Most months, the biggest AI story is a launch. In June it was a switch-off. A US export-control directive led one major lab to disable its most capable models globally, reportedly for eighteen days and counting as the month closed.
Whatever the rights and wrongs of the intervention, the operational lesson landed on every business that had built a workflow on top: capability you don't control can be withdrawn without notice, for reasons that have nothing to do with you.
The concentration-risk rehearsal
We've argued for months that the model should be a swappable component: business logic, data and approval workflows in your systems, with the model called in to do defined work. June was the argument made flesh: organisations built that way changed a configuration setting; organisations built around one provider's particular capabilities had an outage.
This is not a reason to avoid frontier models. It's a reason to architect as though any of them could be unavailable next Tuesday, because one of them was.
The supply chain goes vertical
OpenAI unveiled its first custom inference chip with Broadcom, and released its newest models into a government-gated preview rather than general availability. Google shipped live-translation features and new home hardware, while its next flagship model reportedly slipped its release date for a second month.
Read together: the frontier is consolidating around a few vertically-integrated players, moving at a pace they increasingly don't control themselves. For buyers, roadmap promises deserve more scepticism than ever: build plans around what's shipping, not what's slated.
Content gets licensed, properly
Getty and OpenAI reportedly signed a multi-year deal putting Getty's licensed photo library, hundreds of millions of assets, inside ChatGPT. Expect more of this: the era of quietly contested training data is giving way to commercial licensing at scale.
For businesses, it's a useful precedent. The material your organisation owns (documentation, imagery, data) has licensing value in an AI economy, and contracts you sign should be explicit about who may train on what.
What to do with all this
June's actions:
- Run the thought experiment: if your AI provider vanished tomorrow, what breaks? Fix that
- Favour shipped capability over roadmaps in any AI purchasing decision
- Check your supplier contracts for training-rights language before it matters
- Keep approval workflows in your own systems: they're what let you swap providers calmly
