Offline LLM Device Analysis

A comprehensive hardware comparison for building offline, local-AI appliances.

This repo contains interactive HTML analysis documents comparing edge AI platforms for running large language models (LLMs) entirely offline — no cloud, no internet, no subscriptions. Originally developed for the Nano Boot product concept: a plug-and-play AI appliance for privacy-conscious consumers and off-grid users.

---

The Question

What's the most capital-efficient hardware platform for running 4B–27B parameter LLMs offline, on solar power, in an appliance form factor?

Platforms Compared

Platform	Price	AI TOPS	RAM	Power	Max Model
NVIDIA Jetson Orin Nano Super	$249	67	8 GB	7–25W	~8B
Raspberry Pi 5 + AI HAT+ 2	$225	40	16 GB	15W	~7B
Apple MacBook Neo	$599	35	8 GB	15–30W	~8B
NVIDIA Jetson Orin NX Super 16GB	$599	157	16 GB	10–40W	~13B
Beelink GTi15 Ultra (Core Ultra 9)	$645	99	32 GB	105W	~20B
NVIDIA Jetson AGX Orin 32GB	$999	200	32 GB	15–40W	~27B
NVIDIA Jetson AGX Orin 64GB	$1,999	275	64 GB	15–60W	~70B

Analysis Documents

All analyses are self-contained HTML files with no dependencies — just open in a browser.

analysis/01-jetson-vs-macbook-neo.html

Jetson Orin Nano Super vs Apple MacBook Neo — Head-to-head across 14 categories including AI throughput, cost efficiency, off-grid viability, and appliance fit. The Jetson wins 9/14 for an appliance use case; the Neo wins on brand trust and UX familiarity.

analysis/02-orin-nano-vs-agx-orin.html

Orin Nano Super vs AGX Orin 64GB — The full-stack comparison: raw performance, LLM inference speed, form factor, power, and off-grid viability across 18 categories. Includes the "8x cost / 4x performance" finding that drives the capital efficiency analysis.

analysis/03-cost-per-analysis.html

Cost-Per Deep Dive — Dollar-for-dollar breakdown: cost per TOPS, cost per GB RAM, cost per token/second, 3-year TCO, and annual power costs. Visualized with efficiency gauges and break-even scenarios. Key finding: the AGX delivers 0.5x capital efficiency vs the Nano.

analysis/04-capital-efficiency-shootout.html

6-Way Capital Efficiency Shootout — The comprehensive comparison across all platforms including the Raspberry Pi 5 + AI HAT+ 2. Includes an SVG scatter plot of the price-vs-capability efficiency frontier, marginal cost analysis, community/ecosystem scoring, and a recommendation for a four-tier product line.

landing-page/index.html

Consumer Landing Page — A fully designed, responsive product page for the Nano Boot lineup (Nano Boot Kit / Nano Boot / Nano Boot Pro / Nano Boot Max). Dark theme, mobile-ready, all copy written. Image placeholders with detailed art direction in the companion spec.

landing-page/LANDING-PAGE-SPEC.md

Build Spec & Image Manifest — Complete handoff document for building the final landing page. Includes 19 image assets with exact dimensions, art direction, and ready-to-paste AI image generation prompts for Midjourney/DALL-E/Flux.

Key Findings

1. Best overall value for a single-user offline AI appliance: NVIDIA Jetson Orin Nano Super ($249, 67 TOPS, $3.72/TOPS)

2. Best budget/DIY option: Raspberry Pi 5 + AI HAT+ 2 ($225, 40 TOPS, 15W — lowest cost and power, largest community, but immature LLM stack)

3. Best capital-efficient upgrade to 13B models: Jetson Orin NX Super 16GB ($599, 157 TOPS, $3.82/TOPS — nearly identical efficiency to the Nano with 2x RAM)

4. Best for community/multi-user serving: Jetson AGX Orin 32GB ($999, 200 TOPS, 32 GB — handles 27B models, serves 6-10 users)

5. Skip the AGX Orin 64GB unless you specifically need 70B models — it's the worst capital efficiency in the lineup at $7.27/TOPS

6. The MacBook Neo is not an appliance — it's a laptop. Great consumer device, wrong form factor and locked-down OS for this use case.

Suggested Product Line

Tier	Hardware	Retail	Model Size	Audience
Nano Boot Kit	Pi 5 + AI HAT+ 2	$299	4B–7B	Makers, DIY preppers
Nano Boot	Orin Nano Super	$449	4B–8B	Individuals, families
Nano Boot Pro	Orin NX Super 16GB	$899	8B–13B	Power users, homesteads
Nano Boot Max	AGX Orin 32GB	$1,799	13B–27B	Communities, clinics

How to Use

# Just open any HTML file in your browser
open analysis/01-jetson-vs-macbook-neo.html

# Or serve locally
python3 -m http.server 8080
# Then visit http://localhost:8080

All HTML files are self-contained (inline CSS, inline SVG, Google Fonts via CDN). No build step, no dependencies, no JavaScript frameworks.

Tech Stack for the Analysis

• Pure HTML/CSS with JetBrains Mono + Inter fonts
• Inline SVG for scatter plots and data visualization
• Responsive design (works on mobile)
• Dark theme throughout

Data Sources

• NVIDIA Jetson Developer Documentation
• NVIDIA JetPack 6.2 Super Mode
• Apple MacBook Neo Specs
• Raspberry Pi AI HAT+ 2
• Jetson AI Lab Benchmarks
• Beelink GTi15 Ultra
• MediaTek Genio Pro 5100

License

MIT

---

Built with Claude. Analysis current as of April 2026.