NVIDIA's AI Factory Economics at 60 Million Dollars Per Megawatt

NVIDIA has published analysis of the economics underlying AI factories, the massive data centres built specifically to run AI workloads. According to the company's latest guidance, each megawatt of AI factory capacity costs approximately 60 million dollars to build and deploy. This figure is crucial for understanding the capital intensity of AI infrastructure and how profit works in the business of providing compute resources to companies training and running large language models.

To put this number in perspective: AI factories operate at the scale of tens to hundreds of megawatts. A single large AI factory might be 100 megawatts, which means six billion dollars in infrastructure costs before a single GPU is purchased or a single model is trained. This doesn't include land acquisition, permitting, cooling infrastructure, power transmission upgrades, or the costs of the actual GPUs and other hardware. The 60 million dollars per megawatt figure is a rough average that encompasses these elements, but even so, the total capital requirement is staggering.

NVIDIA frames this in terms of three characteristics: productivity, durability, and fungibility. Productivity refers to the utilisation rate and throughput of the facility. An AI factory that sits idle is losing money. Operators need to keep the GPUs running, training models, or serving inference requests, or they waste capital. Durability means the infrastructure must last long enough to recoup its investment. Facilities typically need to remain operational and productive for at least 5-10 years to justify the upfront costs. Fungibility refers to the flexibility of the infrastructure to handle different workloads and purposes.

This framework reveals the core economics of AI infrastructure. The business model depends on stable, long-term demand for compute. Companies building these factories are betting that AI workloads will remain valuable enough to justify the enormous capital costs. This is why hyperscalers like Amazon, Google, and Microsoft are making these investments: they have both the capital and the certainty of demand (from their own services and customers).

The Cost Drivers and What They Mean

The 60 million dollars per megawatt breaks down into several components. Power infrastructure represents a huge portion. The electrical systems needed to deliver electricity to hundreds of megawatts of GPUs are expensive and require coordination with utilities and governments. Cooling is another major cost. Modern GPUs generate enormous heat, and removing that heat reliably without interruption costs money and requires sophisticated engineering.

The actual computing hardware (GPUs, CPUs, networking equipment) represents the largest share of costs, but NVIDIA's number doesn't isolate that because the company is interested in the total cost of deployment, not just hardware. A GPU might cost 30,000 dollars or more, and a large AI factory might contain tens of thousands of GPUs. Networking infrastructure to connect all those GPUs and allow data to flow at the required speeds is another significant expense.

Building costs, including the physical facility, power distribution, cooling systems, and redundancy infrastructure, make up the remainder. Most modern AI factories are designed with significant redundancy: if one cooling loop fails, others take over. If one power feed is disrupted, backup systems activate. This redundancy increases costs but is essential for meeting uptime requirements.

NVIDIA's emphasis on these three characteristics (productivity, durability, fungibility) is strategic. The company is essentially arguing that customers shouldn't just think about hardware costs, they should think about the total cost of ownership of AI infrastructure. This shifts the conversation from "How much does this GPU cost?" to "How much does it cost to operate a productive, reliable, flexible AI factory for 10 years?"

Implications for the AI Industry

These numbers have several important implications. First, they create a high barrier to entry. Starting a new AI infrastructure company requires billions in capital and expertise in real estate, power systems, cooling engineering, and supply chain management. This concentration of capital and expertise favours large companies that already have access to capital and experience managing large data centres.

Second, they make it clear why hyperscalers are investing so heavily in their own infrastructure. Outsourcing to third-party infrastructure providers is more expensive than building your own. A company like Google that needs 200 megawatts of AI compute will end up building 12 billion dollars in infrastructure, but they can amortise costs across their own AI services and rent excess capacity to customers.

Third, these economics explain why energy has become the bottleneck for AI scaling. You can design faster GPUs and build better algorithms, but if you can't secure enough power and build facilities fast enough to deploy them, you hit a ceiling. This is why tech companies are suddenly interested in power-generation technologies like geothermal energy, nuclear fusion research, and small modular reactors. The limiting factor on AI capability is now often electricity, not hardware.

Fourth, the numbers suggest that smaller AI companies will increasingly rely on leasing compute rather than building their own infrastructure. The capital requirements are simply too high for most companies to justify building their own data centres. This creates a market for infrastructure-as-a-service where companies like Lambda, Crusoe Energy, or NVIDIA's own cloud services provide compute resources.

The Future of AI Economics

NVIDIA's disclosure of these numbers seems calculated to reset expectations about the cost of AI infrastructure. When companies are spending 60 million dollars per megawatt and facing multi-gigawatt build-outs, the conversation shifts from venture-backed startups disrupting markets with superior software to trillion-dollar companies building infrastructure to support their own AI ambitions. It's a sobering reminder that despite the hype around AI capability, the actual business of running large AI systems is fundamentally about capital allocation and infrastructure engineering.