NVIDIA is reportedly paying $12.9 billion for a company that earns $150 million

Hey {{first_name|Investor}} -

NVIDIA ($NVDA) is reported to be buying a company called Hugging Face for about $12.9 billion. Hugging Face makes roughly $150 million a year in revenue.

Nothing has been signed and neither company has confirmed it, so hold the details loosely. But that gap between the price and the revenue is the interesting part, and working out why it exists teaches you something you can use on any company you own.

What NVIDIA is buying there is position, and position is one of the two ways a company builds a moat.

A moat is whatever stops a competitor from taking your business when they'd very much like to.

They come in two broad shapes, and telling them apart is most of the work.

The first kind is built from assets. Factories, mines, patents, networks of pipes. Things you can photograph and put a value on. The competitor's problem is that copying you means spending what you spent and waiting as long as you waited.

The second kind is built from position. Being the default. Being the thing everyone already knows how to use, already built around, already trained their people on. Nothing about it appears on a balance sheet, because nobody bought it. It accumulated.

Two companies at the center of AI have taken one of each, and set next to each other they make the difference obvious.

Taiwan Semiconductor ($TSM) makes chips for other people. Apple designs the chip in your iPhone and TSMC builds it. Same for NVIDIA, AMD, Qualcomm and most of the industry.

Its moat is the most physical kind there is. A leading-edge chip factory costs tens of billions of dollars, takes years to build, and needs a workforce that knows how to run a process so precise that a speck of dust ruins the batch. You can have all the money in the world and still be four years behind.

That's why TSMC held 73% of the pure-foundry market in the second quarter of this year. Samsung had 7%. Everyone else was smaller than that.

When Morris Chang started TSMC in 1987 with $220 million, he made a rule that it would never design its own chips. TSMC could easily have designed them. Customers wouldn't have trusted it if it did.

He'd watched what happened to small chip designers who took their work to IBM or Texas Instruments to be manufactured. Those companies would want the design handed over as part of the deal, and could then sell their own version of it. Chang said his customers wouldn't be second-class citizens whose chips only got made when the big firms had spare capacity.

So TSMC deliberately made itself incapable of competing with the people who depend on it.

A customer worried about a supplier stealing their design doesn't have to trust a promise from TSMC. They can check that there's nothing for TSMC to steal it for. The company has no products of its own to put it into.

TSMC's moat is concrete and capital, plus a structural promise it can't break even if it wanted to.

NVIDIA's moat looks nothing like that, and it's easy to underestimate because there's so little of it you can point at.

CUDA is the software layer that lets programmers actually use NVIDIA chips. NVIDIA has given it away since 2006. Twenty years of university courses, textbooks, tutorials and job listings later, an enormous share of the world's AI programmers know CUDA and don't know the alternatives.

No patent protects any of that. A competitor can write software just as good and still run into twenty years of engineers who learned the other one.

Jensen Huang described the result on the last earnings call, talking about the free AI models anyone can download and run: "Nearly all open models run on NVIDIA. And the reason for that is because NVIDIA's footprint around the world is the highest."

He's saying they're the ones everything already gets built for.

That's a habit, and habits are worth more than they look.

Once you see the moat as position rather than product, NVIDIA's behavior makes more sense. And it works two ways, only one of which involves buying anything.

Some of it, NVIDIA buys. In December it bought SchedMD, the company behind Slurm, the software that decides which job runs on which machine inside a big computing cluster. And now, reportedly, a platform where AI models get published and downloaded.

Most of it, NVIDIA gives away. CUDA has been free for twenty years. In March, NVIDIA published a design for laying out an entire AI data center, and the power and cooling vendors have been building to it since. Nobody paid NVIDIA a cent for that document, and that's the point. A company that builds to NVIDIA's drawing ends up with a building shaped around NVIDIA's hardware.

Both routes end in the same place. A chip is a chip either way. What NVIDIA is collecting is the set of points where somebody makes a choice.

Which is why $12.9 billion for $150 million of revenue isn't as strange as it sounds. Hugging Face hosts around 3 million AI models and has roughly 13 million registered users, and when a programmer writes the standard line of code to fetch a model, it goes to Hugging Face by default. Nobody decided that. It just became true.

The files on it are free and copies exist everywhere. The software is open source and anyone can copy it. What can't be copied is being the place everyone already looks.

There's a well-known idea in tech strategy that companies should want the things sold alongside their product to be cheap and everywhere, because cheap accessories sell more of the main thing. IBM funded Linux for years on exactly that logic. Free operating system, expensive IBM services.

By that logic NVIDIA should want a model platform free and abundant, and for years it acted like it. Buying one is the opposite instinct, and it tells you NVIDIA thinks the ground its habit is built on is worth owning outright.

Moats don't last forever, and the two kinds fail completely differently.

An asset moat breaks when someone spends the money, or the technology moves. You can see it coming in the numbers, and Intel ($INTC) is the case study.

Intel owned its factories for decades and that ownership was its advantage. Then its manufacturing fell behind and the advantage inverted. In 2023 its factory business lost $7 billion while Intel sent about 30% of its own production to TSMC. Last quarter that business booked $5.8 billion of revenue and still lost $2.1 billion, and only $293 million of that revenue came from actual outside customers. The rest is Intel selling to itself.

A company that buys chips from the market can switch when its supplier falls behind. A company that owns the factory can't, because the supplier is itself.

A position moat breaks when the habit changes, and that's much harder to spot. There's no factory going obsolete and no write-down to read about. What happens is that a generation of new engineers learns something else, or the default in the tooling quietly points somewhere new. By the time it shows up in revenue, it's been happening for years.

Which is exactly why a company with a habit-based moat spends real money defending the places where habits form.

Every one of these you can check yourself, without paying for anything.

Ask what a competitor would have to do to take this business. Spend billions and wait five years is one answer. Persuade millions of people to change how they work is a different one. Both are moats. They need different things watched.

Look at capital spending as a share of revenue. High and rising says the moat is physical. Low says whatever protects this company isn't on the balance sheet, and you'll need to find it somewhere else.

Check how much of a segment's revenue comes from outside customers. Intel's $5.8 billion looking like a business until you find the $293 million is the cleanest lesson in this whole email. Internal transfers can dress up as a market.

And watch what a company buys when it doesn't need the revenue. A profitable company paying 86 times sales is telling you which ground it thinks matters, and that's usually more informative than anything in the press release.

They're different bets about what stays scarce.

TSMC bet that making the thing would stay hard, and it's been right for nearly forty years. NVIDIA is betting that being the default is worth more than any single product advantage, which is a reasonable bet when your competitors include several customers now designing their own chips.

When a business starts buying things next to its product rather than more of its product, it's telling you where it thinks the next competition happens. That signal turns up in a press release long before it turns up in a share price, and it costs nothing to notice.

Make Your Own Alpha goes deeper on this: how I decide what's actually protecting a business before I own any of it, and the questions I run before a company earns a place in my portfolio. Get your copy here.

Stay disciplined, Koh

P.S. Two new prompt packs arrive next Thursday, along with a major update to the current prompt pack. One finds prospects worth your time, the other tries to break them. Book owners see them first.

Disclaimer: Nothing in this email, the Reset and Invest newsletter, the Make Your Own Alpha book, or any course or digital product from Starshine Media LLC constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed. Always do your own research and consult a licensed financial advisor before making investment decisions. I hold $NVDA, $TSM, $AMD, and $INTC.