AI Compute
Web3 / ai data
The hardware, infrastructure, and energy resources required to train and run artificial intelligence models at scale. Modern AI workloads fall into two broad categories: training, where models learn patterns from massive datasets through billions of iterative gradient update steps, and inference, where trained models generate outputs in response to individual queries. Graphics processing units (GPUs), particularly NVIDIA's H100 and A100 data center chips, are the dominant hardware for both tasks due to their highly parallel architecture. Training a frontier large language model requires clusters of thousands of GPUs running continuously for weeks, at costs reaching tens or hundreds of millions of dollars. Inference is more modest per query but adds up enormously at scale across millions of daily users. AI compute demand has grown explosively since 2022, creating supply shortages and sky-high spot prices for GPU capacity, making access to compute a strategic bottleneck comparable to energy or capital in prior industrial eras. Example: During 2024-2025, a single H100 GPU rented at spot prices from major cloud providers could cost $2-4 per hour, and large AI labs routinely operated clusters of 10,000 to 100,000 GPUs simultaneously. This created an opening for decentralized GPU networks like io.net and Aethir to offer excess capacity at lower prices, targeting inference workloads for AI startups that could not secure long-term cloud contracts. Why it matters for AI: AI compute scarcity has become one of the defining constraints of the current technology cycle. Nations, corporations, and startups compete aggressively for GPU access, and governments have imposed export controls on high-end AI chips to limit rivals' capabilities. The compute bottleneck shapes which organizations can build and deploy frontier AI, making it a geopolitical and economic issue far beyond its technical dimensions.
Explore the full Web3 Glossary — 2,000+ expert-curated definitions. Need guidance? Talk to our consultants.