Comment by cogman10 | Hacker News Reader

Comment by cogman10 | Hacker News Reader

Parent Full thread

And none of these companies are operating their own fabs, that's the problem.

Fabs are a cutthroat business that's very hard to get into. It costs billions of continual investment to stay a float. That's why there's really only about 3 different companies with cutting edge fabs. TSMC, Micron, and Samsung. Even intel, who built a huge portion of their business on cutting edge fab tech, has struggled to keep funding it. AMD got out of the fab business almost a decade ago (spinning off global foundries) and that spin off is no longer cutting edge. AMD uses TSMC.

Fabs are some of the most expensive factories to operate on this planet due to a constant need for brand new equipment and cutting edge research. That's why there's not an Apple, Google, Meta, or Nvidia fab. That's why there's not an AMD fab. That's why Intel fabs are treading water.

Without the constant investment, you very quickly find yourself in the company of yet another cutthroat industry, the "not cutting edge" fabrication industry. And that, by and large, has already been locked up by about a dozen fab companies.

I've made this same argument so let me make a counter-argument:

There are some ways to get this off the ground much quicker. One or more companies could buy an existing non-leading-edge fab like GlobalFoundaries. That buys a lot of expertise so you're not starting from zero.

DRAM also benefits from being very regular and relatively simple. It used to be what you bring up on a new process node to help prove things out.

It also isn't impossible to reduce reliance on ASML if you're willing to throw money at it. That's definitely a super-long-game move but it could be done.

I'm not going to argue that someone is going to do any of this but if demand is sustained it is possible.

It does help, but I have to wonder how many people are still working at glofo currently who are researching node shrinks. They stopped their research into the 7nm process in 2018 and all the indications are that they aren't really continuing it.

Meanwhile, I believe SOTA is at least 3 or 4 node shrinks beyond that 7nm process. It'll take years for them to catch up to where micron is currently.

That's fair. I had hoped GF was not entirely giving up but was going to wait until the tech for new nodes was proven and cheaper to obtain. It seems that is not really possible at the moment because anyone who is involved in that supply chain doesn't maintain the ability to stand up new processes on older nodes. Way too busy standing up new nodes.

So much for efficient free markets.

And a 64G DDR5 ECC DIMM costs $3K and is backordered. If ths isn't a bubble and demand persists, some new players are eventually going to want a cut of that.

We aren't talking about making new lug-nuts. A company can't just will a fab into existence.

For example, Micron is actively building a few new fabs. One of which has been in progress since Biden (pretty close to my home in fact). It's not going to be completed for another 5 years at a minimum. And this is a company that has the experience and partnerships for producing fabs.

Yes, a new company might decide they want to enter the market, but even if they decided, today, "Yes we'll do this" I'd expect a minimum of a decade before they start spinning out their first chips. That's also at least a $1T investment at this point to get started.

> And this is a company that has the experience and partnerships for producing fabs.

Not even they necessarily have the experience to do it! Intel has a policy called "Copy EXACTLY!" for fab construction where they make every irrelevant detail the same as their last fab, because they don't actually know which of the details matter.

Oh for sure. A story that gets told at the micron factory is that for a long time they'd experience chip failures at increased rates. During the night it was pretty good, but during the day these random failures would creep up.

After spending a lot of time studying the problem what they finally realized is they built the building too close to the interstate and vibrations from the interstate were ultimately making their way into the factory causing errors.

To combat this problem, they spent millions retrofitting shock absorbers onto the building.

It's not shocking that intel would do the same because even the slightest movements and vibrations can spoil the chips. Putting a restroom in the wrong spot might spoil a batch when someone flushes the toilet.

The thing is, pretty much everyone relevant assumes it is a bubble and that eventually large players will end up facing mob justice. That's why the hundreds of billions of $ IOUs are getting passed around like hot potatoes, and that's (in addition to ASML, the key part of anything EUV lithography, being booked out for years) why no one is planning to construct dozens of billions of dollars worth of fabs.

In addition, the know-how is concentrated in Taiwan. You literally can't train enough people in enough time to move everything out of there.

> concentrated in Taiwan

Where are SK Hynix and Samsung located again? Or 95% of Micron's facilities?

So yea, Samsung built a chip factory pretty close to where I live. Number one it is forking gigantic. You don't just slap one of these babies down. Next, the equipment that goes inside of that massive clean room building is a problem in itself. That takes years to get ordered, then years to ensure it works right, with employees that have a very particular skill set.

Again, people might want part of it, but they are also a bit smarter than you are and read history books to see exactly how this is going to play out and then they gladly walk away before they light their money on fire.

Absolutely. "If this isn't a bubble and demand persists" is doing a lot of heavy lifting there, though, and it may be some time before manufacturers decide that's the case.

Reply on news.ycombinator.com