Mac Studio 512GB RAM Option Disappears Amid Global DRAM Shortage

[Ishimura]

You can find negative examples anywhere. Are you choosing to ignore the positive examples? Effectively 100% of Anthropic's code is now written by Claude code.

The dashboard my colleague made isn't slop. It does the job it needs to do. It's an internal tool only. We're not shipping it to anyone else. My colleague is an educator, not a software developer. But they are highly intelligent, and have a builder's mindset. This is a project that would have taken weeks or maybe even months without generative AI, and thus would have never come to fruition. This is the worst AI they'll ever use.

If you focus entirely on the negatives then you can ignore the evidence that completely contradicts your views and allows you to hold onto them as long as possible. Many people would prefer to maintain their initial and very emotional response to the tech rather than stay informed about the actual capabilities and state of the field.

They’d rather kneecap their understanding than concede they were wrong initially, or allow themselves to feel overwhelmed by the prospects of what is to come if the current rate of progress holds.

But it doesn’t matter. None of these discussions matter. The models continue to advance at a staggering pace. Events will overtake them. By which I mean events will continue to overtake them more - clearly they have already.

 

[Rychiar]

The AI models need this memory configuration.

I'm sure, but still, extremely niche audience for an already quite niche product.

 

[CWallace]

You can find negative examples anywhere.

In their defense, they did say "it is only as good as the person using it". So if the person using it is skilled like your colleague, AI coding tools can indeed allow generation of quality code and applications that were previously too difficult to develop in a reasonable timeframe.

As another example, Apple developer Steven Troughton-Smith recently noted that he used XCode's new AI tools to "quickly" (as in within an eight-hour workday) write three new applications that would have taken far longer for him to do without those tools.

I'm sure, but still, extremely niche audience for an already quite niche product.

But that niche is expanding every day and, eventually, could grow to the point it is no longer really just a niche.

 

[mannyvel]

For everyone that likes making software, AI tools are super fun. Once you figure out how to get them to work it's truly a joy. The challenge seems to be that each level of app complexity needs a new structure around it so you can drive the AI effectively.

Which is the same as software - product and project managing a tiny product is a lot different than managing an enterprise-scale product.

The difference is that soon anyone will be able to play. The problem for normal people is that you need to be able to direct the AI, and some (a lot?) of its choices are, well, not great. I suspect many of its choices are biased towards reddit users, marketing copy, and/or bloggers...who tend to be more advocates of technology rather than practitioners.

At some point I'm sure data will become reputational. Until then, someone who knows what they're doing still needs to drive the bus.

 

[Zwhaler]

512GB was $4,000 when it was available.

I am thinking the $4,000 premium price tag is why people were not buying that upgrade, and why Apple eventually removed it. That is pretty steep.

This may have been true in the beginning. Since the openclaw craze began Apple is probably curtailing the demand for these types of machines intended to run local AI models, specifically given the sharp increase in RAM prices. Hopefully this will revert when the next generation of Mac Studio releases.

 

[Virinprew]

I'm sure, but still, extremely niche audience for an already quite niche product.

That’s why it’s a build-to-order. Right? 🤔

 

[John-Paul Hunt]

Chia Coin mining needs at least 512GB of RAM on MacOS as the only way you can plot chia mining plots is to use the RAM for RAMDisk otherwise you need a RTX GPU to plot chia K32 plots at fast speeds for 32TB and soon 50TB HDDs as MacOS does not have EGPU support anymore when it was pushed back in the day before the price of chia coin crypto tanked.

 

[M3Stang]

This is Apple's halo car. Insanity to stop selling it. Ok to do a price correction — all buyers understand the reason.

Apple should instead release M5 Ultra with 1TB RAM. Does not matter how much it costs — $20k would be much better than non-available. It is a significant AI narrative Apple needs.

I see the Mac Studio as the GT3, great but not best, and the Mac Pro as the GT3 RS. Better in some ways (on track / PCI-E lol) and worse for daily driving (cost). Just waiting for them to release their new GT3 and RS at this point 😀.

 

[Ishimura]

As a note, today I probably saved my wife's companies $35-50k a month after a random chat with Gemini about energy storage. Apparently there are renewable energy offset programs that "nobody" knows about in her home country. Makes the $10 gemini pro subscription through Verizon worth it for sure.

The reactions here to posts like this is so strange and telling. This is just flatly a positive use case of AI. The response? No replies with any arguments to the contrary, but two laughter reactions and a thumbs down reaction.

(It’s also the same few people reacting this way to any positive comments about use of the tech)

The negativity towards anything about LLMs - literally any application of them - is hysterical and irrational.

