Reliable and stable, getting better every week
I think it is important to note that many reviews on the Grace Blackwell machines that are from late 2025 or early 2026 may not apply anymore. This runs the Nvidia Ubuntu operating system, and Nvidia has really been great about fixing a lot of problems the system had in running this very different architecture when it was first released.

This is my 2nd system. These systems are getting more and more pleasant to own and use every week – support for them is growing in the opensource world, and NVIDIA is pushing out docker containers that are ready made to do many things well if you choose that path- they offer playbooks with easy step by step instructions that make it seem more like you are getting great software custom made for these new architectures in addition to a great machine for the price you pay.

I was a bit disappointed that adding the $$$ cable and clustering 2 together didn’t really give me the one combined machine I’d hoped for, but I do like my current set up. (It worked fine,and let me run a huge model, but it just wasn’t optimal for my use to do so.) On this machine, I am only using it for inference. I am running VLLM on it, currently with Qwen 3.6 31B but it’s always available. I have ran larger models on it, but this seems to be the sweetspot for me with the agentic systems that connect to it. I hardly ever go past 65% system memory I have room to grow but system availability is WONDERFUL if I don’t consider reboots due to system upgrades.

Do expect FREQUENT UPDATES. Nvidia has been pushing updates almost daily – sometimes multiple times a day, and they all require reboot.

The NVIDIA sync program works great for this machine. I run mine headless – you can set it up headless too if you like. I did, and didn’t have a bit of trouble. I do suggest that you go ahead and make a repair boot-able USB disk – you can get that on the NVIDIA site. This is NOT regular Ubuntu.

I am disappointed by people posting negative reviews that buy these systems for gaming. They are NOT for gaming natively. These are AI machines. This will keep your inference in-house if that’s what you want – one is enough to power a small business with inference with a model that is good at tool call, and vision.

Whether or not you should get a 1tb or a 4tb one depends a bit on how you are going to use it. I have one with 4tb that I use for storing multiple models, but I got this one with 1tb because it is only ever going to run 1 model and vllm. 1tb is more than enough to store a few models and vllm. It depends on your use case. I’d get the 4 if I wanted to do more like generative work with diffusion models, comfy-UI/voice/ music. I do those things on my 4tb and run Ollama too, but if you are going to run multiple services like that, even though you have the HUGE combined memory pool, you still have to think about how you are going to allocate and time resource use.

The graphics look great running generative text or image to video. Video is smooth as can be.

It handles heat well – but inference is HOT. I highly recommend that you do NOT point the backside at you – it can be like having a small space heater right next to you. I learned that the hard way.I have mine on a stainless steel commercial kitchen table and that is VERY helpful if you have more than one of these running. It’s not as bad if they are not always available. Just remember, inference is HOT. If they get too hot, they will go into thermal throttling. I find they really do need a cool room, and good airflow for optimal operation.

There’s more than enough ports, but they are a bit tight. I do like the power button on the front. They use a regular Ethernet port, and have wireless capability built in.