Which means nothing. Apple is extremely good at calculating how much inventory they should produce. They’re not going to make more than they’ll sell and have them sit on shelves. They’re also not going to cut into their profit margin too much by paying their way out of the memory shortage.
If a new premiere high-margin product sells out "within a month" that means there's a disruption in sales and an opportunity for competitors to meet the demand they just generated while manufacturing/distribution tries to catch up.
It would absolutely represent either a mis-estimation of demand or compromise to some supply/manufacturing bottleneck that they couldn't overcome without sending their release calendar into chaos.
I think I agree with the economic sentiment in general but who is the competitor that can even make something comparable? And by that I mean: Matches the memory amount, the memory bandwidth, the form factor, the power draw, the compute and the OS
In it's role as a local AI engine (for inference, training, or fine-tuning) you don't need to compete on all those factors to be a viable competitor if this were sold out.
Those are hard-to-beat features for typical Apple users, for sure, but this is aiming for a different niche and the potential buyers in that niche may not be as concerned with form factor or power draw (for instance), and some may even see the OS as a negative factor. Those are just factors that happen to come along for the ride since it happens to be an Apple product.
> Apple is extremely good at calculating how much inventory they should produce.
That doesn't really matter though for a physical product, unless you are implying they are 1) selling it at a loss or 2) not attempting to maximize profits from the product. If they sell out, it means they hit the target profit from the sale of the product.
Large mixture-of-experts models, like Muse Spark or Kimi K3.
MoE models want a different ratio of memory/compute/$ than dense models like a Gemma4 31B or Qwen3.8 27B.
Since only some of the experts are active each round, you don’t need as much compute to get good tok/sec, but you do need a lot of memory to avoid streaming weights from SSD or system RAM.
You could run a full Muse Spark on one of these for $24k at probably around 50 tok/sec.
You’d need 5x RTX 6000 Pros at $80k for Muse Spark (albeit at much higher tok/sec).
On a $24k budget, you’re looking at a maximum of 96GB from Nvidia or 384GB from Intel or AMD. So you’re bottlenecked by PCIe bandwidth, and all those TFLOPS go to waste.
Other configurations of the Studio have similar logic for other MoE models. For instance, Qwen3.8 flash-next runs at a very acceptable 75 tok/sec on a 128GB M5 Max for about 1/3rd the price of an RTX 6000 or 40% less than 4x B70’s.
Getting an equivalent amount of ram in GPU's is substantially more expensive than this. A single RTX 6000 Blackwell with 96gb of ram is like $16,000 now. Someone please correct me if I'm wrong but essentially with the Mac Studio you're getting the ability to run larger models, at slower speeds, for a lot cheaper.
- Larger MoE models only need compute proportional to the number of active parameter, so Qwen3.8-Flash-Next (180b with 6b active) runs more tok/sec than Qwen3.8-27b on the same hardware.
- Larger models typically have higher quality output and require fewer thinking tokens to get to an answer.
Consequently, measuring time to complete a task with a wall clock, the speed difference is less than you would expect.
1.3TB/s of Unified Memory. I'm not sure today how much distributing dense models across GPUs is ok or bad these dats but you don't have to think about it by having all the memory pool available on the one GPU you have.
still cheap compared to those things. lab down the hall in grad school had an onyx [1] with the fancy polarized shutter 3d glasses for playing what we called "chem tv" or semi-empirical small molecule docking experiments - that was something like $250k or so in the early 90s.
Top spec build is $29,799 + tx. While local AI is certainly an enticing prospect I can’t really imagine myself ever seeing a return on that investment.
Further they’ve already begun hyping the M7 Ultra and even if that got delayed until 2028, this machine only being SOTA for 1.5yrs is a hilariously short window in which to maximize your investment.
Enthusiasts generally don't see a "return on investment" on their luxury cars or their boats either, but they enjoy what they're able to do with them and aren't exactly embarassed about the class signaling of being a high spender.
Add to that the smattering people with genuinely justifiable use cases, and you have more than enough demand for a limit-pushing, highest-margin model of your product platform.
I'd be willing to spend that much if it unlocked the ability to run top-of-the-line open weight models at their best quantization at a decent enough speed for personal agentic use. We're clearly not there yet, and I just can't imagine dropping the price of a brand new car on something that will only get me heavily quantized versions of the smaller models.
