GPUs
What Is an AI Server? Architecture, Parts, and Power Explained
A normal server rack draws about 9 kW. A GB300 NVL72 draws 142 kW, weighs 1,580 kg, and costs about $4 million. Here is what is inside an AI server and how the pieces fit.
ModemGuides BlogGuides for people who would rather own their hardware than rent it.
AI Infrastructure
GPUs, VRAM, mini PCs, and the software stack for running models on hardware you own.
35 articles
Latest from AI Infrastructure
GPUs
Local AI Server Build: Hardware, Budgets, and Used GPU Picks
Memory prices tripled and used GPUs followed. Here is what a local AI server needs, three builds that still make sense, and how to run one without exposing your network.
Reality check
Can You Run Muse Spark Open Locally? Yes on One Mac Studio, Not Through Ollama
Meta's 816B open model runs on one Mac Studio from 176GB. Ollama's version is cloud-only, the license has strings, and smaller models score higher.
Start here
Read these before you buy hardware for local models.
What hardware runs local models?
Local AI Hardware in 2026: GPUs, VRAM, and Honest Build Tiers for the Memory-Shortage Era
Best Local AI Models by VRAM and Memory: 8GB to 512GB (2026)
Best Mini PC for Local AI in 2026: Hardware Tiers From $80 to $3,650
Is the Ryzen AI Max+ 395 worth it?
The Ryzen AI Max+ 395 for Local LLMs: An Honest Reality Check
More from AI Infrastructure
Page 1 of 2
Reality check
Can You Run Mistral Large 4 Locally? Only on 512GB, and Not Until the Weights Ship
Mistral Large 4's weights land at the end of October. At an estimated 400GB for 2-bit, one 512GB Mac Studio clears the bar. Here is the full math.
Reality check
Can You Run Aleph Alpha's Kolibri-1 Locally? Yes, on 48GB — Not in Ollama Yet
Kolibri-1 is a 78B model that thinks like a 3.5B one. Here is what runs it at home today, how fast, and why Ollama and LM Studio can't load it yet.
Comparison
Mac mini M6 vs M5 Pro for Local AI: Which One Runs Your Model?
Apple's 2026 Mac mini refresh puts the newest M6 chip in the cheaper model and the M5 Pro in the expensive one, which makes the choice less obvious than it looks. For local AI the answer comes down to two numbers: how...
Can a Mini PC Run a 70B Model? Strix Halo and Unified Memory Explained
A 70B model needs about 40GB of memory, more than any consumer graphics card has. AMD's Strix Halo mini PCs solve that with up to 128GB of unified memory. Here is what they run, how fast, and why a big MoE model...
Reality check
Can You Run Qwen 4 27B Locally? What the Announcement Changes for Home Hardware
Qwen 4 27B was named on stage September 22, not shipped. The Qwen3.8-27B hardware baseline, what the new architecture could change, and what to wait for.
VRAM
Can I Run DeepSeek V4.1-Flash Locally? Size, VRAM and Hardware Requirements (1017 Update)
1017 is a 510GB download that needs 150 to 200GB in memory, with Engram on SSD. Two DGX Sparks or a 256GB Mac run it; one 128GB Spark does, at a cost.
Buying guide
Best Local AI Models by VRAM and Memory: 8GB to 512GB (2026)
Find your memory number and get the verified pick for it. Six tiers from an 8GB card to a 512GB Mac Studio, sizes sourced, licenses checked, October 2026.
Reality check
Can You Run GLM-5.3-Flash Locally? Real GGUF Sizes and Speeds (Oct 2026)
Mainline llama.cpp now runs GLM-5.3-Flash. Real GGUF sizes, measured 128GB speeds, and why Ollama's version is cloud-only.
VRAM
Run DeepSeek V4-Flash Locally: Hardware Requirements 2026
The 0731 weights are out and the tier map is current: 103GB build on mainline llama.cpp, LM Studio gone local, and the M5 Mac Studio math.
How-to
How to Turn Off Amazon's AI: Disable Alexa for Shopping (Formerly Rufus)
Amazon replaced Rufus with Alexa for Shopping and still offers no off switch. Every working method to remove the AI on desktop and mobile.
Buying guide
Local AI Hardware in 2026: GPUs, VRAM, and Honest Build Tiers for the Memory-Shortage Era
Which GPU and how much VRAM for local AI in 2026 — honest build tiers repriced for the memory shortage, from used RTX 3090 to Apple Silicon.
Buying guide
The Mac Studio M5 Ultra Is Official: What 512GB Actually Runs, and Who Should Buy One
Apple's Mac Studio M5 Ultra is official: 512GB max, not 768GB. Which local AI models fit, the cluster math, and whether to buy, wait, or skip.
Memory prices
Will GPU Prices Go Up in 2026? What Nvidia's Rubin Ramp Means for Your Next Build
Nvidia's Rubin racks are shipping at scale, each pooling 20.7TB of HBM4. The honest buy-now-or-wait call for GPUs, RAM, and SSDs in late 2026.
Reality check
Has Open-Source AI Actually Caught Up? The 2026 Reality Check
Open models now sit in the frontier top five, but half the viral scoreboard is unshipped. The verified ledger, the price math, and what runs at home.
Memory prices
The AI Compute Wall: Power, Chips, and Why Owning Your AI Stack Is the Hedge
The viral RAND chart, explained without the geopolitics: why AI's power and chip limits are already raising your costs, and the local-first hedge.
How-to
Self-Verifying AI Agent Loops: What's Real, What's Hype, and How to Run One
A self-verifying AI agent loop checks its own work against live sources. Here is the real technique behind the viral Kimi swarm, minus the hype.
Reality check
The Ryzen AI Max+ 395 for Local LLMs: An Honest Reality Check
The honest tokens-per-second, the $1,499-vs-$2,200 SKU trap, the Linux memory unlock, and whether a Strix Halo box can replace your Claude Code bill.
Buying guide
When Will RAM Prices Go Down? The 2026 Shortage, Explained
RAM prices tripled, Apple pulled high-memory Macs, and relief isn't coming in 2026. Here's the shortage explained — and how to build local AI anyway.
Buying guide
Best Open-Source LLMs of 2026: Ranked by Capability, License, and Hardware
Twelve open-weight models worth running in September 2026, ranked by what they can do and what their licenses let you do with it. Qwen3.8-27B leads the single-GPU class, GLM-5.3 and Qwen3.8-2.4T lead the field, and the gap to the closed frontier is...
Security
Digital Security Checklist 2026: Personal + Network
AI models now find zero-days in every major OS. Karpathy's 15-step guide plus home network hardening to protect your digital life in 2026.
MCP
MCP Security Guide: Protect Your AI Tools from Attack (2026)
MCP servers connect your AI tools to files, APIs, and system commands — but 43% have at least one vulnerability. Learn how to audit, harden, and secure your MCP setup before attackers do.