NVIDIA AI Factory Cost Per Megawatt: $60 Million
A one-megawatt AI factory runs about $60 million. NVIDIA argues the payback rests on throughput per megawatt, hardware lifespan and reuse across workloads.
In short
NVIDIA puts the build cost of an AI factory at roughly $60 million per megawatt and traces the return on those GPU racks to three variables: earning capacity, useful life and demand.
At a glance
- Build cost: about $60 million per megawatt of AI factory capacity, per NVIDIA.
- Vera Rubin NVL72 claims over 30x the throughput per megawatt of GB300 NVL72, citing SemiAnalysis AgentX data.
- Up to 45x lower cost per million tokens on DeepSeek V4 Pro — a vendor comparison, not independently audited.
- Barkr useful life: five to six years for an eight-GPU H100 system, nine to 10 years for GB300 NVL72.
- A six-year-old A100 still carries roughly 25% of its original cost, according to Silicon Data.
NVIDIA puts the build cost of an AI factory at roughly $60 million per megawatt and traces the return on those GPU racks to three variables: earning capacity, useful life and demand.
The pitch is an accounting argument
Writing on the company blog on October 1, 2026, Shruti Koparkar frames the questions an operator asks before releasing capital at this scale. Earning capacity is the revenue a site would book if it sold every token it can produce. Useful life is how long the installed hardware keeps earning. Demand decides whether the first two matter at all.
At the stated price, a 100-megawatt site represents $6 billion in equipment and construction. That multiplication is ours, not NVIDIA's, and the post never splits the $60 million into silicon, shell and grid connection.
Throughput per megawatt
For the next generation NVIDIA offers specific multiples: Vera Rubin NVL72 is said to deliver more than 30x the throughput per megawatt of GB300 NVL72, with up to 45x lower cost per million tokens on the DeepSeek V4 Pro model. The post credits SemiAnalysis AgentX for the underlying data. Both figures are vendor-published, and we found no independent benchmark confirming them.
Depreciation is doing the heavy lifting
The durability claim leans on rental and resale evidence rather than spec sheets. NVIDIA notes that the A100, shipped in 2020, is still commercially active six years on, and that CoreWeave has extended bookings on that generation into 2029. Analyst firm Barkr puts useful life at five to six years for an eight-GPU H100 system and nine to 10 years for GB300 NVL72.
On residual value the post cites Silicon Data for a six-year-old A100 holding about 25% of its original cost, and Ornn Data for five-year A100 contracts still fetching 80% of the one-month rate. Software and kernel work is said to keep raising what installed silicon can do, but NVIDIA attaches no multiplier to that effect.
Fungible means one fleet, many jobs
The third pillar is workload flexibility: language, vision, biology, physics and robotics, across preprocessing, training and inference, plus non-AI work such as simulation, graphics and scientific computing. NVIDIA credits CUDA, more than 1,000 CUDA-X libraries and over 10 million developers for that portability. Older generations stay economically viable because not every job needs the newest system.
Customer examples carry the point: Lilly runs a 1,016-GPU on-premises cluster for protein and genomics models, Pinterest spreads 14,000 GPUs across Blackwell, Hopper and earlier architectures, and Texas A&M University reports 95% to 98% utilization across 26 projects. Unilever puts the saving on digital-twin product imagery at 50% against conventional photography.
What the post leaves out
This is marketing material with numbers attached, not an audited financial model. There are no token prices, no occupancy assumptions for a named site, and the analyst houses are referenced without links. Read the original NVIDIA analysis and treat the multiples as claims to be tested.
FAQ
How much does an AI factory cost per megawatt?
NVIDIA says roughly $60 million per megawatt. The post does not split that figure between GPUs, building shell and grid connection.
How long does an AI GPU stay economically useful?
The Barkr figures NVIDIA cites are five to six years for an eight-GPU H100 system and nine to 10 years for GB300 NVL72.
Is Vera Rubin NVL72 really 30x faster than GB300 NVL72?
NVIDIA claims over 30x throughput per megawatt using SemiAnalysis AgentX data. It is a vendor number with no independent verification available.
