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

Eon co-founders Ofir Ehrlich and Gonen Stein went on No Priors with Elad Gil to make a straightforward argument: models and compute have near-zero switching costs; historical enterprise data does not. The evidence they point to is real. Google paid roughly $10M for Spirit Airlines' operational data out of bankruptcy, a competing lab bid against them, and Ehrlich says labs are now approaching hedge funds for financial data. The mechanism is that agent training needs traces of actual organizational decisions, messy handoffs, and edge cases that synthetic data can't replicate. Harvey's legal dataset, Ehrlich notes, barely clears the bar because it already looks synthetic.

Two founders selling data infrastructure telling you data is the only moat are pitching, not analyzing. And one distressed-asset acquisition doesn't make a market. Most enterprise data is duplicated CRM exports and Slack logs nobody would pay for. The more interesting question is whether labs are paying because enterprise data demonstrably produces better models, or because they're hoarding out of competitive anxiety and the price will get rationalized when synthetic data closes the gap. Nobody on the episode shows the model trained on Spirit's data outperformed one that wasn't.

The open-source angle is the part Ehrlich and Stein gesture at without spelling out. If models commoditize and data is the differentiator, a fine-tuned Llama or Mistral on your proprietary operational data beats renting a frontier model you share with every competitor. You don't buy that moat; you already own it, sitting on backup tapes. Whether your data is clean enough to train on is the actual work. That's convenient for what Eon sells, but the strategic logic holds regardless.

The security argument is less novel than they frame it. An agent with legitimate credentials that can drop a table is a permissions design failure, not a new threat category requiring new detection infrastructure. Scope agent identities tight, separate read from write, log every non-human action, and put a human gate on destructive operations. Your DBAs already know this discipline. The genuinely hard part is the agent-spawning-agent chain: once a no-code tool built by your marketing team is calling three external agents, the audit trail is gone and the blast radius is machine-speed.

The most actionable line in the episode is easy to miss. Ehrlich says "we're not in the time of token maxing anymore." That's the shift from pilot to P&L. Inference is now a procurement line item somebody defends in a quarterly business review. Long, chatty agent loops that looked clever in the demo get value-engineered down. Retrieval that fetches only what's needed beats dumping the whole knowledge base into context. That reprices architecture decisions your team already made.

The low-regret moves don't require resolving the moat debate. Tighten agent permissions and non-human identity logging this quarter. Start measuring cost-per-task on your agent workflows before procurement does it for you.

The call: By mid-2027, at least one major lab will publicly disclose a paid acquisition of real-world enterprise operational data, beyond web scrapes and publisher licensing, explicitly for agent or model training. Medium confidence. The Spirit deal and the competing bid show the behavior is real and contested. The main path to being wrong is that these deals stay quiet because they're legally messy and competitively sensitive, but the same pressure that made Google outbid a rival also creates incentive to signal to other data-holders that there's a buyer. Disclosure is the likelier outcome.

Your draft

Your draft

Eon co-founders Ofir Ehrlich and Gonen Stein went on No Priors with Elad Gil to argue that in the AI era, your data is the only moat left. Models and compute have near-zero switching costs. Historical enterprise data does not. The proof points: Google buying Spirit Airlines' data out of bankruptcy for ~$10M, labs approaching hedge funds for financial data, and AI agents becoming an insider ransomware threat that moves at machine speed.

This is a briefing, not a decision. So the question for a dev manager shipping AI into production: does any of this change what your team builds next quarter, or is it a vendor pitch wearing a trend?

Reversibility: Mostly Type 2. Trying a data-cataloging layer or an agent-access-control tool is reversible. The one Type 1 lurking here is agent permissions architecture. Get that wrong and you are cleaning up after a machine-speed incident, not rolling back a config.

Forcing function: None hard-dated. But the security claim ("six months ago nobody discussed this, now every leader has a story") is the part with a clock on it, if it is true.

---

The Skeptic

Two founders selling data infrastructure tell you data is the only moat. Of course they do. That is the pitch. The $10M Spirit deal is real and interesting, but one distressed-asset buy does not make a market. Airline pricing and ops data is genuinely rich. Most enterprise data is a swamp of duplicated CRM exports and Slack logs nobody would pay for. The "every leader has an agent horror story" claim is exactly the kind of unfalsifiable urgency a vendor manufactures. Where are the incident reports? The named breaches? For a PM: they are telling you the house is on fire, and they sell extinguishers.

