In 2024, when the AI industry finally started talking about the "GPU memory bottleneck" as a critical infrastructure challenge, Liran Zvibel had been building around that problem for over a decade. WEKA, the company he co-founded in 2013 with Maor Ben Dayan and Omri Palmon, exists precisely because Zvibel understood - earlier than most - that compute without fast, accessible data is a very expensive way to wait.
The company's latest technical leap, the Augmented Memory Grid on NeuralMesh architecture, extends GPU memory capacity by 1000x - moving from gigabytes to petabytes available to a single inference workload. That's not incremental. That's a decade of compound bets on a specific idea paying out simultaneously.
"When you're training a model, you're compute bound. Inference is the opposite: It's memory bound."
- Liran Zvibel, CEO of WEKA
The inference insight is what drives WEKA's current direction. Training captured all the headlines - Nvidia chips, billion-dollar compute clusters, the race to scale. But inference, where AI models actually do useful things for actual users, has a different economics. You can throw money at training without a ceiling. Inference must win on cost-per-token. And cost-per-token depends almost entirely on how fast you can serve data to GPUs.
Zvibel has been making this argument publicly since before it was fashionable. Now it's the operating thesis of a $1.6 billion company powering 11 of the Fortune 50 and more than 300 of the world's largest AI and GPU deployments.
"With training, there's no amount of spend that doesn't make sense. Unlike training, where you need to win on the outcomes, inference must win on economics."
- Liran Zvibel
The Augmented Memory Grid, released in late 2025, is the product-form answer to that thesis. Using RDMA and NVIDIA Magnum IO GPUDirect Storage, it delivers petabyte-scale memory access at speeds that blur the line between RAM and storage - reducing time-to-first-token in AI inference by up to 20x.
1000x
GPU Memory Extension
20x
Faster Inference TTFT
3x
Revenue Growth, 3 Years Running
0.02%
Companies Reach $100M ARR
Origins
From Tel Aviv to Silicon Valley, the Long Way Around
Zvibel's path to WEKA ran through the Israeli military and a storage acquisition that nobody in San Francisco was paying attention to at the time. He graduated from Tel Aviv University with a degree in Mathematics and Computer Science, did a stint as a system programmer at the university's own computational centre in 1998, and then entered the Israeli Defense Forces in 2000 as a software engineer. He left in 2005 as a Captain, having led a project that won the Israeli Defence Award - one of the country's most prestigious honors for defense technology.
After the IDF, he joined XIV Storage Systems as a principal software architect and founding technical team member. XIV built a distributed, self-healing storage system based on commodity hardware - an approach that was genuinely ahead of its time in 2005 when most enterprise storage was still proprietary hardware stack plays. IBM acquired XIV in 2007, and Zvibel stayed on to manage hardware and platform teams inside IBM for several years.
That acquisition was formative. It showed him what an enterprise-grade storage architecture could look like at scale, inside one of the world's largest IT companies, with real Fortune 500 deployment constraints. It also showed him what IBM couldn't or wouldn't build - the kind of software-first, performance-obsessed data platform that the cloud era would eventually demand.
There was a brief detour into consumer tech: he co-founded Fusic in 2011, a social video platform with recording, editing, and mobile capabilities. By 2014, he'd moved on to WEKA. The Fusic chapter is easy to overlook, but it points to something worth noting - Zvibel is willing to work in spaces far outside his comfort zone if the problem is interesting enough.
The Business
What WEKA Actually Does
WEKA is not a storage company in the conventional sense. It is software that makes storage act like memory - specifically for the kinds of parallel, random-access workloads that AI training and inference run. Traditional storage solutions were designed for sequential reads, backups, archival. Modern AI workloads need something closer to RAM at petabyte scale.
The WEKA Data Platform runs on commodity cloud and on-premises hardware, handles multi-tenant environments with enterprise-grade security, and integrates natively with Kubernetes, Docker, and the broader MLOps toolchain. It supports genomics pipelines at life science companies, electronic design automation at chip manufacturers, financial modeling in quantitative funds, and - increasingly - the largest LLM training and inference deployments in the world.
The February 2025 CEO letter Zvibel published was notably specific about what the company achieved: more than $100M in annual recurring revenue, the third consecutive year of more than doubling revenue, and unicorn valuation status at $1.6 billion following the May 2024 Series E round. NVIDIA is a strategic investor. Generation Investment Management - Al Gore's climate-focused fund, which rarely backs pure infrastructure plays - also participated.
The NeuralMesh architecture, WEKA's newest platform layer, integrates NVIDIA BlueField-4 DPUs to create what Zvibel describes as a "neurally adaptive" storage fabric. The architecture lets WEKA move data at speeds that eliminate the traditional boundary between persistent storage and working memory. It's the kind of work that earns NVIDIA's endorsement and storage certifications for NVIDIA Cloud Partners.
"2025 is actually the dawn of artificial general intelligence... essentially the beginning of the fourth industrial revolution."
- Liran Zvibel, speaking at RAISE Summit Paris, October 2025
Whether that framing is right about AGI timelines, it captures his seriousness about the moment. Zvibel has been in infrastructure through enough technology transitions - client-server to cloud, spinning disk to flash, HDD arrays to software-defined storage - to know that platform shifts create winner-take-all windows. He's betting WEKA is positioned at the right junction, again.
Worth Knowing
Six Things Worth Knowing
Captain, IDF. Zvibel left Israeli military service as a Captain in 2005 - a credential that's unusual on Silicon Valley founder bios. He led a project that won the Israeli Defence Award, one of the country's top technology honors.
Early Twitter. He joined Twitter in December 2010 - before it was the de facto tech-industry communications channel. He's had @liranzvibel since the platform was still figuring out what it was.
Three exits, one thesis. XIV (IBM acquisition), Fusic (social video), WEKA. Each company is different. The thread is: build systems that handle data at scale, under pressure, when failure isn't an option.
NVIDIA picked him. NVIDIA invested in WEKA's Series E. That's not a marketing partnership - chip companies don't write checks to storage software companies unless the software makes their hardware meaningfully better.
Centaur status. Getting a startup to $100M ARR is something only 0.02% of companies ever do. WEKA crossed that threshold and kept growing. The company calls it "centaur status" - rarer than unicorn by an order of magnitude.
The D language. Zvibel gave a technical talk at DConf 2018 in Munich on using the D programming language for large-scale storage systems - a niche technical community, not a keynote stage. He still thinks like an engineer.
