With 96 GB+,
here is what really runs.
Experiences reported by people using their machines every day. The figure is the reported setup capacity; multi-GPU and unified-memory layouts can behave differently.
in this range
A HIPAA-style internal LLM platform, built on a cluster of DGX Sparks
Beyond personal daily use, this person stood up a Spark cluster for a client so their employees get an internal, compliance-friendly LLM platform.
A cardiologist's referral system, fully air-gapped
Referral triage that extracts details, standardizes them, and suggests categories and tests — with the final call always left to a doctor, and nothing ever leaving the network.
Construction estimating work moved onto a 7900XTX and two DGX Sparks
After six months of tinkering, Qwen 3.8 27B justified a 7900XTX plus two DGX Sparks, taking over a job the estimating team was spending dozens of hours a week on.
Running sales work on GLM 5.3 Flash across two DGX Sparks, 200M tokens a week
A sales professional dropped every frontier subscription and runs work agents on two DGX Sparks, averaging about 200 million tokens per week.
Internal company tools on a scavenged 1TB ECC server with mixed GPUs
A self-built server with 1TB of ECC RAM and a mix of used 3090s and P100s produced internal tools whose economic impact is reported at ten times the hardware cost.
Eight A100s in tensor-parallel, inside a company with a hundred more
A large enterprise deployment: eight A100 80GBs running Qwen 3.8 under SGLang with tensor parallelism, while the company brings newer B200s online for bigger models.
Four DGX Sparks running agentic work all day
A four-DGX-Spark cluster under vLLM serves several concurrent agentic coding and research sessions, with unquantized weights and honest notes on concurrency limits.
A cloud-free workstation: RTX Pro 6000, vLLM and a DeepSeek harness
A custom Linux workstation with an RTX Pro 6000 and 48GB of RAM runs programming, copy editing and summarisation through vLLM, with no cloud use all year.
A workstation GPU as a drop-in subagent for Codex and Claude
Qwen 3.8 Flash Next on an RTX Pro 6000 takes implementation and review tasks delegated from Codex or Claude, replacing Sonnet and Opus for that slice.
Uncensored Qwen 3.8 Flash Next tops a 19-task hacking benchmark at 80.6% on 2x RTX Pro 6000
On a 19-task cybersecurity benchmark, an uncensored Qwen 3.8 Flash Next solved 80.6% first try, but the run used someone else's 2x RTX Pro 6000 and only 2 runs per task.
A 4x RTX 5090 (128GB) office rig serving a 50-person company: local RAG, chat, diffusion and agentic coding
A custom 4x RTX 5090 build with 128GB of VRAM, an Intel Xeon w5-3423 and 192GB of ECC DDR5 runs DeepSeek V4 Flash, Qwen 3.8 and Gemma 4 32B locally for a 50-person company that wants its data off grid.
Three rigs for freelance work an NDA won't let leave the building
Client confidentiality rules out cloud APIs for this freelancer, who spreads different model sizes across three separate machines.
Building a two-server local AI ecosystem across 17 GPUs
A self-hosted AI ecosystem spans two Linux servers, assigning different GPUs to long-context coding, inference, speech, RAG and media workloads.
ISP NOC automation on two vLLM servers with 4× RTX 3090 each
A small ISP (~7,600 customers) serves its agentic operations work on two identical machines, each with 4× RTX 3090 (250W-capped) and vLLM v0.30.0; llama-benchy measures ~5993 tok/s prefill and ~199 tok/s generation.
Compare with another budget.
Every capacity guide, or start from your own machine and a task.