Comment by apatheticonion

It's really easy to set up a manufacturing process for basically anything if you can spend 100x per unit compared to the big optimized factories and you don't mind the product being a lot slower.

The clean room isn't the hard part about being competitive. It's using advanced lithography to cram billions of cells into a single chip. If you want make DDR2 chips on a 90 nanometer process, that is accessible to a whole lot of companies, but nobody will buy the product. And in the micrometer range you can DIY like this guy.

I'm not really sure what you mean by "certified"; I don't think JEDEC are handing out stickers. Although I am reminded of the Bunnie Huang article about SD cards, where the Shenzen vendor would give you the same SD card with the manufacturer logo + holographic authenticity .. of your choice.

The real problem with the RAM business is that it was commodified; normally manufacturers make a relatively small margin. No incentive to build a factory for that. These are not normal times because (a) someone has bought all the RAM and (b) someone has blown up a whole load of globally critical infrastructure in the Middle East.

The risk the existing RAM manufacturers are being cautious about is the risk of normal: if you start building a factory now, will you be selling into a RAM glut?

> what exactly is stopping any company from doing this themselves and breaking into the RAM business?

nothing, except the terrible yields that they would obtain, and the lack of scale making the entire enterprise not profit generating (as the amount of profit per sale is too low if it even is positive, but you can't set it higher as there's cheaper, "better" ram available from pre-established fabs that do have economies of scale).

You could play the artisanal angle, and market it as home grown, organic ram. Not sure how much real buyers of ram care, but might get a few hobbyists in the market.

You can make eDRAM using logic processes (which are currently less bottlenecked than RAM, at least for non cutting-edge nodes) but the cost is still prohibitive compared to specialty DRAM processes (even when considering the recent increase in DRAM prices). If you were doing that, you'd want to use it for compute-in-memory instead, which basically pushes NUMA to an extreme of having lots of tiny cores each with direct access to its own local DRAM.

So, CXMT?

i assume the reason is that this is a very competitive market and you need hundreds of billions in investment to just start producing at a competitive quality and price, with massive uncertainty that you will be able to make that money back

i mean, you'd think if someone is willing to through 60B (or 10B for an option to buy at a 60B valuation) at Cursor, then other people would give it a shot at starting a SotA semiconductor fab, but apparently the profit expectations are not there

You have to surrender your silicon(e) to the state!