DeepSeek V3 Released — Matches GPT-4o at Fraction of Cost

Chinese AI lab DeepSeek releases DeepSeek-V3, a 671 billion parameter mixture-of-experts model that matches or exceeds GPT-4o on major benchmarks while reporting training costs of only ~2.788 million H800 GPU hours (approximately $5.6 million). DeepSeek claims the model was trained on Nvidia H800 chips (a downgraded version still available in China). The release shocks Silicon Valley, with investors questioning the business case for $100B+ AI infrastructure investments.