Google Paid $10 Million for a Bankrupt Airline’s Data. That’s the Moat.

In August 2026, Spirit Airlines went through bankruptcy, and Google picked one thing out of the liquidation. It did not take airplanes or routes. It paid $10 million for the airline’s data and, according to Ofir Ehrlich, is using that data to train models. Host Elad Gil adds a rumor from the bid process: the other bidder was Mercor, an AI company. A bankrupt carrier’s internal records drawing competing offers from two AI firms is more revealing than the price itself.

Ehrlich and Gonen Stein, co-founders of Eon, join Gil on No Priors to unpack what that purchase means. Their last company, CloudEndure, was sold to AWS and became Application Migration Service. Eon looks like a cloud backup and disaster-recovery business, but the product is built as a data foundation for the AI era: map and classify data across hyperscalers, ingest structured and unstructured sources, store them cheaply for protection and recovery, and then let customers query, search, and run models on top.

Ehrlich’s starting point is blunt. When Eon launched, he jokes, he looked like “a crazy person starting a non-AI company in the AI world.” The tailwind flipped that bet. Models and compute matter, he says, but they are relatively ephemeral, with almost zero switching costs. Whether you run a hotel chain, a food chain, or a technology company, “the most valuable thing that you have is actually your data.”

Gil has heard “data is the new oil” for years and stayed skeptical. Post-training and reinforcement learning changed his mind. Companies such as Applied Computing now fine-tune models against specific datasets, which finally gives proprietary data a clear path into production. The question about Spirit is what anyone would actually *do* with an airline’s records: a customer-support agent, internal documents, something else?

Stein’s answer has two layers. First, Google already operates in travel; the dataset lets it train, understand, and monetize that domain further. Second, Spirit is a large enterprise with hierarchy, middle management, and workers collaborating — the kind of real operating sample that public datasets almost never capture. Hugging Face had just released a legal dataset, he notes, but “you don’t find too many good data sets that [don’t] look like really synthetic data.” Public data does not teach a model how a company works.

That scarcity is why agent companies either go to customers or buy data to train in-house before the first live interaction. Ehrlich hears the same question constantly if you are a “tech data CEO”: are you willing to sell your data? Labs are going to Wall Street to buy hedge-fund data and learn how to map and analyze companies. Years of tapes sitting on a shelf collecting dust have become gold. Everyone has access to the same new tools; people and approved data are the remaining advantages.

Eon’s business starts from a mismatch. Enterprises already have the data; it is locked, expensive to reach, and scattered across business units. Stein walks through the typical scene. The CEO has been playing with ChatGPT. The board has decided AI is real and data is the new oil. They tell the data-team leader to activate it. The data-team leader knows every new Silicon Valley tool. Then they hit the actual question: where is the data?

It lives with business-unit owners who may not even know the data team exists. A unit leader has people running systems for twenty years, including ones nobody fully understands, holding production data and sensitive information. Stein’s image is familiar: there is always a server nobody can explain, still plugged in, that no one will turn off because nobody knows what is on it. Extracting anything means pulling engineers off other work, compromising security, compliance, or uptime, and then parking the extract somewhere inefficient.

The incentives are opposed. The data team is tasked with getting the data. The business owner is tasked with keeping systems running and making sure “the salary of the CEO” does not accidentally end up in someone else’s training set. Eon’s answer is to find, classify, and map data across the organization, add context and a semantic layer, then continuously ingest what is relevant without touching production — with audit trails so sensitive fields are not shared by accident.

Gil summarizes the product as three jobs: aggregate historical and current data, mask PII or set permissions, and expose the result to AI applications. Ehrlich’s punchline is the irony. Customers already have this data. Eon converts locked, expensive copies into a cheaper foundation with mapping, classification, access control, and a hook into AI workflows.

The security section is the densest stretch of the episode, and it starts with a scar from the AWS years. A very large customer was hit by ransomware. Both sides thought the managed disaster-recovery service had them covered. After the fact, Ehrlich learned the customer had never mapped, classified, or tagged resources properly, so the protection never covered what mattered. Sixty percent of the environment was exposed. That failure is one reason they started Eon: backup without inventory is empty.

The threat model has shifted. “Up until now, the concerns came from human threats,” Ehrlich says. The same class of damage now comes from non-human actors: agents with legitimate access and legitimate permissions. A table is suddenly dropped. Detection still looks for irregular write patterns and entropy changes, then supports granular, fast recovery — but the velocity is extreme. Six months earlier, nobody raised this with him. A few months later, every leader he met was either afraid it would happen or had already lived it.