The 256GB M5 Ultra Mac Studio can (AFAIK) run the latest open weight Flash models but only 2 bit quant. This by extension will be able to do 4 and that's a big difference. We're not at the point of running 1T+ parameter models locally yet but we'll be there by 2030 easily.
I personally believe that local LLMs have a huge captive market for people who cannot or should not host that data with a third-party. For example, it's an open question of whether or not using cloud LLMs by lawyers breaks attorney-client privilege. HIPAA with medical records is another one.
But I want to expound on one point that you made, which is really important. You say that this spec only has a SOTA life of ~1.5 years and that's an issue. Take that to its natural conclusion. Why isn't it an issue for hyperscalar data centers, particularly as NVidia wants to move to annual chip updates.
I believe the hardware is going to be commoditized. At that point the only thing that matters is performance-per-Watt. NVidia's R100/200 doesn't make the B100/200 obsolete but if the R series gets 50-100% PPW more and these DCs are using gigawatts of power, that's a massive difference.
What on earth is a "natural conclusion" in this context? I'm going to assume that part of your comment was written by AI.
> Why isn't it an issue for hyperscalar data centers, particularly as NVidia wants to move to annual chip updates.
Because he's not a hyperscaler? He's not operating data centers where compute is fungible, he's buying an individual machine that he wants to use for years. He can't just keep buying a new machine and fob off the old one to someone else's less demanding workload.
Enterprise GPUs have a lot longer effective life than these consumer chips (although even the old 3090s are reasonably useful, especially with the recent driver unlock tech). I'd argue that these consumer chips have only now gotten to the memory capacity/processing capacity to make local LLM use actually reasonable.
How many parameters are you planning to use for your doc editor? Email prose is optimal on a different model. Web parser? You might only be able to do one at a time with switching, which will work, but could be tiresome.
Bought it without looking too closely to make sure I got in before the rush (comes Nov 25-Dec 3 before my M5 Ultra 256GB which is supposedly shipping end of January).
After taking a look the $11.5k premium over the 256GB might not be worth it. Gonna have to mull over the decision and possibly buy two 256GB machines instead (which might be the smarter option).
This is crazy to me as well: half a terabyte (!) of RAM?! How is that even possible lol (I know this is possible, but seems totally overkill, even though I know LLMs eat RAM for breakfast)
The 512 GB ultra seems too expensive. You can buy multiples of the cheaper one. Like 4 of the 128 GB max, or 2 of the 256 GB ultras for less. (And end up with more total bandwidth and storage and compute).
Unless there is some sort of restriction on number per customer.
This is just to make me feel poor for "only" having an M5 Max 128GB model. Which is insane compute for a home computer, but now I already feel behind one month into it.
I've had my eye on the normal Mac Studio M5 Ultra (normal plebeian 96GB RAM) and the release (edit: ship) date there keeps slipping, now into January. I'm surprised Apple is pushing the backlog even more.
I have one. It can run lots of smaller models really well. I’ve been loving Gemma4 especially. My needs more more large scale information extraction and not trying to replace Claude code locally though.
If anything the interesting models that can run in about 64gb of memory are increasing because most people can’t access more memory than that. I’m especially excited for clef and cleric as they continue to improve. Yeah, you don’t get 1mil token context length, but even frontier models can lose focus at that level
I've been holding off because I will need to do experimentation with local models at some point...but not yet and won't be for a bit but I am afraid prices will rise given component cost trajectories.
I've been tempted to just get the M5 Max instead and just innovate with smaller models myself.
512gb being more expensive than just two 256g is a little strange. You mean you can get the same exact total cluster ram but double the compute/storage nodes and save $700?
What sort of jobs are memory bandwidth bound that people are running? Will they actually see 10x speed improvement or hit some other real world bottleneck? Are these jobs "short" on the order of seconds or minutes or is the node being hammered for two weeks straight? I'm just curious about it.
This is why people like to pick up the new Studios from the Apple Store in person - for that short ride back home their car is worth four times what it was on the way there.
Holy price increase, in 2025 the equivalent M3 Ultra (32-core CPU / 80-core GPU with 512GB of RAM) cost $9,499.
Now the equivalent M5 Ultra (36-core CPU / 80-core GPU with 512GB of RAM) is priced at $22,299.
