Alon Horev works where glamour goes to wait for data. In the popular telling of artificial intelligence, the model thinks, the chip sweats and the answer appears. Between those events lies a republic of cables, caches, files, permissions, indexes and failure modes. Horev has spent his career in that republic. The VAST Data co-founder and chief technology officer is interested in what happens when a beautiful demonstration becomes a system that must run on Tuesday, survive Wednesday and still remember what it learned by Friday.
His public biography is almost suspiciously lean. Israeli. Self-taught. Military service focused on software and hardware development. Then database and storage work at IBM's XIV operation, Intucell and Cisco, followed by DigiData, the company he founded. On LinkedIn, he reduces his operating philosophy to six words: “Taking products and teams from inception to reality.” It is the sort of sentence that sounds modest until one considers how frequently technology expires somewhere between those two nouns.
VAST began in 2016 around a question that the storage business had learned not to ask. If flash drives, fast networks and new classes of memory had changed the physical ingredients, why keep the old recipe? Storage had long been divided into tiers: quick and expensive here, capacious and slow there, with software shuffling data between them like an anxious maître d'. The founders wanted one architecture that could be fast, large and economical without requiring customers to choreograph the movement.
An education written in systems
“Self-taught” can be a romantic label, particularly when applied after success has improved the lighting. Horev's record is less decorative. It points to an education conducted through systems that punish vagueness. Hardware either carries the signal or it does not. A distributed election either produces a leader or a quarrel. An input-output path either recovers from an error or turns a very expensive machine into furniture.
The patent trail reads like a syllabus in keeping complex machines honest. Horev is named on work covering scalable storage, redundant paths to storage devices, leader election in distributed systems, recovery from failed IO flows, allocation of work across compute cores, file similarity and storage security. Patents do not supply a personality, but these subjects reveal a durable preoccupation: coordination under stress. The interesting system is the one that continues after a component behaves badly.
That same specificity appears in the way he recruits. In 2020, searching for a senior developer, Horev listed the desired terrain plainly: kernel, networking, C++ and Rust. No manifesto, no ping-pong table, no promise to change civilization before lunch. Just four difficult things and an invitation to help build a fast file system.
“We're going to be in a post-applications world quite soon.”Alon Horev on the shift toward agentic software
The storage company outgrew storage
The first VAST proposition was easy to explain, though decidedly less easy to engineer: collapse storage tiers into a single all-flash system. The company's Disaggregated Shared Everything design separates the machines serving requests from the boxes holding data, then lets servers reach the shared pool over a fast fabric. The objective is not merely speed. It is to remove the copying, locking and administrative theater created when each workload gets its own little kingdom.
AI enlarged the problem. Training devours great datasets in bursts. Inference must answer continuously. Agents add a further appetite: they need context, history, vector search, permissions and state. A model without memory can be impressive in precisely the way a brilliant dinner guest with total amnesia is impressive. Every conversation begins well and none accumulates into a relationship.
One system, more shapes of memory
Horev's thesis is that those shapes should not require a procession of separate systems. Files, objects, tables and vectors ought to inhabit one namespace, with compute brought close to the data. This sounds like tidying until scale makes untidiness punitive. Every extra copy must be moved, secured, governed and eventually explained to a finance department. Every idle GPU is a small furnace burning money while waiting for breakfast.
In one account, he described customers that planned for five petabytes and ran out of space in three weeks. Checkpoints multiplied. Datasets forked. Models retrained. The tidy forecast met the untidy behavior of researchers, which is generally where forecasts go to receive character. His advice is to plan not only for performance but for density, power and the physical room left in the building.
When the app dissolves, responsibility remains
A revealing argument from Horev is not about storage at all. He expects the familiar application to lose its monopoly on how people use software. Instead of opening one program for each task, organizations will deploy agents that watch events, retrieve context and call services. A video system might identify a car crossing a red light, smoke appearing in a building or a person falling. The detection is only the opening sentence.
What follows is policy. Which event deserves action? Who is allowed to view the footage? Should the system notify an operator, an emergency service or nobody until a person confirms it? Which record must be kept? These are not questions a model can answer merely by becoming larger. They belong to the organization deploying it, and someone has to encode them.
That is why Horev does not accept the breezy forecast in which agents dispose of software engineers. He believes the work moves upward, toward translating intent into controlled behavior. Engineers connect identity, data, policy and consequence. They inherit less of the screen and more of the responsibility. The job becomes simultaneously more abstract and more accountable, an occupational combination capable of producing both elegant diagrams and difficult meetings.
“A modern AI factory is a distributed system of compute, data, memory, and models.”Horev at the Mistral AI Now Summit, 2026
A factory is more than a room full of chips
In Paris in May 2026, Horev joined leaders from Mistral AI and NVIDIA to discuss AI infrastructure as critical infrastructure. The phrase risks floating away on conference air, but his explanation tethered it to engineering. An AI factory must scale across sites, handle many forms of data and keep working while inference and agents run continuously. Training was a campaign. Inference is a municipal service.
This reframes the fashionable metric. The important question is not simply how many GPUs sit in the cluster, but whether useful tokens emerge per watt, per dollar and per second. A chip waiting on storage is an expensive monument. Horev has argued that neocloud customers favor all-flash infrastructure because predictable data access matters more than preserving an old distinction between hot and cold tiers. Video training can bring half an exabyte into a single dataset. At that point, moving the furniture is not a strategy.
VAST's 2026 financing put a loud number beside this quiet work: $1 billion raised at a $30 billion valuation. The figure belongs to the company, not to Horev personally, and valuation is no substitute for a functioning data path. Still, it records the distance traveled from the years in stealth, when a small founding group was attempting to persuade physics, software and economics to share a table.
The next seam to remove
The latest problem is trust. Enterprises possess data they cannot send to an outside model service. Model builders possess weights they will not place on machines they cannot trust. Each side owns something valuable, and neither wishes to be the first to set it unattended on the conference-room table.
VAST's DataEnclave, introduced in September 2026, is an attempt to change the room. It uses hardware-isolated execution and cryptographic attestation so a model and protected enterprise data can meet without either party surrendering control of its asset. Horev described the appeal in characteristically operational terms: fewer pipelines, fewer copies and fewer separate policy systems. The model moves closer to the data, while the rules travel with both.
This is the line running through his career. A storage tier is a seam. An idle GPU is a seam. A model without memory is a seam. So is an agent separated from policy, or sensitive data stranded away from the model that could use it. Horev keeps looking for the boundary that creates needless movement, delay or ambiguity, then asks whether architecture can make it disappear.
There is no romance in a data path until it fails. Then everyone becomes a poet of latency. Horev's work sits in the long interval before that revelation, where systems are designed, tradeoffs are argued and invisible machinery is asked to behave. AI may be learning to speak, reason and act. Someone still has to make sure it remembers where it put the keys.