AI Stopped Being a Feature. It Became Infrastructure.

AI used to be the thing you bolted onto a product. A chat box. A summarize button. A feature you could ship and demo. That is changing. The interesting work now happens underneath, in the systems your product will lean on without anyone clicking a button.

Three stories landed at once, and they point the same direction. A cardiac risk model. A respiratory disease nonprofit. A new data layer for AI. Different fields, same move. Let me catch you up.

The model that saw what doctors could not

UC Berkeley researchers trained an AI on more than 440,000 EKGs from Sweden, matched against death certificates, to spot people at risk of sudden cardiac death. The numbers matter. Standard tests flag a high-risk group with a 4.6% annual death rate. The AI isolated a group with a 7% rate, and it caught people who looked low-risk by today's rules.

The part to sit with: the model found a signal in the EKG that doctors did not know was there. Lead author Ziad Obermeyer calls sudden cardiac death one of medicine's stubborn mysteries because people die too fast to study what happened. The AI did not just sort patients better. It pointed at new physiology to go investigate.

That is the shift. The tool stopped being a faster version of the old test. It became a way to find questions nobody knew to ask.

Treating disease like a utility problem

Stripe is funding a $500 million nonprofit called Intercept, with backing from Anthropic, the OpenAI Foundation, and Bill Gates, to go after the common cold and flu. The goal sounds wild on its face: remove respiratory viruses the way cities remove impurities from water.

The framing is what got me. Nan Ransohoff, the Stripe exec running it, says the average person spends 5% of their life fighting a cold or flu, yet drug companies skip it because more than 200 viruses cause it and no single vaccine pays off. So they are betting on broad countermeasures and air-cleaning systems for whole buildings, plus AI-driven protein design.

Note who is paying. The same companies building AI models are now funding the public health layer underneath daily life. The infrastructure mindset is leaking into where they put their money.

The web was never built for this

The third piece is the least flashy and maybe the most useful for you. AI models need fresh data, and the web was not designed for machines to read it at scale. Prices change, inventory shifts, sentiment moves. A model trained on a snapshot gives stale answers.

The gap is real. Gartner says 60% of AI projects without AI-ready data will be abandoned by year's end. Bright Data CEO Or Lenchner puts it plainly: a strong model on top of weak data is "a genius who knows nothing." One survey found 97% of AI organizations depend on real-time web data, and 90% feel boxed in by access rules.

So a whole layer is forming to fetch live, structured, compliant web data at scale. Not the model. The pipes that feed it.

The deep cut

Here is the thing that changes your Monday. In all three stories, the value moved from the visible feature to the layer underneath, and that layer is owned by someone else. Berkeley's model only worked because Obermeyer spent a decade getting hard-to-access medical data. Intercept exists because the data and incentives to prevent disease were broken. Bright Data sells access because building it in-house, in Lenchner's words, "becomes a full-time engineering problem that competes with the actual AI work."

Your product is going to depend on infrastructure you do not control: data feeds, health models, prediction systems. That is fine, until access, pricing, or accuracy shifts under you. Treat your AI data sources the way you treat any critical vendor. Know who owns them, what happens if they change the rules, and where your fallback is. The feature is easy to demo. The dependency is what you have to manage.

Three questions for your team

• Which AI capability in our product actually runs on data or models we do not own, and what is our plan if that source changes its terms or quality?

• Is our AI grounded in fresh, real-time data, or are we shipping confident answers off a stale snapshot? How would we even know?

• Where could a model find a signal in our own data that we have not thought to look for, the way Berkeley found a new EKG pattern, and who on the team owns that exploration?