5 years of (200/month) tokens at that price, meanwhile an rtx 5090 pc is about half that… hmm
but i wonder how much these token costs are sustainable or not, it may be in the long term cheaper to have your own hardware if token costs go up (and hopefully hardware gets cheaper again)
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two heads on.
Thank you for this. I wish Apple focused their silicon design on improving the TTFT metrics but coming from an M3 Pro, it still looks laggard compared to Nvidia's TensorCores in the 5090.
Maybe Apple is an acquisition away from changing that balance.
The rumor on Apple's processor roadmap is that they're skipping other M6 variations (all previous generations had Pro and Max, a few had Ultra) in order to focus on the M7 generation for AI reasons. What exactly the M7 improvements are who knows.
I think that comes down to TSMC. Nvidia apparently booked out the whole A18 or 16 node. Apple is on 2nm right now and M7 will jump right to A14. According to my quick AI research anyway.
The selling point of the M5 Ultra Mac Studio is that you can run much larger models that the 5090 can't without swapping. NVidia aggressively segments the market on VRAM for this reason. That's why a 5090 has an MSRP of ~$2k (but good luck getting one for less than $4k) while a 6000 Pro, which is basically a 5090 with 96GB of RAM has now soared beyond $15k where 3-6 months ago it was more like $10-11k. A 6000 Pro has the same memory bandwidth but slightly more CUDA units (IIRC ~24k vs ~21k).
This advantage won't be apparent with a 27B model. The 256GB MS can probably run the newer Flash models locally, something you can't do on a 5090.
I don't think we'll get a successor to the 5090 until late 2028, maybe even 2029. I'm basing this on the launch date of the 5000 series and that we haven't got a midcycle refresh yet. Rumor has it the chips are ready but the 3GB RAM modules are 3-4x the price of the 2GB modules used on the current cards.
Apple should see a Mac Studio major update in 2028. That might even force NVidia's hand. But it's really impossible to say what the state of the market will be 2-3 years from now. It may have completely crashed. I suspect not however.
The interesting thing will be when the bandwidth demands start forcing HBM memory onto these home/enthusiast solutions.
But what about builds that combine 8 of the 5090 with infiniband between boxes? Wouldn't that be comparable to the mac in terms of price and potentially beat it by a lot in terms of performance for the large MoE? I understand the space/heat/noise considerations, but price wise it may still not make as much sense as people think. (Agreed that it is hard to get the NVIDIA hardware and the 6000 pro are priced less competitively).
While that sounds super awesome,
How many people are actually going to build and maintain that vs a box you can grab at the mall that fits in a lunchbox?
On Apple website it says 512GB memory option is available in October. I guess bumping to that one would cost additional 4-6k US$. So an Ultra with 2TB storage would be north of 15k US$.
That’s like 12 years worth of OpenAI Pro subscriptions
I think we all expect the heavy subsidized subscriptions to end or significantly increase in price at some point, but it could be years from now and I'd rather spend a similar figure on an hypotetical Mac Studio M8 Ultra, or whatever more advanced competitor that will have likley appeared by that time.
A more apples-to-apples comparison would be with API cost in OpenRouter at the same tok/s rate for the same models that you can run locally, maybe.
HN always has these completely contrived counterarguments. What is actually going to realistically happen that will prevent use of an LLM provider? Did you think that the OP literally meant the 12 years or maybe it was just to show how expensive using a Mac Mini as an alternative is?
Yeah, anyone who thinks local AI is going to save them money is likely to be disappointed, at least if they want to run models that are even remotely capable.
Plenty of other reasons to get excited about local AI, but I don't think cost is one of them.
Maybe you are using a local model to go after some Millennium Prize problem and you don't want OpenAI to take your work and use it to win the prize for themselves? $15k might be a bargain.
And, yes, I know a current local model wasn't going to solve the Navier-Stokes problem, but I'm just using it as an example where privacy might be valuable.
Despite being on a site called Hacker News, we seem to often overlook the simple aspect of wanting local AI hardware to hack (not necessarily in the cybersecurity sense) with. I got my local AI hardware because it's an enjoyable hobby for me with bleed over into professional pursuits (but work provides plenty of compute access so this spillover is a very minor factor).
Hard to guess, it can go either way. If you will need to be in a syndicate to use non-sterilized models, that mac makes sense. But if there is mandatory registration of personal cyberarms, you risk going to mines once they check you purchases. You could try to play normie and pretend you simply wanted to show off, by keeping your actual work on external disk, but that leaves traces on system. Counting on someone in the Gap renting you gray iron works as long as you can swap credits. Still, this gear is tiny. Put it in your e-car, with uplink, and leave it at uncle's farm. Discreet.
