华盛顿州贝尔维尤一处安静社区，加里・弗莱克的现代住宅地下室，布置得像一间功能齐全的男士休闲空间：大屏显示器与数字投影仪中间摆着一张舒适沙发，还有一台动感单车划船机。但这间屋子真正的重头戏，是一台造型酷似巨型拼装模型的定制主机，机架上悬挂七块英伟达高性能显卡。
科技爱好者在家中装配高性能 AI 主机，随之而来的是棘手的散热难题
弗莱克曾任职微软，也是多家初创公司创始人。他组装这台机器，用于基于数 TB 脑电图数据训练定制 AI 模型，脑电图用来记录大脑电信号，该项目所属的创业项目他现已不再跟进。
训练 AI 模型，他本可以向云服务商租用数据中心的工业级 GPU 算力。但他说，用折扣配件自己装机要划算得多，整套设备总成本不到 6000 美元，远低于每月付给亚马逊云或者微软云的费用。
即便算上电气改造、全新暖通空调系统的开销（合计 1‑2 万美元），用来给地下室散热，抵消 AI 主机巨大发热量，自己装机依旧划算。 “我想要算力握在自己手里，” 弗莱克说，他计划未来数月重新启用这台机器开展新项目，“我不想租用别人的算力。”
阁楼、地下室、家庭办公室本地跑 AI 负载，已经成为这一轮 AI 热潮之下出人意料的现象。主流 AI 大厂，都在堆满数万片 GPU 的大型远端数据中心训练、运行大模型，满足海量算力需求。
加里・弗莱克家中的 GPU 设备
与此同时，一股本地 AI 浪潮正在兴起。大量小型开发者、硬件发烧友出于成本、隐私、使用便利，选择在自己的家用设备上跑 AI。部分信奉自由理念的开发者认为，本地 AI 可以避免未来政府限制、切断自己使用 AI 技术的权限。OpenClaw 这类本地智能体平台，可以直接调用本机应用、本地文件完成复杂任务，也助推了这股风潮。
有些时候，家用 AI 设备纯粹就是很酷的家居景观。上个月，Shopify 首席执行官托比・吕特克晒出家里三组竖直服务器机柜，机柜缠绕粉色 LED 灯带，在社交平台走红。设备包含一台戴尔主机，搭载英伟达高端显卡，厂商称之为 “终极桌面超算”。吕特克还巧妙解决了设备散热难题。
家用 AI 主机兴起，对硬件厂商来说是意外利好。Mac mini、Mac Studio 在智能体开发者群体中大受欢迎，销量走高；戴尔也表示，家用 AI 爱好者市场迎来爆发式增长。
戴尔高级杰出工程师马克・哈蒙斯称：“这是时代风潮。今年 OpenClaw 问世之后，相关硬件需求直接暴涨。”
在家中放置高性能主机并不是新鲜事，“家庭实验室（homelab）” 爱好者社群早已在 Discord、Reddit 等平台蓬勃发展。21 世纪初，发烧友在家搭建服务器，挂载海量硬盘，存放下载的 MP3 音乐与视频。后来加密货币挖矿热潮，又带动家用高性能 GPU 需求。
希滕・沙阿就是当年的加密货币玩家。2020 年，他在家中服务器机架装上八块 GPU 挖以太坊。如今，软件创业公司 Crazy Egg 的 CEO 沙阿，转向 AI 开发，大幅扩充加州雷德伍德肖尔斯家中的算力军备，一共 18 台设备，其中包含 14 台 Mac mini 与 Mac Studio。
沙阿表示，本地算力让他可以掌控自己开发的网站优化软件运行表现。云大模型时常宕机、卡顿已经是常态。“每当云端更新大模型，接连好几天系统都会不稳定，故障频发。”
但代价是居住舒适度大打折扣。全部算力设备摆在家庭办公室，机器散热让室内温度经常飙升至华氏 85 度（约 29.4 摄氏度），他经常只能穿短裤待在房间。“我该装一台空调。” 沙阿说道。
不少本地 AI 发烧友已经成为圈内网红。一位居住在奥斯汀的 YouTube 博主，要求匿名，对外只用频道名 Digital Spaceport。他在车库搭建了自己口中的 “家庭数据中心”，硬件包含五块 AMD 显卡，外加大量网络、存储设备。目前他在本机运行国内开源大模型通义千问 Qwen‑3.8‑27B，用来剪辑、分析自己的 YouTube 视频。
另一位波兰科技顾问兼 YouTube 博主，同样匿名，使用频道 ID 0xSero。随着硬件越买越多，他不得不把 AI 主机在家中四处挪位置。四块英伟达 6000 显卡最初放在家庭办公室，但噪音巨大、发热量极高，房间空气被烘干，他认为这是去年自己生病的诱因。于是整套设备被搬到阁楼。
0xSero 最早也使用云端 AI 模型，2024 年末转折点到来：AI 编程工具 Cursor 涨价。“那一刻我下定决心，要搞本地 AI。”
他坚持一个观点：AI 系统的所有权、控制权应该掌握在用户手上，未来有可能出现政府管控、限制 AI 访问权限。“我真心喜欢 AI，我希望能够持续拥有使用它的权利。”
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A quiet community in Belviu, Washington State, the basement of Gary Flake's modern home, is set up like a fully functional men's leisure space: There was a comfortable couch between the large screen monitor and the digital projector, and a moving bicycle rower. But the real big scene in this room is a custom-made mainframe that looks like a giant assembly model, with seven British Weeders high-performance GPU on the shelf.
