AI model training

Mercor, an AI startup training AI models, secured $350M in a Series C round led by Felicis, valuing the company at $10B. The funding will expand talent, improve matching systems, and accelerate delivery. Mercor utilizes a network of 30,000 experts, paying them over $1.5M daily to provide nuanced training. The company benefited from market shifts after Meta’s investment in Scale AI, creating an opening for providers like Mercor. However, it faces competition from other players in the data-labeling and AI training space.

Meta and Oracle are upgrading their AI data centers with NVIDIA’s Spectrum-X Ethernet, designed for the demands of large-scale AI. This shift to open networking aims to improve AI training efficiency across massive compute clusters. NVIDIA’s MGX system offers flexibility and scalability, while innovations like 800-volt DC power delivery and power-smoothing technology address power efficiency. Spectrum-X supports scalable connectivity between data centers and integrates with various NOSs, enhancing AI infrastructure accessibility. The upcoming Vera Rubin architecture will further support next-gen AI factories.

Huawei introduces a “digital wind tunnel” to simulate AI model training and inference. This platform, encompassing Sim2Train, Sim2Infer, and Sim2Availability, optimizes model configurations, accelerates inference, and enhances system stability. By virtually testing AI models before deployment, Huawei aims to improve efficiency, reduce costs, and ensure optimal performance on their Ascend hardware.