INSIGHTS
Insights from the owner's side of the table
These are not trend pieces. Each article traces back to twenty years of building high-density infrastructure and more than fifty data-center deployments, plus the public record on how the AI buildout is actually being financed and built. Two threads run through everything published here: heat is the real constraint on AI capacity, and the fastest new capacity comes from power that already exists.
Heat Is the Product
Every watt that enters an AI data center leaves as heat, and the operators who design for that fact first are the ones whose economics work. Rack density has broken the air-cooling playbook, and liquid cooling turns the cooling budget from a facilities line item into the business model.
Read the analysisThe Smelter Play
The pace of AI capacity is set by grid interconnection queues, not by capital or hardware. Dead industrial sites with live power rights are the fastest route around the queue, and a $200M idle smelter becoming a $19B twenty-year campus lease shows the scale of the repricing.
Read the analysis
Data Center Cooling Economics: Liquid vs Air vs Immersion in 2026
Average rack density jumped 69% year-over-year to 27 kW in 2026, and the NVIDIA B200 Blackwell demands liquid cooling at 1,200W TDP. Jon Moen, who deployed liquid-cooled GPU clusters at MIT, Cornell, and Princeton as USA Technical Director at EKWB, breaks down the complete TCO picture — air, direct-to-chip, and immersion — so you stop letting the wrong cooling choice determine your AI ceiling.
adamsilvaconsulting.com · May 2026
Supermicro vs Dell: AI Server Head-to-Head for Enterprise Workloads
In my multi decades of client-server experience, Supermicro is cost-effective when amortized over time, employs great technical support (especially Level 2) and designs AI-mission-specific, highly technical, maximum dense GPU-slotted servers better than the competition — especially when deployed in a dense GPU AI server farm environment. The Supermicro SYS-522GA-NRT supports 13 slots with 10 GPU capacity versus the Dell PowerEdge XE9680 with 10 slots and only 8 for GPUs. Draw your own conclusions.
adamsilvaconsulting.com · April 2026
Liquid-Cooled AI: Why Air Cooling Is Technical Debt
Air-cooled GPU servers look fine on a spec sheet and thermal-throttle at hour six under real AI workloads. Jon Moen, who has deployed liquid-cooled H200 clusters at research universities and enterprise data centers, explains why liquid cooling is no longer optional — and how to right-size the hardware for training versus inference.
adamsilvaconsulting.com · March 2026
The GPU Server Buyer's Guide: H100 vs H200 vs B200 for AI Commerce Workloads
A practitioner's guide to choosing between NVIDIA H100, H200, and B200 GPUs for AI commerce infrastructure — with real thermal data, TCO calculations, and right-sizing recommendations from someone who has deployed these systems at MIT, Cornell, and CrowdStrike.
adamsilvaconsulting.com · March 2026
Read the analysis, or put it to work
The same thinking behind these articles runs every Vulcan engagement, from cooling economics to powered-site acquisition.