[aurareturn]

Their chips weren’t selling because:

1. They’re useless for training in 2025. They were designed for training prior to LLM explosion. They’re not practical for training anymore because they rely on SRAM which is not scalable.

2. No one is going to spend the resources to optimize models to run on their SDK and hardware. Open source inference engines don’t optimize for Cerebras hardware.

Given the above two reasons, it makes a lot of sense that no one is investing in their hardware and they have switched to a cloud model selling speed as the differentiator.

It’s not always “Nvidia bad”.

[arisAlexis]

They have 10x inference speed. Try Mistral and get awed.

[aurareturn]

They're roughly 2.5x faster. https://www.cerebras.ai/press-release/maverick

Not 10x. But at greater cost and lower scale.

[arisAlexis]

Sources for greater cost and lower scale?