There's a lot of expensive GPU hardware sitting idle between workloads. If providers can make that capacity easier to access without sacrificing reliability, it could open up a much larger market than traditional GPU rentals alone.

We have an EPYC build coming for 4x servers with each having 4x RTX Pro 6000 SE each and the AMD Epyc 9965 and 768GB Memory!!!

DHS has banned DeepSeek from federal civilian government systems, including third-party implementations. This raises a bigger concern for local AI adoption. I'm really curious to see what path the government takes from here.

Think RAM prices are bad now? The memory shortage could get WORSE in 2027 and potentially last until 2029–2030. This is getting even uglier for PC builders.

I'm curious to know what percent of the general population knows about local AI versus the cloud (Claude/ChatGPT). We do live in a bubble here on X.

The build with the 4x RTX Pro 6000 MaxQ water-cooled and the AMD 9985X has been running for about 10 hours now with 100% load on the CPU and GPU, and these are the temps we're looking at while the system is near whisper quiet. The goal is not to run as cool as possible, but also to make it comfortable to live with by your desk. It's a pretty hot day here in California, and the warehouse ambient temperature is 88°F, so that has some effect as well.

All 4 of the RTX Pro 6000 MaxQ are fully working, and integrated into the water cooling loop with the AMD 9985WX. Installing Linux now to begin our testing

The 4x RTX Pro 6000 MaxQ water-cooled build is coming together, almost ready to go into burn-in testing!

All 4 of the RTX Pro 6000 MaxQ have had their water blocks installed and are ready to go into the system.

Who's looking for some RTX 5090? 😎

All 4 of the RTX Pro 6000 MaxQ cards that we are currently water cooling will be going into this build with a fully water-cooled loop for the CPU as well.

RTX Pro 6000 Max Q fully disassembled.

First RTX Pro 6000 MaxQ water block done. 3 more to go.

RTX Pro 6000 Max Q watercooled build incoming.

Even if individual AI models become cheaper to use, the number of things people want AI to do keeps growing.

NVIDIA and AMD are taking different approaches to AI infrastructure, but both are moving toward open hardware standards through OCP. NVIDIA is contributing its Vera Rubin rack and compute tray designs, focusing on cooling, power delivery, and easier deployment. AMD's Helion targets agentic AI with CPU-heavy and hybrid CPU+GPU rack configurations, including designs with up to 32,768 EPYC cores. As AI infrastructure scales, the focus is expanding beyond GPU performance to how entire systems are designed, built, and maintained.

“Cheaper per token” does not automatically mean “cheaper to serve.” The model that wins is the one that completes your workload efficiently.

A cheaper AI model doesn't always mean lower costs.

“Software is dead” is a pretty bold statement, but with how quickly AI coding agents are improving, I can see why people are starting to make that argument.

Software isn't dead. The economics of creating it are changing dramatically.