DeepSeek-V3 Trained With Only 180K GPU-Hours, Slashing MoE Compute Costs
teortaxesTex · x · 2026-08-01
A comparison of DeepSeek's generational training compute reveals massive efficiency gains: V1 required 300K H800 GPU-hours per 1T tokens, while V2 and V3 dropped to 173K and 180K respectively. The author estimates their next model (V4-Flash) might need only around 66K hours, totaling roughly 2M GPU-hours.
This implies DeepSeek's current best model is smaller and cheaper to train than its version from 26 months ago. The author praises this as the ultimate validation of Scaling laws for MoEs, contrasting DeepSeek's transparent inference unit economics with the secretive myth-making of Western frontier AI companies.
Related event: DeepSeek Drastically Reduces Training Compute Costs Across Models(2 posts)→
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16