eRacks Systems

09/04/2026

New Post: Hard Drive Prices Jumped 40 to 90 Percent Since March: Six Months of Our Own Tracking

Open Source Experts Since 1999

09/03/2026

Ubuntu has built-in AI dictation now, and the interesting part is where it runs: on your own machine, not in someone's cloud. Your voice never leaves the desk.

That is the pattern worth watching on the Linux desktop. The useful AI features are arriving as local ones, which means the hardware under the desk decides how well they work.

We build Linux desktops and laptops for exactly this: enough memory and enough GPU to run local models, set up and tested before they ship. https://www.omgubuntu.co.uk/2026/08/testing-ubuntu-ai-voice-dictation

09/03/2026

This matches how we actually work: a frontier model for the genuinely hard problems, a local model on your own hardware for the constant everyday volume.

Most AI requests are routine. Paying per-token cloud rates for all of them, and shipping the content out of your network to do it, is the expensive way to run the routine 80 percent.

We build the server that handles that side, and set it up so both halves work together. https://www.xda-developers.com/i-use-claude-and-local-llms-together-now-and-it-costs-half-as-much-while-being-twice-as-fast/

09/02/2026

Someone built an open-source tool that scans your hardware and tells you which local AI models will actually run on it, and how well. Genuinely useful, and worth a look before you spend anything.

We will tell you the same thing for free, from the other direction: name the model you want to run and we will tell you what it takes to run it well, or tell you that the machine you already own is enough. That answer is the same whether or not you buy from us. https://share.google/bQ9HTR3AJlGul6hnW

09/01/2026

Canonical is steering Ubuntu toward AI agents: the tooling, the packaging, the assumption that a server will be running models and agents rather than just web apps.

Good news for anyone running private AI, because it is the distribution most of that work already sits on. Ours ships on Ubuntu LTS with the stack installed and tested: container runtime, GPU drivers, model serving with an OpenAI-compatible API, a chat interface your whole office can use, and a written record of the configuration so it can be rebuilt.
https://www.theregister.com/software/2026/06/08/canonical-sends-ubuntu-into-the-ai-agent-era/5252373

09/01/2026

People told a chatbot things they would not tell a colleague. Some of it surfaced later in court filings.

That is not a scandal, it is how hosted services work: your conversations sit on someone else's servers, under their retention policy, reachable by their legal process.

The alternative is not giving up AI. It is running the model on hardware you own, where the transcript never leaves the building and the only retention policy is the one you set. We build those servers and set them up so they work on day one. https://www.yahoo.com/news/us/articles/confided-chatgpt-secrets-ended-court-090000568.html

08/31/2026

Cloudflare launched Kitesurf, a browser engine built for AI agents instead of people: no Chromium, V8 isolates on their Workers platform, 3 to 7 times less CPU and memory. The engineering is genuinely good, and a full Chrome per agent really is absurd.

Notice where it runs, though. Entirely on Cloudflare's edge. Not self-hostable, and closed for now. So the company that sells websites protection FROM bots now also runs the bots' browser, which means it can see exactly which traffic is an agent, from both sides of the transaction.

That is not a conspiracy, it is a strategy, and it is the cloud pattern arriving one layer up. The counter is the same as it has always been: run the stack on hardware you own. https://www.infoq.com/news/2026/08/cloudflare-kitesurf-browser/

08/28/2026

Intel's board partners are squeezing the 32GB Arc Pro B70 into a single slot: 160W, blower-cooled, shown at Computex. Eight of those in one chassis is 256GB of GPU memory (VRAM, the limit on what models fit), enough to serve 200B-parameter-class models.

This is the card we build our value AI tier around: two of them put 64GB of GPU memory in a server for less than the list price of a single 96GB datacenter card. Every AI server we ship leaves the bench benchmarked, with measured tokens-per-second on your actual models.

https://wccftech.com/sparkle-squeezes-intel-32-gb-arc-pro-b70-into-a-single-slot/

08/28/2026

New Post: We Benchmarked a Dual Intel Arc Pro B70 AI Server: Real Numbers, Real Gotchas

08/27/2026

A PBS station is suing to regain access to 70 years of its own archival footage - locked up in a cloud dispute.

We've been saying this for years: we like on-prem NAS. And if you must use the cloud, always keep an on-prem backup - even near-line. Your archive should never be one contract dispute away from unreachable.

https://gizmodo.com/pbs-station-sues-to-regain-access-to-70-years-of-archival-tv-history-2000798323