Keep pulling the thread on Andrew Feldman.
The Cerebras wafer-scale chip is 58 times larger than any other chip ever made.
Cerebras systems are 15 times faster than the fastest GPU for AI inference.
On certain problems, Cerebras systems can be 50, 100, or even 1,000 times faster than graphics processing units.
In December, Cerebras signed a contract with OpenAI valued at over $20 billion.
In March, Cerebras signed a deal for AWS to deploy Cerebras systems in its data centers.
Following its IPO, Cerebras's market capitalization is valued at $64 billion.
Cerebras's systems produce fast tokens at a much lower cost and with a fraction of the power compared to GPUs.
Cerebras avoids key industry supply constraints by not using HBM memory, TSMC's CoWoS process, or TSMC's 3-nanometer process.
The entire AI industry is currently constrained by the availability of powered data centers, a situation expected to last for at least the next 15 to 18 months.
Cerebras provides compute power for both OpenAI and Cognition.
Andrew Feldman asserts that NVIDIA's CUDA software stack is no longer important in the AI landscape and has no role in inference.
Two of the three leading frontier AI models, Google's Gemini and Anthropic's Claude, were not built using NVIDIA's CUDA.