When Your AI Company Gets Banned: How Delphi Survived the OpenAI Crisis

There was a day last year when Dara Ladjevardian thought Delphi was over.

The company had been banned from OpenAI and Anthropic simultaneously. Everything was breaking. The infrastructure was down. He couldn’t reach anyone to explain what happened or negotiate a path forward. For a few hours, Dara genuinely believed the company was done.

What saved it wasn’t luck. It was a philosophical stance he’d been studying: anti-fragility.

“There was one time last year where we got banned from OpenAI and Anthropic. And that was a tough day because I thought the company was literally over,” Dara recalls. “So it’s just like everything in the product is breaking. I can’t get in contact with anyone from OpenAI and Anthropic. We’re done.”

The reason for the ban was specific. Delphi had helped a political figure create a digital version of themselves, which violated the terms of use of major model providers. This is the kind of edge case that catches startups off guard. You didn’t set out to help political figures. You built a platform for experts to share their knowledge. But you didn’t have explicit rules that said “no political figures,” and a request came in from someone with resources and intent.

The Moment of Clarification

Crisis is clarifying. In that moment, with the company broken and no contact to the providers, Dara had to answer a fundamental question: Is Delphi actually dependent on these platforms, or is Delphi just temporarily using them?

The answer determined everything.

If Delphi was just a thin wrapper around OpenAI’s API, then the ban was curtains. But if Delphi had some independent moat—some structural reason to exist beyond “we use the best LLM”—then the ban was painful but survivable.

What saved the company was that Delphi’s core innovation isn’t the LLM. It’s the knowledge graph. The architecture that represents how a human thinks. That architecture is Delphi’s IP. The LLM is the execution layer, and while the best execution layer is table stakes, it’s not the defensibility.

“We found a lot of clever ways to get back up and eventually get unbanned from those platforms,” Dara says. But the real learning was deeper.

The Anti-Fragility Principle

Dara had been studying Nassim Taleb’s “Anti-Fragile,” which argues that some systems don’t just resist stress—they get stronger from it. A muscle breaks slightly, then heals stronger. A bone fractures, then ossifies. Antifragile systems benefit from chaos.

The intuition is to prevent chaos. Build redundancy. Plan for failure. Those are all good. But true anti-fragility goes further: it’s building systems that improve because of chaos.

For startups, this means asking: what hard things would make us better? What pressures would force us to make architectural decisions that are good anyway?

Being banned from OpenAI forced Delphi to think about infrastructure independence. It forced them to evaluate whether they’d built a product or a feature on someone else’s platform. It forced them to ask whether the knowledge graph was actually defensible or whether it was just a nice layer on top of a commodity LLM.

“This is actually the norm. And this is going to happen many more times,” Dara now believes. Regulatory pressure. Dependency chain breaks. API changes. Bans for unexpected reasons. These aren’t anomalies for AI startups. They’re the operating environment.

The companies that survive aren’t the ones that planned perfectly. They’re the ones that built systems that get stronger when things break.

What Anti-Fragility Looks Like in Practice

For Delphi, anti-fragility meant several things:

First, it meant having architectural optionality. A knowledge graph representation of a person doesn’t inherently require OpenAI or Anthropic. It could theoretically use other models, open-source models, or even custom infrastructure. When you’re forced to diversify, you discover that your core value isn’t dependent on any single provider.

Second, it meant having institutional credibility. A young company getting banned by major AI providers and then unbanned requires someone to vouch for you or some reason to reverse the decision. Delphi had built enough positive intent with the AI community that when they explained the situation, providers were willing to reconsider.

Third, it meant having a clear policy for the future. Instead of hoping this never happens again, Delphi now explicitly disallows political figures from the platform. This isn’t a feature limitation—it’s a structural boundary of the company. It’s saying: “We made a mistake that violated our values, and we’re committing to not make it again.”

Notice what didn’t happen: Delphi didn’t say “we’ll never allow controversial figures” or “we’ll be super permissive and hope for the best.” They found the actual principle that matters to them—human-verified expertise, not political manipulation—and enforced it.