Just commenting here because you're discussing hardware: I thought the test results from the SSD published in this article [1] were pretty interesting. Maybe that's old news though.
"It also happens to be a Mac, with an operating system that looks nice and doesn’t suck"
Yes Apple has some of the best hardware out there, albeit overpriced. But the software is such a hindrance and I can't take anyone that states otherwise seriously. If only it had proper Linux support (and the Asahi people do an amazing job but you can reverse-engineer only so many stuff with limited funding, and then you have to do it again for new models). MacOS is good if you just want to have a standard experience, which to be fair is most people. It's good for just setting up an LLM server I guess since the hardware is a perfect fit. I wouldn't touch it otherwise.
This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.
I'm also curious about any new low hanging optimization opportunities in the kernels for this new hardware.
It's already clear to me that M5 Mac Studio is more cost-effective than anything you can run on open router, assuming decent utilization.
The M5 Mac Studio will be the most cost effective way to run uncensored cyber capable open agents.
An exciting tipping point will be if programmers can get an Astra-Ultra like experience all week with this hardware. That would be a real sense where this hardware exceeds the value of even 20x cloud subscriptions.
This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.
Local models are definitely not as productive as SOTA, sadly it's not close yet. I do think someday they will be "good enough" to use, but they aren't today. Even the SOTA models barely code well, with Opus 4.5 being the first, good coding model.
That being said, I think it's absolutely imperative that we keep pushing local model performance. We need to continue to advance technology there and ensure that the model labs don't do regulatory capture in the name of "safety" (or anything else).
While I know it's not apples to apples, the target comparison right now is 2x DGX Sparks. Similar price, 256gb. The conversation has focused on memory bandwidth vs. compute in agentic loops, so for most people the raw numbers will mean less than the "time per task" in coding benchmarks.
This is a great article and bodes well for the M5, but we should expect more like this comparing to other platforms before we truly understand where it fits.
It's not the M5 Ultra itself, but the M7s or M9s that will do the damage.
99% of people will use whatever AI is free. The sophisticated, heavy users that are willing and able to pay a lot of money the ones that will be interested in controlling their inference bills.
Today, the sweet spot where an M5 Ultra makes sense is tiny. But we might expect that to grow a lot.
Ordered one for OpenFOAM. Excited for it. Will be live to not have my laptop running CFD 24 hours a day, but my M1 Max is currently my fastest machine… I’m expecting about 3.5x from the M5 Ultra.
I think Apple is really sleeping on making this run a Linux server. These things are very capable and draw very little wattage when idle. It would make an excellent homelab device, but MacOS currently holds it back in this regard.
When people benchmark MLX related quant models, they really need to publish numbers on benchmarks. You cannot take this as it is what you get of the original models. MLX uses pretty simple quantization methods so at lower bits without QAT, it is just not as good quality as llama.cpp ones.
$12,299
5 years of (200/month) tokens at that price, meanwhile an rtx 5090 pc is about half that… hmm
but i wonder how much these token costs are sustainable or not, it may be in the long term cheaper to have your own hardware if token costs go up (and hopefully hardware gets cheaper again)
The numbers I was most interested in are tucked away in a chart towards the bottom - the speed comparison of the Mac Studios v.s. a RTX 5090:
A whole bunch more comparison numbers in this section: https://www.macstories.net/stories/m5-ultra-mac-studio-revie...
Those are some incredible graphs, that leap in prompt processing going from M3 to M5.
Also: ~30 token/s on GLM 5.3-flash, locally. (That's roughly Opus 4.8-tier. I think).
/meta Here's a CSS filter that stops those nuisance chart animations,
A dense 27B doesn't really make sense for the Mac. A MoE makes way more sense when you have modest bandwidth but lots of memory.
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two heads on.
Those RTX 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.
can confirm.
I dont' know why people spend huge money on these and Spark. The 5090 is running qwen 3.8 at 200+ tps!! That's 1-2 orders of magnitude faster.
same reason they spend huge amounts of money on rolexes when seikos work better (the tech crowd isn't immune from vanity).
Both are probably single-token decode performance, which is reasonable to show. Otherwise agree RTX 5090 should shinebetter with NVFP4.
Thank you for this. I wish Apple focused their silicon design on improving the TTFT metrics but coming from an M3 Pro, it still looks laggard compared to Nvidia's TensorCores in the 5090.
Maybe Apple is an acquisition away from changing that balance.
