NVIDIA Pitches Fungible AI Infrastructure

The chipmaker argues software flexibility shields data center investments from rapid obsolescence and custom silicon competition.

On October 1, 2026, NVIDIA published a technical and financial paper titled “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment.” The document argues that modern computing facilities operate on fungible AI infrastructure. Shared on the company’s official blog and corporate X account, the paper addresses mounting pressure from Wall Street over returns on massive artificial intelligence spending.

Building an advanced data center currently costs roughly $60 million per installed megawatt, according to NVIDIA’s estimates. With industry infrastructure budgets surpassing hundreds of billions of dollars annually, cloud operators worry that expensive clusters could sit idle if model architectures change quickly.

What fungible AI infrastructure means in practice

In economics, fungible assets can be exchanged without losing utility or value, like currency or commodities. NVIDIA applies the term to data center hardware that can switch workloads without physical reconfiguration of server racks.

According to the company, the CUDA software layer allows the same hardware cluster to move between model pre-training, fine-tuning, and inference. If training demand drops, operators can redirect servers toward standard enterprise computing. Those non-AI tasks include accelerated data processing through cuDF, mechanical engineering simulations with Dassault Systèmes, and graphics rendering. NVIDIA cited real-world deployments, such as Eli Lilly running 1,016 GPUs for genomics and Pinterest managing 14,000 GPUs for visual recommendation.

Custom chips versus hardware longevity

The argument serves as a counterweight against custom silicon. Cloud giants like Google and Amazon invest heavily in application-specific integrated circuits (ASICs), including TPUs and AWS Trainium processors. As Larry Dignan at Constellation Research noted, investors frequently question the capital return on those outlays. NVIDIA counters that fixed-function chips face higher obsolescence risks when new model architectures emerge. Programmable GPUs, by contrast, adapt over time through software updates.

Secondary market tracking by research firms Sprout and Silicon Data provides historical backing for this longevity claim. NVIDIA A100 systems, introduced in 2020, remained commercially active in 2026, six years later, retaining roughly 25% of their original acquisition value on the resale market. Specialized provider CoreWeave extended customer reservation contracts for A100 capacity through 2029, while Microsoft ran its V100 GPU fleet for 8.4 years, well beyond the standard six-year accounting depreciation window.

Efficiency claims and financial reality

For upcoming hardware generations, NVIDIA promises significant efficiency improvements. An independent projection from research firm SemiAnalysis estimates that NVIDIA’s future Vera Rubin NVL72 systems will deliver over 30 times the throughput per megawatt and up to 45% lower cost per million tokens on the DeepSeek V4 Pro model compared to GB300 NVL72 racks.

Financial analysts remain cautious about treating AI clusters as eight-to-ten-year capital assets. Projecting a decade of utility for hardware that has barely entered server rooms is an assumption, not an established fact. That model also depends on continued spending from venture-backed AI labs operating at steep financial losses. For enterprise customers outside cloud hosting, token consumption remains an operational expense rather than a proven driver of net profit.

NVIDIA CEO Jensen Huang is scheduled to present further details on the architecture and economics of these facilities on October 21, 2026, during his keynote at GTC Berlin.