Jeff Denworth

The through-line

Denworth's consistent argument is that a company only breaks through when a genuinely new architecture arrives at the same moment as a new class of application. Cool technology on its own is not enough: "cool technology in and of itself doesn't compel like a market disruption but on the flip side if you have cool technology at a time where there's an application disruption then you can actually combine these two to to work towards building a large independent business" [1]. He reads his own market in generational waves, where NetApp arrived with open systems and Isilon arrived with large volumes and web 2.0 content, each earning a run measured not in hardware refreshes but in whole administration team generations, and he positions VAST Data as the wave defined by AI changing how customers want to relate to their data [1]. Alongside that sits an equally consistent view of how to run the business: sell very large systems to very large customers, stay capital efficient, and refuse to be measured by the metrics the industry says matter.

What has shifted is scope. The early framing is a storage company out to make flash cheap enough to end tiering [1][11]. By 2023 the framing is a platform of a data store, a database and a forthcoming data engine [2], and by 2025 it is a distributed runtime with serverless compute and vectors as a native data type, which he calls an AI operating system [4]. The underlying claim never changes: the architecture was always meant to be a programmable, massively scalable computer, and the storage layer was simply where they started [4].

On why customers wanted cheaper, not faster

The founding observation was that the flash market had turned into an arms race for speed that nobody had asked for. "none of them said we needed something faster every one of them said we wanted something cheaper" [1]. Two conclusions followed. The first was operational: every day a customer has to choose a point on the spectrum between performance and capacity is a day spent making a compromise [1]. The second was about what was coming, since GPU computing, AI, machine learning and big data algorithms only improve when exposed to more data, and a pyramid of infrastructure that parks the largest datasets on the slowest media defeats that. As he puts it, an AI model "just kind of randomly trolls through all sorts of data and if that data isn't on flash your your next generation processors are going to have a really bad day" [1]. Hence the blunt statement of intent: VAST is "a extinction level event for the hard drive" [1], and the corollary that a single scale-out pool of flash becomes a consolidation opportunity where all applications and all data can live together [1][11].

He also stresses that this was deliberately an unfashionable bet. VAST launched an on-premises storage appliance at a moment when customers were saying everybody was going to cloud, and rather than picking off one segment the ambition was to "steamroll the whole thing" [3]. A prominent investor spent the mid-2010s telling the market that storage was dead, which he treats as part of why the space was open: "we just got going in a very unsexy time in the market and started building" [4]. The design brief that came out of that was to study where cloud was disrupting things "at the architectural level not at the delivery model level" and build something different enough that a customer had no analogous cloud experience to compare it to [1].

On architecture as the thing competitors cannot copy

Denworth argues that almost every modern data system descends from the Google file system paper of roughly twenty years ago, and that modern AI needs something far more transactional and sophisticated than that lineage allows [2]. VAST's answer is a disaggregated shared everything design, which he contrasts with shared nothing, and which he claims is "the first embarrassingly parallel system, distributed system that's ever been built": a cluster where "no two machines have to talk to each other at all to do anything" [4]. He connects this to a longer preoccupation with parallelism, noting from reading Thinking Machines that there had not been a successful parallel computer until CUDA was coupled with the GPU, and that CPU cores were managed monolithically by comparison [4]. The payoff, he says, shows up hardest outside storage, in data types historically crippled by intermachine chatter, vector databases being the clearest example [4].

He refuses the hardware-versus-software framing. "it's not hardware or software, it's how the two come together" [4], and he compares the current moment to early Apple town halls where the processors themselves mattered [4]. The same instinct shaped the go-to-market: he interviewed someone from a software-defined storage company who struggled with the permutations customers wanted, and concluded that Dell, NetApp and Pure were still "selling circles around these technologies" because customers vote with their wallets for simplicity [1]. Concentrating QA on a single hardware platform while letting the manufacturer sell the appliance was the resolution [1]. On defensibility he is unsentimental: new systems architectures only arrive on a generational basis, people who do not study distributed systems may not appreciate what has been invented, and once you have momentum, "who's going to pay to kind of try to catch us" [4].

