Lesson 11/12Advanced14 min read·3 diagrams

Economics & Financing

$/MW capex, opex breakdown, neocloud vs hyperscaler vs colo business models, depreciation cycles, and the eye-watering capex acceleration that's redrawn the entire infrastructure industry.

1 · Capex — what the money buys

Approximate Capex Breakdown — 100MW AI Data Center (~$1B)GPUs / accelerators50%Networking (IB/optics)12%Power infra (UPS, gen, gear)10%Cooling (CDU, chillers, towers)8%Building shell + civil7%Storage5%Software / licenses4%Other (cabling, security, etc.)4%GPUs dominate. At ~$30k–$40k per H100 and 30k+ GPUs in a 100MW cluster, accelerators alone exceed $500M.
Approximate breakdown for a 100 MW AI cluster (~$1B in IT). Silicon dominates; building shell is a smaller share than people expect.

A purpose-built 100 MW AI campus runs $2.5-5B turnkey including silicon. The shell-only cost is closer to $800M-$1.2B; the GPUs and networking add the rest.

2 · Opex — what it costs to run

Power
$3-7M/MW/yr
Wholesale ~$0.04-0.08/kWh × 8760h
Cooling fluid + maintenance
~$200k/MW/yr
DLC adds vs air
Staffing
$1-2M per 50 MW
24/7 NOC + DC eng + facilities
Spare parts + repairs
~3% of capex/yr
Mostly GPU replacements
Connectivity
$0.5-2M/MW/yr
Carrier circuits, peering

Power dominates opex. At $0.05/kWh wholesale, 1 MW continuous = $438k/year. A 100 MW campus = $44M/year just on electricity.

3 · The three business models

Hyperscaler (Microsoft, Google, AWS, Meta)

Build for own use. Capex is internal. Customers pay per-token (Azure OpenAI, Bedrock) or per-instance (EC2, Compute Engine). Highest margins, highest moat — they own the silicon, the infrastructure, and the customer relationship.

Neocloud (CoreWeave, Lambda, Crusoe, Nscale)

Buy GPUs (often financed with debt collateralized by the GPUs themselves), build small-to-medium DCs, rent compute by the hour or contract. CoreWeave's IPO in March 2025 raised $1.5B at a ~$23B valuation — proof the model works at scale.

Risk: GPU obsolescence. An H100 bought in 2023 at $35k is worth significantly less when B200 ships at higher perf/$.

Colocation (Equinix, Digital Realty, Iron Mountain)

Build the building, sell rack space + power + cooling + cross-connects. Don't own the IT. Steady cash flow, lower margins, very long-lived assets (30+ year useful life on the shell).

4 · The depreciation puzzle

How long does an H100 last in service? It's a real accounting question that affects $billions of corporate income.

  • Microsoft — extended useful life of server hardware from 4 to 6 years in 2022 (impacted depreciation by $3.7B).
  • Meta — extended to 5.5 years in 2024.
  • Amazon — 6 years for servers.
  • CoreWeave — uses 6 years on GPUs (heavily scrutinized by short sellers).

The longer the useful life, the lower the annual depreciation expense, the higher the reported earnings. The trade is real depreciation rate vs reported. Skeptics argue Hopper-class GPUs will be commercially obsolete in 3-4 years given Blackwell's 2-3× efficiency.

5 · The 2024–2026 capex explosion

Aggregate capex from the four US hyperscalers (Microsoft, Google, Meta, Amazon):

2022
~$140B
Pre-AI boom
2023
~$165B
ChatGPT year
2024
~$235B
GPT-4 / Llama-3 buildouts
2025E
~$320B
Stargate, Hyperion, Rainier
Cumulative '24-'27
$1T+
Conservative analyst estimates

Source: Hyperscaler 10-K filings; Synergy Research Group quarterly capex reports; IEA 2025 outlook; Bloomberg analyst consensus.

Lesson 11 — TL;DR

  • • 100 MW AI campus capex: $2.5-5B turnkey. GPUs are ~half.
  • • Opex dominated by power. ~$44M/yr per 100 MW at $0.05/kWh.
  • • Three business models: hyperscaler (own use), neocloud (GPUs-as-a-service), colo (real estate).
  • • Depreciation life is contested — every extension boosts reported earnings.
  • • 2024-2027 cumulative hyperscaler capex will exceed $1T. Unprecedented.

Useful? Share so the next engineer learns this faster.

Share: