How NVIDIA AI Factories Maximize Return on Investment
AI factories are being built at an unprecedented scale, with some facilities requiring hundreds of megawatts and the largest reaching into the gigawatt range. At roughly $60 million per megawatt, the capital required to build and operate these facilities is substantial. For operators making investments of this magnitude, understanding the potential return is critical.
Three factors largely determine the economics of an AI factory:
- Earning capacity — how much revenue a factory could generate if it sold all the tokens its infrastructure can produce.
- Useful life — how long the underlying AI hardware can continue generating value.
- Demand — how much market demand exists for the computing capacity and tokens the factory produces.
These factors are closely connected. A factory with high theoretical capacity delivers limited returns if much of that capacity sits idle. Likewise, strong demand cannot compensate if hardware becomes economically unproductive after a short period. And a system with a broad range of capabilities can access more workloads, helping maintain utilization as market requirements change.
NVIDIA’s approach is designed around maximizing all three dimensions. The company’s AI factory platform is engineered to be productive, durable and fungible — characteristics intended to improve utilization, extend infrastructure life and broaden the range of workloads a system can support.
Productive: More performance from every megawatt
Power is one of the most important constraints in modern AI infrastructure. As a result, the amount of AI work a facility can perform within its available power envelope has a direct impact on its earning potential.
For AI factories, one important measure is tokens per second per megawatt. More tokens generated within the same power envelope can increase revenue potential, while a lower cost per token can improve margins.
According to SemiAnalysis AgentX data cited by NVIDIA, NVIDIA Vera Rubin NVL72 systems deliver more than 30 times the throughput per megawatt of NVIDIA GB300 NVL72 systems, while delivering up to 45 times lower cost per million tokens on the DeepSeek V4 Pro model.
Such gains are the result of optimization across the entire technology stack rather than improvements in a single component.
NVIDIA’s approach combines advances in GPUs, networking, memory, software and AI models. This full-stack codesign allows each layer to work together, helping maximize overall system performance and efficiency.
Greater efficiency can also create additional demand. When the cost of generating AI output falls, applications that were previously too expensive to run can become economically viable. As more use cases become practical, overall demand for compute can increase rather than decline.
Durable: Keeping installed infrastructure productive
Rapid advances in AI hardware raise another question: what happens to existing systems when a newer generation arrives?
For NVIDIA, the answer is that newer hardware does not necessarily make previous generations obsolete. Different AI workloads have different performance requirements, and many applications can continue to run economically on earlier systems.
The NVIDIA A100, for example, launched in 2020 and continues to be used commercially years after its introduction. CoreWeave has also extended bookings for systems first introduced in 2020 through 2029.
Industry data suggests that the productive life of AI infrastructure can extend well beyond traditional depreciation assumptions. A September 2026 analysis from Sprout, The Productive Life of a Data Center GPU, tracks how major operators have extended server-life assumptions over time.
Other market analyses point to continued resale and rental value for previous-generation NVIDIA GPUs. Barkr estimates a useful life of approximately five to six years for an eight-GPU H100 system and nine to 10 years for a GB300 NVL72 based on resale values.
The continued value of installed hardware is supported in part by software compatibility. NVIDIA’s CUDA platform spans multiple GPU generations, allowing operators to continue using existing infrastructure as new architectures are introduced.
Software optimization can also improve the performance of deployed hardware over time. New libraries, kernels and software updates can help existing systems handle workloads more efficiently, potentially extending their economic usefulness.
Fungible: One platform, many workloads
Productivity and durability become even more valuable when infrastructure can support a broad range of applications.
A factory designed for a single workload makes a long-term bet that demand for that workload will remain strong. A more flexible system can shift between applications as market requirements change.
NVIDIA AI factories are designed to support different AI models and workloads across language, vision, biology, physics and robotics. They can be used throughout the AI lifecycle, from data processing and pretraining to post-training, inference and agentic AI.
They can also be deployed across different environments, including hyperscale data centers, AI clouds, enterprise infrastructure, sovereign AI programs and edge locations.
The same infrastructure can support workloads beyond AI, including scientific computing, simulation, graphics and other forms of accelerated computing.
At the foundation is CUDA, NVIDIA’s programmable computing platform. Its ecosystem includes more than 1,000 CUDA-X libraries, supporting applications ranging from deep learning and vector search to computational lithography, quantum simulation and climate modeling.
This combination of programmability and specialized hardware allows NVIDIA GPUs to support a wide range of workloads without being limited to a single application.
Tensor Cores and the Transformer Engine provide AI-specific acceleration, while the broader programmable architecture retains the flexibility required for other computational workloads.
The result is infrastructure that can potentially maintain higher utilization by moving between different types of work.
Flexibility in the real world
The breadth of NVIDIA’s platform is reflected in deployments across industries.
Lilly, for example, is using a 1,016-GPU on-premises cluster for protein, small-molecule and genomics models, while also supporting chatbots and agentic workflows.
Pinterest has used NVIDIA infrastructure to post-train and deploy a vision-language model across a hyperscale cloud environment involving thousands of GPUs spanning multiple NVIDIA architectures.
Revolut uses NVIDIA cuDF to process billions of transaction records before training and deploying a foundation model through an AI cloud.
Runway has trained a world model on NVIDIA Hopper systems and serves it on NVIDIA Blackwell infrastructure through cloud resources.
At Texas A&M University, NVIDIA-powered infrastructure supports molecular simulation and AI-driven drug discovery, with reported utilization of 95% to 98% across 26 projects and seven institutions.
The flexibility extends beyond AI as well.
Cosm uses NVIDIA-powered infrastructure for high-resolution video playback, streaming, real-time graphics and synchronized displays. Dassault Systèmes uses NVIDIA technology for virtual-twin simulations supporting applications including aircraft certification and vehicle development. Unilever uses digital twins to generate product imagery, reducing reliance on traditional photo production.
These examples illustrate the potential value of a general-purpose accelerated computing platform: infrastructure can serve multiple applications rather than being tied to a single workload.
Maximizing AI factory returns
The economics of an AI factory ultimately depend on more than peak performance.
A productive system can generate more output within a constrained power envelope. A durable system can continue generating value for years. A fungible system can adapt to changing workloads and demand.
NVIDIA’s AI factory strategy brings these three characteristics together through full-stack codesign, a broad software ecosystem and a common programmable architecture.
As AI workloads continue to evolve — from generative AI and reasoning to agentic systems, scientific computing and physical AI — flexibility can become as important as raw performance.
For AI factory operators making investments measured in millions of dollars per megawatt, the objective is therefore not simply to build the fastest infrastructure available today. It is to build infrastructure capable of generating value efficiently, remaining useful over time and adapting to as many sources of demand as possible.
That combination of productivity, durability and fungibility is central to NVIDIA’s approach to maximizing the return on AI infrastructure investment.
Source Link: https://blogs.nvidia.com/