Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
You guys must be too young to remember how the PC scared the crap out of white collar workers back in the 70s and 80s. They wanted to type everything out with carbon paper, or hand write into their ledgers. They were so scared because they didn't understand it, looks like the AI haters are just the same. Can you imagine not having Excel or Word to use? That's what the 70s and 80s workers wanted. Now AI is helping cure cancer, but yea, keep hating on it.
There was a weekly PBS TV show teaching you how to use VisiCalc. I was just a kid, but it was still mind blowing, and I could see how people could be afraid of it because if they couldn’t learn it they might lose their jobs.

And surprisingly, there were people later who were against WYSIWYG because they had learned all the shortcuts and macros of wordperfect for their specific industries and didn’t want to relearn skills that “weren’t necessary.”
 
The 512 GB Mac Studio is perfect for on-premise corporate LLM servers.

The economics really favor these Mac Studios for enterprise customers.

Token costs are really expensive, and corporations can split a single server for a hundred employees.

A single $200/month individual Claude Code Pro Max 20x subscription might actually cost $2000/month in tokens at API token prices that enterprise accounts have to pay, since Anthropic doesn't let corporations use individual plans. So already it might pay for itself over year or 2.

And that $200 Claude Pro Max 20x subscription has 5 hour and weekly limits that an on-premise server isn't going to have, and it can be shared across multiple users. So it's REALLY favorable for enterprises.

Put four M5-Ultra 512GB machines in a tower and you can run the largest open-weight models at reasonable quantizations. It doesn't just make sense for the enterprise looking to save money. It is the difference between using AI or not for a business that must guarantee their data never goes off site.
 
Last edited:
This is already the case. Apple uses LPDDR RAM. This kind of RAM is not used in AI data centers. The data center computers use RAM that is both “registered" and ECC. It is not at all interchangeable.

What creates the shortage is that the same semiconductor manufacturing equipment is used to make both kinds. The bottleneck that creates the shortage is "space in the factory". It takes many years to create more factory space, and the shortage rolls all the way back to a company in the Netherlands that makes the equipment used to make the chips. In other words, you have to wait in line to build the factory because the machines that make the RAM are sold out.

But the RAM itself is already very different.
It's not exchangeable, but it has the same effect. Apple should absolutely get into this business. Who cares if it takes a few years to build facilities when the end result is independence from the ghouls running the RAM industry right now.

Apple has rightfully made itself independent from Intel, GPU makers, networking components makers and so on. All are cases where the outside makers were either incompetent (Intel) or the components were too expensive (GPUs, networking).

But 'when' may not be in our lifetime. Just saying.
It doesn't take a lifetime to build some RAM factories. Just saying.
 
  • Like
Reactions: Omega Mac
Put four M5-Ultra 512GB machines in a tower and you can run the largest open-weight models at reasonable quantizations. It doesn't just make sense for the enterprise looking to save money, it is the difference between using AI or not for a business that must guarantee their data never goes off site.
Turns out Apple's On Device compute philosophy scales up to enterprise nicely.
 
  • Like
Reactions: vantelimus
Does this story sound odd to anyone else? Apple hosted "a 'Business at the Park' event involving executives from major companies Ford, Disney, and Anthropic" and yet also "did not possess... staff focused on developer relations..."

How does Apple even create and put on an event like that without such staff? Do they want us to believe that this was completely created by consumer-facing staff only?

Make it make sense, The Information.
Apple Foundation Model 3, announced in June - ballpark timing of the gathering. Lots of stories of AFM and enterprise going back a year.
 
Sounds like Businesses don’t trust cloud AI either. They want the power in their hands. All these datacenters are wasted investments: The People what private power.
Any business claiming trade secrets — especially in pharmaceuticals, health care, robotics and consumer tech, to name a few — should be very wary of OpenAI and Anthropic tech. It’s the AI cloud version of Amazon building cheaper Amazon-branded products based on sales info from its storefront to undercut successful third party suppliers.
 
This will unleash a whole lot of new video customers as thousand of Ai film and video makers start offline AI studios. Of course there's going to be an enormous demand. Who isn't tired of the credit-sucking and moneygrubbing of online AI generators?
 
Apple should’ve rebuilt everything, not just processor and modem but memory too. Then got behind that so by the time a shortage like this happen, Apple would be so proprietary that it wouldn’t even matter, it would still be Apples to Oranges.
Without their own fab facilities it does not matter. The cost of which are non-trivial. Dell did this with their CAMM memory but their machines don’t see any pricing benefit because of limited fab capacity.

This is already the case. Apple uses LPDDR RAM. This kind of RAM is not used in AI data centers. The data center computers use RAM that is both “registered" and ECC. It is not at all interchangeable.

What creates the shortage is that the same semiconductor manufacturing equipment is used to make both kinds. The bottleneck that creates the shortage is "space in the factory". It takes many years to create more factory space, and the shortage rolls all the way back to a company in the Netherlands that makes the equipment used to make the chips. In other words, you have to wait in line to build the factory because the machines that make the RAM are sold out.

But the RAM itself is already very different.
This.
It's not exchangeable, but it has the same effect. Apple should absolutely get into this business. Who cares if it takes a few years to build facilities when the end result is independence from the ghouls running the RAM industry right now.

