When you think of an AI data center, most people picture giant servers packed with multi-million dollar NVIDIA graphics cards. But one Chinese startup has shown that there is a completely different way. Instead of dedicated AI servers, it has built an infrastructure of around 1,000 computers Mac mini with chip Apple M4The goal is simple – to run artificial intelligence cheaper, more efficiently and without anyonefees for cloudemail services.
Apple Mac mini The M4 was certainly not designed as a data center server, but it's gaining traction among AI developers thanks to its combination of low price, high performance, and extremely low power consumption. Basic Mac Configuration mini The M4 costs $599 and uses about 10 to 30 watts of power under typical AI workloads. That's significantly less than dedicated servers with powerful GPUs, which typically draw hundreds of watts per card. In the case of the Chinese startup, this resulted in a farm of about a thousand Macs. mini, which together serve the requirements for operating large language models.
You might be interested in
It's not about training AI, it's about running it
It is important to note that such infrastructure is not used to train the largest AI models. This area is still dominated by NVIDIA graphics accelerators such as the H100 or the newer Blackwell generations. Mac mini but it has a different mission. It is ideal for the so-called AI inference, i.e. a situation where an already trained model answers user queries or performs specific tasks. This is where high energy efficiency can be much more important than absolute performance. One of the biggest advantages of the chip Apple M4 is a Unified Memory architecture. The CPU and GPU share the same operating memory, so there is no need to constantly copy data between individual parts of the system. This brings greater efficiency when working with large language models and reduces the energy consumption of the entire system. This is precisely why Mac mini has gained popularity among developers using tools for local AI operations.
One of the main reasons why companies build such projects is the rising costs of cloude-services. Each query sent to a remote AI model means an additional fee. The more users use the service, the higher the monthly costs. However, if the company owns its own hardware, the situation changes. After paying the acquisition costs, it mainly pays for electricity and regular infrastructure management. The costs for individual AI queries can therefore be significantly lower.
View post on Instagram
Modern open-source tools are also contributing to the expansion of similar solutions. One of the most famous is Ollama, which allows running large language models directly on local computers. Individual Macs can be connected into a single unit and together serve requests over a local network. Developers thus obtain their own AI infrastructure without having to rely on remote cloudservices. In 2025 and 2026, there will be an increasingly frequent trend where companies stop renting computing power from large companies on a long-term basis. cloudcompanies and start investing in their own hardware. The reason is simple. Once the infrastructure is paid for, the bill for cloud growing every month. Pro For companies with a high number of AI queries, this can be a significant financial saving. Mac mini Moreover, it scales very well due to its size. In the space that would normally be occupied by several large servers, dozens of small computers can be placed, whose power consumption remains relatively low.
Although Apple never presented its computers as a solution for data centers, it is the advent of local artificial intelligence that opens up new possibilities for their use. Mac mini is gradually becoming an interesting alternative for companies that need to run their own AI models efficiently, cheaply and without dependence on cloudservices. It is unlikely that Macs will be replaced by the most powerful servers with NVIDIA graphics accelerators in the foreseeable future. But they show that for many tasks, extremely expensive hardware is not always needed. Sometimes it can be more profitable to build an infrastructure from a large number of small, energy-efficient computers. And that is exactly what a Chinese startup has demonstrated. Instead of a few monstrous AI servers, it created a data center consisting of a thousand Macs miniIf a similar approach proves successful in other companies, it could be one of the most interesting directions artificial intelligence will take in the coming years.
