When Weka CEO and cofounder Liran Zvibel has to describe what his company does in simple terms, he says, "it's like an alchemist's effect in the Middle Ages. In effect, it solves the bottleneck in the AI industry known as the "memory wall." To allow AI to quickly access the data needed for calculations, memory components are required, but in an era where graphics processors perform billions of operations per second, memory has become a critical challenge. So AI is leaping forward, but memory is lagging behind."
Zvibel explains that the new era, in which the emphasis is shifting to operating AI agents, and with it the increasing need for training models, has led more and more enterprises to discover Weka, as well as the entire category. "When generating inference, many more 'tokens' - the basic processing units of AI - are required than can fit in the regular memory of a graphics processor," he explains.
"We are expanding memory and in the process greatly reducing latency, the time gap in which data reaches the chip, to the point where it doesn't notice whether it came from the memory component, or from a Weka system. The result is that our customers can suddenly process many more tokens."
The memory problem, which is rarely discussed, has sent component manufacturers' stocks soaring. SanDisk’s share price has risen 3,550% in the past year, SK Hynix has jumped 677%, and Micron has jumped 102%.
Under their noses, a huge industry has sprung up to help Nvidia, which is still responsible for the vast majority of processors in the new technology era, deal with the memory wall problem. Weka is one of the companies doing this.
The segment leader, VAST Data, last month became the most valuable privately-held Israeli company at a valuation of $30 billion, after a $1 billion secondary share sale.
Another rival is DataDirect Networks (DDN), owned by French-Jewish duo Alex Bouzari and Paul Bloch, which raised $300 million last year at a $5 billion valuation.
Weka received a valuation of $1.6 billion in its most recent financing round two years ago, and the three companies are currently Nvidia's only close partners in memory and storage management. These companies gain access to Nvidia's engineering and sales teams, as well as a steady stream of customers from the chip giant.
"Ultimately, this industry operates under severe constraints," explains Zvibel. "There is a ceiling on the amount of GPUs that Nvidia can produce and a ceiling on the amount of electricity that companies can purchase, even the biggest tech companies, so the place where something can be changed is to optimize the way AI data centers work, and the key to that lies in the memory components and their management."
With 5 hours to run a calculation, we created a buzz in Silicon Valley"
Zvibel (46) was born and raised in Tel Aviv and has been in front of a computer, "more or less since I can remember." He studied at Rothberg High School, where for a period he led a program for gifted youth, in which he compressed four years of high school into two years. "They canceled it after they realized that for most graduates it screwed up their lives," says Zvibel. "I'm perhaps one of the best examples on the sane side."
When his friends graduated from high school, Zvibel began a bachelor's degree in mathematics and computer science, enlisted in the IDF Intelligence Corps, underwent training in a prestigious track and later served in a special tech unit. When he was discharged, his friends from the military service hired him for the company they had founded, storage technology startup XIV - a company named after the 14th class of the Talpiot program, five of whose graduates founded the company.
He was one of the first employees they hired, and about a year after it was founded, it was sold for about $350 million to IBM, which was then acquiring storage companies in Israel. He spent four years until he left in 2011.
"I learned there how to found a startup, and how big companies work, and how enterprises buy products. I couldn't have built Weka if IBM hadn't bought XIV. It was a time when almost every company was acquired and there were many people who worked in large companies. It provides a lot of experience that you can't get working in a small private company or in the military."
To a large extent, companies like VAST Data and Weka would not have been able to exist without a glorious chapter in Israeli high-tech, which in the first decade of the 21st century excelled in data storage companies even more than in cybersecurity.
In the acquisition fever at the time, companies like Kashya, Diligent and Storewise were sold for hundreds of millions - most of them to the two companies that were then competing for the most innovative enterprise storage systems - EMC and IBM. Extreme.io, which was sold for $450 million to EMC in 2012, was the most recent chapter in the wave of exits in the Israeli storage industry. Its VP development at the time of the sale was Israeli Renen Hallak, cofounder and CEO of VAST Data
Weka was founded by Zvibel together with his former colleagues at XIV and in the army, Omri Palemon and Maor Ben Dayan. The three made sure to bring with them many former employees from XIV, Extreme.io and Scale.io and Weka began operating 13 years ago.
This happened mainly due to new technology that made it possible to quickly connect a standard computing server to flash-based memory - a memory component that was mainly known as data storage on a disk-on-key or phones, and was also starting to reach the world of computing servers - but access to it was cumbersome and slow. "Until then, the network was a hundred times slower than the server, and storage was based on a more outdated method such as storage based on hard disk controllers," says Zvibel.
In the company's first year of operation, 2014, no one imagined that eight years later, OpenAI would launch ChatGPT and usher in a new era of creative AI. "We knew we wanted to give the most efficient system to customers who needed high-volume computation, but that was before the AI era that we know today," he notes.
"At IBM, we saw increasing use, so we believed in this trend, and in the meantime we tried our luck with customers who required fast processing such as hedge funds, algorithmic trading companies and pharmaceutical companies, but it took us years to reach real deep tech companies. Starting at the end of the 2010s, we noticed that the customers who came with the greatest interest were those who worked with graphics processors, back in an era when AI was still called machine learning. In 2022, when ChatGPT was released to the world, we were ready with a product that gained enthusiasm: there were large companies that took 14 days to run a calculation on flash memory - with us they finished in five hours. It created a buzz by word of mouth in Silicon Valley."
