SysAdmin Weekly #26: The Off-Switch You Don't Own
A frontier model went dark worldwide in 72 hours this month. Here is why that is an argument for owning your hardware...
TL;DR
· A frontier AI model can be switched off worldwide in 72 hours, and not by you. Anthropic’s Fable 5 and Mythos 5 proved it this month.
· Clearly it’s not a Claude problem. It is a dependency problem, and it is the strongest case yet for owning the hardware your inference runs on.
· SysAdmin Weekly Podcast Episode 048 is out: the AI doom narrative held up against the actual data with Eric Siron.
· The WEF Future of Jobs numbers say net +78 million jobs by 2030, and the fastest-growing skill categories behind AI and big data are networks and cybersecurity. The work that keeps infrastructure running is on the growth side.
· Tool of the Week: Ollama, the second time around. Same recommendation as Issue 24, sharper reason to act on it now.
From the Console
I stripped my lab down a few weeks ago. Pulled the always-on gear I was not really using, consolidated a couple of roles onto hardware I trust to keep running, and rebuilt the rest with longevity in mind instead of raw horsepower. I expected to give something up for that.
Turns out…..I have not, in fact, given up something for those changes…. Everything I actually depend on is running as well as it ever did, and my power bill came back noticeably lighter. Turns out a lot of what I was running was running because it was already on, not because I needed it on.
Given the recent news about Anthropic… it makes me thing about something. Every box in that pared-down lab is one I control. It does not get a worse model pushed to it overnight. It does not get switched off because of a directive passed down from three companies upstream of me (or the US Government for that matter…). Those cases sounded like a philosophical point a month ago. This week it stopped being philosophical.
And now, back to our regularly scheduled programming.
The latest on the SysAdmin Weekly Podcast
Episode: 048 - The AI Doom Narrative vs the Data: Layoffs, Energy, and Jobs in 2026 Topic: Andy and Eric Siron take the AI fear stories one at a time, the layoffs, the data center energy and water panic, the “office workers are gone in eighteen months” predictions, and put each one next to what the actual data says.
Why this one matters:
· The WEF Future of Jobs numbers, the post-COVID overhiring correction, and the hyperscaler nuclear and cooling commitments, read against the headlines instead of in place of them
· What AI actually looks like at a SysAdmin’s keyboard versus what the press would have you believe it looks like
· Plus the nerd hour: Linus Torvalds on AI-generated bug reports drowning the kernel security list, Edge caught storing passwords in clear text, and a Forgejo and Restic lab cleanup
Listen on Spotify
Watch on YouTube
The Take
A model you build your work around can be turned off in three days, worldwide, and you will not be the one who decides it.
That is not a hypothetical. Anthropic released Claude Fable 5 to the public on June 9, with the wider Mythos 5 class going to a narrow set of partners the same week. By the evening of June 12, both were dark. Not deprecated, not rate-limited; switched off for every customer on the planet. The trigger was a US export-control directive barring access by any foreign national, and because nationality cannot be filtered on a live API in real time, the only way to comply was to pull the models for everyone. Anthropic’s own foreign-born staff included.
The part relevant for us mere-mortal SysAdmins? None of that had anything to do with how we were using the model. You could have had a clean, paid, well-behaved workflow built on it, and it would still have gone dark on Friday night because of a conversation between an executive and a cabinet secretary that you were not in.
And that conversation is its own story. The reporting says Amazon CEO Andy Jassy is the one who flagged the concern to Treasury Secretary Scott Bessent, citing Amazon researchers who got Fable 5 to cough up cyberattack-useful output. Amazon is Anthropic’s investor. Amazon is Anthropic’s cloud host. And Amazon, this week, is the party that walked a competitor’s three-day-old model release into a regulator. I am not going to tell you what to read into that. I am going to tell you those three roles sitting in one company is exactly the kind of thing you want to know about when it comes to the supply chain under your tooling.
To be fair to Anthropic, they pushed back on the severity, noting the same capabilities are already reachable through other public models, and Opus 4.8 and Sonnet kept running the whole time. So this is not me telling you Claude is the problem. Claude is not the problem. The dependency is the problem.
