Analysis and engineering notes on AI data center infrastructure — power, cooling, racks and the build-out of compute capacity worldwide.
A line-by-line cost model for a behind-the-meter gas plant in Canada: fuel, TIER carbon, O&M, capital recovery and the provincial levy, in CAD, plus the four feasibility gates and the risk register that decides whether the project survives.
Read article →Above 1,000 m air density falls fast: servers lose 7 C of inlet allowance by Mexico City, turbo gensets give up 12.5%, UPS 6%, switchgear needs a 1.28x BIL correction - and open tower pumps lose 3 m of NPSH. Full derating stack with worked numbers for a 2 M...
Read article →RTX 5090 racks at 575 W per card: air cooling tops out near 15-20 kW/rack at PUE ~1.55, warm-water direct-to-chip liquid runs 40 kW+ at PUE 1.15-1.20. Full 1 MW pod math: ~3.06 GWh and ~$306k saved per year.
Read article →Container data center shells live or die by their paint. The application process - spray parameters on corrugated panels, stripe coats, recoat windows - sets the starting quality; the washdown, inspection grading (ASTM D610/D714), and spot-repair rhythm dec...
Read article →Field welding requirements for outdoor cooling tower and condenser-water piping: ASME B31.9, ASME Section IX qualifications, fit-up tolerances, galvanized coating removal and A780 repair, cold-weather preheat, spot RT and the 1.5x hydro test.
Read article →A reproducible two-year ROI model for US GPU compute rental: the H100/B200 rate ladder, $0.49 per GPU-hour of real operating cost, utilization sensitivity, and the residual-value debate that decides the return.
Read article →Stage-by-stage working principle of the solid-state transformer - cascaded H-bridge rectifier, dual active bridge isolation, direct 800 VDC output - with the Faraday arithmetic, chain efficiency, capex and the honest open problems.
Read article →Why 54 V racks hit a wall at 200 kW, what the 800 VDC chain changes between grid and GPU, and the arithmetic behind the 5% efficiency, 45% copper, and 70% maintenance claims - plus the four problems it leaves unsolved.
Read article →ISO Class 8 particulate limits, ANSI/ISA-71.04 G1 gaseous limits, and the five ways contamination kills servers: heat sink fouling, zinc whiskers, creep corrosion, fibre end faces, and filter bypass. With computed filtration and energy numbers.
Read article →ANSI C84.1, the ITIC curve, and IEEE 519 define clean power on three time scales. Harmonic spectra, K-factor arithmetic, neutral sizing, and the mitigation toolbox for AI-era halls.
Read article →ATS vs STS in data centers: where each transfer switch belongs in the power chain, why an STS needs two live sources, what the 4 ms transfer means against the ITIC window, and the SCR loss the ATS never pays.
Read article →What to run on a GPU server before it carries production work: DCGM diagnostic levels, thermal soak, XID and ECC acceptance limits, NCCL bus bandwidth targets, and the exit criteria that define a pass.
Read article →