Hacker News Reader

Comment by simonw | Hacker News Reader

The headline here under-serves the article in my opinion: this is a fascinating, deep explanation of how the memory market works and why increased demand for HBM (used by big GPU racks) hurts the availability of wafers for DDR and LPDDR (used by laptops and phones).

View on HN

This also put a lot into perspective:

> these memory makers have learned a very particular lesson from the unforgiving history [deep drops in demand] of their industry: always leave demand unmet

I can't blame them for keeping some reserve demand ready so they keep having customers over the years.

I'm no expert analyst on this topic, but I'm worried this time might be different for them. China finally has the technological capability to challenge this time around and they will gobble up the unmet demand, allowing them to get more experience and more money to catch up.

I wonder if this is actually true in the long-term though. If they were to flood the market with lots of high capacity memory, then I think our programs would start using more memory too. As a result we might end up needing more memory faster compared to if they keep demand unmet.

Just consider that a chat application today takes more memory than a full 3D game with thousands of users (including chat) + the operating system used to ~20 years ago. If 128 GB of RAM were the norm then there's a chance we might expect to buy 128 GB of RAM too.

But I suppose it's really a question of how many dollars we expect to spend on memory rather than the specific amounts of memory.

I think that Slack's memory usage is "not bad" at ~300MB just because Discord exists, and somehow that seems to get into the gigabytes before even looking at a message.

Too much of IRC has been lost to Discord :(

LiberaChat FTW!

Yeah everything is a browser app now gotta ship a whole browser engine even for a simple messaging app.

Lots of people have bought baseline 32/64 GB ram in their laptops for a long time (the last few laptops).

Applications can be lazy and grow to the amount of resources using, but the real need for maximum ram is using it for these new purposes.

"I wonder if this is actually true in the long-term though. If they were to flood the market with lots of high capacity memory, then I think our programs would start using more memory too. As a result we might end up needing more memory faster compared to if they keep demand unmet."

It's a gambler's ruin problem. Future profits are worth zero if you go out of business first.

For older hacker news readers who have lived through multiple RAM price spikes - it has held true for decades so far. Making memory is one of the most vicious tech markets to compete in.

The last one I remember was after the 2011 tsunami. I think things are worse these days because memory isn’t just a chip that you toss on a board anymore, it’s now frequently delivered in the same package as the CPU.

There was the 2007 spike following the Windows Vista release when a huge chunk of the corporate market had to buy new machines, there was also the Hynix fire in 2013, to pick a few recent ones. The 2007 spike was pretty rapidly killed by the famous global financial crisis that followed.

SSDs have spiked too. The 5tb backup spinning rust drive I bought for CAD$150 in September 2024 now retails for CAD$260.

They are running fabs at 100% capacity. New fabs take many years to build, cost billions, and are out of date within a few process moves.

One cannot simply "flood the market" or the Chinese would have done it ages ago.

The Chinese memory makers have been sabotaged by USA, otherwise they would have already "flooded the market" and reduced the prices.

Even so, the production of cheaper Chinese DDR5 memory is increasing despite not being able to import all the tools and materials that are needed, so eventually they will benefit from this memory shortage that allows them to gain market share that they would not have gained otherwise.

I'd like to learn more about how they've been sabotaged. What happened?

Google US semiconductor sanctions on China. Basically US (and allies pressed by US) aren't allowed to export their most cutting edge semiconductor tools and materials to China, so China can only make chips using older-gen tools on older process nodes than western-aligned competitors.

The other part of the sabotage is that western companies then have restrictions on what chips they're allowed to buy from China. IDK if RAM chips are also on that list.

> IDK if RAM chips are also on that list.

Apple cancelled their deal with YMTC (Chinese RAM company) after the US sanctioned the latter. I don't know whether that is directly because of the sanctions, or indirectly (e.g. if Apple thought sanctions would hamper YMTC's ability to supply the goods), but they have had the same effect.

https://www.lightreading.com/business-management/chinese-chi...

https://www.business-standard.com/article/international/appl...

I dont know if trade sanctions counts as "sabotage" but point taken.

Well it kinda counts as self-sabotage now, at least in effect. We shot ourselves in the foot so China couldn't compete as easily.

Is any of that relevant to RAM? Half the DRAM is non-EUV and pretty much all DRAM has been stuck on 10nm for the last decade

That isn't "sabotage". Those are export restrictions.

China also runs adverse competitive industrial policies (e.g., industrial spying, flooding markets). Not making value judgment here.

You have quite the post history BTW.

> and are out of date within a few process moves.

The fabs themselves (especially for memory) are being retooled for process moves. At least that's what micron has been doing.

But otherwise I agree. Micron has a new fab being built that was started around 2020. It's nowhere near finished.

We already had that. Memory prices got really low for a while. If there are too many companies with too much capacity they’re forced to race to the bottom until one of them collapses and reduces supply.

Yup, see my previous boom-and-bust comment. In particular vendors will remember the price crash in 2022/23, the industry only recovered once the AI mania set in. There'd been another boom around 2020/21 when supply-chain panics meant hyperscalers bought up huge inventories, and just before that another low around 2019. So the current situation is standard pre-bust behaviour.

> always leave demand unmet

And then allow small low cost manufacturers to get the rest of the market... like China has been doing this whole time.

It's more than that though. Memory is a series of boom-and-bust cycles. Also, it costs around $15 billion and takes 2-3 years to build a fab. So if you were to start now, you'd have a pretty good chance of hitting the next bust right after spending $15 billion. It's not necessarily some sinister conspiracy, it's because execs are reluctant to be the ones to take the blame for setting fire to $15 billion. Looks kinda bad on your CV.

How does that work in practice though? There’s a central “RAM Oversight Council” that checks that no one sells too much?

There are only 3 memory manufacturers. They just quietly follow the same strategy. They learned not to talk about it from the last time they were fined for antitrust price fixing in the 2000s.

ouch. That’s indeed a problem.

This author has outstanding articles regularly

Agentic workflows and computer use. You still need regular servers for the agents to use.

Reply on news.ycombinator.com