On the platform becoming an AI operating system

The layering is deliberate. The original idea was a massively scalable computer that could be programmed against, then years went by building capability upward: storage first, database services about three years in, then the data engine of functions and triggers designed to convert unstructured data into something structured that businesses can work with, which he described as "kind of this thinking machine that we're building" [2][4]. The 5.4 release is the point at which that closes the loop, with roughly 50 to 60 features and, critically, a full distributed runtime and serverless environment supporting at-scale eventing, so that when data lands you add a few functions and the machine contextualises it and stores the metadata where it can be queried [4]. He frames the whole thing through Jensen's AlexNet story: a raw data product, the context that used to come only from annotation, and the compute that refines the raw into something queryable, which maps onto a storage system, a database and a computer [4].

Vectors as a native data type are the second pillar, because embedding models produce an expression of a piece of data that AI tools can use, making vectors the key element for structuring unstructured data through interfaces like LangChain or LangGraph, or plain SQL over vectors [4]. He explains retrieval as nearest neighbour and similarity search, distance calculation between the query and what already exists, which lets teams avoid needing fully omniscient models and instead point a tool at a supplemental source, for instance to hunt a bad clause across a contract set [4]. He justifies the OS label concretely: an SDK to program against, capabilities taken down to the device driver level, storage, database, indexes, eventing and runtime combined into a computer that scales to a data centre and across data centres [4]. His analogy is the RAID controller, a capability nobody buys separately anymore because Intel folded accelerators into the microprocessor, and that folding-in is what VAST is doing [4]. He is candid that the ambition is the point, an idea "that we shouldn't be constrained by how people have built and kind of categorized and segmented their products up until now" [4].

On the new data stack and where the growth is

He sees a whole stack forming with large language models on top of a data platform on top of new AI clouds on top of AI processors, all of it doubling, tripling and quadrupling while a Guggenheim note argued the hypergrowth phase was over for the incumbent big data players stuck below thirty percent [2]. VAST's position is the data layer between the hardware, the clouds deploying it and the applications above [2]. His sharpest distinction is with Snowflake and Databricks, which he characterises as focused on big data, business intelligence and reporting: "you can't take like a a genome and easily put put it in a data warehouse you can't take a video and put it in a data warehouse" [2]. Since unstructured data is roughly twenty times bigger than what fits naturally in a database or warehouse, a deep learning data platform addresses an opportunity twenty times larger than classic big data, and only neural networks and GPUs make dormant archives newly interrogable [2].

He also describes the demand cycle honestly: models are trained on data, then inference infrastructure generates more data in prompts, user data and logs, and the whole thing has to run as a constant process [2]. Asked for a 2024 prediction he simply bet that the hype would not slow down, pointing to the investments customers were planning [2].

On where the AI workloads actually land

Customer selection follows a rule he states plainly: "let's go where the data sources are large, and let's go where the GPU density is pretty high because people will need scale and scalable systems" [4]. That puts VAST behind the world's largest AI model builders and model executors, large enterprises retrofitting generative AI into products and services, webscale vendors launching AI services, and the new GPU clouds such as CoreWeave, Core 42 and Lambda, which he describes as organisations with a tiger by the tail on GPU deployments needing something secure and multi-tenant [2]. He has also cited work with xAI as an example of an organisation making massive investments [4].

On applications, he points to a large government smart city project running sensor and video surveillance data through a RAG engine so an agent can do the work of a security guard watching monitors, with other agents deciding actions such as alerting authorities or writing a report [4]. What made this newly possible is reasoning: perception AI could identify a cat or a ball or a person in a video but could not understand the context in between, and reasoning on GPUs gives far better situational understanding [4]. The same logic drives media work with the NHL and, more broadly, sports leagues and news broadcasters sitting on vast volumes of captured live feeds they did not script, which vision language models can now summarise or turn into personalised fan experiences [4]. He notes similar work in Homeland Security that he cannot discuss [4]. His personal favourite is the vector database side, where customers want two things never previously available: very high performance writes into a vector space, sometimes millions of events per second, and scale without paying "king's ransoms" for memory-heavy indices, against a target of trillions of vectors searchable in constant time [4]. He cites one of the largest technology companies in the world having tried twenty vector databases without finding one scalable or performant enough [4], and an AI lab that built its own event streaming infrastructure for reinforcement learning because Kafka clusters fell over at scale [4]. He expects internet companies to matter here, since AI labs are becoming internet companies and internet companies are becoming AI labs, and anyone handling social content will have enormous volumes to enrich [4].

