WEKA

A parallel file system: why classic storage is not enough

WEKA was founded in 2013 by Liran Zvibel and Maor Ben-Dayan. Training a model reads millions of small files over and over in random order. Classic storage systems, designed for sequential, predictable reads, handle this pattern worst of all. A parallel file system designed from the start for this kind of load keeps the accelerators highly utilized instead of leaving them waiting for input.

Not just AI: genomics, engineering and finance

Alongside AI, the technology is also used in more traditional high-performance computing fields: genomic research, engineering simulations and financial data analysis, anywhere the volume of data and access to it, rather than computing power, is the bottleneck. The company offers the product both for installation in the customer's data center and as a service at cloud providers.

November 2022: $135 million at a $750 million valuation

In November 2022 WEKA raised $135 million at a valuation of $750 million. This was the base from which its valuation jumped less than two years later.

May 2024: official announcement, a unicorn with $140 million

According to WEKA's official announcement, in May 2024 the company raised $140 million in an oversubscribed Series E round led by Valor Equity Partners with participation from NVIDIA, Generation Investment Management and Atreides Management, at a company valuation of $1.6 billion, crossing the threshold that makes a private company a "unicorn". "The continued backing of this preeminent group of investors speaks volumes: WEKA is building something truly transformative that modern data-driven organizations need now, more than ever," said Liran Zvibel, the company's co-founder and CEO, in the announcement. According to the same announcement, more than 300 of the world's largest AI and GPU deployments were running on the WEKA platform at the time.

VAST Data and WEKA: one field, two Israeli companies

The field WEKA works in is very close to that of the Israeli company VAST Data. Both illustrate how concentrated the Israeli industry is in the infrastructure layer beneath the models, rather than in the models themselves. Often it is actually the speed of access to data that determines performance, not the processor or the model.