Rethinking Legacy Data Infrastructure with Eon Co-Founders Ofir Ehrlich and Gonen Stein

Summary

• The most tradeable frame in the episode: models and compute are commodities, accumulated enterprise data is the moat. Gonen Stein’s formulation — models and compute are “relatively ephemeral, almost zero switching cost,” while “the most valuable thing that you have is actually your data.” Exhibit A: Google just bought bankrupt Spirit Airlines’ data for $10M (“They didn’t buy airplanes. They bought the data”), with Elad Gil noting the rumored rival bidder was Mercor — multiple AI companies bidding on a bankrupt airline’s dataset.
• A structural bid for “dead” corporate data is forming because good real-world training data is scarce. Stein says tech CEOs “constantly get questions — are you willing to sell your data?”, labs are “going through Wall Street” to buy hedge fund data, and people still use the public Enron data as real company data for lack of alternatives. Data that sat “on a shelf collecting dust” is repricing; expect more bankruptcy-estate data buys.
• Eon’s arbitrage: customers already own the gold, but it’s locked, scattered, and expensive to access. Eon’s pitch is a cloud “data foundation” that maps and classifies data across hyperscalers, ingests it continuously without compromising production or compliance, controls access to PII and other sensitive data, and exposes it to AI workflows — solving the misaligned incentives between data teams tasked with AI and business-unit owners guarding 20 years of systems “everyone’s afraid to turn off.”
• AI agents are the ransomware threat “on steroids” — legitimate access and permissions, extreme velocity. Stein recounts an AWS-era customer who was 60% exposed to ransomware because resources weren’t mapped, classified, and tagged; now the same threat comes from “non-human actors… with legitimate access… legitimate permissions” and “all of a sudden a table is dropped.” Stein says six months ago nobody would discuss it — now every leader he meets either fears it or has lived it. Ehrlich adds: “We need to assume breach, whether it’s malicious or not.”
• Non-human identity security is exploding as a category, and dashboards multiply rather than die. Stein’s contrarian call: agents activating agents make chains of responsibility “almost impossible” to track — hence an “infinite amount” of NHI security startups, endpoint security’s return, and more dashboards as “the only way to figure out what they have going on.” Non-technical builders using tools like Lovable can create “a complete set of actors inside the organization not bound by the rules of the organization… but handling sensitive data.”
• The AI transition is the cloud migration replayed faster — and fear has flipped from enabler to inhibitor. Ofir Ehrlich, who lived through the CloudEndure/AWS migration era, says AI is “like that, but on steroids”; customers are “losing control to a point where that’s becoming an inhibitor” — pausing deployments over leak and IP risk. Meanwhile GTM is being rewritten: forward-deployed engineers went from a Palantir oddity to everyone’s motion, PLG suddenly works for dev tools (Cognition cited), and Long Lake proposes buying slow companies outright to convert them into AI companies — “arbitrage.”
• The hedged bottom line: “we just started — most companies still don’t use AI.” Legacy plumbing (Fivetran, dbt, Monte Carlo) was built for single-purpose questions; token costs mean “we’re not in the time of token maxing anymore,” and Databricks is “reinventing themselves… if you can’t beat them, join them.” The buildout ahead is the thesis.

Deep dive

1. Data is the moat — and bankruptcy court is the new data marketplace

• Stein’s opening frame, delivered as self-deprecation: “I’m this crazy person starting a non-AI company in an AI world” — until the tailwind made data the most important asset. Models and compute carry “almost zero switching cost”; whether you’re a hotel chain or a tech company, “the most valuable thing that you have is actually your data.” Proof point from two days prior: Google bought bankrupt Spirit Airlines’ data for $10M — “They didn’t buy airplanes. They bought the data.”
• Gil’s addition: the rumored other bidder was Mercor — multiple AI-world companies competing in a bankruptcy process for an airline’s enterprise dataset. He asks whether out-of-bankruptcy data buys become a pattern; Stein says they’ve already seen multiple use cases and expects the trend to grow, including labs “going through Wall Street” to buy hedge fund data.
• Why the scarcity bid: real-world training data barely exists. Harvard just released a legal dataset, but people still use the public Enron data as “real data from a company [to understand] how a company works.” Spirit’s data works both as airline data and as a large-enterprise corpus — “hierarchy, middle management, top management and workers working together.” AI “flattens the playing field,” leaving people and accrued data as the durable edges.