The rumor on Apple's processor roadmap is that they're skipping other M6 variations (all previous generations had Pro and Max, a few had Ultra) in order to focus on the M7 generation for AI reasons. What exactly the M7 improvements are who knows.
https://www.macrumors.com/2026/06/25/2027-macs-m7-chips/
I think that comes down to TSMC. Nvidia apparently booked out the whole A18 or 16 node. Apple is on 2nm right now and M7 will jump right to A14. According to my quick AI research anyway.
That sounds a lot like AI fantasy slop.
Apple just shifted to N2. They’re not going to be doing another major shift right away.
And TSMCs own roadmap would put your hallucination years away at best for a a product that follows a roughly annual cadence https://www.tomshardware.com/tech-industry/semiconductors/ts...
The selling point of the M5 Ultra Mac Studio is that you can run much larger models that the 5090 can't without swapping. NVidia aggressively segments the market on VRAM for this reason. That's why a 5090 has an MSRP of ~$2k (but good luck getting one for less than $4k) while a 6000 Pro, which is basically a 5090 with 96GB of RAM has now soared beyond $15k where 3-6 months ago it was more like $10-11k. A 6000 Pro has the same memory bandwidth but slightly more CUDA units (IIRC ~24k vs ~21k).
This advantage won't be apparent with a 27B model. The 256GB MS can probably run the newer Flash models locally, something you can't do on a 5090.
I don't think we'll get a successor to the 5090 until late 2028, maybe even 2029. I'm basing this on the launch date of the 5000 series and that we haven't got a midcycle refresh yet. Rumor has it the chips are ready but the 3GB RAM modules are 3-4x the price of the 2GB modules used on the current cards.
Apple should see a Mac Studio major update in 2028. That might even force NVidia's hand. But it's really impossible to say what the state of the market will be 2-3 years from now. It may have completely crashed. I suspect not however.
The interesting thing will be when the bandwidth demands start forcing HBM memory onto these home/enthusiast solutions.
But what about builds that combine 8 of the 5090 with infiniband between boxes? Wouldn't that be comparable to the mac in terms of price and potentially beat it by a lot in terms of performance for the large MoE? I understand the space/heat/noise considerations, but price wise it may still not make as much sense as people think. (Agreed that it is hard to get the NVIDIA hardware and the 6000 pro are priced less competitively).
While that sounds super awesome, How many people are actually going to build and maintain that vs a box you can grab at the mall that fits in a lunchbox?
No, $40K is not comparable to $10K.
Not forgetting of course that an RTX5090 is what 600W+ ? And the Mac is probably half that at most ?
sure. so is 2x power worth 10x perf? I think it is in most cases.
That's a dense model. Of course it will do worse.
Now try running that Qwen 3.8 Next model on the 5090 and tell me what TPS you get (hint: it's near 0 since it doesnt fit).
On Apple website it says 512GB memory option is available in October. I guess bumping to that one would cost additional 4-6k US$. So an Ultra with 2TB storage would be north of 15k US$.
That’s like 12 years worth of OpenAI Pro subscriptions
Agreed.
Specially since one can pay half right now to OpenAI and sign a 12 year iron clad contract for uninterrupted service delivery of OpenAI Pro.
I hope that was sarcasm.
"iron clad" :)
I think we all expect the heavy subsidized subscriptions to end or significantly increase in price at some point, but it could be years from now and I'd rather spend a similar figure on an hypotetical Mac Studio M8 Ultra, or whatever more advanced competitor that will have likley appeared by that time.
A more apples-to-apples comparison would be with API cost in OpenRouter at the same tok/s rate for the same models that you can run locally, maybe.
HN always has these completely contrived counterarguments. What is actually going to realistically happen that will prevent use of an LLM provider? Did you think that the OP literally meant the 12 years or maybe it was just to show how expensive using a Mac Mini as an alternative is?
Mass revolts of the peasantry burning down data centers and cutting fiber lines.
Yeah, anyone who thinks local AI is going to save them money is likely to be disappointed, at least if they want to run models that are even remotely capable.
Plenty of other reasons to get excited about local AI, but I don't think cost is one of them.
Maybe you are using a local model to go after some Millennium Prize problem and you don't want OpenAI to take your work and use it to win the prize for themselves? $15k might be a bargain.
And, yes, I know a current local model wasn't going to solve the Navier-Stokes problem, but I'm just using it as an example where privacy might be valuable.
Agreed, plenty of other reasons to get excited about local AI.