This thing will sell out within a month
Which means nothing. Apple is extremely good at calculating how much inventory they should produce. They’re not going to make more than they’ll sell and have them sit on shelves. They’re also not going to cut into their profit margin too much by paying their way out of the memory shortage.
If a new premiere high-margin product sells out "within a month" that means there's a disruption in sales and an opportunity for competitors to meet the demand they just generated while manufacturing/distribution tries to catch up.
It would absolutely represent either a mis-estimation of demand or compromise to some supply/manufacturing bottleneck that they couldn't overcome without sending their release calendar into chaos.
I think I agree with the economic sentiment in general but who is the competitor that can even make something comparable? And by that I mean: Matches the memory amount, the memory bandwidth, the form factor, the power draw, the compute and the OS
In it's role as a local AI engine (for inference, training, or fine-tuning) you don't need to compete on all those factors to be a viable competitor if this were sold out.
Those are hard-to-beat features for typical Apple users, for sure, but this is aiming for a different niche and the potential buyers in that niche may not be as concerned with form factor or power draw (for instance), and some may even see the OS as a negative factor. Those are just factors that happen to come along for the ride since it happens to be an Apple product.
Sounds like a great way to generate interest and FOMO.
Did you know the McRib is back?
But only for a limited time so you better act fast!
> Apple is extremely good at calculating how much inventory they should produce.
That doesn't really matter though for a physical product, unless you are implying they are 1) selling it at a loss or 2) not attempting to maximize profits from the product. If they sell out, it means they hit the target profit from the sale of the product.
It could also mean they're supply-constrained, and cannot set or reach future targets without accounting for surge pricing.
The 256 GB is backordered by 13–15 weeks, if you click through to that.
What would be the use case for this over a dedicated 8xgpu server?
Large mixture-of-experts models, like Muse Spark or Kimi K3.
MoE models want a different ratio of memory/compute/$ than dense models like a Gemma4 31B or Qwen3.8 27B.
Since only some of the experts are active each round, you don’t need as much compute to get good tok/sec, but you do need a lot of memory to avoid streaming weights from SSD or system RAM.
You could run a full Muse Spark on one of these for $24k at probably around 50 tok/sec.
You’d need 5x RTX 6000 Pros at $80k for Muse Spark (albeit at much higher tok/sec).
On a $24k budget, you’re looking at a maximum of 96GB from Nvidia or 384GB from Intel or AMD. So you’re bottlenecked by PCIe bandwidth, and all those TFLOPS go to waste.
Other configurations of the Studio have similar logic for other MoE models. For instance, Qwen3.8 flash-next runs at a very acceptable 75 tok/sec on a 128GB M5 Max for about 1/3rd the price of an RTX 6000 or 40% less than 4x B70’s.
Getting an equivalent amount of ram in GPU's is substantially more expensive than this. A single RTX 6000 Blackwell with 96gb of ram is like $16,000 now. Someone please correct me if I'm wrong but essentially with the Mac Studio you're getting the ability to run larger models, at slower speeds, for a lot cheaper.
Right. Couple things to note though:
- Larger MoE models only need compute proportional to the number of active parameter, so Qwen3.8-Flash-Next (180b with 6b active) runs more tok/sec than Qwen3.8-27b on the same hardware.
- Larger models typically have higher quality output and require fewer thinking tokens to get to an answer.
Consequently, measuring time to complete a task with a wall clock, the speed difference is less than you would expect.
Exactly, optimizing for VRAM where you don’t need the RTX 6000 performance.
Data centers outfitted with Ikea furniture instead of the usual 19" racks.
1.3TB/s of Unified Memory. I'm not sure today how much distributing dense models across GPUs is ok or bad these dats but you don't have to think about it by having all the memory pool available on the one GPU you have.
Makes me feel like we're back in the era of SGI workstations in the late 20th century! I got to play with a few of those in college.
still cheap compared to those things. lab down the hall in grad school had an onyx [1] with the fancy polarized shutter 3d glasses for playing what we called "chem tv" or semi-empirical small molecule docking experiments - that was something like $250k or so in the early 90s.
[1] https://en.wikipedia.org/wiki/SGI_Onyx
Yes but they were heavy.
They were built like a brick shithouse and felt expensive. You always knew they belonged to someone else.
