Perspectives
Gimlet raises $300 million to make GPUs spend less time waiting
Gimlet has secured a $300 million Series B. The opportunity to make different chips work together lies in AI’s waiting time—and in the engineering of the data center.
Translated from 简体中文
As AI starts doing work for people, waiting for it to finish thinking may become a small but recurring part of the day. Researching, reading code, calling tools and checking results: a task often cannot advance until its previous step ends. Getting one answer faster and finishing the whole job sooner are two different experiences.
On September 4, Gimlet Labs announced a $300 million Series B led by a16z. Bloomberg reported that the round valued the company at $3 billion. Its business is to organize chips with different architectures into an inference service, letting each handle work better suited to it. Funding announcement[1] and Bloomberg report[2]
Image source: Gimlet Labs' announcement of this funding round.
The approach raises a question worth considering: once models can do the work, will the next investment go toward stronger models or toward systems that make them spend less time waiting?
Even expensive chips have tasks they do not do well
Intuitively, making AI faster might seem as simple as buying more high-performance GPUs. But computation, reading data and passing intermediate results within an inference task place different demands on hardware. Putting every stage on the same kind of chip makes management simpler, but can also leave some expensive components waiting for other stages to finish.