I'm definitely leaning solidly towards clustering more ASUS GX10's, although I haven't set up a Strix Halo cluster yet to compare. I love my Strix Halo laptops for the price, portability, and low power consumption of each individual machine, but the GX10 is made to cluster, with the fast built-in ConnectX-7 network interface. The GX10s have also got legitimate CUDA cores, which are so much more performant during pre-processing, compared to all the other lower cost inference options. That's really important for speed on big agentic tasks which involve large contexts being sent to the LLM on each inference call. Also, some models require CUDA to run at all.
Honestly, time to first token would keep me from buying any Mac Ultra, even if Apple comes out with models that have more than 512GB shared RAM. Preprocessing is just so slow on Apple silicon.
For a long time, I had considered building a system with multiple RTX 6000s, using the max-Q version of that card, which tops out at only 300 watts, while maintaining most of that card's dramatically fast performance. The upper max, though, for a system in a normal home within the US, with a normal electric grid connection, tops out at 3-4 RTX 6000s, even with the lower power Max-Q edition.
So even though RTX 6000 performance is much faster than DGX Spark systems that have a GB10 GPU (such as the Asus GX10), RTX 6000 systems will realistically be limited to 384GB VRAM at most, which doesn't satisfy my desire to run trillion+ parameter models.
And the GX10 has been tested in that regard. In the video below, 8 GB10 units (all different brand units combined together) are configured to handle the biggest LLMs, and the entire cluster ran at about 1000 watts:
https://www.youtube.com/watch?v=uYepcMoqvKQ
You can comfortably expect ~112-118GB of net usable VRAM per machine in a configuration like the one in that video. That yields 896-944 GB of aggregate VRAM dedicated strictly to loading model weights, KV caches, and processing content. A 1.6 Trillion parameter model should run comfortably at FP4 on that configuration. And that whole system leverages the native speed of onboard NVIDIA ConnectX-7 SmartNICs to run standard cluster networking pipelines (such as Tensor Parallelism over RDMA). What a beast of a setup for the money.
In comparison, even if you wanted to run smaller models with RTX 6000s, the price is still just so high for what you get. A complete system would cost between $43,000 and $55,000, depending on vendor. Here's an expense breakdown by Gemini, for a system with only 384GB VRAM on RTX 600 Max-Q GPUs:
- GPUs (4x RTX 6000 Pro Max-Q): ~$34,000 – $40,000 (Retails between $8,500 and $10,000 per unit through authorized partners like Micro Center or Wiredzone).
- Processor (CPU): ~$3,500 – $5,000 (e.g., AMD Threadripper PRO 7975WX or 9000 equivalent for 128 PCIe Gen 5 lanes).
- Motherboard: ~$1,000 – $1,500 (Server-grade WRX90 or enterprise dual-socket EPYC board).
- RAM (256GB - 512GB ECC DDR5): ~$1,500 – $2,500 (Must be 8-channel ECC server memory to prevent AI model corruption).
- Power Supplies (PSUs): ~$800 – $1,200 (Requires a specialized chassis or dual 1600W/2000W digital titanium PSUs to handle the transient power spikes safely).
- Chassis & Cooling: ~$1,000 – $1,500 (High-airflow cases such as the Silverstone Alta D1 or Phanteks Enthoo Pro 2 Server Edition).
- Storage (High-speed NVMe RAID): ~$800 – $1,500 (e.g., 8TB to 16TB of PCIe 5.0 enterprise storage to feed data to the GPUs quickly).
Total DIY Cost: ~$43,000-$48,000
For turnkey pre-built workstations of warranted systems from professional enterprise builders like Exxact Workstations or builds modeled after tech deployments (like Andreessen Horowitz's personal AI workstation guide), you should budget more like $48,000-$55,000.
That just doesn't make sense for most personal home deployments which will involve only a few users and normal inference tasks.
I'll also repeat what I've said in many other posts. Capable models such as Qwen 3.6 are already available, to perform real software development work and agentic tasks on much smaller systems. I've built amazing software with a Q4 quant of the MOE version of Qwen 3.6 35a3b, on an old laptop which I bought used for $800, that has only 16GB VRAM (on the built in mobile RTX 3080). It's mind blowing how good that model is, compared to the likes of GPT-OSS:120b, which led the pack less than a year ago (released August 5, 2025). And clearly, performance and quality are going to continue to improve in small models.
So I'm holding off buying any more hardware right now. An 8 unit Asus GX10 cluster is my current 'AI prepper' pick, but with Deepseek-4-pro costing $.85 per million tokens, options like my $20/month zip file workflow with ChatGPT, and local models like Qwen 3.6 MTP already doing amazing things on small machines, I don't see any reason to invest more in hardware at the moment.
As more of my clients decide to purchase their own LLM servers for agentic tasks, especially in environments where HIPAA and other compliance requirements must be enforced, I'll recommend a single Asus GX10 with Qwen 3.6, Gemma 4, and Nemotron 3 Super models. That's fast and capable enough to handle most in-house agentic tasks at a small-medium sized business. A single GX10 is a small investment for unlimited local secure inference, which will quickly beat the cost of using a HIPAA compliant LLM API such as BastionGPT, and the upgrade path is clear and straightforward if more hardware capability is required at any point.