Renen Hallak opened VAST Forward – the company’s first global customer event – with a simple Hebrew phrase: “Know where you came from and where you are going.” While nostalgic, it was a deliberate attempt by Renen to communicate VAST’s next act – its “AI Operating System” ambition – in a decade-long sequence of architectural bets, customer-driven engineering, and a view of the world that says the next phase of AI will be defined less by data, control planes, and operational resilience.
Hallak’s keynote went back ten years and the founding vision of VAST’s origin story. He put forth the early backdrop of the modern AI boom, going back to Google’s acquisition of DeepMind, OpenAI’s founding, when the DGX era was not yet mainstream. At the time he observed AI systems could do narrow pattern recognition but lacked the scale and context to behave like general problem solvers. The premise, as Hallak laid out, was if algorithms could be fed more data faster, capability would rise. He said VAST didn’t start as an “AI company,” but the company concluded early that data infrastructure would sit at the center of whatever AI became. That conviction shaped the vision to build an AI OS.
VAST is a company built around first principles, not rules
Hallak’s keynote emphasized culture. He said VAST employees are mission-oriented builders, innovators, and work on a set of “principles not rules.” The operational implication of this philosophy, as presented, is a willingness to attack problems that are more meaningful and structural rather than incremental. He emphasized this is especially when large customers and prospects described recurring friction that conventional architectures couldn’t resolve without forcing ugly tradeoffs.
One early example came from a hedge fund use case segregating “new” versus “old” data without sacrificing either capacity economics or performance. Hallak positioned that problem as emblematic of a broader industry constraint, where large-scale analytics projects were being held back by architectural compromises. VAST’s response was rather than add a feature, focus on creating a new architecture. Specifically, a disaggregated, shared-everything design, a different approach to metadata organization, and platform bets on ecosystem building blocks (he called out Mellanox and Intel as early partners). Notably, Hallak emphasized that VAST had to develop new maths, because the needed algorithms weren’t readily available and then he made an important claim that the system scales “super linearly” as underlying resources grow.
While such stories and statements from founders are often polished for audience effect, it nonetheless underscores VAST’s ethos to operate on first principles versus incrementalism.
Product cadence as a journey to the “OS for thinking machine” northstar
Hallak laid out VAST’s last decade as a product-and-capability sequence that manifests in a vertically integrated stack. Each year’s milestone was positiioned as a response to real customer demands, but also as another layer in what becomes an operating model for AI-era data and compute.
- 2017 – VAST Data Store: Multi-protocol access, “universal storage,” consistent performance across workloads, and the ability to mix and match hardware across generations – positioned as foundational to keeping “all data accessible to AI workloads.” He referenced early engagements with NASA, Boston Children’s, General Dynamics, and Square Point.
- 2018 – Media & entertainment: Pixar as an early beta customer (watch our CUBE interview with Pixar), with a use case centered on unifying data across geographies to enable collaboration. Hallak added other proof points (NHL, Jane Street, CZ BioHub) and introduced the idea of what we call a “supercloud” (not his term) – i.e. data access independent of physical location, extending to distributed inference at the edge and reinforcement learning workloads.
- 2019 – The “scan the filesystem” problem: Hallak argued that at massive scale, traditional approaches break down because by the time you scan a billion files, the state has changed and consistency is compromised. VAST’s response according to Hallak was “a new type of database,” combining storage and database ideas, and bringing structure to unstructured data. The strategic intent being to reduce database sprawl and types required to support modern workloads by putting the data together in a way AI can actually use.
- 2020 – A “data engine” and a new interface: Hallak described rethinking the interface between data and applications, including a new language “based on streams,” and a split from the old logic that tightly binds apps to data plumbing. He tied this to reasoning and agentic workflows, plus reinforcement learning at large scale.
The culmination of this arc or product innovation was described as an “Operating system” – with everything vertically integrated into one stack. Whether one agrees with the label or not, the intent is VAST wants to define an execution environment for AI-era workloads where storage, database semantics, data engines, and performance patterns are orchestrated as a single system rather than assembled as a chain of brittle parts.
The customer list as a proof points – From enterprises to AI clouds
Hallak used customer references as a way to show VAST’s expansion from enterprise and research environments into AI-centric infrastructure providers. He cited a mix spanning media, finance, research, and GPU cloud platforms (including names like CoreWeave and Lambda). By 2024, he emphasized “more AI clouds” and “more supercomputers,” and broadened the list again (including mainstream enterprise brands and security-focused names), positioning VAST as increasingly attached to “AI factories.”
One notable product highlight from 2024 was an “Insight Engine” with a chatbot interface – less as a UI gimmick and more as an example of a system that treats data understanding and interaction as first-class. He also called out a local KV-cache capability aimed at improving GPU efficiency – an important nugget, because KV-caches sit at the intersection of inference performance, memory hierarchy tension, and security concerns.
The underlying message: This era requires a persistent AI OS
Hallak’s forward looking talk moved from company history into a future-state narrative he called a “thinking machine,” tied to older ideas like perpetual learning machines, machines adapting to humans, and recursive learning systems. The point wasn’t to romanticize AI, rather it was to describe an infrastructure requirement in that AI systems must operate continuously, adaptively, and safely, at scale.
