Hugging Face published a report on July 30, 2026, detailing how idle graphics processing units create ongoing expenses that mirror grounded aircraft in the aviation industry. Author Gabriel Pimenta de Freitas Cardoso wrote that hardware expenses accrue continuously by the calendar hour through financing, depreciation, power, and cooling, while hardware output only accumulates during active compute hours.
Compute Scarcity
The report tracks how compute constraints expanded between 2020 and 2026. Microsoft built OpenAI a supercomputer featuring more than 10,000 GPUs and 285,000 CPU cores in 2020 to train GPT-3, which ranked among the five largest systems globally at the time. By 2026, Anthropic established multi-gigawatt commitments across four hardware platforms: Amazon, Google, Microsoft, and AMD. Meta completed a comparable multi-gigawatt deal during the same period.
Enterprises are shifting from API services to owned hardware to avoid linear per-token usage fees as workloads reach production volumes. Owning infrastructure trades variable operational costs for fixed capital costs, but forces organizations to size clusters for peak demand rather than average usage.
Hardware Allocation
A single cluster now hosts training, fine-tuning, quantization, real-time inference, batch inference, embedding generation, and model evaluation. Real-time inference requires minimal latency, whereas batch workloads tolerate delays of several hours, creating scheduling conflicts across identical hardware racks.
Unlike commercial aircraft, which can transfer to different geographic routes, idle GPUs can only take on workloads matching their exact memory, duration, and latency parameters. The report highlighted domain-specialized systems, citing tests where Dharma AI's DharmaOCR model surpassed Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese text. Cardoso noted in the report's discussion section that Dharma AI will publish a follow-up article next week focused on job scheduling software.
