Deep Dive: How to Train a 3-Trillion-Parameter Model
inductionheads · x · 2026-07-18
This piece discusses the technical realities and infrastructure bottlenecks of training ultra-large models in the 3T (trillion) parameter range.
- Parameter scale isn't the main bottleneck: More players are capable of training 10T-parameter models than the public assumes; total parameter count alone isn't the core limit.
- Sparsity and expert parallelism: Simply stacking experts won't generate commercial value. High-margin, high-sparsity models are key to recouping compute investments.
- Physical hardware limits: Infrastructure hits physical constraints when parameters exceed 1T. For instance, the weights of 896 experts cannot fit into a single node; even with only 16 activated during inference, weight sharding is mandatory, inevitably driving up network communication latency.
More from Infra
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- Gavin Baker argues Nvidia may be one of open source AI’s biggest supporters — GavinSBaker · 2026-07-22
- AI Power Demand Exposes US Energy Gap, Urging Shift from Scarcity to Abundance — bradneuberg · 2026-07-22
- Gavin Baker says Nvidia’s $630B figure would be system revenue, not all Nvidia’s — GavinSBaker · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22
- Report says Nvidia could build 1,000 Vera Rubin racks a day, implying $630B quarterly at system level — GavinSBaker · 2026-07-22