All posts·[ Business ]

What a 5 MW AI compute hall actually costs

Capex, opex, PPA, lease — we break down the real numbers for a 5 MW AI compute hall running NVIDIA H100s on z1power battery-backed power.

YJ
Yasir Jahangir
Co-founder & COO
Mar 24, 2026
8 min read

When customers ask what a 5 MW AI compute hall costs, they usually want a single number. The honest answer is: it depends on structure, term, and how much risk you're taking on the power side. Here are the real ranges.

The capital stack

A 5 MW AI compute hall running NVIDIA H100s has roughly these capital components:

  • GPU compute: 200-300 H100s at current pricing. ~$8-12M.
  • Networking: InfiniBand NDR fabric, switches, cabling. ~$1-2M.
  • Server chassis, storage, racks. ~$1-2M.
  • Cooling: liquid cooling loop, CDUs, facility water. ~$2-4M.
  • Power: z1power 5 MW BESS, switchgear, transformers. ~$4-6M.
  • Building shell, generators (backup-only), security. ~$3-5M.

Total capex: roughly $20-30M for a turnkey 5 MW hall. That's about $4-6k per GPU at the H100 price point.

Operating cost

Operating cost is dominated by power and depreciation:

  • Power: ~$2-4M per year at average US industrial rates, with battery arbitrage.
  • Maintenance + AURA + insurance: ~$1M per year.
  • Depreciation on $25M of capex over 5 years: ~$5M per year.
  • Total opex: ~$8-12M per year.

Per-GPU economics

All-in cost per H100 per hour: roughly $1.50-2.50, depending on utilization and financing structure. That's competitive with major hyperscalers, and you get the grid-independence story.

Want a custom TCO model?

Email tco@smarttec.dev with your workload profile (training vs inference, utilization, term) and we'll send a detailed model within a week.