digital advertising attracted intermediaries once the budgets arrived. There was always another network, another buying relationship, another way to count the money flowing through the system.
AI data has its version of that excitement now.
A company can report spectacular monthly revenue while most of the money passes straight through to experts, source owners, contractors, compute providers and delivery operations. It may still be a very good business. But the headline tells me surprisingly little.
I want to see the economics underneath it.
How much is actually collected? How much goes out to fulfill the order? How much work is rejected or redone? How long before the buyer accepts it? What does it cost to win the second order? Does each new delivery become easier, or does every project require rebuilding the workforce and the process?
Imagine a business that invoices $10 million and spends $7 million fulfilling the work. There may be $3 million left before other expenses. That could support a profitable company. It does not, by itself, establish software-like margins or recurring value.
Multiply one unusually strong month by twelve and you have an annualized pace, not twelve signed months of repeat demand.
Accounting treatment matters, too. Gross billings, recognized revenue, gross profit and cash in the bank are not interchangeable. I would rather see a smaller number I can understand than a larger number that hides the actual economics.
This isn't skepticism about the demand. The demand is real enough for sophisticated buyers and suppliers to invest heavily in the category.
It is skepticism about what we're measuring.
Ad tech taught me that markets eventually stop paying a premium for the privilege of moving money from one party to another. They pay for outcomes, reliability and useful infrastructure.
The first purchase order is a promise. The second is proof.
I've always believed that the real test of a company starts after the sale.
A customer can be convinced by a good demonstration. The harder question is whether the product performs when it leaves the demonstration.
In AI training, the difference is even more unforgiving.
A lab can buy a task set, run it against private evaluations and discover that the model improved on the vendor's benchmark but not on the problems the lab actually needs to solve. Or it can discover that an agent passed the tests by exploiting the grader instead of completing the intended work.
That's not a useful training outcome, regardless of how polished the dataset looks.
I expect researchers to become much more demanding about this. They will want evidence that the task was genuinely difficult, the environment could be reproduced, the grader resisted shortcuts and the resulting improvement transferred to problems the vendor didn't select.
That changes who wins the relationship.
The vendor who shows up with a catalog has one kind of conversation. The vendor who understands the research team's failures, helps design a useful intervention and can prove the result has a different one.
The researcher's confidence becomes the real currency.
Procurement can negotiate the commercial terms. It cannot make an irrelevant task valuable to a model.
I have a simple rule for this business: Never confuse internal acceptance with customer acceptance. And never confuse customer acceptance with measured model improvement. Those are separate milestones.
If a supplier can repeatedly move through all three, it has a reason to be invited back.
AI data is about to have its own fraud and provenance reckoning.
Anyone who lived through programmatic advertising remembers that scale created an uncomfortable question: Was the traffic real, and did the seller have the right to sell it?
We dealt with bots, opaque supply paths, questionable inventory and measurement disagreements. Better standards didn't solve every problem, but the market became more serious about verifying where impressions came from and who was authorized to sell them.
AI data is approaching a similar moment.
Some contractors are reportedly using AI models to complete tasks that buyers believe are original human work. An answer may be excellent, but if a lab paid for an independent expert judgment and received an undisclosed model-generated response, the product has been misrepresented.
There is another layer to this problem. A company can supply data it possesses without necessarily owning the rights needed to turn that data into model-training material. Anonymizing names does not automatically erase confidential business processes or intellectual property.
These questions are going to become commercial, not merely legal or technical.
Who originated this work? Which systems and tools were used? Was AI assistance permitted and disclosed? Was the outcome independently checked? What transformations were applied? Which uses are licensed? What happens when a source must be excluded or recalled?
I expect an entire set of standards and operating practices to emerge around those answers, much as ads.txt, sellers.json and supply-chain transparency helped buyers understand digital-ad inventory.
But we should learn from ad tech's mistakes as well. A badge on a dashboard is not proof. An automated detector is not a complete provenance system. And recording everything without a clear legal basis is not a responsible shortcut to a defensible dataset.
The answer is better evidence, tied to the underlying work and enforceable rights.
Synthetic data is not the enemy. Misrepresented data is. Unverifiable quality is. A source that cannot explain its permissions is.
Trust will become a differentiator precisely because trust is expensive to earn.
The next product isn't a bigger dataset. It's somewhere the model can practice.
This is where I believe the opportunity becomes much more interesting.
Think about a retailer facing an inventory shortage. A real employee might have to read a supplier email, reconcile purchase orders, check warehouse stock, look at promised customer delivery dates, calculate the financial impact and decide which approvals are required.
That is a business task. It involves several systems and a chain of decisions.
A static dataset can describe what happened. A well-designed training task can ask an AI agent to work through the problem. A reliable RL environment can let it take actions in a controlled version of those systems, encounter realistic constraints and receive a trustworthy result when it finishes.
The difference is enormous.
The environment can reset. It can vary the starting conditions. It can introduce new exceptions. It can reveal whether a model can handle a longer workflow, not merely produce a persuasive paragraph about one.
But I don't think building a simulator is automatically the moat either.
If its tasks are unrealistic, the model learns the wrong behavior. If its grader can be exploited, the model learns shortcuts. If it keeps producing trivial variations of the same exercise, its value decays as models become better.
The real business is maintaining useful difficulty.
That requires authentic source context, people who understand the domain, engineers who can reproduce the work and evaluation methods that reward the right outcome.
