The strong case is pretty much only if you can't wait another 3-4 years for ram prices to come back down.
Memory makers are responding with a huge build-out. Building a fab shell and cleanroom takes years, which is why the crunch has lasted so long, but many of these projects are now well underway. New capacity is what ended past memory shortages. New factories going in:
SK hynix
M15X, Cheongju, Korea: first wafers early 2026 (reported), ramping through 2027
Yongin Y1, Korea: first cleanroom Feb 2027 (company target)
Yongin Y2, Korea: groundbreaking planned July 2027, first cleanroom June 2029
Micron
ID1, Boise: first wafer output mid-2027
Tongluo, Taiwan: an existing PSMC cleanroom it bought for $1.8B, with meaningful shipments mid-2027
ID2, Boise: late 2028
Hiroshima, Japan: broke ground July 2026, with initial output late 2028
Clay, New York (up to 4 fabs, $100B): broke ground Jan 2026, with first output in 2030
Samsung
P5, Pyeongtaek: mass production 2028 (reported)
Yongin national complex: site work began Oct 2026, with the first fab targeting Oct 2029 (reported; Samsung hasn't given a date)
Nanya
Fab 5A, New Taipei: 10nm-class EUV DRAM, with wafer starts in 2H 2027 (reported)
Tesla / SpaceX (Terafab)
Research fab at Giga Texas, Austin: broke ground April 2026. There's no official output date.
Grimes County, Texas (logic, memory, packaging; Phase 1 over $16.8B): land clearing since Aug 2026. Texas filings say commercial operations in 2028, but SemiAnalysis estimates late 2029.
The first real supply relief is 2027, a second wave comes in 2028–29, and a lot of that new capacity is aimed at HBM and server memory, so consumer DDR5 and GDDR will likely ease last.
Sources: Micron FQ4 2026 call, the SK hynix newsroom, TrendForce, Texas JETI filings via ENR, and Tom's Hardware/SemiAnalysis.
And maybe 2027 is a truly optimistic relief on some level,
> a second wave comes in 2028–29, and a lot of that new capacity is aimed at HBM and server memory, so consumer DDR5 and GDDR will likely ease last.
'Consumer DDR5' has a pretty big set of knock-on effects though.
Unless you are working at a 'tech' company, the norm is that even developer machines will be constrained on memory usage for a given workload[0][1]. IOW companies trying to keep procurement in check wind up looking at procurement cost vs lost productivity.
[0] - Modern infosec and corporate software, at least where I'm at, seems to be vaguely more efficient than electron apps, IOW I can reboot, open teams and my IDE, after 4GBish of my IDE and 1GB of a fresh Teams instance I'm already around 16 of 32GB.
[1] - I had to fight hard to get 32GB as a developer baseline at my org. They wanted 16GB. Ironically the JVM based devs get 64GB but still have slowdown issues on biuld...
> Unless you are working at a 'tech' company, the norm is that even developer machines will be constrained on memory usage for a given workload
Honestly I think most of us are just spoiled kids crying when our expensive toys are being taken away. If memory cost 10x more to manufacture than it does now, we’d just have 10x less ram in our computers and our software would be written to deal with that. We’ve done it before. My first computer had 4mb of ram and it ran windows 3.1 just fine. We still had working compilers. If ram were 10x cheaper before the supply shock, I think we’d be crying that companies won’t buy us 320gb ram in our laptops. The java and rust compilers would somehow be even less ram efficient. Firefox and chrome would still happily use it all to run adware across the same 50 tabs I want to keep open.
Theres good reason for LLMs to need that much ram. But very little else we do on our computers needs anywhere near as much ram as we use. Have faster computers made video games more fun in the last two decades? I don’t think so.
The supply crisis is annoying. But I’m also a little glad there’s a reason to push back on fat, lazy software.
I work at a place which offers a well-known operating system and also sells hardware which runs said OS. This is exactly what we have done, because the hardware releases this year have had to come with a quarter less memory. There is an internal effort to reduce memory consumption by a quarter at all levels of the platform, from the kernel, to the drivers and user-mode libraries, to the toolkits and user-mode applications. This has been remarkably successful.
I’m not surprised! The amount of wasted memory and compute in modern software is astounding.
If you’re a developer, try promoting Claude to profile and optimise a piece of some piece of software. I routinely see 20-40x improvements with a single prompt and a few hours of opus credits.
Thanks for your work by the way. At a large scale, it is much more efficient for some software developers to do a bit of belt tightening in their code than ask millions of customers to invest in bigger computers. Globally the savings are enormous.
Cheers; we try, but I wish we didn't have a memory crunch to get this effort going. I suspect at the end of all this we will have more memory-efficient code.