On selling before you have a product

The most repeated piece of practical advice is that market research and selling are the same activity. With an idea on a piece of paper, no product and himself as the only go-to-market person in the United States, he started picking up the phone, and across the first two and a half years ran something like 500 sales calls, writing every conversation down and funnelling it back to R&D [1]. The result was market research, a large pipeline of beta customers, and an outsized amount of revenue waiting at general availability because they had been selling the whole time [1]. The pitches were tested in public: hyper converged, backup, data lake, and the one that landed was a disruptive scale-out file and object system extensible to other protocols and access methods from VMware to containers to big data [1]. He describes the same period from the culture side, holding sales kick-offs for the engineering team before there were any customers or salespeople, flying out to spend half a day with early prospects, and coming back to a room where half the people looked at him like he had three heads, which he credits with instilling a deep sense of customer centricity [3]. He did two sales kick-offs without a single salesperson [3]. Selling before exiting stealth was unorthodox, and by launch they had sold more than any storage company had in a first year [3]. His marketing emphasis later shifted from technical deep dives aimed at technical buyers toward customer reference marketing as the company scaled [6][7][8], and the early motion was always aimed at customers with petabytes or exabytes of data whose needs he took time to understand deeply [6][7][8]. He also identifies the ChatGPT launch as the accelerant that brought a flood of AI cloud service providers and venture-backed large language model companies [6][7][8].

On rewriting the storage business model

Gemini came out of two problems at opposite ends of the customer range. VAST's minimum system is around 700 terabytes, and for half of the customers he speaks to that is all of their data, so they resist buying it just to get a taste; slicing systems into licences to serve a 100 or 200 terabyte customer meant contorting a hardware-driven P&L and "that kind of didn't feel right" [1]. At the top end, the organisations he expects to be left standing are those that treat infrastructure as a competence, buy a lot, and negotiate below the standard P&L ratio [1]. The answer was to take the hardware off the books, compensate the sales team on software, and decouple the software from the hardware lifecycle, so customers buy software on their own agenda and amortise and refresh hardware on theirs while deployment stays simple [1]. He is emphatic this is not software-defined storage [1]. The pitch is that you can buy infrastructure like a hyperscale company, even below cost by aggregating the buying power of the whole VAST customer community, while deploying it like an enterprise [1]. He describes the customer personas as those buying from Amazon and those who want to be the next Amazon, with the programme built for the latter: "you don't have to necessarily buy from Jeff Bezos in this case, you can become the next Jeff Bezos" [3]. The transition was made as a hard cut on 1 February rather than the two or three years public companies take, the sales team proved surprisingly receptive, and the numbers went up fivefold year over year [1].

Underneath this sits a refusal of the standard scorecard. "we never worry about gross margin but we do worry about what is the cash that we get from a customer experience," he says, citing Amazon running some of the worst gross margins in retail history while dominating the industry, and noting that a thin deal is usually 10, 20 or 30 petabytes, so volume covers it [1]. The conclusion is that focusing on what actually matters rather than what people say you should be measured by means "you don't have to build a business like everybody else you just have to build your business the right way" [1].

On raising money you do not need

VAST reached cash flow positivity by the end of 2020, while comparable storage players had spent between half a billion and a billion dollars getting there, and at equivalent revenue levels VAST's headcount was about half the size of its peers [3]. Customers frequently pay several years of support up front, so the company is not burning cash even while tripling annually [1][2]. That makes fundraising a communications decision rather than a survival one. He recounts a competitor observing on Twitter that VAST talks about customer momentum, which creates investor enthusiasm, which creates a large valuation and a very cheap raise, which loops back into customer enthusiasm, and he embraces the description: "we actually use investment as a marketing vehicle more than you know a sustaining function for the company" [1]. Concretely, going into the Series D with $140 million already on the balance sheet, another $80 million cost about two percent of the company and tripled the valuation to $3.7 billion, leaving VAST valued alongside companies nearly twenty times its size by headcount and revenue [1]. The Series E followed the same logic, with no good application for the money but real value in signalling that something new is happening and in having Fidelity, NEA, Bond and Drive Capital validate it, at a time when most of the market was down and VAST had almost tripled its valuation since 2021 [2]. He values investors who bring more than capital, citing Next 47's investment thesis of providing business development assistance as the investment arm of Siemens [3], and advises founders to choose investors offering strategic value and long-term partnership rather than money alone [6][7][8]. He is also aware of timing luck: the Next 47 round closed right at the beginning of COVID, and it turned out they never had to dig into the rainy-day fund at all [3].

