Fastino Bets on Gaming GPUs to Shake Up AI Model Training

Startup Levels the AI Playing Field

Fastino, a stealth-mode startup, has just secured $17.5 million in seed funding led by Khosla Ventures with participation from Real Ventures and iNovia. Its goal? Upend the AI hardware space by making high-performance model training vastly more affordable. Instead of relying on expensive and power-hungry data center GPUs like Nvidia H100s, Fastino harnesses low-cost consumer gaming GPUs to train massive models. The Montreal-based company has quietly built a distributed training system that aggregates performance across hundreds of off-the-shelf graphics cards—effectively democratizing access to powerful AI training infrastructure. This economical approach could greatly benefit smaller AI startups and academic researchers priced out of mainstream options.

Gaming GPUs in the AI Arena

Fastino’s system is optimized not for brute force, but for coordination. By cleverly distributing workloads and building custom software to manage training across a network of consumer cards, Fastino says it delivers performance comparable to top-tier enterprise GPUs at a fraction of the cost and energy consumption. CEO André de Bellefeuille emphasizes that their solution isn’t just cheaper—it’s also greener and more scalable, pointing to their success in training the Llama 2 7B model on a fraction of the typical carbon footprint. As cloud giants scramble for Nvidia chips, Fastino’s contrarian strategy could rewrite best practices for AI infrastructure.