We have been building this case in different shapes for a while. Issue 21 was about compute consolidating into a handful of hands. Issue 25 was about the hardware squeeze landing on your doorstep. Both of those were about cost and supply. This adds a third axis, and it is the sharpest one: availability you do not govern. A model running on hardware you own cannot be revoked by a directive you never saw. That is the whole pitch for local inference, and it just got handed the cleanest example it has ever had. The full timeline is here if you want to read it cold.
None of this means rip the hosted models out of your stack. They are still better than what most of us can run at home, and that gap is real. It means know which of your workflows would simply stop if a model went dark on a Friday, and have an answer that does not require anyone’s permission to switch on.
Question I Got Asked (and the Real Answer)
The question: “Be honest, is AI going to delete my job?”
The common answer: Some executive’s eighteen-month countdown, repeated back with the serial numbers filed off. Half the room nodding, the other half quietly updating their resume.
The real answer: The most-cited primary source on this is the WEF Future of Jobs Report, and the people repeating the doom number rarely seem to have read it. The report projects 170 million jobs created and 92 million displaced by 2030, for a net gain of roughly 78 million. That is real churn, about 22% of the jobs studied, and it is going to HURT in specific places; clerical and administrative roles, cashiers, and data entry are on the wrong side of it.
The more useful number in that same report is the skills outlook: the fastest-growing skill categories through 2030 are AI and big data, networks and cybersecurity, and technological literacy. Those are not the skills of a job being deleted. Those are the skills of the work that keeps the rest of it running.
The displacement is real and worth taking seriously. The version where it deletes the people who keep systems running is not what the data says. Again, Eric Siron and I spend a full episode of the podcast (above) on this distinction, and it is worth the listen, because the gap between the headline and the report is wide enough to drive a career decision through.
Community Signal
James Kilby - “Self-Hosting an AI Stack Using vSphere, Docker and NVIDIA GPU” - The practical version of everything above. Kilby walks through standing up a containerized inference stack on a single NVIDIA Tesla P4 inside vSphere, with the install commands and, the part I appreciated, the real power numbers: roughly 7 watts idle, 50 to 60 under active inference. If you read From the Console and wondered what “own your inference” actually costs to run, this is a concrete answer from someone who measured it. Updated June 2026.
Tool of the Week
Ollama - Open-source local LLM runtime. Pull a model, run it on hardware you own, expose an OpenAI-compatible endpoint any existing tool can talk to.
If you read Issue 24 you have already heard this pitch. The reason it earns the slot a second time is that the argument for it just got sharper. For issues 24, the case was privacy, cost, and not being told you cannot use AI at all. The Fable 5 takedown adds the axis I did not have a clean example for in May: availability you govern. Again…. a model running on hardware you own cannot be pulled by a directive you never saw.
Honest scope has not changed since Issue 24. Local model quality at consumer hardware tiers still trails the frontier hosted models, and for greenfield engineering work you will still reach for the hosted option more often than not. For the workflows that actually need to keep running no matter what happens upstream, this is the fallback The Take is asking you to have. The Quick Win below is “stand it up and prove it works.” An hour of your time, a permission slip from nobody.
Quick Win of the Week
Make a one-page list of every tool and workflow in your shop that calls a hosted AI model, and mark which ones would simply stop if that model went dark on a Friday night. For the one that matters most, pull an open-weight model in Ollama, point the tool at the local OpenAI-compatible endpoint, and confirm it actually runs. You are not switching off the cloud. You are proving you have a fallback that needs nobody’s permission to turn on, before the week you actually need it.
Fun Retro SysAdmin Fact
This is not the first time Washington export-controlled software. Through the 1990s, strong encryption was classified as a munition under the same export rules as weapons, and PGP author Phil Zimmermann spent roughly three years under federal criminal investigation for “exporting munitions without a license” after his code spread overseas; the case was dropped in 1996, and the idea that publishing software could be an arms-export crime has felt absurd ever since. Felt.
Until Next Week
Know which of your systems would keep running if the company upstream of you had a bad Friday.
Stay Frosty,
Andy
SysAdmin Weekly