 

[MacBH928]

This is the configuration people are buying to run local AI models. A cluster of 4 Mac Studios with 512GB can run the largest models

I think that will cost you like $40K , or about 166 years of claude subscription. just get the subscription

 

[Nugget]

I think that will cost you like $40K , or about 166 years of claude subscription. just get the subscription

The economics depends on the use case. For people using the API, Claude isn't flat-rate like the subscriptions are. API usage can blow way past even the $200/month Pro tier (flat-rate Cowork/Claude code use). Heavy use of opus or sonnet models on an API plan might justify even a $40k cap exp outlay for private, behind the firewall LLM use especially if you're looking at amortizing/depreciating over 2-3 years.

 

[rezwits]

We are not talking about regular RAM here. This unified memory on apples chips is more comparable to vRAM on GPU’s. This demand for high memory is mainly due to running local AI. Regular RAM is just isn’t fast enough. Flagship GPUs have even higher bandwidth but they don’t ship with that much of vRAM. Apple with its M chips is uniquely positioned in the middle - it’s fast enough and you can have loads of it. M ultra chips is a hot buy for local inference.
...

I said this when the M1 Pros came out. The unified memory is essentially DOUBE. And I will "double down on this again"!
If you want to have a PC with 512GB to run AI (now in this instance) first of all you kinda can't with one machine (unless you talk about H100s etc). But lets look at what you could do:

You could get 2x NVIDIA 5090 ($3500x2=$7000) with 32GB that's 64GB, but then you would need to get 64GB of RAM for your CPU, (actually I would do 128GB of RAM so you can breathe).

With this configuration how much RAM do you have? 128GB, 64GB+64GB. But it's split, but the arguments that say well what could you do with even more CPU RAM? True, true, there are some applications but for most people and AI (only GPU offload). But here is the thing, a MacMini M4 Pro (last year!, actually Late 2024!) at $2500 could have 64GB of RAM. Now in THIS AI instance, that machine can "DO" pretty much what the PC can 64GB for 64GB. For inference, but it would only get really 30 token/s where as the PC would be like 4 times faster, but at today's prices you're talking about $8,000 for the PC and $2,500 for MMM4P, 4 times the cost 4 times the speed!

But a PC that has 32GB CPU RAM and 32GB of GPU RAM (1 card, $4,000) PC

Would have to run GGUF models and offload to CPU, but that's still a $4,000 PC! Why because the MacMini M4 Pro having 64GBs of RAM, IN THIS CASE is like having 128GBs of RAM!

Now what does this mean in regard to the MacStudio with 512GBs of RAM when the use case is AI?

You figure it out... FFS

There are Law Firms and Medical Companies Buying (should I say who bought!) 20 of these and set them up to WORK actual real work that instantly justified letting 20 employees go... if not one machine per employee, does it pay for itself? Of course!

The thing is people and some say LOCAL AI isn't ready yet, well the only thing from LOCAL AI companies want or really wanting to improve is for CODING. But as for written word, transcripts, transcribing, legal forms, etc etc, OH LOCAL AI WORKS HARD!

UNIFIED RAM is equal to 2x the RAM! And please put your HAHA emojis in the bottom ya FOOLS... oh and Apple is so far behind in AI once again OMG Fools...

That's why they have the HOTEST SELLING machine in AI today at $10,000! That they couldn't even keep in STOCK!! FFS, (<== I like that FFS, haha)

 

[rezwits]

They are readying the

That makes no sense.
Why prevent people from buying ?
Just raise the price to whatever is necessary.

Apple is getting ready to release the MacStudio M5 Ultras with 512GBs (M5s aren't getting any RAM increase the Log2 math doesn't work, M6 maybe), but OMG the M5 Max is insane CPU/GPU/NPU speed increase over the M3 AND M4 variants. The Ultra is just gonna be off the charts! (for AI)

 

[rezwits]

The reactions here to posts like this is so strange and telling. This is just flatly a positive use case of AI. The response? No replies with any arguments to the contrary, but two laughter reactions and a thumbs down reaction.

(It’s also the same few people reacting this way to any positive comments about use of the tech)

The negativity towards anything about LLMs - literally any application of them - is hysterical and irrational.

That's the thing, this my 3rd "frustration post at the idiocy" and lack of knowing/ignorance at what the TECH INDUSTRY is going thru, I didn't eveb mention the OpenClaw use cases. This thread is SO LAUGHABLE. It's sad tho...

 

[museic]

You could get 2x NVIDIA 5090 ($3500x2=$7000) with 32GB that's 64GB, but then you would need to get 64GB of RAM for your CPU, (actually I would do 128GB of RAM so you can breathe).

If you're looking at 2 5090s for AI, you're better off with the RTX Pro 6000. 96GB memory on one card and you don't even have to get into tensor parallelism.

 

[marco114]

thats actually cheap. 512GB DDR5 RAM costs $19k for a Dell Server. 64GB DIMMS are $2408. Look at memory.net

512GB was $4,000 when it was available.