Despite being on a site called Hacker News, we seem to often overlook the simple aspect of wanting local AI hardware to hack (not necessarily in the cybersecurity sense) with. I got my local AI hardware because it's an enjoyable hobby for me with bleed over into professional pursuits (but work provides plenty of compute access so this spillover is a very minor factor).
Hard to guess, it can go either way. If you will need to be in a syndicate to use non-sterilized models, that mac makes sense. But if there is mandatory registration of personal cyberarms, you risk going to mines once they check you purchases. You could try to play normie and pretend you simply wanted to show off, by keeping your actual work on external disk, but that leaves traces on system. Counting on someone in the Gap renting you gray iron works as long as you can swap credits. Still, this gear is tiny. Put it in your e-car, with uplink, and leave it at uncle's farm. Discreet.
I gotta have some of what you had :)
I thought it was a fun bit of cyberpunk fiction. Those who downvoted him seem to have taken it at face value?
I appreciate the reference to RUSH: Red Barchetta in the final line.
Just commenting here because you're discussing hardware: I thought the test results from the SSD published in this article [1] were pretty interesting. Maybe that's old news though.
[1] https://www.macworld.com/article/3238319/mac-studio-m5-max-r...
The model being tested is 18k as configured.
I didn't expect this to make the 5090 to look like a good deal.
Yes Apple has some of the best hardware out there, albeit overpriced. But the software is such a hindrance and I can't take anyone that states otherwise seriously. If only it had proper Linux support (and the Asahi people do an amazing job but you can reverse-engineer only so many stuff with limited funding, and then you have to do it again for new models). MacOS is good if you just want to have a standard experience, which to be fair is most people. It's good for just setting up an LLM server I guess since the hardware is a perfect fit. I wouldn't touch it otherwise.
This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.
I'm also curious about any new low hanging optimization opportunities in the kernels for this new hardware.
It's already clear to me that M5 Mac Studio is more cost-effective than anything you can run on open router, assuming decent utilization.
The M5 Mac Studio will be the most cost effective way to run uncensored cyber capable open agents.
An exciting tipping point will be if programmers can get an Astra-Ultra like experience all week with this hardware. That would be a real sense where this hardware exceeds the value of even 20x cloud subscriptions.
Local models are definitely not as productive as SOTA, sadly it's not close yet. I do think someday they will be "good enough" to use, but they aren't today. Even the SOTA models barely code well, with Opus 4.5 being the first, good coding model.
That being said, I think it's absolutely imperative that we keep pushing local model performance. We need to continue to advance technology there and ensure that the model labs don't do regulatory capture in the name of "safety" (or anything else).
Ehh, the actual elephant in the room is:
"why bother with local AI at all when you can lease a GPU for $5/hr?"
To which the answer is you shouldn't bother, unless you have a bunch of money to throw at hobby projects.
While I know it's not apples to apples, the target comparison right now is 2x DGX Sparks. Similar price, 256gb. The conversation has focused on memory bandwidth vs. compute in agentic loops, so for most people the raw numbers will mean less than the "time per task" in coding benchmarks.
This is a great article and bodes well for the M5, but we should expect more like this comparing to other platforms before we truly understand where it fits.
Imagine spending a trillion dollars on data centers and then reading this article. Nightmare fuel for OpenAI
For 99.99% of people, spending 15 grand on a Mac Studio just to run Qwen 3.8 locally is a non starter.
It's not the M5 Ultra itself, but the M7s or M9s that will do the damage.
99% of people will use whatever AI is free. The sophisticated, heavy users that are willing and able to pay a lot of money the ones that will be interested in controlling their inference bills.
Today, the sweet spot where an M5 Ultra makes sense is tiny. But we might expect that to grow a lot.
This dream machine costs over $15k (not including the Apple Studio Display)? Nah, that dream is SO OUT OF TOUCH!
If you are buying expensive hardware to run LLMs "on your own machine" you will soon find your ladder is on the wrong wall.
Ordered one for OpenFOAM. Excited for it. Will be live to not have my laptop running CFD 24 hours a day, but my M1 Max is currently my fastest machine… I’m expecting about 3.5x from the M5 Ultra.
I think Apple is really sleeping on making this run a Linux server. These things are very capable and draw very little wattage when idle. It would make an excellent homelab device, but MacOS currently holds it back in this regard.
[delayed]
When people benchmark MLX related quant models, they really need to publish numbers on benchmarks. You cannot take this as it is what you get of the original models. MLX uses pretty simple quantization methods so at lower bits without QAT, it is just not as good quality as llama.cpp ones.
At this point I think I will get the DGX gb300 workstation though I will wait a bit more for the cold season