The Day After: AI Companies Prepare for the Worst (While Others Say It Won't Happen)

Imagine waking up to find your bank isn’t responding. The app won’t open, the ATM is out of service, the internet keeps dropping. Then the power goes out. On the news, when you finally manage to watch it, someone says the word everyone was dreading: AI.

It sounds like the plot of a dystopian series. Yet according to an Axios report published on October 9, it’s one of the scenarios the companies building artificial intelligence are rehearsing. In private.

And the most interesting part isn’t the scenario itself, but the reactions it triggered.

What the Axios report actually says

The Axios report claims that senior executives at Anthropic, OpenAI and other companies in the industry are gaming out the “day after” a catastrophic AI-related event. Not the catastrophe itself, but what would come next.

The scenario considered most likely is a large-scale cyberattack that could shut down:

• access to financial services
• internet connectivity
• essential services such as power and water

And after the event, what are these companies afraid of? Not just the direct damage, but:

• a violent public backlash against AI and the people who build it
• a political revolt, with calls for much stricter rules
• a Congress that, in the heat of the emergency, acts quickly and with little technical expertise

To prepare, according to Axios, these companies are red teaming worst-case scenarios (putting themselves in the attacker’s shoes to find weak points) and racing to brief lawmakers, so they can shape the rules that will be written after the first disaster. The report adds a sobering detail: many industry insiders believe a major incident could happen within the next 6 to 12 months.

The official responses were cautious. OpenAI confirmed it runs preparedness exercises across a range of scenarios, adding that these scenarios are not treated as inevitable. Anthropic declined to comment.

💡 Note: running worst-case simulations doesn’t mean expecting them. Banks, hospitals and power grid operators have done it for decades. If anything, the news is that AI companies have started doing it with an eye not only on technical damage, but on the political and social reaction.

The TV response: “a huge exaggeration”

A few days later, on Fox Business’s The Claman Countdown, the report was discussed by two guests: John Sculley, former CEO of Apple, and David Steinberg, CEO of Zeta Global.

Both said they hadn’t read the report and weren’t aware of it. And both were skeptical:

• Sculley called it “a huge exaggeration”, adding that we’d need to know who is making these claims.
• Steinberg said he would be “shocked” if things ever got to that point.

Then they shifted the conversation: the real discussions about AI, they said, are about how to make people’s lives better, not how to destroy them.

Nothing unusual so far. But there’s one detail that barely came up on air: Sculley and Steinberg aren’t outside commentators. They co-founded Zeta Global together in 2007, and Sculley is still the company’s vice chairman emeritus. They were on TV, in practice, also to talk about Zeta’s new product, unveiled the day before the report.

I’m not saying they’re acting in bad faith. I’m saying that when people who sell AI to businesses comment on a report about AI risk, it’s worth keeping in mind.

There’s another point too: neither of them works at Anthropic or OpenAI. So their reaction doesn’t refute the report: it only tells us that they didn’t know about it.

The paradox: those who prepare and those who downplay

Let’s put the two together:

Here’s the paradox: the companies building the most powerful models are preparing for the worst-case scenario, while those using those models to build products tend to play it down.

It’s not necessarily a contradiction. Frontier model developers see up close what their models can do, including the things they shouldn’t. Companies building products on top of those models mostly see the use cases that work. They’re two different perspectives on the same technology.

But the question we should all be asking remains: how do we harness AI’s positive power without putting sensitive data, system security and people at risk?

A real-world case: Zeta Global and AthenaOS

While we’re at it, let’s look at what Steinberg was talking about, because it’s a good example of how AI is making its way into business processes.

Zeta Global describes itself as an “intelligent AI infrastructure company”: it uses AI and a large consumer data asset to help businesses understand their customers and personalize marketing.

