Your AI Data Center Is Not an Energy Story. It is a Server Depreciation Problem.

Your AI Data Center Is Not an Energy Story. It's a Server Depreciation Problem.

The most important number in AI infrastructure right now is not gigawatts. It's years.

A cost model published May 14 by Epoch AI — a research organization that tracks AI capability and economic data, whose senior researcher Ben Cottier is listed on Epoch's team page — quantifies something the industry has known qualitatively for years: servers are the dominant cost of owning a large AI data center, and their lifespan is the variable that moves the most money. Extending a server's useful life from three years to seven years cuts the annual cost of a one-gigawatt facility by $5 billion, swinging total annualized cost from $12 billion to $7 billion. Energy, by contrast, runs $0.6 billion per year at the same facility scale. The gap between hardware and power is not close: an 8-to-1 ratio.

The finding comes from GovAI research scholar Amelia Michael and Epoch senior researcher Ben Cottier, whose model — available as a public Google Sheets cost model — assumes 100% NVIDIA GB200 NVL72 systems — the current generation of rack-scale GPU hardware — at current list pricing. A one-gigawatt facility built to their specifications requires roughly $38 billion in upfront investment; servers account for $5 billion of the $8.5 billion annual total cost of ownership, or 60%, according to Epoch AI's data insights. Energy is the largest operating expense line but comes in at $0.6 billion. The facility, cooling, land, transformers, and grid connection are secondary to what gets bolted into the racks.

The Epoch model is a stylized estimate, not a verified facility filing. Bernstein Research's separate estimate for one gigawatt of AI data center capacity — $35 billion — is in the same ballpark, suggesting the numbers are directionally robust even if the exact figures carry model risk. Epoch publishes its full spreadsheet with sources and assumptions. Anyone can interrogate the inputs.

The specific new contribution is the depreciation curve. Epoch modeled three IT equipment lifespan scenarios: three years, five years (base case), and seven years. At three years, the annualized cost of a one-gigawatt facility runs $12 billion per year. At five years, it falls to $8.5 billion. At seven years, it reaches $7 billion. That $5 billion swing between worst and best case is the number that matters for capital allocation decisions. The engineering work that extends a server's useful life — better thermal management, predictive failure modeling, liquid cooling that doesn't degrade over time — is now the highest-return problem in AI infrastructure.

What this means competitively: hyperscalers are not primarily competing on power purchase agreements. They are competing on hardware procurement, refresh cadence, and the durability engineering that determines whether a GB200 rack delivers value for five years or three. NVIDIA's pricing leverage compounds when servers represent 60% of total cost rather than the 30% typical in conventional data centers. Whoever solves hardware longevity controls the single largest variable in AI economics.

This reframes the conversation around AI infrastructure investment. Recent coverage — including NVIDIA's $2.1 billion stake in IREN for power infrastructure — has focused on power as the scarce resource in AI buildouts. That framing is not wrong, but it is incomplete. The Epoch model shows that power is expensive but bounded; server depreciation is expensive and sensitive to engineering decisions that can be made faster than new power infrastructure can be permitted and built.

The hardware cost trajectory reinforces the depreciation story. Epoch separately published analysis showing that NVIDIA's Vera Rubin NVL72 racks — the successor to GB200 — will cost $6 million to $8.8 million per rack depending on configuration, per Tom's Hardware's pricing analysis. The next generation pushes capital intensity higher.

A secondary finding in this week's Epoch Brief covers Claude's benchmark profile: Anthropic's models consistently overperform on software engineering benchmarks relative to their general capability index and underperform on mathematics. The pattern is consistent across model generations, per Epoch AI's Claude ECI analysis. Recent Opus 4.6 and 4.7 show the math gap narrowing — within one point of their general ECI score.

Epoch AI, May 14, 2026.