I’m spitballing here, but one way it might be possible for TMSC (or whoever puts up the capital, which doesn’t have to necessarily be TMSC) to mitigate risk, given that this is Nvidia allegedly talking to them, is to build some kind of new memory specialized for AI use that would displace existing memory for AI use. Like, for neural nets, you don’t have the same constraints that you do on conventional DRAM. If you can have someone who builds neural net systems and a memory maker work together, you might be able to produce a new sort of memory that would be a better buy for neural net users.

A few things I can list off-the-cuff:

• Access patterns on neural nets are predictable long in advance. That means latency doesn’t matter much (so having far higher latency than current memory may be fine) and requesting large blocks of memory might be valuable. Accesses will be sequentially requesting memory over a large range, so it also may mean that one could change how cooling works in the memory, if one has some guarantees on a single spot not being frequently re-requested a lot. A given byte of memory won’t be re-written, once written, for a long time, so you could relax how rapidly the hardware needs to be able to sequentially do writes to a given byte.

• You have a lower ratio of writes to reads. I read one article talking about how HBM isn’t yet optimized for AI use on write/read bandwidth availability grounds. That’s talking about the memory interface, not the memory chip itself; they were just arguing for changing the ratio of bandwidth reserved for writes and reads. However, I would bet that you could also optimize the memory chip around this.

So you’d basically come up with, I dunno, “neural net memory (NNM)” or something like that. It’s not that you have to use it for neural nets, but that its characteristics would be optimized for neural net use. That only makes sense if there’s enough scale of demand for a single application to put engineering time in, expend the fixed costs, but there is presently a lot of demand, and might be more down the road.

That doesn’t make risk of demand dropping go away, but…it’d tend to shift that risk to conventional memory manufacturers, since if demand doesn’t meet expectations and you’re doing neural nets, a buyer would probably buy the specialized memory chip before the non-specialized memory chip. Still risk, but it’d be shifted to Micron/SK Hynix/Samsung.

At the start of the AI boom, we had hardware that was not optimized for AI use. It was generic random access memory with most of the hardware stack designed for rendering 3D graphics. I’d guess that it’s very likely that that hardware stack is not optimized for neural net use, and I’d bet that Nvidia would be the most-likely party to be at the heart of attempts to re-engineer the hardware stack to optimize it, as that’s somewhere that they’re uniquely-well-positioned to make changes.

EDIT: I remember a while back when John Carmack was tossing the idea out there of using a loop of fiber-optic cable as a memory for AI use. That may not be the optimal design, but one does see that there is a lot of room to think about specializing the memory to the very different demands that neural nets place on memory rather than the demands that conventional RAM users have.

The kind of memory the current manufacturers are focusing on is specialized for data-center-grade GPUs. These GPUs are insanely power-hungry and require data-center-grade cooling. When the bubble pops and data centers go under, you’re not going to see consumer-grade RAM flood the market. So if TSMC wants to enter the market conservatively with the assumption the AI surge will not last, going into the consumer RAM side would be a better bet.

We had this issue during the crypto boom and the manufacterers that added capacity then are glad they did.

What else could use up that RAM? How about standalone VR? We could get mainstream adoption by 2030.

The point is building factories sometimes pays off