The Researcher

The durable claim here is synthetic data hitting a ceiling for agent training. Ehrlich's line on Harvey's legal dataset ("you don't find too many good data sets that don't look like really synthetic data") points at something real: agents need traces of actual organizational decisions, messy handoffs, edge cases, not clean generated examples. That is why labs are bidding on operational data specifically. But note what is NOT claimed: nobody says a model trained on Spirit's data beat one that wasn't. The moat thesis is asserted, not measured. For a PM: the labs think real-world process data is scarce and worth money, and they are voting with acquisitions.

The Compute Pragmatist

Ehrlich buried the most actionable line in the episode: "We're not in the time of token maxing anymore." Translation for your team: the era of stuffing giant contexts and letting the agent wander is over because every token costs money and enterprises now count. This is the shift from pilot to P&L. Your inference bill is now a procurement line item somebody defends in a QBR. That reprices architecture decisions you already made. Long, chatty agent loops that looked clever in the demo get value-engineered down. Retrieval that fetches only what is needed beats dumping the whole knowledge base into context.

The Builder

Forget the moat philosophy. What ships Tuesday? The agent-permissions problem is the one real thing here. An agent with legitimate credentials that can drop a table is a permissions design failure you can fix now. Scope agent identities tight, separate read from write, log every non-human action, put a human gate on destructive operations. You do not need Eon to do this. You need the discipline your DBAs already have, applied to service accounts that now think for themselves. The "agents spawning agents" chain is the genuinely hard part: once a no-code tool built by marketing calls three external agents, your audit trail is gone.

The Open-Source Advocate

The whole moat argument quietly hands the win to open weights. If models commoditize and data is the differentiator, then a fine-tuned Llama or Mistral on your proprietary data beats renting a frontier model you share with every competitor. That is the strategic read the founders gesture at but do not spell out: Elad Gil's aside about open-source models fine-tuned against proprietary datasets is the actual playbook. You do not buy the moat. You already own it, sitting on backup tapes. The question is whether your data is clean enough to train on, which is, conveniently, what Eon sells.

---

Where they disagree

The Skeptic and the Researcher split on the moat. The Researcher sees labs paying real money for real-world process data, which is evidence. The Skeptic notes nobody has shown that data produced a better model, so the price could be labs hoarding out of FOMO, not proven value.

The Builder and the vendor pitch split on the security fix. The founders frame agent threats as a new category needing new detection infrastructure. The Builder says most of it is old-fashioned least-privilege access control that your team can implement without buying anything.

And the Compute Pragmatist quietly undercuts the whole "deploy AI immediately" board panic: if every token now has to justify itself, the forward-deployed, ship-it-fast motion collides with the ROI discipline the same founders admit has arrived.

What this actually hinges on

Is real-world enterprise data a durable moat, or a temporary scarcity labs are overpaying to fix? If it is durable, the open-source-plus-your-data play is the correct long-term architecture and you should start treating your data pipeline as a strategic asset now. If it is temporary, this is a land grab that gets repriced the moment synthetic data closes the gap.

For your team, the low-regret moves are Type 2 and don't depend on resolving that: tighten agent permissions and non-human identity logging this quarter, and start measuring cost-per-task on your agent workflows before procurement does it for you.

---

Prediction: By the time OpenAI, Anthropic, or Google ship their next frontier model class in the first half of 2027, at least one major lab will have publicly disclosed a paid acquisition of a real-world enterprise operational dataset (beyond web scrapes and licensed publisher/media content) explicitly for agent or model training.

Confidence: Medium. The Spirit deal and hedge-fund approaches show the pattern is already forming.

Why: Google already paid ~$10M for Spirit's data out of bankruptcy and a competing lab reportedly bid against them, which means the behavior is real and contested, not one-off. The stated mechanism is that agent training needs traces of actual organizational decisions that synthetic data cannot fake, and labs are running short of fresh web text. The opposite outcome (labs staying quiet about such buys) is plausible because these deals are legally messy and competitively sensitive, but the same competitive pressure that made Google outbid a rival also creates incentive to signal to other data-holders that there is a buyer, so at least one disclosure is the likelier path.

Revisit by 2027-06-30: We're right if a major lab confirms a paid deal for proprietary operational/transactional enterprise data for training. We're wrong if the only disclosed data deals through that window remain web content, publisher licensing, or media archives.

One more thing worth saying plainly. The most useful line in the episode wasn't about moats. It was Ehrlich admitting the token-maxing era is over. That is the shift that actually lands on your team's roadmap, and it arrived while everyone was still arguing about who owns the data.