<ul><li>Microsoft’s Satya Nadella tackled the AI scaling challenge at Microsoft Ignite.</li><li>The scaling laws in AI are not physical laws but empirical observations.</li><li>Efforts in areas such as model architectures, data regimes and system architectures are needed to push innovation.</li><li>Microsoft is working closely with NVIDIA on scaling test-time computing for OpenAI.</li><li>'Tokens per watt plus dollar' is Microsoft's new metric to reflect the efficiency of generating tokens.</li><li>Data centres are now viewed as a product by NVIDIA rather than just GPUs.</li><li>Electricity, supply chain and capital are major challenges when it comes to scaling computing.</li><li>OpenAI plans to partner with TSMC and Broadcom to launch its first in-house AI chip by 2026.</li><li>Major inference chip makers like Groq, Sambanova and Cerebras have improved performance.</li><li>NVIDIA is working with its suppliers to cool overheating racks.</li></ul>

AI Scaling Laws Crumble Under Token Pressure

Discover more