Stein describes the organizational mess around that fear. CIOs are pushed by boards and CEOs to enable AI now and not block anyone. Non-technical employees build apps with tools such as Lovable and upload company data with no sense of security, compliance, or who will use it. Their agents call other agents — OpenClaw talking to WhatsApp, the internal network, and other applications. The result is “a complete set of actors inside an organization, not bound by the rules of the organization and not necessarily running within the premises,” still handling sensitive data.

The security market is already voting with its feet. Stein sees an “infinite amount” of cybersecurity companies in non-human identity, because NHI has become a number-one or number-two problem. Endpoint security is repeating an older cycle: agents now run on laptops, straddling internal networks and external services, and CIOs cannot see the chain of responsibility. Everyone in the org can become a builder — social media, legal, finance — which he calls both amazing and the reason “we live in very interesting times.” Ehrlich’s design rule is assume breach, malicious or not, and be ready to act.

Gil asks whether a decade of warehouses, dashboards, ETL, and BI — a stack built for humans asking predefined analytical questions — still holds when agents reason over larger sets and hop across SaaS apps. Stein’s answer on dashboards is the opposite of the usual take: there will be *more* of them. Coding agents already write much of the world’s code, agents activate other agents, and humans cannot interrogate every non-human actor. Dashboards become the only way to see what is going on.

The plumbing underneath is too limited, he argues, because each tool solved one niche. He walks downstairs for coffee, pays by card, and a transaction lands in some database. Today someone extracts it, someone else processes it later, and nobody shares context. That was acceptable when you only did with data what you intended at write time. Now the value sits in combination: one person holds New Yorkers who love burgers, another holds New Yorkers who love pizza, and neither knows the other exists. Together, they have a list you can post-train on.

Ingestion volumes are “absolutely insane,” much of it generated by agents, so noise and value arrive together. Tools have to understand data from many locations, clean the noise, and keep it usable. Fivetran, dbt, and Monte Carlo were excellent, Stein says, but very specific. Databricks, which he calls “one of the most incredible companies on the planet,” is reinventing itself: more data, more sources, its own agents and databases, trying to own the path from creation to use. Three years ago, nobody used data this way.

Cost is the new constraint. “We’re not in the time of token maxing anymore,” Stein says. Pay millions if the value is there, but store efficiently and spend tokens only where they return something. Ehrlich’s product goal matches that world: automatically show builders and agents what exists, ingest without a manual pipeline per app, and keep control over what those systems create.

Before Eon, the founders lived through the last infrastructure shift. CloudEndure helped enterprises move thousands, tens of thousands, or hundreds of thousands of servers onto AWS, Azure, and GCP; the product was OEM’d into consoles, then became Application Migration Service after the AWS sale. That work was heavy on both technology and people. Ehrlich’s comparison is that AI is “like that, but on steroids.”

The difference is speed and loss of control. Transformations happen much faster, and customers lose the plot so quickly that fear becomes an inhibitor: they pause because systems might break, data might leak, or IP might escape. Cloud was abstract — “someone else’s computer,” hard to explain to a grandmother. AI is not. After the ChatGPT moment, everyone had felt what it could do. CEOs, boards, and shareholders now push: use AI, or become irrelevant. People adopt it for the value and for the fear.

How companies buy software is changing with that pressure. Forward-deployed engineers used to look like a Palantir specialty nobody else understood; now everyone does it. Large enterprises want AI immediately, cannot wait a year or two for internal process, and import Silicon Valley engineers with playbooks to reshape the org. Sales cycles shrink. Cognition is Ehrlich’s example of stacking motions: start with product-led growth, add an FDE motion into banks, and have the vendor’s engineers do the work customers do not want to do.

A more radical path is buying the company instead of waiting for it to adopt. Ehrlich points to Long Lake: acquire a slow traditional business, turn it into an AI company, raise margins, and make everyone win. He stays sober about the label “revolution.” Most companies still do not use AI. They know data matters, know existing processes are mundane, are scared, and still have to move. “I know it’s a very wild ride,” he says, “but I think… the world… is going to be better because of that.”

The episode is about large-company balance sheets, but the arithmetic is the same for a smaller team. Models depreciate. Compute is rented. Data is yours — if you can find it, classify it, and keep it from leaking into an agent you cannot see. Spirit’s records were worth $10 million because they were complete enough to identify and bid on. A company that cannot list what it holds is not ready to talk about moats.

The AWS ransomware story is the other practical test. Protection that was never mapped or tagged did not exist when it was needed. The same discipline now has to cover agents: separate identities, least privilege, audit logs, and a rule that non-technical builders cannot upload company data into Lovable-class tools by default. Unlocking archives for training only after PII and financial fields are handled is not a legal afterthought. Bare data is a liability; governed data is the asset Google just paid for.