I am thinking the $4,000 premium price tag is why people were not buying that upgrade, and why Apple eventually removed it. That is pretty steep.

 

[Hammerd]

Like most RAM requirements, whether 16GB, 32, 48, 64, or whatever: If your workflow needs it, it needs it. For most people, anything beyond 16GB is "overkill"—meanwhile, earlier this year, I upgraded earlier than I intended to go from 32 to 64, because my work as a designer means swapping back and forth between Photoshop, InDesign, and Illustrator—with large projects, I blow past 32GB easily. 64 buys me breathing room and avoids pageouts which, even with Apple's hyper-fast SSDs, slows me down.

Likewise, there are use-cases fort 128, 256, even 512GB of DRAM. Running local AI models is the one you hear about most lately, but any kind of intensive data analysis—medical, financial, longitudinal population studies, and so on—requires massive dataset manipulation. Time is money, so the more of that data you can page into RAM at a time, the more you can get done, faster.

I have a computer with 64 gigs of RAM, 8TB pcie 5.0 SSD and a 5080 that I use as a PlayStation 3 emulator and some MS Office. And I’m not even rich.. go figure

 

[reyesmac]

Apple is just hoarding the chips. They'll probably do an update to the studio and tout it as a ai beast of a system, enticing people to buy the 512 option now that they can't compare it to the current system that would handle 90% of what the newer version could. If prices go higher they can simply charge more for chips they purchased at much lower prices already.

 

[MacBH928]

There are Law Firms and Medical Companies Buying (should I say who bought!) 20 of these and set them up to WORK actual real work that instantly justified letting 20 employees go... if not one machine per employee, does it pay for itself? Of course!

I really would like to know how is llm replacing workers, because i have yet to see an ai do human work. in fact, a lot of the time it messes things up and needs guidance. What is a real use case? don't tell me writing letters, you could use a template for that.

 

[Kirkster]

I really would like to know how is llm replacing workers, because i have yet to see an ai do human work. in fact, a lot of the time it messes things up and needs guidance. What is a real use case? don't tell me writing letters, you could use a template for that.

Your avatar is 💯 on point by the way…

There are a bunch of places where AI is reducing headcount. My company let go of 20% of the product dev team. Even in the Medical field you will see the clerical staff reduced because of the doctor being able to dictate their patient notes into AI and it doing all of the back end paperwork/coding with a staff member just reviewing to ensure correctness.

Now where it won’t work out is the trades electricians, plumbers, hvac guys, it will however simplify their work in places.

It’s not going away, but it is in a lot of cases a tool that can aid smart workers and make them better, it’s already proved that, but it’s just a tool at this point.

 

[MacBH928]

My company let go of 20% of the product dev team.

what was the product dev team doing and how did AI replace them? i would really like to know. Maybe its just regular downsizing?

 

[Kirkster]

All the “regular” downsizing happened last year. These were in our product development team. Our CEO is head over heels about AI helping people be more effective and using the “tools”. “One guy using AI cut dev time from two years to 4 months saving tens of thousands of dollars”. Like it or not it’s happening.

 

[poematik13]

I really would like to know how is llm replacing workers, because i have yet to see an ai do human work. in fact, a lot of the time it messes things up and needs guidance. What is a real use case? don't tell me writing letters, you could use a template for that.

There are companies like premsys.ai that make customized LLM’s for many different industries to automate repetitive admin tasks, and they keep it physically on premises not on cloud. These types of mac studios are key to that

 

[Netrox]

I think it's a good move. It makes no sense to get 512GB RAM and get little in return, performance wise. It does not scale up well beyond 256GB RAM. Now it's likely that we'll very soon see M5 Ultra which is expected to have a much faster bandwidth and that is where 512GB RAM can make economic sense but will likely cost a fortune.

But again, we've seen how LLMs have been optimized in the past few years and it was no longer necessary to have massive amount of RAM thanks to improved fine-tuning algorithms.

 

[museic]

I think it's a good move. It makes no sense to get 512GB RAM and get little in return, performance wise. It does not scale up well beyond 256GB RAM. Now it's likely that we'll very soon see M5 Ultra which is expected to have a much faster bandwidth and that is where 512GB RAM can make economic sense but will likely cost a fortune.

But again, we've seen how LLMs have been optimized in the past few years and it was no longer necessary to have massive amount of RAM thanks to improved fine-tuning algorithms.

It's not about "scaling up well", it's about being able to run the more intelligent models like Kimi K2.5, WITH enough space left over for a usable context window. Yes, the smaller models are getting smarter than they used to be, but the larger ones of the same generation will ALWAYS produce superior output. Of course, the best local models available are still not as good as the frontier models that run on massive server farms.