The numbers around AI look enormous right now. Cursor, the AI coding tool, has reportedly seen its yearly revenue pace jump past 2 billion dollars, and the company itself expects to end the year near 6 billion. Salesforce says its AI agent product, Agentforce, is now bringing in 1.2 billion dollars a year.
These are the kind of numbers that make AI look like it has already won. But a recent analysis makes a point that is easy to miss under all that growth: fast adoption is not the same thing as a mature, stable technology.
Electricity was proven to work commercially in 1882. It still took almost four decades before factories, regulators, and equipment makers settled into the roles that let electrification spread everywhere. AI is closer to that early, messy phase than most people realize.
Here is the plain version of the argument. For a technology to change entire industries, businesses need to know the rules of the game: what they can safely build on top of it, who they can count on for what, and what protection they have if a bigger company copies their idea for free. Right now, those rules for AI barely exist.
The clearest proof is what happened to OpenAI's store for custom chatbots. It launched with millions of custom tools and a promise to pay the people who built them. That payment program never happened, and OpenAI quietly walked away from it.
Around the same time, SAP told outside companies that AI agents can no longer freely plug into and act on data stored inside SAP's own software, unless it goes through SAP's approved path. Salesforce, similarly, bills its AI agent work in units that only Salesforce defines. Each of these is a large company setting the rules for everyone else on its own terms, which is the opposite of a stable, shared system.
This matters for any business buying AI tools right now, for two reasons. First, using the same AI model as a competitor gives no real edge, since that competitor can rent the exact same intelligence. Second, whatever advantage a business builds has to live in the layer above the model itself: its own customer data, its trusted relationships, its specific way of doing the work.
That is part of why many hospitals run their AI tools through Epic's medical records software rather than a flashier outside product. Epic already has decades of trust built in that a newcomer cannot copy overnight, and that trust matters more than which AI model sits underneath it.
The practical move right now is not to chase the newest model. It is to use this unsettled period to learn cheaply, using low cost AI models to find out exactly where employees still beat the machine on real tasks, and to write that down before the gap closes. That record becomes an asset that survives no matter which AI company wins the bigger fight.
There is a real risk sitting underneath all this. The world is set to spend more than 1 trillion dollars on AI data centers this year alone, and that spending only pays off if this shared foundation actually forms and businesses build durable value on top of it. If it keeps taking years to settle, a lot of that money will look like it arrived far too early.