eRacks Systems

Name: eRacks Systems
Address: 2100 S Bascom Avenue Suite 1 #516, Hayward, CA, 95008, US
Telephone: +14084550010

Home
United States
Hayward, CA
eRacks Systems

eRacks Open Source Systems was founded in 1999 with a vision to make open source software available on the finest hardware, from rackmounts to ultrabooks.

This is the official page for eRacks Open Source Systems. News, views, and juice you can use will be regularly published here. eRacks was founded in 1999 with a vision to make open source software available on the finest hardware, and is the leading provider of rackmount servers, laptops, notebooks, ultrabooks, and desktops pre-configured with the latest Open Source software, including Linux, BSD, Apache, and cloud software such as OpenStack or CloudStack.

10/29/2026

There's an interesting shift in how AI inference (running a trained model, as opposed to training it) actually works under the hood. It splits into two distinct phases: prefill, where the system reads and understands your prompt, and decode, where it generates the answer one token at a time. Those two phases have different hardware needs, one is compute bound, the other is memory bandwidth bound.

That's part of why you see chipmakers starting to build specialized silicon for each phase instead of one general purpose GPU (graphics processing unit, the chip that does the parallel math for AI) trying to do both.

We've been building custom AI servers at eRacks for a while now, and this is exactly the kind of thing that shapes how we spec a box: how much of your workload is prompt processing versus generation, and what that means for the GPUs, memory, and interconnect you actually need. If you're planning an on-premise AI server, it's worth thinking about your workload mix before you buy.

https://share.google/UMh8qVcXsp9erGmtn

10/29/2026

We came across a solid rundown of newer Linux distros (distro = a packaged version of Linux with its own tools and defaults) that are worth watching: AnduinOS, blendOS, and a handful of others built around one clear idea instead of trying to do everything.

This is the same spirit we've built eRacks on since 1999: pick good open-source parts, keep them stable, keep improving them release after release. Whether it's a desktop distro or the Linux stack running on one of our rackmount servers, the distros that last are the ones with a clear audience and consistent ex*****on, not hype.

If you're curious what's out there beyond Ubuntu and Fedora, this is a good starting point.

https://share.google/tbJFoYAQuaUcGttcr

10/28/2026

DapuStor's 245TB-class SSD (solid state drive) is a good marker of where enterprise flash storage is headed as AI workloads push hyperscale data centers to rethink storage priorities. Capacity per drive keeps climbing, and that changes the math on density, power draw, and rack space for anyone storing large training sets or model checkpoints.

We've been building dense storage platforms for 25 years, and this is the same trend we watch when we spec eRacks/FLASH arrays: fewer drives, same or more capacity, less power per terabyte stored.

If you're planning storage for AI workloads, the drive-level trend matters even if you never touch a 245TB SSD directly. It's a sign of where the whole flash market is moving.

https://share.google/x19REzd8lN7YeKG2f

10/28/2026

HexOS takes TrueNAS (an open-source storage operating system) and wraps it in a simpler web interface, aimed at people who want NAS (network-attached storage, a box of drives shared over your network) without learning enterprise storage tools first. A year of steady updates has it closer to that goal.

It's a good example of the gap between "powerful" and "usable" closing in open-source storage software. That gap matters whether you're running a self-built box or something off the shelf.

If you'd rather start from hardware that's already built, tuned, and ready for whatever OS you choose, including TrueNAS, that's exactly what eRacks/NAS systems are for.

https://share.google/sMPiXDQzs4qgSwUUt

10/24/2026

AMD's AM4 socket, the desktop platform that's been around since 2016, keeps getting new motherboards. Gigabyte put out four more. That means DDR4 (the memory generation before DDR5) is still very much a supported, buildable platform, not a dead end.

For anyone with DDR4 sitting around, or just trying to build a solid box without paying DDR5 prices, this is good news. The platform is mature, drivers are settled, and parts are easy to find.