The investor who compares the solution to a Ferrari
One of the significant growth drivers of AI agents, and Weka in particular, is the meteoric progress in the automated software development industry, where AI engines take on a significant portion of the software development work in an enterprise, supervised by experienced programmers.
"The memory wall is becoming an acute problem in the AI industry due to the increasing shift to operating AI agents in organizations, and first and foremost due to the proliferation of autonomous software development agents, such as CloudCode, Windsurf, OpenAI's Codex and Cursor," says Zvibel. "This is the first big 'killer app' of the agent era, and they need endless memory because they require a lot of context and use many tokens to perform their computational activity."
This is where Weka’s added value comes in, which may be less helpful in chat engines, but in the area of "coding" agents manages to increase the performance of Nvidia processors by as much as tenfold.
After VAST Data, Weka is probably the fastest-growing company in the world in the field of memory management for Nvidia's graphics processors. According to executives in the AI market, its annual recurring revenue (ARR) is estimated at $100-150 million with customers such as Oracle and Tesla, which trains its cars and robots with the Israeli system, alongside pharmaceutical giants such as Novo Nordisk and Eli Lilly.
The major cloud companies, including Nebius, CoreWeave, Cohere, Stability and Hugging-Face, are working with it at an increasing pace, not least due to the success of competitor VAST Data, and due to referrals from its loyal partner Nvidia. The Israeli system is installed in the Sphere events hall in Las Vegas and allows video to be projected on its giant screens at a natural pace. The company declined to comment on its revenue or identify its customers and asked not to comment on the matter.
Among investors in the company are: Valor Fund, one of the early investors in SpaceX; Atreides Management, which became familiar with Israel through its investments in Avigdor Willenz's chip companies; Generation IM Fund of former US Vice President Al Gore; and the Israeli Key1 Fund.
Danny Akerman, managing partner at Key1, is one of the early investors in Weka. "We invested in the company three times and the company has grown exponentially," he recounts.
On the huge fundraising by rival VAST, Akerman says, "The fact that investors are investing in it at a valuation of $30 billion means that they expect a return of at least 2-3 times on their money, meaning that they believe that only one company in this market will reach a valuation of $100 billion, which implies a future market value of hundreds of billions of dollars, which will also be shared by competitors."
Akerman says "Weka has always focused on high-performance computing, it is an expensive and boutique solution, a Ferrari that not everyone can drive. But this field is gaining increasing demand with the expansion of AI use."
Memory shortage and "resting processors": Nvidia's ceiling
Has VAST Data paved the way for Weka and other competitors in the field? Zvibel refuses to comment directly on VAST Data, but notes, "The market shows that Israeli companies can play in the big leagues. Wiz showed this earlier this year, but these trends give us confidence that we are doing the right things from Israel. This is also proof that the market is growing and understands that there are AI infrastructures that go beyond Nvidia processors.
"Now, as the focus of the AI market shifts from chat applications to real agents that run critical tasks within companies, the market feels how much it needs systems that can expand memory, one that manages to produce many more tokens for each Nvidia processor."
Zvibel is sure that the rapid growth of AI consumption is also leading to greater openness to new technologies that are not necessarily sold by giant companies, but also by agile startups: "I see many who want to experiment with new products that were not there before. The market is facing crazy growth, the serious players understand this, and small players also understand that their current size is not important, but what they can achieve."
It's no coincidence that the category has become one that produces hundreds of millions of dollars and is already worth tens of billions - it grew where Nvidia was stuck at the ceiling. "You can't use AI properly if the graphics processor memory is relatively small and the processors find themselves resting activitly and waiting for data to arrive," explains Akerman.
"When you ask a chat, like Claude or ChatGPT, based on a previous question or a document you uploaded earlier, it uses memory to refer to past answers. When its memory runs out, it simply moves on to employ the next graphics processor. The problem is that there is a shortage of graphics processors, but if there is software that intelligently allocates additional memory for you, it saves you the need to work with several graphics processors at the same time and this is a significant boost to memory. Companies like Weka and VAST Data have entered this niche."
Zvibel explains that in recent months, the memory shortage has been shaping the industry in ways no one had predicted: "The shortage of the type of memory that Nvidia is embedding next to its chip, high-bandwidth memory (HBM), has led to other types of memory suddenly becoming necessary, even though the market was not at all prepared for it.
"The cloud giants simply started buying memory from whatever came close, and when the cheaper types of memory were not available, they went for the high-end products, such as flash memory (NAND), on the assumption that 'if there is no bread, eat cake'. Thus, there is a pincer movement on non-volatile memory, which is more expensive, when on the one hand it is more in demand, and on the other hand flash manufacturers converted production lines to more expensive types of memory to increase their profitability. This explains the sharp increase in memory stocks - companies like SK Hynix, Western Digital and Micron. I believe this will last two years and then there will be a series of factories that will work hard to reduce the shortage."
Will the "memory wall" disappear one day when Nvidia, AMD or Amazon overcome the memory shortage and manage to provide their own competing solution that may make an entire industry redundant? "The memory pyramid won't go anywhere," says Zvibel. "The laws of physics dictate a permanent memory limit because the faster the chip, the faster it needs access to more data. In fact, every time Nvidia launches a new generation of processors, it becomes more competitive, but that doesn't contradict the fact that every piece of silicon is limited by the amount of memory that can be placed on it and the more they increase it, the more data we will always allow them to access."
Published by Globes, Israel business news - en.globes.co.il - on May 7, 2026.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2026.