It's why I expect specialists to have a real advantage. A legal expert can recognize the flaw in a superficially convincing contract review. An experienced software engineer can distinguish a real fix from a patch that simply silences a test. A finance operator knows when a workflow was completed correctly and when someone skipped the approval that mattered.
General infrastructure will be valuable. So will domain-specific judgment.
The companies that bring those together will be far more difficult to replace than the ones merely reselling access to either.
I don't think the future is simply an exchange for AI data.
It is tempting to watch ad tech evolve and conclude that the same winner will emerge here: an exchange connecting every data supplier to every lab.
I'm not convinced it will be that simple.
An ad impression is perishable. Once delivered, it's gone. Training data can be reused, transformed or licensed repeatedly, depending on its rights. The value of a particular task depends on what the buyer's model already knows and what its researchers are trying to improve.
That makes data harder to price as a standard commodity.
There are also far fewer frontier-lab buyers than there were advertising buyers. Many sensitive datasets and custom environments will continue to be negotiated directly because they demand specialized security, legal terms and evaluation.
What I do expect is a stronger infrastructure layer.
Better provenance formats. Clearer descriptions of training rights. Reproducible task and environment packages. Standard ways to report accepted versus rejected material. Reliable interfaces for graders. Auditable records of what was delivered and why it passed.
Some of those capabilities will become common protocols. Some will become standalone businesses. Others will live inside the suppliers that have earned the deepest relationships with researchers.
DoubleClick and The Trade Desk became important not because they simply took another fee between buyer and seller, but because they became part of how customers operated. Google and Facebook built enormous businesses in part because they owned differentiated consumer-facing products and the inventory those products created. These are different forms of advantage, and they won't translate into one identical AI-data business model.
The right question isn't, “Where can we put a toll booth?”
It's, “Which problem becomes harder for the customer if we disappear?”
What I expect the next three years to look like.
First, the easy pricing power will weaken. More vendors will learn to recruit experts, package data and build basic task environments. Buyers will compare suppliers more closely. Some businesses will consolidate, some will specialize, and some will remain profitable services companies without becoming giant platforms.
Second, the financial story will mature. Large revenue announcements will still attract attention, but serious buyers and investors will pay much closer attention to retained revenue, repeat orders, fulfillment margins, accepted-unit economics and working capital. A data vendor can grow while destroying cash if it pays its workforce today and waits months for customer acceptance.
Third, source rights and provenance will become central to product quality. Companies that can document legitimate access to unusual, useful operating data will have an advantage. Companies that cut corners on confidential sources or describe AI-generated work as independent human judgment will have a problem they cannot solve with marketing.
Fourth, research integration will matter more than generic catalog size. A supplier that sees a buyer's failure mode, builds a useful task family, tests it honestly and improves its next batch has a better chance of renewal than one delivering another interchangeable thousand prompts.
And fifth, some of the best data businesses will move closer to where work happens. They may build useful developer tools, operate specialist networks or complete enterprise workflows themselves. That creates a continuing source of real problems, failures, corrections and outcomes.
That last opportunity is powerful, but it carries responsibilities. Owning a product does not give a company unlimited rights to every customer's code, messages or proprietary process. Durable supply has to be built on legitimate permissions and real utility.
I expect this market to keep growing. I also expect many of its current companies to discover that growth and durability are not the same thing.
Both things can be true.
Why I'm building AuraOne around the result, not the record.
My path through advertising shaped how I think about the business I'm building now.
At ClickAgents, I learned why distribution matters. At BlueLithium, I learned why understanding the signal is more valuable than counting the impression. At RadiumOne, I learned that a connected signal is only useful when it leads to a better decision.
At AuraOne, I want those lessons applied to AI training and real enterprise work.
That means putting expert judgment, source context, execution history, verification and accepted outcomes in the same system. It means making it possible to understand what was requested, what happened, which decisions mattered, why an output passed review and what the customer actually received.
It also means recognizing the difference between the useful work a product performs and the data that work might generate. A coding tool should earn its users because it helps them build. An enterprise service should earn its clients because it completes the work. If permitted traces and outcomes can then be turned into valuable training assets, that opportunity should be earned through careful production, validation and rights—not assumed because the activity exists.
We are building toward a system where accepted work can inform better tasks, clearer evaluations and more reliable outcomes. We still have to prove the quality, repeatability and economics of that system in the real world.
I see that as a strength of the thesis, not a reason to avoid the test.
The next era of AI data will not be won by whoever produces the biggest number on a slide. It will be won by the companies that make the next research decision or the next real-world workflow measurably better.
Build for the market after the shortage.
I have lived through a period when simply having access to internet inventory was enough to get a meeting.
That period ended. Advertising didn't end. It became a larger, more sophisticated industry.
The same distinction matters now.
The AI data boom is creating real opportunity. It is also creating companies that may be mistaking a temporary shortage for a permanent advantage. Some will build lasting infrastructure. Some will become excellent specialist businesses. Others will discover that the buyer never needed them as much as the buyer needed the supply they happened to control.
I don't want to build for the easy part of the cycle. I've already seen what happens when that part ends.
I want to build something that remains useful after every competitor has access to the same models, the same workforce marketplaces and the same generic tools.
Here's the question I would leave every founder with:
When your customer can buy the same data somewhere else, what will they still trust you to do better?
Build that. The rest is just volume.