In other words you couldn't mistake them for an $800 Mac Mini.
Now try imagining a SGI workstation with 512GB of SOLDERED DOWN RAM. You can't because no one would have bought them.
To be fair, these are significantly larger than a Mac Mini. More like one of the old Macintosh Quadras or Performas.
But your main point still stands.
Quadras and Performas were in full mini-tower or desktop configurations.
This thing looks like the ashtray you'd find next to it.
Anyway I'd get the same feeling every time I opened a Sun server. The metal work alone felt like fine industrial craftsmanship.
Top spec build is $29,799 + tx. While local AI is certainly an enticing prospect I can’t really imagine myself ever seeing a return on that investment.
Further they’ve already begun hyping the M7 Ultra and even if that got delayed until 2028, this machine only being SOTA for 1.5yrs is a hilariously short window in which to maximize your investment.
Enthusiasts generally don't see a "return on investment" on their luxury cars or their boats either, but they enjoy what they're able to do with them and aren't exactly embarassed about the class signaling of being a high spender.
Add to that the smattering people with genuinely justifiable use cases, and you have more than enough demand for a limit-pushing, highest-margin model of your product platform.
> Enthusiasts generally don't see a "return on investment" on their luxury cars
I hate to break it to you but the same thing has happened with cars. Many in-demand performance cars like the GT3 RS have been appreciating.
I'd be willing to spend that much if it unlocked the ability to run top-of-the-line open weight models at their best quantization at a decent enough speed for personal agentic use. We're clearly not there yet, and I just can't imagine dropping the price of a brand new car on something that will only get me heavily quantized versions of the smaller models.
It is not exactly an investment, but I guess a lot depends on what you actually would want to do with it.
> It is not exactly an investment
At $30,000 it's sure an expensive hobby. You could get into Caterham racing for that...
Those are corporate dollars making that spend, thats the only people.
Wait, where is Apple hyping the M7 Ultra?
They're not. It was a bit of spurious guesswork by Mark Gurman, and we all know how those guesses end up.
The 256GB M5 Ultra Mac Studio can (AFAIK) run the latest open weight Flash models but only 2 bit quant. This by extension will be able to do 4 and that's a big difference. We're not at the point of running 1T+ parameter models locally yet but we'll be there by 2030 easily.
I personally believe that local LLMs have a huge captive market for people who cannot or should not host that data with a third-party. For example, it's an open question of whether or not using cloud LLMs by lawyers breaks attorney-client privilege. HIPAA with medical records is another one.
But I want to expound on one point that you made, which is really important. You say that this spec only has a SOTA life of ~1.5 years and that's an issue. Take that to its natural conclusion. Why isn't it an issue for hyperscalar data centers, particularly as NVidia wants to move to annual chip updates.
I believe the hardware is going to be commoditized. At that point the only thing that matters is performance-per-Watt. NVidia's R100/200 doesn't make the B100/200 obsolete but if the R series gets 50-100% PPW more and these DCs are using gigawatts of power, that's a massive difference.
What really is the effective life of AI hardware?
> Take that to its natural conclusion.
What on earth is a "natural conclusion" in this context? I'm going to assume that part of your comment was written by AI.
> Why isn't it an issue for hyperscalar data centers, particularly as NVidia wants to move to annual chip updates.
Because he's not a hyperscaler? He's not operating data centers where compute is fungible, he's buying an individual machine that he wants to use for years. He can't just keep buying a new machine and fob off the old one to someone else's less demanding workload.
Enterprise GPUs have a lot longer effective life than these consumer chips (although even the old 3090s are reasonably useful, especially with the recent driver unlock tech). I'd argue that these consumer chips have only now gotten to the memory capacity/processing capacity to make local LLM use actually reasonable.
Is this good enough for light web browsing, emails and word editing?
Nothing “light” about web browsing nowadays…
How many parameters are you planning to use for your doc editor? Email prose is optimal on a different model. Web parser? You might only be able to do one at a time with switching, which will work, but could be tiresome.
> Is this good enough for light web browsing, emails and word editing?
How many Chrome tabs do you have open?
And 2 electron apps.
Might even work as a basic Facebook machine.
Bought it without looking too closely to make sure I got in before the rush (comes Nov 25-Dec 3 before my M5 Ultra 256GB which is supposedly shipping end of January).