We build on whatever platform fits the job and the budget. A mature, well supported socket like AM4 is often the practical choice for a quiet desktop or a basic workstation.

https://share.google/pr2xB64uxPb204ku0

10/24/2026

We run a lot of boxes that live their whole life at a command line, no desktop, no GUI, just SSH (secure remote login) in and work. So a browser that runs entirely in a terminal and still renders full web pages caught our eye.

Brow6el uses Sixel, a graphics format that lets terminals display bitmap images, to show real rendered pages via the Chromium Embedded Framework. Bookmarks, ad blocker, JavaScript console, Vim-style keyboard navigation, the works, all without a window manager.

It's also a quiet rebuttal to the current trend of browsers shipping AI assistants nobody asked for. Sometimes the right tool is just a browser that browses.

https://share.google/7Noyr7ig4Eo3RWB4b

10/23/2026

AMD's Ryzen AI Halo is a mini-PC platform built around the Ryzen AI MAX 300 chip, aimed squarely at running AI models locally instead of in the cloud. AMD promised full ROCm (their software stack for running AI workloads on AMD hardware) support, and it's expected to come with 128GB of memory.

The interesting part is the Windows support. NVIDIA's DGX Spark, which this is clearly built to compete with, runs Arm architecture and needs emulation for a lot of software. Ryzen AI Halo just runs Windows natively, so it can dual-boot and do normal desktop stuff too.

This is part of a bigger pattern we track closely: local AI hardware is getting more capable and more flexible, not less. We build private on-premise AI servers for customers who want their models and data staying in-house, and platforms like this are exactly the kind of hardware that trend is built on.

https://buff.ly/edULVUr

10/23/2026

Minisforum's MS-S1 MAX is one of several small-form-factor workstations built around AMD's Ryzen AI Max+ 395 "Strix Halo" APU (a single chip that combines CPU and GPU cores). What makes it worth noting is the up to 128GB of LPDDR5X memory shared between CPU and GPU, which is the kind of unified memory setup that matters for running AI models locally instead of only in the cloud.

We've been building Linux and open-source rackmount and desktop systems since 1999, and this is the same trend we see across the AI hardware market generally: platforms keep getting denser and more capable per watt.

If you're looking at on-premise AI hardware, whether that's a compact workstation or a rackmount GPU server, our AI server line and AIPROV provisioning service are built for exactly this kind of local deployment.

https://share.google/Bcq7yGZk6EKoXMEyA

10/21/2026

Here's an interesting OS design idea we came across: Loss32. Instead of running a normal Linux desktop and launching Windows programs through W**E (a compatibility layer that lets Windows apps run on Linux), Loss32 flips it: build the whole desktop environment on top of W**E itself, sitting right on the Linux kernel.

It's a different philosophy than projects like ReactOS, which tries to reimplement Windows from scratch, or older efforts like Lindows. Whether or not Loss32 ships as real software, the idea is worth thinking about if you care about how these compatibility layers actually work under the hood.

We build and support Linux systems all day, every day, so stuff like this is right up our alley. Read the article here: https://share.google/EX42ZLs8zHGJ5nQkj

10/20/2026

There's a tool in the Claude Code ecosystem called Ralph Wiggum, named after the Simpsons character, that takes a brute force approach to agentic coding (AI that works through tasks autonomously instead of waiting for step by step prompts). It feeds its own failures and stack traces back into itself and keeps grinding until the task is done.

It started as a 5 line Bash script from a developer who got tired of manually re-prompting every error. Instead of fixing the bottleneck with more reasoning, he just automated the retry loop.

We run a lot of boxes doing exactly this kind of unattended, always on compute work, whether it's agentic coding, local AI inference, or batch processing. If you're building out a server to let something like this run overnight without babysitting it, that's the kind of workload our AI server line is built for.

https://share.google/BV8mO04avcVMEGoxb

Address

2100 S Bascom Avenue Suite 1 #516
Hayward, CA
95008

Telephone

+14084550010

Website

http://eracks.com/, https://twitter.com/eRacks

Alerts

Be the first to know and let us send you an email when eRacks Systems posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.