He offered a scenario-driven view of 2036 where agents close the gap with what people do today, extreme personalization, “unimaginable abundance,” and domain-specific agentic systems (genomics, synthesis, toxicology QA). He painted a world where content becomes ephemeral and personalized (“movies generated once, viewed once and disposed”), and where “agentic scientists” can test vastly more hypotheses. In our view, the important part of this section wasn’t necessarily the speculative outcomes, rather it was the infrastructure implication that enabling, coordinating, and controlling all those activities is a bit opportunity that VAST is uniquely pursuing.
This is where the “OS” metaphor becomes more tangible. Hallak described an OS that triggers inference and fine-tuning, sits “in the middle of every arm movement of every robot,” intermediates thought creation, and “keeps track of all memories.” He emphasized persistence, real-time action, mastery of skill sets, and 100% uptime expectations. Whether 100% is aspirational or a proxy for “no planned downtime,” the message is that agentic systems will raise the reliability bar beyond what many enterprise systems are designed to deliver today.
The stack is being rebuilt, and VAST wants to sit in the middle
Hallak described the infrastructure shift with a layered model (he referenced Jensen’s “five layer cake” of apps, models, cloud, hardware, energy) and then grounded it with the physical realities of racks moving from ~10kW to 500kW, old CPU + 10GbE + HDD assumptions giving way to bandwidth-heavy fabrics, flash-centric systems, and DPUs. His macro claim was that in every revolution there’s a primary ingredient – e.g. PCs were compute, the internet was networking, and AI is data (with GPUs and CUDA as the accelerant). Convenient positioning but frankly not dubious in our view.
He also emphasized governance and observability saying everything must be “audited, observed, and queryable.” In our opinion, that’s a subtle but important statement because as agentic workflows expand, the operational requirement to move from from “can it run?” to “can it be controlled, proven, and explained?”
2026: “The VAST AI Operating System” and what’s missing
Hallak positioned 2026 as a point in time – i.e. the kernel is in place, but substantial work remains. He described three big pillars already forming inside of VAST – unstructured data management via the data store, compute via the AI engine, and the beginnings of a kernel-like control plane. Then he listed missing features that effectively define VAST’s product agenda:
- Database functionality to consolidate structured data forms
- A lightweight VAST instance for edge deployment (robots and beyond)
- Large-scale reasoning up and down the memory stack with secure KV-caches and “no leakage”
- Enabling every agent to fine-tune its own model
- A “tuning engine” so agents can create IP and monetize
- Secure distribution and monetization – leasing models without being “ripped off”
- Agentic ID verification
- Policy setting and enforcement as a core function of the AI OS
In our view, that list is was one of the keynote’s most interesting parts. He basically gave away key parts of his roadmap saying he’s confident of VAST’s lead on the competition. Moreover, this wasn’t just a roadmap. It was a statement that the next platform battle is about control + security + economics layered on top of data and compute. If agents are to operate autonomously, enterprises will demand identity, provenance, policy enforcement, and mechanisms to prevent model and data exfiltration. And if agents are to become economic players, monetization and IP protection become fundamental, not bolt-ons.
Engineering to solve hard problems as strategy
Hallak closed with a business update that framed VAST as “built for this time,” not only because of architecture but because it is “a company of pioneers” tackling hard engineering problems. He cited $100M in incremental cash generation per quarter, “triple” revenues year-over-year, and – most strategically – VAST being “attached to large AI factories,” which he said gives the company favored position.
We believe the point is VAST wants to be the data-and-control substrate of AI factories, not just a storage vendor participating in GPU buildouts. If the company executes on that strategy, VAST’s TAM expands beyond traditional enterprise storage and into the emerging category of AI systems infrastructure where the winners will be those that can prove performance, resilience, observability, and policy control at scale, across core and edge environments, while simplifying operations for organizations that don’t have armies of AI engineers and SecOps specialists.
Is VAST really an OS for AI?
In our view, VAST’s “AI Operating System” positioning is both viable and novel, but easy to overstate if one considers an “OS” in the classic Linux/Windows sense. Today, VAST looks less like a full operating system that schedules compute and enforces end-to-end process fencing, and more like an OS for AI IO. Meaning we see it as more of a data-plane control layer that governs how massive, multimodal data is stored, indexed, secured, moved, cached, and made observable (and trusted) for models and agents.
Why the distinction? The “system of intelligence” comprises much more than VAST’s domain and touches models, a cognitive surface, transactional engines, agent frameworks, orchestration, etc.). These live above VAST in other parts of the emerging AI software stack. Where VAST has a credible right to the OS metaphor is in becoming the persistent data substrate – i.e. unifying structured and unstructured data, collapsing sprawl, tightening GPU feeds (including KV-cache optimizations), and making the entire pipeline auditable and queryable. This is important and an enormous opportunity – with a TAM north of $100B in our view. However, the “AI OS” claim becomes fully defensible only if VAST expands from data-plane dominance into a broadly adopted system of intelligence player. This would bring the company into new competitive domains and from our point of view take it out of its comfort zone.
Bottom line
In our opinion, Hallak’s keynote was a bold declaration that AI’s next phase is a data-and-control-plane problem. Models and silicon will capture headlines, but the platform advantage will accrue to the systems that can persist memory, feed data to coordinate agentic action, secure and audit behavior, and deliver reliability at enterprise scale. VAST is betting it can turn a decade of vertically integrated data infrastructure into an “AI Operating System” and then make that narrative real in 2026 and beyond.
Watch our CUBE interview with Renen Hallak at VAST Forward 2026 where we discussed his vision, the company’s culture and the future of VAST.
