Comment by stego-tech
19 hours ago
This is the biggest memory repricing cycle I've ever seen in my life; some degree of high price/limited availability and "free RAM with purchase of Doritos" cycle is always expected, but this has been the worst one yet.
As other commenters have pointed out but I might have missed in the article, compute maturation is amplifying memory constraints right now and making it worse. Device upgrade cycles are getting longer because most compute-based products have matured, with CPUs not seeing substantial gains and memory usage really only expanding at the absolute top end of workloads pre-LLMs (3D and HPC in particular). An iPhone 14 still has almost all the features of the iPhone 17, because the compute capabilities are remarkably similar; Geekbench shows a performance delta of ~25-30% between the 14 and 17 Pro Max models, which is pretty paltry considering the devices are separated by four years of manufacturing improvements. This extends into desktops, laptops, tablets, STBs, and more, with only VR devices and larger ARM/RISC-based kit seeing more substantial uplifts as general designs improve.
So with compute stagnating and memory constrained, my money is on vendors taking this as an opportunity to gradually shift away from a yearly release cadence and slow down to a biennial cycle that alternates between budget and flagship launches every other year. Even if LLMs fail spectacularly and all that memory capacity becomes available, HBM memory likely isn't to find its way into many consumer devices (just ask AMD how it worked out for them on consumer GCN GPUs).
The name of the game, especially for consumers, is efficiency - "potato builds", as I've been calling them. Software and services optimized for lower power, smaller-specced devices of increasing age instead of pandering to flagship devices with poorly optimized code or engines for the sake of new shinies (like Raytracing). Between the memory shortage, shifting geopolitics, rising costs, and stagnant wages, consumer purchasing power is going to be squeezed like a vice for the foreseeable future, and businesses will need to adapt around that reality.
> Even if LLMs fail spectacularly
Haven't they already proven to be extremely useful? In some areas they are definitely here to stay, coding/software and search (retrieve and summarize information). There's a bunch of places where they are surely shoehorned in, overhyped, and don't belong, but there's also equally many places where they might still be transformative but aren't used yet.
But overall I think the technology is well proven.
I always leave room open for failure, and that approach has generally served me well personally and professionally. I have never been punished for having an exit strategy.
Besides, the marketplace is still in its infancy for LLMs, with a lot of unanswered questions. A lot of those questions surround the commercial viability of frontier models on bespoke hyperscaler data centers with limited usage outside of LLMs specifically should those economics be non-viable. Since that's where the memory is being tied up into, that means it's a critical question to answer in order to determine long-term investment needs into further memory fabrication.
They work well functionally, but financially anything can still happen.
> Haven't they already proven to be extremely useful?
Most certainly not. The accuracy issues mean that they can't really be used effectively for coding or search, the two things you mentioned.
My bet is that vendors will simply discontinue their low margin phones, which are usually the budget phones.
For example, Apple might make fewer iPhone 18 and let it sell out frequently. They’ll use their RAM supplies mostly for the Pro phones.
I don’t think Apple will stop releasing new iPhone Pros every year. The business is too big.
> For example, Apple might make fewer iPhone 18 and let it sell out frequently. They’ll use their RAM supplies mostly for the Pro phones.
Apple has a second option that may not be open to most other vendors - as they've just demonstrated with the MacBook Neo, they could cut the RAM in half on the budget models. One good cycle of optimising the hell out of their (almost entirely native) software stack, and iOS would once again sing on a 4GB SKU.
I doubt they'd consider that because it's a downgrade. If things get truly desperate, then maybe.
2 replies →
I mean, that's the correct short-term read, but if LLMs in hyperscalers remain commercially viable to the point of tying up memory for several years, and if that necessitates an expansion of memory fabrication to satiate unmet demand, and if that demand ends up getting hoovered up by AI companies again due to their unmet or delayed demand from technological adoption, then Apple et al may not have much of a choice but to adopt such a profound strategy change.
There's a lot of 'ifs' there to be sure, but they'd be fools not to at least discuss the possibility internally and understand their options.