Apple has rightfully made itself independent from Intel, GPU makers, networking components makers and so on. All are cases where the outside makers were either incompetent (Intel) or the components were too expensive (GPUs, networking).


It doesn't take a lifetime to build some RAM factories. Just saying.
Apple is not immune for GPU, CPU or modems. They are still built in the same places. I am not sure whether a memory fab facility would be the answer as it requires different processes to build. 2nm vs other nm as examples.
 
  • Like
Reactions: ikir
Without their own fab facilities it does not matter. The cost of which are non-trivial. Dell did this with their CAMM memory but their machines don’t see any pricing benefit because of limited fab capacity.


This.

Apple is not immune for GPU, CPU or modems. They are still built in the same places. I am not sure whether a memory fab facility would be the answer as it requires different processes to build. 2nm vs other nm as examples.
I agree with you, true, I just want Apple to have everything to themselves, and of themselves. So we, as a community, don’t have to deal with outsider’s problems.
 
According to energy consumption tables at published by the U.S. Energy Information Administration (eia.gov), a staff of 5 people working for one week pulling together and formatting compliance reports will use approximately 5x133=665 kWh of electricity at work. According to Figure 2.1 in Energy and AI, a report published by the International Energy Agency (iea.org), a moderate agentic task consumes about 50Wh of electricity. Taking these figures as rough orders of magnitude, rounding up to the nearest power of ten for the agentic query (100Wh) and down for the humans (100kWh), we see the humans will burn about 1,000 times the electricity doing this job than the AI agent. So, what were you saying about the efficiency of electric-power usage?
You missed the part where, unless we’re unaliving everyone whose job gets taken by AI, those folks will still consume energy *and* the ai platforms that would do their job will too

Also the energy for someone to check the AI didnt hallucinate its way into doing the job incorrectly
 
  • Like
Reactions: noraa
The truly ironic thing with all this AI crap is that the more the big AI companies build out their data centers and suck all the available RAM available, the harder and more expensive it’s going to get for consumers to afford computers to utilize these companies AI.

The level of shortsightedness is astounding, albeit hardly surprising.
 
According to energy consumption tables at published by the U.S. Energy Information Administration (eia.gov), a staff of 5 people working for one week pulling together and formatting compliance reports will use approximately 5x133=665 kWh of electricity at work. According to Figure 2.1 in Energy and AI, a report published by the International Energy Agency (iea.org), a moderate agentic task consumes about 50Wh of electricity. Taking these figures as rough orders of magnitude, rounding up to the nearest power of ten for the agentic query (100Wh) and down for the humans (100kWh), we see the humans will burn about 1,000 times the electricity doing this job than the AI agent. So, what were you saying about the efficiency of electric-power usage?
So you’re saying humans require energy to live? Mazel tov! You just solved all of our energy problems! In the words of Bender Bending Rodriguez “Kill All Humans!”
 
  • Like
Reactions: seek3r
No you don’t. The whole US economy is in this rabbithole so bad that it will likely collapse the intire economy with the current defficit.
You're right... I heard that private equity is holding most of the cost for AI... but they also hold a lot of equity in US bonds although how much they own is not known.... If AI crashes they would probably take out their holdings from US bonds to shore up their loses... that could cause an increase in interest rates for everyone....
 
Right, those sold so well and would make so much sense now.
My point was that they *would* sell well now. An xserve with either an extreme version of the M5 or dual or quad M5 Ultras with a fast internal interconnect would be a dream to anyone considering picking up several studios to wire up with TB for RDMA.
 
The same people hoping AI to fail is like Steve Ballmer laughing at a $500 iPhone and self driving cars which use like 300GB of RAM and Tesla Optimus type robots probably going to make the RAM crisis even worse come 2030 "you'll own nothing, and you'll be happy"
 
  • Like
Reactions: norbinhouston
Executives aren’t developers, completely different audience.
Full quote, bold is mine:

The company reportedly did not possess an engineering team dedicated to business customers or staff focused on developer relations, and lacked an enterprise AI strategy.

The story does not make sense from that perspective. They make it sound like Apple is some naive startup that has never done something like this before. Apple dwarfs the other companies in terms of revenue, and has been working with them all.
 
The Mini's are not even good for AI though..

It's simply all the stupid YouTuber's hyping them but at this point we should know better than to trust them.

You're way better off building a desktop with RTX.
Even a cheap 12gb vram RTX will destroy an M5 Pro.

Even then your local AI options are extremely limited in reality.
24gb vram gets you into a bit better league but still very limited.
After about 30gb vram You're shooting straight up to around 128gb to be in the next tier but that's a whole different price level and makes sense for very few people to do that rather then just paying for cloud.

We need to reach the point that most people have come to this realization after testing and then the bubble of buying this hardware like Mini's will pop.
The more tech savy have already gotten there. It's just the less tech savy and more gullible people believing what they've heard.
A lot of them won't even end up using local AI after buying. Another good chunk will not realize how much better of they would be with using cloud.
True bubble and if the tech doesn't catch up to the lie of local AI fast enough then it will pop.

Then there's the whole issue of cloud AI pricing that could change a lot.
Another time bomb. Already been bursting with many companies seeing the massive bills for API usage and canceling or extremely restricting budget for AI.
Providers need to very quickly figure out how to make their Frontier models run more efficiently.
 
  • Love
Reactions: turbineseaplane
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.