Getting AI Into Production Is No Longer The Hard Part

Ciarán Cosgrave, CEO, Nearform.

​On the evening of June 12, 2026, chats froze, active prompts failed and automated pipelines broke as two of the most capable AI models on the market seemingly disappeared.

Citing national security, the U.S. government ordered Anthropic to block all foreign nationals from accessing Claude Fable 5 and Mythos 5. With no way to do that selectively on short notice, everyone lost access, including Americans. There was no technical issue, and the models worked perfectly until they were turned off. Access came back about two-and-a-half weeks later, but the lesson is that access can disappear with little warning.​

What strikes me most is this hit businesses running AI in production, not sandbox experiments. For the last few years, the conversations I’ve had with enterprise CTOs and other tech leaders have been about escaping the pilot “graveyard,” and, in the meantime, an entire consulting market has emerged promising the answer. But with every consultancy now making the same promise, getting AI into production has become a baseline expectation rather than a prize.

Instead, enterprise teams—including their boards and investors—are increasingly focused on what happens after go-live. I keep hearing the same two questions from them:

1. Can we afford to keep running these systems at scale?

2. Can we depend on what we’ve built still being there next month?

Engineering alone can’t answer either, because no amount of clever code fixes a price you can’t predict or a government intervention. They’re questions about risk and volatility, which is why they’ve moved from the CTO’s desk to the boardroom.

The Bill Nobody Modeled

And the economics aren’t fixed, either. Open-weight models, which can be downloaded and run on your own infrastructure rather than accessed through a vendor, shift more of the running cost into your control while reducing your exposure to a vendor suddenly cutting off access.​