Every quarter, I like to do a small update to give our readers a behind-the-scenes look at what is happening. Often, there is a big difference between what folks see publicly and the inner workings of STH, so I like to peel back the layers. The rise of agentic AI this quarter will have a significant impact on the industry, and so I wanted to spend time on that.
Previous Updates
If you want to check out how this series has evolved, here are the links to the previous ones:
- STH 2019: Q1 – Q2 – Q3 – Q4
- STH 2020: Q1 – Q2 – Q3 – Q4
- STH 2021: Q1 – Q2 – Q3 – Q4
- STH 2022: Q1 – Q2 – Q3 – Q4
- STH 2023: Q1 – Q2 – Q3 – Q4
- STH 2024: Q1 – Q2 – Q3 – Q4
- STH 2025: Q1 – Q2 – Q3 – Q4
It is amazing how time flies.
STH Q1 2026 Letter AI Got Scary Good
Something happened this quarter, and it was palpable. Around Christmas 2025, AI was getting very good, but in Q1 2026, it felt like it flipped. We now have agents handling basic setup tasks, coding scripts we need, and many more ambitious projects planned. Just to give you a quick preview, between OpenClaw and Claude Code, in February, we set up an 8x-node NVIDIA GB10 cluster, including RDMA networking on the MikroTik CRS804 DDQ switch, the 10GbE network, NAS storage mounts, and so forth. By the end, we were running vLLM in TP=8 mode with Kimi-K2.5 across all eight nodes. Sure, it was slow, but it was also remarkable. We did not even bother to reset a few of the nodes that had been used for other projects. Instead, we just let the agents figure it out and get it done. Meanwhile, OpenClaw has been building a monitoring solution that integrates per-PDU power consumption as well for the cluster. It is also optimizing performance and trying different tuning to make models faster.
That all may not seem exciting at first, but it is also a capability that was not really usable/ useful six months ago. If you take a derivative of this trendline, it is on a crazy slope.
One of the strangest parts is seeing the STH user base fragment. We have folks who work at very large companies saying they use Claude or another AI coding tool daily. We have folks building clusters and running local LLMs. There are also folks who think that AI is just slop. To be fair, there is a lot of AI slop, and it is just going to get worse. It is just that longer-duration tasks are happening without intervention. That GB10 cluster task is not a trivial one. It is not officially supported by NVIDIA (NVIDIA currently supports up to four nodes.) Being frank, a few folks have 8x NVIDIA GB10 clusters, but it is still a relatively niche setup, and there are differences among them. This is not just a chatbot regurgitating. This is sending an agent on a task just like we would send someone on our team (or realistically, I would think it is fun and I would do it.)
A great example of the evolution we saw this quarter. The GB10 video we did, using the Dell Pro Max with GB10 to Profit within 12 Months, which used n8n and gpt-oss-120b, went live in January, based on a great workflow in November/December 2025. By mid-February, we are using gpt-oss-120b very little, as Qwen3.5-122B tends to be a much better tool calling model.
You might be feeling OpenClaw’d out at this point. There is a lot of hype around it, without a doubt. If you have not tried it, there are very few things I could recommend that you do more than that. The reason it is hyped is that it is very good and very useful. It was a complete security nightmare, but there is also a lot of investment in making it better. Still, it will end up being a security nightmare almost always because it is good enough to start treating like an “agent”. The tasks we are using it for are not “research and tell me about XYZ”. Instead, they are things like, “Read Google’s TurboQuant paper, and implement it for the KV cache on vLLM so we can use it in our NVIDIA GB10 clusters. Make a plan that includes a test harness, and also plan to set up an NVIDIA GB10 node to test this on.” The prompts are a bit more than that, but the idea is there. It might take the agent hours of trial and error, going through test loops. Ultimately, it gets done. This is dispatching complex tasks, not just using it as a glorified chatbot.
Something that shifted in Q1 2026 is that, in the “homelab” space, I no longer find myself needing another Ubuntu or Proxmox VE host to run more VMs to learn on. Sure, we have a ton of systems. More storage is good, but prices have gone up a lot. Instead, even at (much) higher prices, the marginal utility we get from adding a new GPU server or an LPDDR5X unified memory server running another model is crazy in comparison. When we did the Minisforum MS-02 piece, I have to say I love that system. On the other hand, the Minisforum MS-S1 Max is having a huge impact since it has been running embedding models, Whisper AI, and more. Making another Docker container for another self-hosted service is great. Being able to just say “hey, I want to self-host this, figure it out, and get it running” is much more impactful.
Recently, I was at NVIDIA GTC, and I noticed that many of the analysts I have known for some time had been posting articles during the event. They looked strange. These are people I have known for a decade or more. I asked a few if they are using AI to write it, and the answers were some form of “yes, but I guided and edited it.” That is really scary to me.
