Compete where nobody else is

What Everyone Says

Go where the demand is. Serve the popular use case, the biggest segment, the models and categories everyone already searches for. Winning a slice of a huge market beats owning all of a small one.

It sounds obviously right. The demand is provable, the market is already educated, and you can point at competitors to prove the category exists.

Why That's Wrong

Provable demand is also visible demand, which means everyone can see it. You end up as one option among many, competing on price and speed against companies with more capital.

Eugene Cheah puts the maths plainly. On the most popular AI models, "if you go to the number one model, there are like 10 providers. We are just one in 10."

Being one of ten in the obvious market is a worse position than being the only one somewhere less obvious. The crowded segment is crowded precisely because it was easy to identify.

What Eugene Did Instead

Featherless AI hosts more than 40,000 open source AI models, heading for a stated target of all three million on Hugging Face. Most inference providers carry fewer than 100.

The economics of the tail are why nobody else bothers. Eugene explains that for models that will not receive more than a million requests a month, "most providers will not bother hosting" because it cannot justify standing by a GPU around the clock for them. His team could serve them because they had built hot-swapping, bringing a cold model online in about five seconds instead of the thirty minutes standard tooling takes.

That turned an unservable segment into exclusive inventory. Beyond the top 100 models, "typically we are the only provider."

Eugene has done the sizing. The top hundred models are roughly the top 60th percentile. The bottom quarter is made of slices so small he says he cannot see the individual colours on a pie chart. His conclusion, stated as a question: "would you rather compete with 100 other providers at the top 50% or would you rather compete with no one at the 25% slice?"

The Principle Underneath

An uncontested segment is only worth owning if something structural stops others from entering. For Featherless that is cost structure, not preference. Competitors are not ignoring the tail because they missed it. They are ignoring it because their architecture makes it unprofitable.

This is the difference between a niche and a moat. A niche is a segment you chose. A moat is a segment your competitors cannot profitably serve even after watching you succeed there.

The payoff compounds through distribution too. Being the only provider listed against a model on Hugging Face means discovery happens without you paying for it.

Should You Do This?

Do this if you can name the specific reason competitors cannot follow you, and that reason is built into how your product works.

Skip it if the segment is empty because nobody wants what it offers. Eugene had evidence of real usage, and enterprise customers fine-tuning their own models, before betting on the tail.

The question to ask: if your largest competitor decided tomorrow to serve this segment, what would stop them?

If the answer is nothing, you have a head start, not a position.