AI Inference Workloads to Overtake Training by 2033, Hitting 46 GW by 2035
Event summary
- AI inference workloads will grow at a 42% CAGR, surpassing training workloads by 2033 and reaching 46 GW by 2035.
- Fine-tuning workloads will surpass foundation model training by 2032, growing to 21 GW by 2035.
- Code generation will dominate inference workloads, scaling to 24 GW by 2035, while audio generation grows fastest at 42% CAGR.
- Neocloud providers will nearly catch hyperscalers in inference capacity consumption by 2035, reaching 15 GW vs. 16 GW.
The big picture
The shift from training to inference workloads signals a structural change in the AI market, driven by enterprise adoption and the need for scalable, cost-effective AI solutions. This transition will reshape competitive dynamics, with neocloud providers gaining ground against hyperscalers and enterprises prioritizing inference-optimized capacity. The market is moving towards a more heterogeneous compute landscape, shaped by autonomous agentic systems and multi-modal workloads.
What we're watching
- Inference Optimization
- How cloud providers will balance performance, latency, cost, and compute utilization to deliver inference at scale.
- Competitive Dynamics
- Whether neocloud providers can sustain their growth and challenge hyperscalers in inference capacity consumption.
- Workload Diversification
- The pace at which fine-tuning and multi-modal workloads will reshape AI infrastructure strategy.
Related topics