---
title: "A Law Firm Put AI on a Gaming GPU. The Bigger Question Is Control."
description: "5 source-backed AI claims from A Law Firm Put AI on a Gaming GPU. The Bigger Question Is Control., with key statistics, context, and the original AI Adopters…"
url: "https://kbanc.com/claims-library/a-law-firm-put-ai-on-a-gaming-gpu-the-bigger-question-is-control"
source: "https://aiadopters.club/p/a-law-firm-put-ai-on-a-gaming-gpu"
date: "2026-09-24"
topics: ["strategy", "implementation", "business"]
generated: "2026-09-25"
---

# A Law Firm Put AI on a Gaming GPU. The Bigger Question Is Control.

By Kamil Banc | September 24, 2026

## Claims

1. **Sydney Firm's Local Setup** (source summary): An anonymized Sydney boutique law firm ran an 8-billion-parameter open-weight model on its own RTX 4090.
2. **Vendor-Reported Time Savings** (source summary): The vendor reports email catch-up time fell from about two hours to 25 minutes.
3. **Local Does Not Mean Independent** (Kamil's interpretation): Local AI removes one outside AI processor from a workflow without making the information chain independent.
4. **Client AI Handling Tiers** (source summary): Madgett Law's engagement letter offers clients no AI, in-house AI only, or hybrid choices.
5. **North Dakota Bill Summaries** (source summary): North Dakota's Legislative Council used Llama 3.2 1B Instruct and estimated 15-25% time savings.

## Evidence

### Quote
> "The question is whether you're choosing it, or whether your business has quietly lost every other workable option." - Kamil Banc

### Key Statistics
- **2 hours to 25 minutes**: Vendor-reported reduction in a partner's post-leave email catch-up time at an unnamed 38-person Sydney law firm
- **210 emails per hour**: Vendor-reported summarization throughput of the RTX 4090-based system pulling from Google Workspace
- **15-25% time savings**: Meta's own estimate for the North Dakota Legislative Council's Llama-based bill-summary pilot during the 2025 session
- **25 legal hours per session**: Meta's estimated staff time saved across a legislative session; a plan-based figure, not an observed result

## Context
The article triangulates three cases rather than proving one: a vendor-reported Nexus deployment at a Sydney boutique firm, Meta's account of North Dakota's locally fine-tuned bill-summary pilot, and Madgett Law's client-facing AI handling tiers. The author deliberately flags missing evidence throughout, noting that no case includes independent measurement, security audits, cost disclosure, or error rates. His methodology is practitioner-oriented: test a draft-producing task on 20 real, permission-cleared examples against human, cloud, and local paths with full cost accounting, including checking time and maintenance. The core insight is that intelligence independence rests on four ownership questions covering data visibility, model versioning, permitted actions, and exit paths. None of the cases supports running every job locally; they demonstrate three distinct forms of control a business can deliberately choose.

## Source
- Original: [A Law Firm Put AI on a Gaming GPU. The Bigger Question Is Control.](https://aiadopters.club/p/a-law-firm-put-ai-on-a-gaming-gpu)
- Cite: kbanc.com/claims-library/a-law-firm-put-ai-on-a-gaming-gpu-the-bigger-question-is-control

## Primary Evidence
- [anonymized case study from Nexus Technology Consulting](https://nexustechconsulting.com.au/case-studies/local-llm-email-digest.html) (nexustechconsulting.com.au; supports claims 1, 2)
- [Madgett Law describes a different use of local AI](https://www.madgettlaw.com/practice-lab/run-it-yourself-then-let-the-client-choose/) (madgettlaw.com; supports claim 4)
- [Meta’s account of the Legislative Council](https://dev.meta.ai/llama/resources/case-studies/north-dakota-legislative) (dev.meta.ai; supports claim 5)