Varun
Ammanagi.

What we watch and read is filtered by the same process. Look up a book on a topic, and the recommendations are sorted by sales and average rating. Watch recommendations on YouTube are a result of an algorithm that shows you videos that are optimised for maximum viewing time on a given population. The question that should be asked but very rarely is – whose opinion are we hearing? The answer is the median. Book rankings, Amazon lists, and recommendation algorithms are calibrated for the maximum number of people that are located somewhere in the middle of the curve. It is not a conspiracy, but just how the process of optimisation works. Platforms optimise engagement, and the median user is the standard of what is considered engaging. The content is too complicated – users will stop watching it; too simple – they get bored. What's left is a vast middle band – readable, accessible, reliably unsatisfying to anyone trying to go deeper. YouTube makes the pattern easiest to see. A video with hundreds of thousands of views is, by construction, one a broad audience found accessible and entertaining — not necessarily one that survives scrutiny. The video with 800 views and a wall of equations in the comments might be the more rigorous treatment. The interface implies that view count is signal. Often it's the opposite. None of this makes popular content bad. It makes it a starting point. The problem isn't consuming it; the problem is mistaking it for an endpoint. This is where AI search, used carefully, earns its keep. A well-constructed prompt can bypass the popularity filter entirely. Asking not for the most-recommended book, but for the current state of expert consensus, the strongest counterarguments, or the papers practitioners in a field actually cite. The operative word is carefully. A lazy prompt reproduces the same bias, because the model was trained on the same internet that ranks by popularity. Precision in the prompt is what shifts the output distribution toward the tail. Another subtle trap exists as well. People realise how shallow the popular narrative is in the domain of their own expertise; in the realm you are knowledgeable about, the shallowness of the popular narrative is quite evident. However, the trap is assuming the awareness transfers. Most of us are running on surface-level exposure in the majority of our interests, without the expertise to know it. The book that felt comprehensive probably wasn't. The YouTube explainer that clicked was probably written for someone in ninth grade.

Operating at the median is not bad. But remaining there is.

The equity market increasingly feels detached from basic economic sense.

A company raises real money only occasionally—through an IPO, an FPO, or another issuance. After that, most of the market is just old shares changing hands. That is not inherently absurd. A share is ownership in a business, and ownership should become more valuable when the business earns more, reinvests well, and increases what it can eventually return to shareholders.

But eventually is doing a lot of work there.

At some point, the business has to produce something real for its owners: dividends, buybacks, acquisition proceeds, liquidation value—cash, in one form or another. It does not have to pay today. A good company may retain earnings for years if it can compound them at attractive rates. Fine. But the entire justification for retention is that it creates a larger pool of distributable cash later.

That link now appears optional.

Look around and much of the market is no longer priced on what a business can plausibly return to shareholders. It is priced on what the next buyer might be willing to pay. The current investor is not buying a realistic cash yield. He is buying the expectation that someone else will accept an even worse one.

That is the game.

The company earns a little more. The stock price rises a lot more. The gap is explained away with “quality”, “optionality”, “TAM”, “scarcity”, or whatever phrase is fashionable that quarter. Then the next buyer arrives, pays an even higher multiple, and calls it long-term investing.

Capital gains are not inherently fake. If a company doubles its sustainable earnings and distributable cash per share, the share should be worth more. But when the price rises far faster than the economic claim underneath it, the return is no longer coming from the business. It is coming from multiple expansion—from the hope that tomorrow's investor will be even less demanding than today's.

And multiples cannot expand forever.

A company cannot distribute more cash than it eventually earns. A market cannot permanently deliver returns far above the growth of the businesses underneath it. At some point, price and economic reality have to meet. Either prices fall, earnings spend years catching up, inflation destroys the real return, or the most speculative parts of the market get wiped out.

It does not always end in one glorious crash. Sometimes the punishment is worse: a decade of dead money while everyone insists the thesis is still intact.

Cheap debt can delay it. Low rates can delay it. Passive inflows, momentum funds and a constant supply of fresh money can delay it. But delay is not repeal. None of these changes the underlying arithmetic.

The equity market is not automatically a greater-fool game. A share is a legitimate claim on the future economics of a business.

But once the price of that claim becomes impossible to justify through any realistic future cash flow, you are no longer investing in the business.

You are betting on the next fool.

And the only remaining question is how long he keeps showing up.

I do not know who will win the AI race.

It could be Claude. It could be Gemini. It could be some company that does not exist yet.

But I do not think it will be OpenAI.

OpenAI's greatest advantage may turn out to be the worst thing that ever happened to it: it got there first.

ChatGPT did not merely launch a successful product. It became the product category itself. Millions of people who could not explain what an LLM was suddenly began using one to write emails, summarise PDFs, complete homework, generate captions, debug code and ask questions they were too lazy to Google.

That sounds like an unbeatable lead.

It may actually be a trap.

OpenAI built the audience before it built the economics. It created an enormous population of users who now expect frontier intelligence to be instant, unlimited and preferably free. Every casual prompt costs real money to process, but most casual users contribute very little revenue in return.

OpenAI is therefore stuck subsidising the habit it created.

And what exactly does it receive in exchange?

The standard answer is data. More users mean more interactions, more feedback and therefore better models.

Sure—up to a point.

There are only so many new insights to extract from the ten-millionth request to rewrite an email, solve a homework problem or generate an anime portrait. Scale produces data, but scale also produces mountains of repetition. After a certain point, another million casual users may add far more computing cost than intelligence.

OpenAI has the largest crowd.

That does not necessarily mean it has the most valuable one.

Claude occupies a very different position. The average person may still barely know it exists, but the people who actively seek it out tend to be the ones already interested in AI, coding, research, writing or serious knowledge work. A smaller user base is not automatically a weakness when those users are more intentional, more demanding and potentially more willing to pay.

Anthropic does not need to entertain the whole internet.

It only needs to serve the part of the internet that is worth serving.

Then there is Gemini.

Google does not need Gemini to financially justify its existence tomorrow. It can place it inside Search, Android, Workspace, YouTube, Cloud and nearly every other surface it controls. More importantly, Google can fund the fight from businesses that already produce absurd amounts of cash.

OpenAI has to raise money to finance the race.

Google finances the race while running the rest of the stadium.

That difference matters because the AI contest may not be won by whoever briefly has the best model. Model leadership can change within months. The winner may instead be whoever can afford to keep training, subsidising, distributing and integrating models after the novelty disappears and the industry begins competing on price.

OpenAI has brand recognition, distribution and a large head start.

It also has the burden of defending all three.

It must remain the default chatbot. It must keep free users satisfied. It must prevent paying users from leaving. It must match every Claude improvement, every Gemini integration and every open-source release. It must do all of this while carrying infrastructure costs that rise every time usage grows.

That is not a flywheel unless the economics eventually turn.

Until then, it is a treadmill.

The first mover usually gets to define the market. But it also commits first, spends first, makes the first mistakes and teaches every competitor what not to do. OpenAI paid to educate the world about generative AI. Anthropic, Google and whoever comes next get to compete in a market OpenAI already created for them.

OpenAI proved that people want AI.

It has not yet proved that the company serving the most people will capture the most value.

My bet is that the eventual winner will either have better users, deeper pockets, superior distribution—or all three.

OpenAI has the most users.

That may be precisely the problem.

And yes, the irony is obvious.

I am writing this on ChatGPT.

Sam Altman is still running OpenAI.

And OpenAI is still paying to let me explain why it will lose.