Technology lovers are home-based with high-performance AI hosts, and with them are the challenges of heat dispersion.
Fleck had served as Microsoft and was the founder of several start-up companies. He assembled the machine to customize the AI model based on digital TB EEG data, which is used to record brain telecommunications, and he is no longer following up on the entrepreneurial project to which the project belongs.
Training the AI model, which would have leased the data centre ' s technical GPU algorithm from cloud service providers. He states, however, that it would be much more cost-effective to install his own machines with a discounted spare, and that the total cost of the equipment package was less than $600,000, which was far lower than the monthly payment to Amazon or Microsoft.
Even if the cost of the new HVAC system (totalling US$ 10.2 million) is included in the electrical modifications, it will be cost-effective to install the machine to heat the basement to offset the huge heat flow from the AI host. "I want to figure it out in my own hands," Flick said, "I don't want to rent someone else's calculations."
Attics, basements, home office local running AI loads have become unexpected in this round of AI booms. Mainstream AI large plants are training and running LLM in large remote data centres with tens of thousands of GPUs to meet mass computing needs.
GPU equipment at Gary Flick's house.
At the same time, a local AI wave is rising. A large number of small developers and hardware burners have chosen to run on their own home devices for reasons of cost, privacy, ease of use. Some of the developers of freedom believe that local AI can avoid future government restrictions and cut off permission from using AI technology. These local AI agent platforms, such as OpenClaw, can directly call on their own applications, local files to perform complex tasks and contribute to the tide.
Sometimes, home-based AI equipment is just a cool homescape. Last month, Chief Executive Toby Luttec got out of his home of three vertical server cabinets, entangled in pink LED lamps, red on social platforms. The device consists of a Dale mainframe carrying a high-end GPU in Weiday, known by the manufacturer as “the ultimate desktop overcalculation”. Luttec also skilfully solved the equipment heat-dissemination problem.
The rise of the home AI mainframe is an unexpected gain for hardware manufacturers. Mac Mini, Mac Studio is popular among AI agent developers, selling highs, and Dale says that the market for home-based AI lovers has seen an explosion.
Dell's senior and distinguished engineer, Mark Hammond, said, "This is the tide of times. This year, following the appearance of OpenClaw, there was a direct surge in hardware demand.”
The placement of high-performance hosts at home is not new, and the family lab (homelab) community of lovers is already flourishing on platforms such as Discord, Reddit. At the beginning of the twenty-first century, fever friends built servers at home, mounted big hard drives, stored downloaded MP3 music and videos. Later on, the trend of encrypted currency mining led to high-performance GPU demand.
Heaten Shah was the encryption money player of the year. In 2020, he put eight GPUs on the server rack at home to dig the Ether. Today, CEO Shah of the software start-up company Crazy Egg turned to AI and significantly expanded the calculator arms at Redwood Shaws, California, with 18 devices, including 14 Mac Mini and Mac Studio.
According to Shah, local computing allows him to control the performance of the software on the website he has developed. The cloud LLM is a regular occurrence. "When the clouds update LLM, the system will be unstable and malfunctioning for days."
But the cost is a significant loss of comfort. All the computing equipment was placed in the home office, and the machine heat spread so that the temperature in the room often soared to 85 degrees Fahrenheit (approximately 29.4 degrees Celsius), and he was often confined to his room in shorts. "I should put an air conditioner." Shah said.
A lot of local AI fever friends have become inner net red. A YouTube blogger living in Austin has asked to be anonymous, using the channel name only. He built a “family data centre” in his own mouth in the garage, hardware containing five pieces of AMD GPU plus a large number of networks and storage equipment. He is currently running his own open-source LLMQwen Qwen #3.8 #27B for editing and analysing his YouTube video.
Another Polish scientific consultant and producer of YouTube, also anonymous, uses channel ID 0xSero. As more hardware was purchased, he had to move the AI mainframe around the house. Four British widas, 6000 GPU, were initially placed in the family office, but the noise was high, the heat was very high and the room air dried up, which he considered to be the cause of his illness last year. So the whole equipment was moved to the attic.
0xSero was also the first to use cloud AI, and the turning point at the end of 2024 came: the AI programming tool Cursor raised the price. "I made up my mind at that moment to get a local AI."
He maintains the view that ownership and control of the AI system should be in the hands of users, and that there may be government controls and restrictions on access to permission in the future. "I really like AI, and I want to be able to use it continuously."
Auto-fetched and translated by AI Atlas. All rights belong to the original author.