On October 8, at Zeta Live 2026, the company unveiled AthenaOS, which it calls an “enterprise intelligence operating system”. The idea in a nutshell:

• You start from the objective, not the application. You tell Athena (the conversational interface) what you want to achieve, and the system brings together the data, applications and actions needed.
• Every decision improves the next one. The system runs in a continuous loop: know, decide, act, learn from the result.
• Shared context across different systems, including partner tools such as Palantir Foundry and OpenAI.
• Athena MCP, a “headless” version that brings Zeta’s intelligence into ChatGPT, Claude, Gemini and other environments.

On air, Steinberg described the goal as connecting inventory, marketing, logistics and production in real time. He also claimed that 51% of Fortune 100 companies trust Zeta with their data, and named BMW among its customers.

One detail that didn’t come up on TV: according to the official press release, AthenaOS is still in limited internal beta, with a public beta expected by the end of the year and general availability planned for the first quarter of 2027. So we’re talking about a promise, not yet a product you can buy.

The economic knot: how much do tokens really cost?

The most interesting part of the interview, at least for anyone who builds with AI every day, was the part about costs.

If you use models through an API, you pay per token, that is, based on how much text is processed. The more the AI works, the more you pay. For a company that wants to put AI agents inside every process, that means variable costs that are hard to forecast. And that’s exactly why many businesses are cautious about these services.

The two answers were different:

Sculley: the market will sort it out. Token costs are a temporary phenomenon. Open source and competition will drive prices down.

Steinberg: take the problem off the table. Inside AthenaOS, customers don’t pay for tokens. Zeta uses its own inference models (the Athena Inference Models, developed with Fireworks on NVIDIA Nemotron) and offers a flat price instead of a variable one. Customers know in advance how much they’ll spend.

On Sculley’s point I have few doubts: it’s already happening. I wrote about it in my article on DeepSeek V4: open models with top-tier performance and much lower costs are already putting pressure on prices.

Steinberg’s proposal is more interesting, but it needs to be read carefully. A flat price doesn’t eliminate the cost of tokens: it shifts it from the customer to the provider. Zeta can afford it because it uses smaller, specialized models, and it’s betting that its compute costs will fall faster than customer usage grows. For the customer, it’s convenient. For Zeta, it’s a bet on the future cost of inference. Not by chance, the company also announced it has started working with partners on a dedicated chip.

The long-term vision

Steinberg wrapped up by quoting NVIDIA CEO Jensen Huang: in his words, “the real internet economy is about to begin”.

The reasoning goes like this: large models will become increasingly similar to each other and increasingly cheap. Value will move to the layer that sits on top of the models: the one that connects them to a specific company’s data and processes. And Zeta’s advantage, according to Steinberg, is a data asset that offers an understanding of consumer intent no other company in the world has.

Here the conversation comes full circle, even if nobody in the studio noticed. An AI that knows better than anyone what millions of people want to buy is exactly the kind of system where data security and governance aren’t a detail. To be fair, the AthenaOS press release places a lot of emphasis on identity, permissions, governance and data sovereignty. But it’s precisely the kind of system that, on the “day after”, would end up under regulators’ scrutiny.

And the market, meanwhile? Zeta’s stock, according to figures cited on air, is up 65% year to date and 70% over the past year. Investors, evidently, are closer to Steinberg’s optimism than to Axios’s caution.

Conclusion: who decides how AI risk is told?

To sum up:

• The leading AI labs are quietly preparing for a disaster they consider likely, and for the political reaction that would follow.
• Those selling products built on AI consider it an exaggeration and prefer to talk about growth, efficiency and costs.
• Both things can be true at the same time.

What puzzles me is that the story of AI risk is almost entirely in the hands of people with an economic interest, one way or the other. The labs have an interest in appearing responsible and in shaping future rules. Companies building on top of the models have an interest in reassuring customers and investors. And we, users and developers, are stuck in the middle, trying to figure out how worried we should be.

Maybe the most sensible answer is also the most boring one: preparing for the worst isn’t alarmism, it’s good engineering. We do it with backups, disaster recovery plans and security testing. I don’t see why AI should be an exception.

Which side are you on? Do you find those preparing for the “day after” more credible, or those saying it will never come? Let me know in the comments.

Sources: Axios, Zeta Global press release on AthenaOS, Fox Business interview.