Have the GB10 devices become the current "best value" for LLMs?
I want to buy some real hardware because I feel like I'm falling behind. 3090s are >$1000 on ebay, and building out the server would be very expensive with current memory and storage prices. Macs are backordered for the next 5 months. I have no idea on the status of AMD products or Intel, but I don't want to fight driver and compatibility issues on top of trying to get models and harnesses running.
Are the GB10 variants the best value if you want to buy now? Is it better to try to wait on the M5 releases in 2-4 months? That seems like forever in today's fast-moving environment.
9 comments
I was considering spark as a second device next to my 3090s but I have impression these devices are SLOWER not faster
I would not buy a GB10. It can fit big models and process prompts quite fast but token generation is slow for big models as memory bandwidth is low. On that last point the price vs performance is just wrong.
I would wait for M5 ultra. Or go for a multi gpu rig.
Under 2000 USD? For anything other than inference you really are going to want 1:1 DRAM to vRAM which in this economy is going to byte even with DDR4
I have access to a hardware disposal dump from a mnc, they dump all stuff other than GPUs (sometime they slip through), so for anything that is not soldered to the machine, I should be able to pull off, i already found two sets of DDR5-6000 (2x16GB), im trying to build from junkyard so gpu is the only thing im trying to buy
I think the GB10 has the best price/value for local LLMs at 2 nodes (2x $3,400 Asus GX10 1TB, $80 QSFP56 cable). Thanks to 200GbE, adding a second node nearly doubles speed and model size that will fit.
Two nodes run Intel/Qwen3.5-397B-A17B-int4-AutoRound at around 1,500t/s PP and 30t/s TG on vLLM, perfect for agentic work. They are also great at stable diffusion with ComfyUI (worker node for half of the images, and half of the upscaling), fine-tuning small models.
ebay is expensive, I got my 3090 for less than 800 bucks few months ago at local web marketplace. 24gb was a must for me but 4090 is over 2000 bucks, I'm not doing that.
B70 hahaha. How many TPS are you getting? The memory bandwidth on that card is a joke. $1k USD to fit a Q6 of Gemma 4 31B and you still have to quantize KV cache and it isnt even human read speed at 0 context.
That card is not good. Maybe it will be better months from now but it is too slow.
Built by @aellman
Tools
Directories
Models & Pricing
Endpoints
Rankings
- All Rankings
- All Benchmarks
- Best LLM for Coding
- Best LLM for Math
- Best LLM for Writing
- Best LLM for RAG
- Best Local LLM
- Best LLM for OpenClaw
- Best LLM for Cursor
- Best LLM for Windsurf
- Best LLM for Cline
- Best LLM for Aider
- Best LLM for GitHub Copilot
- Best LLM for Bolt
- Best LLM for Continue.dev
- MMLU-Pro
- GPQA
- LiveCodeBench
- Aider
- AIME
- MATH (Hard)
- Big-Bench Hard
2026 68 Ventures, LLC. All rights reserved.