Renen Hallak Saw the Bottleneck Before the Boom

The useful thing about being taught that a technology does not work is the surprise when it does. At school, Renen Hallak encountered neural networks as a curious dead end. Computers were exact. Brains were mysterious. Asking one to imitate the other seemed charmingly impractical, the computational equivalent of teaching a filing cabinet to recognize a cat.

Then, around 2015, Hallak watched a video of a computer doing precisely that. The machine recognized cats in pictures. The algorithms were not unrecognizable to a computer scientist trained in complexity theory and cryptography. What had changed was the volume of data available to train them. Hallak did what he says he likes to do: take a little journey through what changed, how it worked and where the limiting factor sat.

The journey led below the model, below the fashionable demonstration and into the machinery that feeds it. If learning systems improved with more data, then the decisive constraint would be how quickly enormous datasets could reach enormous pools of processors. Storage, that least theatrical part of computing, had acquired a starring problem.

“I like to go on these little journeys: what changed, how it works, what the limiting factor is.”Renen Hallak

The first engineer learns to scale

Hallak was born in Jerusalem in 1983 and earned bachelor's and master's degrees in computer science, both summa cum laude. His graduate specialization was computational complexity. In 2009, he co-authored a research paper about privacy-preserving approximations for clustering and vertex cover. The subject sounds austere because it is: how close can a computer get to a useful answer, and what must it reveal along the way?

His early career translated theory into systems. He worked with the CTO team at the education-technology company Time To Know, then served as chief architect at Intercast, building a content-distribution system from its beginning through initial deployment. In 2009 he joined flash-storage startup XtremIO as its first engineer. After EMC acquired the company in 2012, he rose to vice president of research and development.

By the time he left Dell EMC, Hallak had managed an organization of more than 200 engineers. The all-flash array he helped architect had crossed $1 billion in revenue. XtremIO was an education in both technical leverage and the strange arithmetic of success: solve a difficult infrastructure problem and, if the timing is right, a product hidden deep inside the data center can become a very large business.

A tradeoff mistaken for a law

After XtremIO, Hallak spoke with potential customers about what their infrastructure could not do. The answer kept returning in different clothes. Fast storage was costly and comparatively small. Large storage was slower. Companies coped by arranging data in tiers, forever moving the useful portion toward speed and the dormant portion toward capacity.

That compromise was tolerable when an application needed only a known slice of information. Machine learning made it absurd. A model may need to inspect the whole archive, and hungry processors are expensive company to keep waiting. Hallak and his co-founders, Shachar Fienblit, Jeff Denworth and Alon Horev, started VAST Data in 2016 to remove the tradeoff rather than improve the choreography around it.

Their design, called Disaggregated Shared Everything, separated compute from the devices holding data while presenting them as one system. Flash capacity, processors, metadata and networking could scale independently. VAST stayed in stealth for roughly three years. When it appeared in February 2019, it arrived with $80 million in disclosed funding and a product called Universal Storage.

There was no guarantee that elegant architecture would become a company. Hallak has been disarmingly direct about his earlier attempts at startups and fundraising. “I did not know how to do it,” he once said. “I didn't understand the game.” He had assumed investors judged ideas mainly on their merit. In practice, he learned, they also judge history, trust and who is willing to vouch for a founder.

VAST's founding team supplied complementary experience. Hallak and the engineering group concentrated on architecture; Denworth took on the commercial story and go-to-market work. The division was not a retreat from technical leadership. It was an admission that a difficult product must also become legible to customers.

The first buyers made that translation possible. VAST sought organizations whose data was large enough to expose the old compromises, then used their objections to sharpen the product. Hallak has described the early commercial climb as moving from zero to $1 million in annual recurring revenue, then to $10 million, in about 18 months. The sequence mattered. A young infrastructure company cannot ask a customer to trust a clever diagram; it must earn the right to become the place where valuable data lives.

For Hallak, that also meant changing jobs without changing titles. The engineer who could reason about metadata structures had to learn sales, hiring and capital. He has spoken about searching for a co-founder who could own go-to-market work because he knew it was not his native discipline. There is a tidy lesson in the move: the confidence to challenge a technical orthodoxy can coexist with enough self-knowledge to hand another person the microphone.

The plumbing becomes the platform

The company's timing now looks enviable. Its first customers included hedge funds and life-science organizations with unusually large datasets. Then came generative-AI builders and specialist AI clouds. Enterprise adoption followed. Each wave asked more of the system: model training, then inference, then streams of interaction from software agents. The product expanded upward from storage into a database, a compute engine and a global namespace.

By April 2026, VAST said it had surpassed $4 billion in cumulative bookings and ended its previous fiscal year with more than $500 million in committed annual recurring revenue, along with positive operating margin and free cash flow. Its Series F financing, a mix of primary and secondary capital worth about $1 billion, valued the company at $30 billion. That was more than three times its late-2023 valuation.

Numbers of that size encourage mythology. Hallak's more interesting claim is architectural. He argues that applications, models and infrastructure are ceasing to behave like separate layers because data connects them all. VAST says its platform already supports environments spanning millions of GPUs. The original nuisance of moving information between slow and fast tiers has become a larger coordination problem: where data lives, how models use it, which agent may see it and whether anyone can reconstruct what happened.

Millions of minds need house rules

Hallak's ambition now reaches well past the storage market. He imagines not one general machine intelligence but millions of agents, each building a slightly different understanding through its experiences. They will talk to one another, call different models and eventually act through robots, vehicles and other machines. A civilization of software assistants sounds delightful until one asks who remembers their conversations and who decides which cupboard they may open.

His answer is an operating system for AI: an infrastructure layer that supplies memory, compute, security, identity, observability and reproducibility. An agent working for one person might be allowed to read private information but forbidden to share it. Two agents could have access to the same record without being permitted to know that the other does. Policies must follow the data, and the system must be able to answer a forensic question months later: why did this agent behave that way?

The phrase “operating system” invites comparison with Windows, macOS or Linux, but Hallak is describing a different scale of coordination. Personal-computer operating systems made unruly hardware usable by applications. His proposed layer would make a sprawl of GPUs, models, data streams and autonomous programs behave like one governable environment. The difficult part is not simply letting components communicate. It is preserving context while keeping boundaries intact, so useful collaboration does not become a universal permission slip.

“They underestimate the magnitude of the change that's coming, and they overestimate how fast it'll happen.”Renen Hallak on enterprise expectations

This is characteristic Hallak: a sweeping destination reached by way of an unromantic constraint. He speaks of thinking machines, then returns to permissions and persistent metadata. He is optimistic without pretending the engineering is finished. “I don't know if we can build a thinking machine yet,” he has said. The uncertainty is not a decorative disclaimer. It keeps the journey open.

He also resists a common confusion about technological change. Enterprises, he says, underestimate its eventual magnitude while overestimating its immediate speed. Ten years is enough to rearrange every kind of work, in his telling, but production adoption still proceeds one low-risk use case at a time. The statement contains both an entrepreneur's appetite and an engineer's calendar.

Hallak began with a useful act of skepticism. He looked at a compromise the storage industry treated as natural and asked whether it was merely inherited. VAST's future will depend on whether its broader operating-system thesis proves as durable as that first question. For now, the company beneath the AI spectacle has made itself difficult to overlook. The cat video was the glamorous bit. The career that followed has been about making sure the machine never has to wait for the picture.