Скачать или смотреть ⚡️Raising $1.1b to build the fastest LLM Chips on Earth — Andrew Feldman, Cerebras

Andrew Feldman, CEO of Cerebras, joins us to discuss their massive $1.1B fundraise at an $8.1B valuation and their mission to build the fastest AI inference chips on Earth. He breaks down how Cerebras achieved 20x faster performance than NVIDIA's B200 GPUs through their revolutionary wafer-scale architecture, featuring a chip the size of a dinner plate with 2,625x more memory bandwidth than traditional GPUs.

Andrew shares the journey from meeting Sam Altman and Ilya Sutskever at OpenAI in 2016 when they were "just PowerPoint and an idea," to now serving major customers like Cognition, Mistral, Meta, and IBM. He explains the fundamental architectural decisions that enabled Cerebras to excel at both training and inference, including their choice to use SRAM over HBM memory and their focus on sparse linear algebra rather than specialized convolutions.

The conversation covers the explosive growth in AI inference demand, with enterprises increasingly replacing closed-source models with fast open-source alternatives. Andrew discusses trends in the 10-30B parameter model space, the importance of speed for AI adoption (citing Paul Graham's observation about ChatGPT's slowness), and the often-overlooked complexities of building AI infrastructure - from data centers that consume as much power as small cities to the critical routing and caching systems that make AI applications work seamlessly.

Key Topics:

• The $1.1B fundraise led by Fidelity and Altus

• Cerebras' 56x larger chip design and SRAM architecture

• 20x faster inference than NVIDIA B200 GPUs

• The shift from AI novelty to daily utility

• Building data centers that pull gigawatts of power

• Why speed is critical for AI's promise to be fulfilled

• The ongoing dominance of transformer architectures