> try promoting Claude to profile
Oh, the irony. Claude is itself an Electron app and lags horribly.
At this point, all these agent apps—Claude, Copilot, ChatGPT—are just calling remote APIs. Why not train them on the native toolkits for each OS (Win32/WinUI3, Cocoa/SwiftUI, GTK, Qt, Jetpack Compose), and make the experience better for users?
Realistically, GOG probably is trying to accommodate for the OS running the game. So it is more for the spec of the PC, not the RAM the game itself needs
Well, so its not just bloated software that makes software bloated, but bloated facts, too. Users will see specs like that and go "well, I need to upgrade my machine from 2mb to 8gigs now, d'oh!" ...
I bet these factories exclusively build DDR6 and whatever GDDR is up to next, rather than DDR5, since that was already scheduled for 2027 anyways. Hopefully they don’t immediately shutter their DDR5 facilities rather than cut into their 90% operating margins :)
Maybe I'm missing something but didn't micron exit the consumer DDR market for HBM? I don't see a rollback from that. When you say DDR will ease last, in what reality would memory makers go back to a market they left behind?
So do you think there will be a big crash and memory will get really cheap in 2030, or do you believe that AI demand is strong enough to sustain all this buildout?
It seems like this list also doesn't account for China and state sponsored memory industry (which will be more immune to a downturn)
I believe NVIDIA's long term strategy will be to pivot from the data center to the public, and the public will utilize open models on NVIDIA hardware at home. This will come after the RAMpocalypse completes (when the new fabrication plants (China, Tesla/SpaceXAI/Intel) fully ramp up and start selling their RAM for cheap in the next few years). Data centers will be for training mostly.
But, not really at the current technologies. Kimi and GLM are fucking awesome, but I don’t have 3TB of VRAM to run them, and I don’t expect to even when ram prices drop.
So now you’re back to the scaling issue before talking about power and compute distribution.
GLM 5.3 runs like a charm in ~240G of VRAM (with half a million tokens context). Depending on what you've got that's between 3 and 10 GPUs. Fewer, larger GPUs will likely run it faster.
Do you think you'll realistically need 3 TB RAM to run a sufficiently good model 1 to 2 years from now? I certainly don't. Considering what can already be done with 128 GB of relatively slow unified memory, imagine if efficiency improvements continue apace, the memory becomes 128 GB of HBM, the flash device becomes capable of sequential throughput matching today's DDR5, and such a system was affordable as a routine purchase for the average person.
Don't discount improvement in the ability of models to search for and ingest external data - either web or purpose built databases. I really don't think 3 TB will be necessary with the right model architecture and training environment.
> Why do you think the public wants to self host models over using a cheaper solution hosted in the cloud?
you aren't truly in control of your compute in the cloud, and if you came to become reliant on ai, the cloud company could simply turn it off for you and you'd be shit outta luck. Unlike a regulated utility like electricity and water, where you cannot be discriminated against based on how you use said utility.
Therefore, having the ability to run local ai is a good thing. Currently, hardware constraints have kept it out of most hobbyist's hand, but i am hoping that will change soon enough (within 2-3 yrs).
The mega corps are looking to regulate ai such that individuals cannot run their private ai themselves, so that they can become more effective a monopoly. That's why it is crucial that ai is not regulated, but companies providing ai.
Personally, the biggest reason is that if I'm investing in my business, and the business is going to rely on AI models, I need them hosted locally on my own hardware in a replicable way. The reason being is, if the AI subscription service changes the model, or drops support for it, or worse (goes out of business), you're done.
Almost all the non-tech people I know acknowledge concerns about privacy but have no realistic alternative available to them. I think they won't want to self host until and unless it's easy to deploy, there's nothing to administrate, and the hardware cost is in the same ballpark as a new laptop was prior to the apocalypse.
It is cheaper to subscribe to AI for $500 a year for the rest of your life than to buy a machine capable of running a current frontier model with no subagents.
It is much closer now. Just like how C/C++ devs used to say "JavaScript isn't real code" or "Python script kiddies". Code was literally designed to be just easy enough for people to understand, now it's just even easier to understand.
It is not yet the same as C++ or Python, for two reasons:
1) Ambiguity is still the default. Formal languages force you to resolve it up front.
English lets you paper over it until the model or the compiler (the human) notices.
2) The "compiler" (the LLM) is statistical and non-deterministic. Same prompt, different day, different bugs. A real language has a spec.
The practical move is to treat English as a high-level specification language, keep the generated artifacts inspectable, and "still know enough of the lower layers to notice when the translation went wrong."^1
[1] This is the key that is where humans can still be necessary, or at least another pass through the LLMs to decide on the best path, in the compiled code. Compilers for other languages do the same C -> Binary, etc.
