“More than 34 hours have passed just for this project to take off And it’s still just a model. This is less than a third, “said the software engineer Francielle Dellamora, who has been riding a calculation station with several programmers – Henri, who published from his personal account the images, at 03:47 in the morning, next to Victor Taelin and others. And it seems that what they are saving is not money.
In a few hours, the post began to viralize, and rightly. What they had built was not a geek curiosity, it is a declaration of principles on the bases of distributed computing, about how to think in terms of performance, reliability and scalability from another perspective. That is, not that of the traditional data center, but that of the self -taught engineer with a screwdriver and an idea. And many Mac Mini M4, stacked as if they were books.
If we talk about gross power, the choice makes no sense
The direct comparison is inevitable: Why use Mac Mini M4 if you can have GPUS such as the NVIDIA H100 or the new H200, much more powerful, for a similar or even lower cost per unit of calculation? We go with some key data and a table to understand it:
| Características | 96 Mac mini M4 | Nvidia H100 / H200 |
|---|---|---|
| Performance | Medium (average of 2.5 pflops fp16) | Very high (average of 4–6 pflops fp16) |
| availability | High (immediate purchase) | Low (difficult to acquire, long waiting) |
| versatility | Very high, general use, development, AI, etc. | Low (optimized for AI) |
| Total estimated cost | 55,000 euros approx. (96 × 570 euros*) | 25,000 euros of media for each H100 GPU 35,000 euros for each GPU H200 |
| Ease of replacement - | High, simple substitution and repability higher than ever | Low (complexity in replacement) |
| Scalability | High (modularity with Mac Mini) | Average (requires specific infrastructure) |
| estimated energy consumption | 3.000 W aprox. (96 × 31 W) | About 700 w per gpu |
| Use in other tasks | Yes (Office, Development, Multimedia) | Limited (mainly ia) |
Because this station, mounted in collaboration with Sage Networks, offers more possibilities. An alleged X user, Trek Badr – with since November 2024 -, wielded a series of arguments that serve to understand what Apple has achieved with its pocket computer. The Mac Mini M4 are extremely small. –5 cm wide and less than 700 grams of weight for each M4 model675 gr to be exact – silent and functional.
An H100 delivery up to 700 tops (billion operations per second) in INT8 inference. A Mac Mini M4 offers 38 tops thanks to its Neural Engine. It is not much. Now multiply by 96: We are talking about about 3,648 theoretical tops. Impressive, but far from what a cluster would achieve with four H100, which already exceeds 2,800 tops with less consumption and less space.
So why choose this architecture? Because the theoretical calculation is not everything. H100 are not available for direct consumption. You can’t introduce yourself to a store and take a full box. Require licenses, contacts, approvals, even export regulationsif it is the case of this team, which has collaborated with a Brazilian company. It is not precisely hardware plug & play. It is not for everyone and it is not easy to replace if something fails. And in view of the problems that Nvidia is having for distribution, due to tariff restrictions, it does not look better in the coming months.
Instead, a Mac Mini M4 is bought with a click. In days you have it at home, in hours if you go to the closer Apple Store. If one breaks, you return it and now. There is guarantee, sure, Each unit is autonomoushas CPU, GPU, neuronal motor, local storage and a complete operating system. You do not need external units to start. And sa modular independence is key.
It is not about the price. It’s about control
The choice of these 96 Mac Mini has nothing to do. It has to do with operational freedom. You can use those same machines to compile code, make renderstrain models or serve content. You can redirect resources, reassign nodes, isolate failures. You have real, physical, virtualization.
In addition, we said it before: the M4 is a silent chip, but silent zero noise level. If you have one you have noticed it: The density of calculation by decibel and by watt is remarkable and unmatched. And they are also low consumption computers, with thermal efficiency far superior to many alternatives. This allows a complete rack without extreme refrigeration systems. And the Ecosystem Macos, together with tools such as Ansible, Docker or even Apple Remote Desktop, allows friction without friction.
This assembly is not just an eccentricity, it is an interesting and alternative vision that reminds of Folding Home, that project where Sony had thousands of PS3 to create a mega simulation system of protein folding. It is true, they are teams that are not designed for this, but that work. This has been 96 decisions made with criteria, patience – judging by the number of hours of connection, testing and what they have left– And perspective.
Images | Victor Taelin captures ups to x
In applesfera | Apple passwords: What is it, what is it for and how safe is the new iOS 18 and Macos 15 app
In applesfera | New iOS 19 – Everything we think about the new operating system of the iPhone