Given how good AI is, especially in larger models, we are seeing more publishers just using AI-generated articles. I think that is the “AI slop.” As someone who has been doing publishing for many years, I get the appeal. A former Editor-in-Chief at Tom’s Hardware told me many years ago that doing hands-on reviews of enterprise products was silly and we should just focus on news because it is cheaper to produce. I did not listen. At the same time, the way I have been explaining it, and did so in our Substack that I will link below, is that there are three main modes in which AI will occur:
- Hyper-Scale AI to run frontier models fast
- Local AI to keep data local and use it for proprietary workflows
- On-device/ physical AI, where the compute must happen on the device
Several folks have also heard me talk about the fact that there are necessarily three different versions of AI that we need:
- AI agent to AI agent for fully automated and fast workflows
- Human-to-AI agent to keep humans in the loop. These will necessarily be slower workflows
- AI agents that do not communicate for tasks that will happen offline or largely offline
My thought is simply this. If we used AI to generate content, we are really asking you to read something one of our AI agents wrote. Flipping this around, I would not be happy if I were constantly served AI-generated content from a site. Realistically, I could just have my agent make the same. If I am going to spend the time to go somewhere and read, then I want it written by a human. If the content is so low-value that no one is going to take the time to write it, it does not make sense to spend my time reading it.
Does that mean we are not going to use AI? Absolutely not. Writing new scripts to automate lab functions and data parsing, fixing little bits on the site and in the back-end, setting up nodes, and doing some of the more repetitive back-end tasks, we will use AI for. Setting up the GB10 cluster is a great example of something I could have done, but it probably would have taken a day or a day and a half, maybe longer, depending on how many meetings would have interrupted me. Heck, while Sam takes most of our photos these days, every so often we need to remove a scratch, a giant bit of dust, or a strange reflection. You better believe we are using Adobe’s tools to do that since it saves so much time. The fact is that not using AI at this point just means you are closing your eyes to better tools. That is all the back-end, though.
I have told the team that if we are writing for our readers, I expect it to be written by humans. That is really based on my personal preference at this point. I want to read what Ryan and our team write. With that said, I get the appeal of doing it the other way, since if your content costs 1/1000th to produce, that will lure many to just make AI content/ slop. Within the next 6-18 months, I expect AI-generated content to be largely indistinguishable from human-generated content. At the same time, I want STH readers to know the plan is to still do it the old-fashioned way because a person had a thought and took up a keyboard.
As our long-time readers know, my standard is that STH should be a website I want to visit and read in an unmodified format. So that is what we are going to keep making. As a type, I also know that it is a leap of faith to assume that folks would prefer reading imperfect human-written content over AI-generated content. We shall see.
The Axautik Group Update and Substack
In Q1 2025, we rolled out our new analyst arm, the Axautik Group. This is the “Analyst” part of our Analyst-Influencer-Other model. We had quite a few popular pieces this quarter.
The Axautik Group will produce content designed more for the financial and executive communities. Some folks will be upset that AG content will be priced more for those communities. This is not a replacement for STH in any way. Instead, it will be designed to take much of what we learn by doing STH and turn it into formats we can monetize.
There is a question of whether this will be more of a subscription model or selling one-off reports. We started doing a Substack in August 2024. That Substack has been growing really well and is becoming an important revenue source in 2025. It is now running at a clip of over 100K views/ month this quarter, which is not bad.
In 2026, we are going to have a lot more AG Substack content, so get ready for that.
STH Labs: The Shorts Channel Update
As a 2024 project, we have the new STH Labs shorts channel here. That has been in a slow growth mode for the last few quarters.
In the next quarter, we are likely going to try a few different content formats on the STHLabs channel. A big constraint/ question we have with the main channel is what do we do with topics that are 5-10 minutes worth of video, not 15-20 minutes.
Subscribe to the STH’s Newsletter and YouTube
Did you know STH has a free weekly newsletter that comes out on Saturday with curated “Top 5” pieces from the week? We know you cannot visit every day, so we can deliver our picks for weekend reading directly to your inbox. Subscribing to the newsletter is easy. Here is the form.
We are not selling your e-mail addresses and MailChimp is managing everything at this point so you can subscribe and unsubscribe from the list as you want.
We are not promoting the newsletter via overlays and pop-ups. Those are very effective, but they are bad for readers. I do not like them, so as long as I have a say, we are not going to have newsletter signup overlays. I run STH as something I would want to visit daily, even if I did not work on it.
Finally, subscribe to our YouTube and check it out here. Since that is a big focus at this point.
Subscribing to STH’s newsletter helps you see my favorite pieces of each week and a preview of the next week. Subscribing to the STH YouTube channel also helps us demonstrate our reach beyond just the website. Sometimes, doing things like showing power consumption or fan noise is easier than writing about it. This has also been something important for STH over the past few months. We have to demonstrate reach, and simple things like subscriber counts help explain reach.
Final Words
Thank you to all of our readers, Substack subscribers, clients, the STH team, and our advertisers for making everything possible.
In Q2, travel is about to pick up significantly for me, with multiple international trips and attendance at several major vendor and industry trade shows. The interesting part will be managing the remote agents in the studio while I am on the road. To the cohort of our readers who think that AI is not going to light fires in many parts of our industry, Q1 2026 is when that ship has sailed. It is now moving into other industries, including publishing, such as STH, where it is set to ignite rapid change.
