It's the first time in the history of the world.
Wait a minute, the AI lab has a Mac mini, and it's going to hit the button!
The Information just went viral and OpenAI has acquired it.Tens of thousands.Mac mini and Mac Studio are still tracking Apple's shipments.
Anthropic went the other way, renting Mac computing from Amazon AWS. According to netizens, one starts at $1 per hour, from $500 to over $900 per month.
It's not cheap to rent, and it's not cheap to buy.
What's that?
These machines are used for reinforcement training and are trained by Computer-use Agents who can operate their own computers.
Tens of thousands of small metal boxes pulled into an AI lab
To train the agent, the goal is to get the model to practice on a real computer.
Thousands of machines are running at the same time, and the models point to the inside of the mouse, open the app, look at the screen, make mistakes, and come back.
You can buy any card with a calculator, but there are some that only a Mac can buy.
This is the rule of apples.
MacOS virtual machines can only run on Apple's own hardware
On a Mac with Apple chips, a machine can run at most two macOS virtual machines simultaneously.
AWS's EC2 Mac is still a "dedicated host" format, and a lease is an entire physical Mac, charged at least 24 hours a day. The reason given in the document is that Apple's license agreement requires this.
That's why.If you want 10,000 macOS environments, you have to put down thousands of virtual machines.。
Doug Brooks, senior manager of Apple's chip products, has given an explanation.
Agent wants a fully self-contained machine, separate from the mainframe, that doesn't turn off for days or nights, and the Mac mini is such a wonderful system.
We will see how bad this wave of demand is, if we look at the profits.
Apple's latest season of Mac revenue$10.4 billionIt's the same.29%This is a small square box.
The cost is the shortfall of the few configurations that are most popular among AI developers, which have been broken up for months.
Some people on X have been lucky enough to order their own Mac Studio early.
Unified memory, plus a fan that won't drop frequency
The macOS environment is not enough, and the machines themselves have to run large models.
Apple's chips not only play GPUs, but they're also capable of delivering high-definition graphics.Unified MemoryThis is a structure that people can't learn in the short term.
The CPU and GPU share a memory pool without having to cut a large model into pieces and stuff it into multiple graphics cards and move it back and forth.
Running a large model is often more important than running fast.
M5 Ultra Highest512GB of unified memoryBandwidth1.2TB/sThe M6 Mac mini in the L.M. Studio shows the speed of processing, which Apple says is 4.8 times that of the M4 and 13.5 times that of the M1.
The second advantage is thatIt's been a long time since I've been able to stay full.。
A lightweight MacBook runs for a few decades and reduces the frequency, while the Agent is a 7×24 that runs for several days.
The heat dissipation here directly determines whether or not it will work.
EXO Labs' open-source program allows multiple Macs to be clustered and run on-premises on a 671B scale.
In the community, the "Mac mini as a Resident Home Server" curriculum has been rolled out, and even hourly rental Mac computing services have emerged.
How does it feel to run Agent on a Mac?
The factory bought it to build the engine room.
It's hard for ordinary people to buy it back and really use it.
The developer, Octavio Herrera's older brother, spent a week working on hisM4 MacMini 32GB basic feeThe last breath tested six models.
Qwen3-30B-A3B, MoE, Q4 quantified 24-30 tok/s, which is 17.3GB.
GPT-OSS-20B:28-29 tok/s, which is 11.3GB
Gemma 4 26B-A4B:20-22 tok/s, which is 14.6GB
Qwen 3.8-27B:4.4 tok/s
The control group, DeepSeek v4-flash runs API: about 57 tok/s
The result is obvious. With the entry-level equipment run-site model, the MoE is essentially pinned to the board.
The chip's computing power is on the one hand, and memory is on the other.
JetBrains, which runs entirely on-premises programming agent, requires 64GB to start with a self-contained 27B quantum model.
The PCWorld version is much simpler.64GB is the floor.。
Someone on X has counted every file.
$3,499, M5 Max 64GBQwen 3.6-27B runs 30-40 tok/s. The 4K cue word is about 5 seconds long, changing a 500 token patch from 13 to 17 seconds, end to end 20 seconds.
$5099, M5 Max 128GBThe bandwidth is constant, the model pool doubles, and the Qwen3-Coder-Next running 80B on the MLX can reach 75-90 tok/s.
$5,499 and M5 Ultra 96GBIn addition to memory, the chip is faster, pushing the same 27B to 48-80 tok/s, end-to-end pressing for 10 seconds.
It's $11,299, and the M5 Ultra is 256GB.The first Mac to have DeepSeek V4 Flash installed in a single, full 4-bit Mac, expected 80-90 tok/s.
 |  |
According to Artificial Analysis, the GPT-5.6 Sol cloud service is 75-79 tok/s, while Opus 4.8 is around 58.
In other words,One is on the table.MacIt's the fastest front-end model service in the world.。
Recently, Cook said at a financial meeting, "We are in the midst of a 100-year flood of memory pricing, with prices rising exponentially".
The official price of the Mac mini was $4,499 over a year ago.
The current M6 Mac mini is already priced at $6,999.
Apple doesn't take orders, old yellow is already at the door
There were tens of thousands of orders in front of us, and Apple didn't accept them.
It has been withdrawing from the corporate market in recent years.
Xserve was shut down in 2011, macOS Server was shut down in 2022 and Mac Pro was shut down in March of this year.
It also built its own server with a Mac chip, but only for Private Cloud Compute.
Some firms tried to buy the calculator but were denied it.
The Information also digs deeper into the details. Apple Press does not have an engineering team dedicated to serving enterprise customers, nor does it have a single developer relationship.
Todd Dailey, the former Apple's AI product marketing manager, who stepped down in April this year, took the initiative.
It's a joke that Apple has any team that really wants to embrace corporate AI.
One of the hottest demands for AI was finally outsourced to two startups, webAI and Mount Thor, founded by Peter Voell, a former OpenAI computing infrastructure employee.
If that's the case, don't blame the old yellow for not being hospitable because the DGX Spark is the soft rib of the Mac.
At CES in January, it also released a version of the software update, with performance up to 2.5 times better than when it was released, and could even be hung next to the MacBook Pro as an external accelerator.
It has been repeatedly mentioned by businesses and developers, the reason being that it is simple and always available.
And it was Apple that reacted.
Last year's Mac end-of-year update was in October or November, and on August 25th, the new Mac mini and Mac Studio were released, with the first publicity written directly to the local front-end models and multitasking clusters.
Today, the AI lab uses the Mac mini as an infrastructure and Apple uses it as a desktop machine.
I mean, what's going on with your mini?