The economics that make this possible are a deliberate choice of customer. Rather than blanketing the market with hundred thousand dollar systems sold to small and medium businesses, VAST went after the biggest companies writing the biggest cheques, with a select number of salespeople earning good commission but a far lower total cost of sale, and initial customer investments approaching $10 million [1]. The results he cites include a $150 million run rate exiting a quarter after 350 percent year-over-year growth, and nearly $100 million of product sold in the first two years, which he says had never been done in the industry [3].

On building the company

His advice to founders is unglamorous. "start-ups are just so much work," and while a good idea is necessary, "execution is the number one element," because "Every single competitor that you enter into the market to displace had more resources than you do, so you have to work quicker and harder than all of them" [3]. He calls himself a consummate startup guy who cannot picture himself in a hundred thousand person company, valuing the agility, autonomy and the ability to break things [3][6][7][8]. The corresponding lesson from the early days is that convention is optional: "you don't necessarily need to toe the line that everybody else does, if the dynamics are different you can carve your own path" [3], and that not rushing worked out fine [3].

On operating discipline the instruction is compressed into two words: "don't trip." Specifically, do not over hire in a way that forces a correction, because corrections have negative effects on the team, and instead run a very consistent, cash conservative business that never has to make quick reversals from an engineering perspective [4]. He is proud of the density of the team, roughly a hundred people hired in a twelve month stretch during the pandemic, an engineering team doubled and slated to double again, and colleagues like Howard Marks whose depth of knowledge he describes as off the charts [3]. He tells the story of a board member seeing a multimillion dollar first order and saying it was impossible, which turned into a challenge not just to prove him wrong but to prove history wrong [3]. The company name itself is an amalgam of fast and big, arrived at because the alternative acronym did not work [3]. Growth then extended geographically, with an Asia Pacific expansion beginning in India under a new sales leader [2], and product portability became a goal in its own right through partnerships with the largest cloud builders and server vendors, so that VAST can function as the data centre data operating system wherever a customer wants to put their data [2].

Takeaways

• Every customer asked for cheaper storage, not faster, which is why VAST aimed at being "a extinction level event for the hard drive" rather than winning a flash speed race [1].
• Technology alone does not create markets: pair a new architecture with an application disruption, or there is no business [1].
• Sell before you have a product. Roughly 500 sales calls over the first two and a half years produced the market research, the beta pipeline and revenue waiting at general availability [1].
• Gemini took hardware off the books and paid the sales team on software, letting customers buy infrastructure like a hyperscaler and refresh hardware on their own schedule; the switch was made as a hard cut and revenue went up fivefold [1].
• Gross margin is the wrong scorecard; cash generated per customer is the right one, and volume covers thin deals the way it does for Amazon [1].
• With cash flow positivity, funding rounds function as marketing: roughly two percent dilution tripled the Series D valuation to $3.7 billion [1][2].
• The disaggregated shared everything design is claimed as the first embarrassingly parallel distributed system, where no two machines need to talk to each other, which is what makes trillion-scale vector search and real-time inserts feasible [4].
• 5.4 added a full distributed runtime with serverless eventing plus vectors as a native data type, completing the storage, database and compute triad Denworth calls an AI operating system [4].
• Deep learning data platforms address an opportunity around twenty times larger than classic big data, because genomes and video never fit the data warehouse [2].
• Operating rule for a fast-growing company: "don't trip", meaning do not over hire into a correction and stay cash conservative so engineering never has to reverse course [4].