2. The irony Eon monetizes: enterprises already own the gold, locked in silos

• Stein’s product overview: a cloud data foundation that maps and classifies data across hyperscalers (“what they have, where they have it, what’s sensitive”), ingests structured and unstructured sources, and serves double duty — cost-effective protection/recovery plus making the data queryable and usable by LLMs. The kicker: “customers already have this data… it’s locked. It’s not accessible and usually it’s very very expensive.”
• Ehrlich’s incentives parable: the data team leader is tasked by CEO and board — “even the boss is playing with ChatGPT” — but the data lives with business-unit owners with opposite incentives, spread across 20 years of systems, including “this server that no one knows what it’s doing… that everyone’s afraid to turn off,” with the risk of “the salary of the CEO” leaking into a training set.
• Eon’s resolution: classify, build a semantic layer, continuously bring in relevant data “without compromising production, without compromising security compliance,” with an audit and classification/access controls so sensitive PII and financial data are not accidentally shared. Gil’s summary of the stack: aggregate historical and current data, mask or permission it, and expose it to AI models.

3. Agents are the insider threat “on steroids”

• Stein’s AWS-era scar tissue: a very large customer he thought was protected by the disaster-recovery service was 60% exposed to ransomware because resources weren’t properly mapped, classified, and tagged — a founding pain point for Eon. The new version: “non-human actors… that essentially have legitimate access to the environment with legitimate permissions… and all of a sudden a table is dropped.” Detection methodology carries over (irregular write patterns, entropy changes), “but the velocity of that happening is extreme.”
• Stein on how fast sentiment turned: “six months ago, no one would even discuss with me” — now every leader either fears it or it happened to them personally. External attacks are easier for adversaries, and approved internal agents add a second front: “I no longer decide what’s really running on my data.” Ehrlich adds: “We need to assume breach, whether it’s malicious or not.”
• Stein on the builder problem: social-media, finance, or legal employees can become builders; non-technical users may build with Lovable and put company data there, while the agents they create may use other agents they cannot evaluate — “a complete set of actors inside the organization not bound by the rules of the organization and not necessarily running within the premises of the organization, but handling sensitive data.” His verdict: “It’s a good thing and a bad thing that everyone inside an organization can become a builder.”

4. The agentic stack: more dashboards, NHI boom, and rebuilt plumbing

• Gil asks whether dashboards go away in an agentic world; Stein’s answer is the reverse — more dashboards, because tracking non-human identity chains “becomes almost impossible,” which explains the “infinite amount” of NHI security companies and endpoint security’s second act: agents on laptops, OpenClaw connected to WhatsApp and internal networks simultaneously.
• Why old tooling breaks — the coffee example: Ehrlich describes Gonen buying coffee, with the transaction hitting a database; someone extracts it, another process later processes it elsewhere, and the context is lost. Fivetran, dbt, and Monte Carlo were “incredible” but purpose-specific. Then the pizza example: one team holds New York’s burger-lovers, another its pizza-lovers, and they don’t know they can find the people on both lists — with clean, contextualized data “you can start asking your data intelligent questions.”
• Volumes are “growing out of proportion,” much of it agent-generated noise with “a lot of value in the noise as well.” Ofir points to Databricks — “one of the most incredible companies on the planet,” in his opinion — “reinventing themselves all the time”: “if you can’t beat them, join them — we’ll build our own agents.” And cost discipline is back: “we’re not in the time of token maxing anymore” — pay millions if needed, but extract value per token.

5. Cloud migration redux — faster, scarier, and rewriting go-to-market

• Ehrlich’s comparison from CloudEndure — later sold to AWS, after supporting large-scale migrations including through Azure and GCP OEM integrations — is that AI is “like that but on steroids.” These transformations are happening faster, and customers are “losing control to a point where that’s becoming an inhibitor, not an enabler,” pausing over fears of data leaks and IP escape.
• Stein on why adoption pressure is stronger this time: cloud was “just someone else’s computer… hard to explain to my grandmother,” but since the ChatGPT moment everyone understands AI — so C-levels, CEOs, boards, and shareholders push from both value and fear, “otherwise we’re irrelevant.”
• Ehrlich flags his own reversal — he used to argue PLG doesn’t work for dev tools; now it is “super hot.” Cognition first used a PLG motion — “we use that approach at Eon” — and then added FDEs going to banks; forward-deployed engineers went from a Palantir curiosity “no one really did understand” to everyone’s motion, shrinking sales cycles. Long Lake goes further: buy the slow company, transform it into an AI company — “capture arbitrage, make higher margins.”
• Stein’s closing hedge: “we just started — most companies still don’t use AI”; adoption is scary, “but you have to do it.” Ehrlich says everyone will eventually go through it and, “in my opinion,” the world will be better because of it.