NVIDIA argues AI factory returns hinge on power efficiency, software longevity, and workload “fungibility”

NVIDIA is framing the next phase of the AI buildout around economics, not benchmarks. In a company blog post published Oct. 1, the chipmaker argues that AI factories are capital-intensive enterprises, built at a scale measured in megawatts and priced accordingly, and that returns come from a specific mix of performance, durability and flexibility.

The post puts a point estimate on the magnitude of the bet. NVIDIA says each megawatt AI factory costs roughly $60 million, so operators will commit capital only if they can see a credible return on investment. NVIDIA then outlines three factors it says jointly shape those returns: the maximum earning capacity of the hardware, whether demand keeps the systems running at high utilization rather than idling after a short cycle, and whether the factory can be used for many different workloads instead of one narrow use case.

Under NVIDIA’s logic, “strength” in one area cannot fully compensate for weakness elsewhere. A factory that can produce a lot, but sells only part of that output, will not generate strong returns. Likewise, even if demand is high at first, the economics deteriorate if the system stops being competitive at full capacity after a relatively short period.

To engineer for those tradeoffs, NVIDIA says its AI factory approach is grounded in codesign across the full stack, from model and workloads down to compute, networking and memory, plus continuous software optimization after deployment. The company points to CUDA-X libraries, describing them as enabling a factory to run accelerated workloads broadly, and it ties its standardized architecture to a “validated reference design” that can be deployed by different operators.

Power, NVIDIA says, is the binding constraint on an AI factory. That turns attention to what it calls tokens per second per megawatt, a metric that links output directly to the energy envelope that limits scale. In the same framework, NVIDIA argues that higher throughput per megawatt increases revenue potential within a fixed power footprint, and that lower cost per token supports margins on each token produced.

NVIDIA’s post includes comparative performance and cost claims attributed to SemiAnalysis AgentX data, comparing systems it describes as NVIDIA Vera Rubin NVL72 and NVIDIA GB300 NVL72. NVIDIA says the Vera Rubin configuration delivers over 30 times higher throughput per megawatt than the GB300 NVL72, and it says it can cut cost per million tokens by up to 45 times on DeepSeek V4 Pro. The post attributes those gains to “extreme” end-to-end codesign across the model and workload levels, software, and compute and system components.

The post then turns to two policy questions operators face: if each generation makes tokens dramatically cheaper, will demand for compute shrink, and what happens to older systems when new ones arrive? NVIDIA answers both by arguing that cheaper tokens expand the addressable set of use cases, which in turn increases total token demand rather than reducing it. On durability, it says not every workload requires the newest platform, so older generations can continue earning by matching the “complexity and shape” of the work to the appropriate system.

To illustrate durability, NVIDIA points to the NVIDIA A100 GPU, which it says shipped in 2020 and is still in commercial service six years later. It also cites an example of CoreWeave extending bookings for units introduced in 2020 through 2029, and it references third-party analysis of how operators extend depreciation schedules over time. NVIDIA also includes market estimates of useful life and resale or leasing value, stating that Barkr assesses useful life at five to six years for an eight-GPU H100 system and nine to 10 years for GB300 NVL72, while additional data providers it names suggest older A100 value can remain significant and that rental pricing depends on contract length.

NVIDIA ties its durability thesis to software portability. The post says CUDA, NVIDIA’s GPU programming platform, runs across generations, so existing purchased hardware is not stranded when architectures change. It also argues that continuous kernel and software optimization can improve what existing hardware does, and it emphasizes “fungibility” as the longer-term utilization engine: the more kinds of AI and non-AI workloads a system can run, the longer it stays valuable as demand shifts. The company says its platform supports everything from language, vision and reasoning models to agentic AI and physical AI, spanning data processing, pretraining, post-training and inference, and that the same infrastructure also supports non-AI uses such as simulation, graphics and scientific computing.

What remains less specific in NVIDIA’s post is the measured impact of these ideas across different operator business models, such as how quickly each operator’s mix of workloads adapts, or how utilization evolves when new generations enter the market. NVIDIA does not provide new, company-specific financial guidance tied to the framework, and its performance comparisons cite third-party analysis rather than presenting audited results under a single test protocol. Still, for operators planning multi-year deployments, the central message is clear: the industry’s ROI problem is inseparable from power-constrained throughput, software-managed longevity, and the ability to sell multiple kinds of computation from the same installed base.