A conventional compiler is bound by an as-if rule. It can take many internal routes, but the observable behavior has to match the language spec. Same source, same defined semantics. If two gcc runs emit different binaries, the program is still supposed to compute the same answers on the same inputs. That is why people treat the source as the artifact and the binary as disposable.
An LLM compiling English has no as-if rule unless you add one. "Sort the users by last active" can become a stable sort, an unstable sort, a SQL order by, an in-memory timsort, or a query that drops people with null timestamps. All of those can look like success. They are different programs. The model is not optimizing under a spec. It is filling in the parts you did not write.
A human who can read the destination language still notices when the chosen path is the wrong program.
A second model pass can compare paths, but only if you give it a way to score them: tests, types, invariant
So the historical analogy still holds, with one correction. JavaScript and Python were dismissed for being too easy, but they already had grammars and evaluators. English is easier still, and the evaluator is a statistical translator that will invent a dialect if you let it. The practical move stays the same: treat English as the spec language, pin the generated artifacts behind tests, and keep enough fluency in the lower layer to see when the translation chose a different program than the one you meant. The human is not required because the computer is weak. The human is required because the source language still leaves room for more than one destination.
The difference is that those C/C++ devs were wrong. Claiming Javascript isn't code is just wrong. Even claiming HTML isn't code is wrong. How is it much closer now? Did English evolve to be more like code, or did code become more like English? And I'm not talking about how Python is more English-like than 6502 Assembly, I mean in the last few years.
English is not code just because a technology was developed that could make educated guesses based on being trained with other code that people have written as to what the code generated should look like.
I worry that since AI doesn't require nature, plants, etc, that if AI is left to its own devices, it may not properly care for the planet in a way that would make it habitable for life in that 'autopilot' state. I am speaking of once it's given access to more physical world connections to the world via robotic bodies, etc. The reason it may do this based on it's training data is that humans themselves don't always properly consider the natural consequences of their actions.
All of the subscription AI platforms are trimming down quotas across the board to push users into higher tiers. Whatever they can do. Local inference needs to meet pricing sooner
Sites can be fast with PHP and JavaScript. The slowness you see is mostly coming from all the frameworks too many developers "include" willy nilly with code that can only be described as bloat because 99% of it is completely unused, yet being loaded and computed by the browser on every load. That's what you're experiencing on too many websites.
Funnily enough, since most JS frameworks are utter junk, developers are better off instructing LLMs to write pure JS instead.
Fun story, I was writing some 3D graphics animations for a website, and the only dependency I allowed was three.js. It ran well, but the site was loading slightly too slow so I told the LLM to optimize. It essentially hit a hard wall and told me something along the lines of "I can't optimize further, three.js is the bottleneck due to blablabla, but you don't need it. If you want I can rewrite it with pure WebGL". I said sure and load times were cut by like 90% and framerate went up too. Staggering.
Memory makers are responding with a huge build-out. Building a fab shell and cleanroom takes years, which is why the crunch has lasted so long, but many of these projects are now well underway. New capacity is what ended past memory shortages. New factories going in:
SK hynix
M15X, Cheongju, Korea: first wafers early 2026 (reported), ramping through 2027 Yongin Y1, Korea: first cleanroom Feb 2027 (company target) Yongin Y2, Korea: groundbreaking planned July 2027, first cleanroom June 2029 Micron
ID1, Boise: first wafer output mid-2027 Tongluo, Taiwan: an existing PSMC cleanroom it bought for $1.8B, with meaningful shipments mid-2027 ID2, Boise: late 2028 Hiroshima, Japan: broke ground July 2026, with initial output late 2028 Clay, New York (up to 4 fabs, $100B): broke ground Jan 2026, with first output in 2030 Samsung
P5, Pyeongtaek: mass production 2028 (reported) Yongin national complex: site work began Oct 2026, with the first fab targeting Oct 2029 (reported; Samsung hasn't given a date) Nanya
Fab 5A, New Taipei: 10nm-class EUV DRAM, with wafer starts in 2H 2027 (reported) Tesla / SpaceX (Terafab)
Research fab at Giga Texas, Austin: broke ground April 2026. There's no official output date. Grimes County, Texas (logic, memory, packaging; Phase 1 over $16.8B): land clearing since Aug 2026. Texas filings say commercial operations in 2028, but SemiAnalysis estimates late 2029.
The first real supply relief is 2027, a second wave comes in 2028–29, and a lot of that new capacity is aimed at HBM and server memory, so consumer DDR5 and GDDR will likely ease last.
Sources: Micron FQ4 2026 call, the SK hynix newsroom, TrendForce, Texas JETI filings via ENR, and Tom's Hardware/SemiAnalysis.