System76 Thelio Mira AI pairs dual RTX Pro 6000 GPUs with 192 GB of GPU memory

System76 pitches a local alternative to cloud GPUs

System76 has introduced the Thelio Mira AI, a Linux workstation built specifically for developing AI models on local hardware. According to the company's product listing, which reached the Hacker News front page, the machine targets developers who want to train, fine-tune and iterate on workloads without paying recurring fees for cloud GPU instances. The top configuration combines a 16-core AMD Ryzen 9000 series CPU with dual NVIDIA RTX Pro 6000 Blackwell GPUs for a combined 192 GB of GPU memory, alongside up to 192 GB of dual-channel DDR5 system RAM. Listed use cases include multi-GPU training, fine-tuning, computer vision and image generation, simulations and inference.

The spec sheet

Beyond the headline numbers, the listing lays out a fairly complete workstation package:

• Liquid cooling, which System76 says supports sustained peak performance during long workloads
• Two M.2 PCIe Gen5 NVMe slots, one M.2 PCIe Gen4 slot and space for up to two 2.5-inch SATA drives, with the caveat that one PCIe x16 slot is disabled when the third M.2 slot is populated
• Dual 5GbE LAN ports, Wi-Fi 7 and Bluetooth 5.4
• A 750W or 1000W SFX 80+ Gold power supply, with a dual-PSU arrangement (1000W plus 750W) in dual RTX Pro 6000 configurations
• An aluminium chassis with stainless steel mesh and a tempered glass front, designed and assembled in Denver, Colorado

Buyers can also choose AMD Radeon Pro GPUs instead of NVIDIA parts. The machine ships with a choice of Pop!_OS 24.04 LTS with the COSMIC desktop, Ubuntu 24.04 LTS or Ubuntu 26.04 LTS, and System76 notes that other Linux distributions, as well as Windows, can be installed.

Reliability and expansion

Two parts of the pitch stand out for AI work in particular. First, System76 highlights the ECC GPU memory, arguing that silent memory errors are a real risk during long training runs and that error correction helps protect hours of compute time. Second, the dual-GPU configuration runs the two PCIe 5.0 x16 slots at x8/x8, which the company says speeds up memory transfers between the cards when training large models.

The workstation is also designed to be opened and upgraded. System76 says swapping RAM, storage and GPUs is straightforward, with documentation provided, and buyers get lifetime access to its in-house support team. Hardware carries a one-, two- or three-year limited warranty, while batteries and chargers are covered for a single year. The company also presents the machine as open source hardware.

One thing the listing does not include is pricing. Prospective buyers will have to configure a system to find out what the dual-GPU setup actually costs.

Why it matters

The market for local AI development hardware has grown quickly as models and fine-tuning workloads have outgrown consumer GPUs. Memory capacity is usually the gating factor for what can be trained or served on one machine, and 192 GB of GPU memory in a single supported workstation is a meaningful pool by that standard.

System76's argument is partly economic, since rented cloud GPU time adds up over months of iteration, and partly about control. Running sensitive or proprietary models locally, on an operating system the company says collects no user data and offers encryption at setup, appeals to teams with privacy or compliance constraints. For developers already working in Linux, a warranty-backed machine with open hardware designs and US assembly is a differentiator against building a dual-GPU rig from parts.

The open question is value. Without published pricing, the case against cloud rental depends on how heavily the hardware gets used, and dual RTX Pro 6000 configurations sit at the expensive end of the market regardless. Still, the Thelio Mira AI is a sign that Linux-first vendors now treat local AI development as a mainstream workstation category rather than a niche.