Keep pulling the thread on Rodrigo Liang.
SambaNova completed a first close of a $1 billion fundraising round at an $11 billion valuation.
SambaNova has raised a total of $2.5 billion in the history of the company.
A single 10-kilowatt SambaNova SN40 rack outperforms a 130-140 kilowatt rack of NVIDIA GPUs for inference tasks.
A single SambaNova rack can run a trillion-parameter model, a task that would require dozens of racks of competing hardware.
The number of chips deployed for AI inference will be orders of magnitude greater than the number of chips used for training.
The next generation of large language models are heading towards 10 trillion parameters, with some open source models already at 1 to 2 trillion parameters.
The AI infrastructure market will likely consolidate to only 2 to 4 major chip providers.
SambaNova's latest funding round was led by General Atlantic, with participation from Seligman Ventures, T. Rowe Price, and Capital Group.
SambaNova has taped out 6 chips in the last 7 years and will tape out its 7th chip next year.
SambaNova's Generation 5 chip is scheduled to begin shipping later this year.
The minimum deployment size for running a large model like DeepMind's 1.5 trillion parameter model on some providers' hardware is 10 to 20 racks, while SambaNova's minimum is a single rack.
In the emerging world of AI agents, low latency is critical because the response times of multiple agents in a chain are cumulative, making slow individual responses unacceptable.