The extraction phase of AI has begun, and it's a healthier place to be than the hype ever was.

For a few years, the entire AI conversation ran on a single axis: can it do the thing? Can it write the essay, pass the exam, fold the protein, drive the car. That question is basically answered now, and something more interesting has replaced it. Consumer spending on AI apps has more than doubled year over year even as download growth has cooled — meaning the novelty tourists have mostly moved on, and the people still paying are paying for real use, not curiosity. Standalone AI tools now count over a billion monthly users, and that's before you fold in the assistants riding shotgun inside WhatsApp, Office, and Snapchat, which push the real number well past two billion. The land-grab is ending. The extraction phase — the one where a product actually has to earn its keep — has started. I think that's the healthiest thing to happen to this industry since the original ChatGPT launch.

The agents are the real story, not the chatbots. The most telling number I've seen this year isn't a benchmark score, it's a workflow stat: OpenAI's own usage data shows a fivefold jump in active coding-agent users in the first half of 2026, with a tenfold increase in agent tasks that run longer than eight hours unsupervised. That's not a chatbot answering a question. That's a piece of software being trusted to go away and work for a day. Whether or not you find that exciting, it's the actual shift underway — a move from AI-as-oracle to AI-as-employee, with all the trust, verification, and liability questions that second framing drags in behind it. A model that's impressive for thirty seconds is a demo. A model that's reliable for eight hours is a coworker. Most of the public conversation is still stuck evaluating demos.

Money is where the real disagreement lives. The bull case and the bear case are, for once, looking at the exact same numbers and drawing opposite conclusions, which is usually a sign that something genuinely uncertain is happening rather than something obvious being denied by fools. The bear case: valuations have pushed the market's cyclically-adjusted P/E ratio above 40 — a level touched only once before, right before the dot-com crash — while hyperscalers are set to spend roughly three-quarters of a trillion dollars on AI infrastructure this year on returns that remain, by Deloitte's own enterprise survey, "transformative" for barely a third of the companies actually deploying it. The bull case: unlike 2000, the companies at the center of this boom already have staggering real profits, and even a name as richly valued as Nvidia trades at a fraction of Cisco's price-to-earnings ratio at the height of the dot-com peak. Both of those things are true simultaneously. That's not a contradiction — it's just what a genuine, high-stakes bet on a real technology looks like from the inside, before anyone knows how it resolves.

My honest opinion, for what it's worth: the mistake right now isn't believing AI is transformative — it clearly already is, for coding, for research prep, for a growing list of specific bounded tasks. The mistake is assuming "transformative" and "correctly priced" are the same claim. They aren't. The internet was transformative and the dot-com bubble still happened, because being right about the technology and being right about the valuation are two separate bets, settled on two separate timelines. I'd bet on agentic AI reshaping how knowledge work gets structured over the next five years with real confidence. I'd bet on this specific market cycle unwinding at least once before that happens with almost as much confidence, and I don't think those two beliefs are in tension.

The unglamorous trend I'd actually watch is the boring one. Not the flagship model launches — this month alone saw three major labs ship new frontier models within twenty-four hours of each other, which tells you competition is fierce but tells you almost nothing about durable advantage. The trend worth watching is governance quietly becoming a product feature rather than a compliance afterthought: audit trails, data handling, and human checkpoints are starting to decide who gets to sell into healthcare, finance, and legal work, not just who has the best model. That's a quiet, unsexy shift, and it's exactly the kind of shift that outlasts a hype cycle.

If I had to summarize where AI actually stands in July 2026, it's this: the "is this real" argument is over, and the industry's own actions have settled it — you don't spend three-quarters of a trillion dollars a year building infrastructure for a fad. What's still wide open, and worth genuine skepticism about, is the "who captures the value, and at what price" argument. That's not a verdict on the technology. It's a verdict on the market around it. Those are different questions, and I think a lot of the current noise comes from people answering one while arguing about the other.

This piece reflects one perspective on a fast-moving, genuinely contested topic — reasonable people are drawing very different conclusions from the same data right now, and that's worth sitting with rather than resolving too quickly.