The Gym Metaphor

Dara’s favorite way to think about this comes from his book recommendation: “It’s like going to the gym, but for your mind. So that the next time you experience something hard, you’re already used to the 50 pound weight and now you can handle the 60 pound weights.”

A company that’s never faced a crisis doesn’t know its own breaking point. A company that’s faced one crisis and survived knows it can survive crises. Each crisis, if survived, builds institutional muscle.

This doesn’t mean seeking chaos. But it means not treating close calls as aberrations to prevent. It means asking: what did this teach us? How do we get stronger from this?

The Longer-term Implication

For founders in the AI space, this is not abstract. Your company likely depends on:

• API access from major model providers
• Cloud infrastructure from a handful of vendors
• Data that might face legal challenges
• Regulations that are still being written

Any of those dependencies could be disrupted. The question is: does your company have structural resilience, or is it just running on borrowed time with borrowed infrastructure?

Delphi’s lesson is that the knowledge graph matters more than the LLM. The temporal reasoning matters more than the model size. The human verification matters more than the API access. If you can orient your company toward defensibility that lives in something you built, then external shocks become tests of resilience, not existential threats.

“Now I’m prepared as you know, knock on wood,” Dara says. He’s not saying nothing bad will happen. He’s saying the company can absorb it, learn from it, and come out stronger.

FAQ

Why would OpenAI and Anthropic ban a platform for helping a political figure?

Both platforms have explicit terms of service prohibiting use of their APIs for political campaigns or political figures. The intent is to avoid being seen as taking sides or enabling manipulation at scale. A political figure using Delphi to reach voters would be a clear violation. The providers took action to enforce their terms.

How did Delphi get unbanned?

Dara doesn’t go into specifics, but it likely involved: (1) clarifying that the political use was an edge case, not intentional; (2) establishing clear new policies to prevent it; (3) demonstrating that Delphi’s actual use cases (experts, founders, investors) are legitimate. Major platforms usually prefer to work with companies willing to enforce policy over banning them permanently.

What does “anti-fragility” mean for a startup?

Nassim Taleb’s anti-fragility means systems that benefit from volatility and stressors. For startups, this means: knowing your actual dependencies vs. your perceived dependencies, building architectural redundancy in critical areas, and treating crises as opportunities to strengthen the business, not just as problems to survive.

Are other AI startups at risk of similar bans?

Yes. Any startup relying on API access from OpenAI, Anthropic, Google, or other providers is potentially at risk if their use case violates terms of service or if political/regulatory pressure increases. This is why architectural independence and policy clarity matter.

Should startups avoid using closed APIs?

Not necessarily. Using the best available infrastructure is rational. But you should understand which parts of your defensibility live in your code versus which parts live in your API access. If 100% of your moat is being allowed to use someone else’s API, you’re vulnerable.

What’s the difference between anti-fragility and just getting lucky?

Anti-fragility is structural. It’s built into the design. Delphi got unlucky and then survived because its architecture is robust to provider changes. A company that’s lucky and then gets unlucky in a different way will fail. Anti-fragility is preparation for the class of problems, not the specific problem.

How do you intentionally build anti-fragility?

Ask: What could break this company? Then, instead of just preventing it, ask: If this happened, would we be forced to build something better? If the answer is yes, you might intentionally do it now.

What about other dependencies—data, talent, regulations?

Same principle. If your data comes entirely from one source, you’re fragile. If your product only works in one regulatory jurisdiction, you’re fragile. Anti-fragility means building options and decoupling where possible.

Did the ban change Delphi’s product strategy?

Yes. It forced them to examine whether they were a knowledge graph company or an LLM wrapper. They realized they were the former, which meant they could be more independent. It also prompted the human-verification-only policy, which is now core to the brand.

Is Dara’s optimism about handling future crises justified?

Partially. He’s correct that having survived one, the company is more resilient. But each crisis is unique. Better to have institutional confidence and clear policies than to assume you can handle anything.

What’s the lesson for founders outside AI?

Anti-fragility applies everywhere. SaaS companies reliant on one distribution channel, hardware companies reliant on one factory, mobile apps dependent on app store policies—all vulnerable. The question is: where is your actual defensibility, and what could break it?