After taking a look the $11.5k premium over the 256GB might not be worth it. Gonna have to mull over the decision and possibly buy two 256GB machines instead (which might be the smarter option).
As someone also considering getting one, what would make it worth it in your opinion?
The fact that I can get this in my hands up to two months earlier than the 256GB version (Nov 25-Dec 3 vs Jan 15-29) is a major factor.
Wish I had your economy for sure.
It's for my startup. I wouldn't have bought it otherwise.
I saw "512GB" and knew deep down it means RAM, but SURELY nobody needs the same amount of RAM as I have total SSD space on my desktop, right?
...Right?
This is crazy to me as well: half a terabyte (!) of RAM?! How is that even possible lol (I know this is possible, but seems totally overkill, even though I know LLMs eat RAM for breakfast)
These are very well-suited for running local AI models, which must be held in memory, and are very large.
Nobody is buying these to write Word docs and play MP3s.
sorry, I guess I forgot my /s.
chinese memory providers can't come online soon enough.
prices shoot by just choosing the 128gb option and even worse the 512gb option.
Eye watering prices as we expected.
The 512 GB ultra seems too expensive. You can buy multiples of the cheaper one. Like 4 of the 128 GB max, or 2 of the 256 GB ultras for less. (And end up with more total bandwidth and storage and compute).
Unless there is some sort of restriction on number per customer.
This is just to make me feel poor for "only" having an M5 Max 128GB model. Which is insane compute for a home computer, but now I already feel behind one month into it.
Two things I wish I knew:
1. Will the RAM shortage turn into a RAM glut & crash in the next year?
2. Even if that happens, will Apple just keep prices sky high anyway?
I've had my eye on the normal Mac Studio M5 Ultra (normal plebeian 96GB RAM) and the release (edit: ship) date there keeps slipping, now into January. I'm surprised Apple is pushing the backlog even more.
Release date? I know someone that has one right now.
I have one. It can run lots of smaller models really well. I’ve been loving Gemma4 especially. My needs more more large scale information extraction and not trying to replace Claude code locally though.
If anything the interesting models that can run in about 64gb of memory are increasing because most people can’t access more memory than that. I’m especially excited for clef and cleric as they continue to improve. Yeah, you don’t get 1mil token context length, but even frontier models can lose focus at that level
I've been holding off because I will need to do experimentation with local models at some point...but not yet and won't be for a bit but I am afraid prices will rise given component cost trajectories.
I've been tempted to just get the M5 Max instead and just innovate with smaller models myself.
ALL the Studio configurations are cheaper than a Lenovo P4 with AMD consumer CPU, no GPU and shit storage.
* 96GB - $6,799
* 256GB - $10,799
* 512GB - $22,299
(M5 Ultra with just 1TB of storage)
512gb being more expensive than just two 256g is a little strange. You mean you can get the same exact total cluster ram but double the compute/storage nodes and save $700?
Which is an especially interesting option given built-in RDMA support with reasonable performance, including fully-connected graphs up to 6 machines…
Didn't take me long to find a nice proof of concept, 4x mac studio cluster in this case:
https://www.jeffgeerling.com/blog/2025/15-tb-vram-on-mac-stu...
you're still limited to TB5's 120Gbps speeds vs 1.2Tbps memory bandwidth on M5 Ultra
What sort of jobs are memory bandwidth bound that people are running? Will they actually see 10x speed improvement or hit some other real world bottleneck? Are these jobs "short" on the order of seconds or minutes or is the node being hammered for two weeks straight? I'm just curious about it.
> What sort of jobs are memory bandwidth bound that people are running?
CFD, for me. Depends on how you segment the domain, but the speed hit can be significant.
It was the same with the M3 Ultra. Two 256GB is a better deal if you want to bother connecting them with thunderbolt.
The 512GB is likely targeting people who want 1TB RAM (2x cluster).
yea but who needs that clutter on their desk
I guess I'll drive my 15 year old Honda CRV for a few more years.
This is why people like to pick up the new Studios from the Apple Store in person - for that short ride back home their car is worth four times what it was on the way there.
Armored truck would be better
Been waiting for this!
The fully loaded version costs $29,799.00!!!!
Laughed at the 2499 starting price until I saw the 5499