AI startup Magic claims a pretraining recipe roughly 50x more compute-efficient than DeepSeek's
Trained for about $0.5M on GB200 hardware, roughly half GPT-3's compute budget, then scaled 10x further to beat public open base models on held-out perplexity. Magic credits many incremental changes.
AI startup Magic published a technical update on September 8, 2026 claiming its pretraining recipe now matches DeepSeek V4 Pro Base's quality using roughly 50 times fewer FLOPs. It trained for about $500,000 on GB200 hardware, roughly half the compute budget originally used for GPT-3. Scaled 10 times further, the same recipe beat public open-weight base models on held-out perplexity across 167 domains, per the company's own account.
Many small gains
Magic says it can "match DeepSeek V4 Pro Base using ~50x fewer FLOPs," and attributes the gain to many incremental architecture, optimizer, and data changes accumulated over time. No single innovation is credited. That matters: one dramatic technique is easy to dismiss as unrepeatable, while compounding small efficiency gains are how most cost reductions in this industry happen.
A data point for anyone modeling training costs
If the claim survives independent scrutiny, a 50x cut in the compute needed to reach a given quality bar matters to anyone forecasting near-term training costs or judging how defensible a frontier lab's compute-spend advantage is. The figure is self-reported by a company with an obvious incentive to look efficient. Wait for third-party replication before you build it into a forecast.