Stop asking whether AI is a bubble. Do the work that pays off either way.

Whether AI is a bubble is the wrong question for anyone running an organization or leading AI enablement.

The answer will ultimately be determined by what an entire economy’s worth of knowledge workers decides to do en masse. No leader running a 200-person organization can control that. What you do control is how your own organization uses these tools, and the work worth doing is the same whichever way it goes.

The case for and against

If you haven’t followed the debate, here’s the short version. Most of these clips run under three minutes. The Klein episode is an hour, and worth it.

The case that it’s a bubble:

• Too much money chasing a race. Paul Kedrosky, on The Ezra Klein Show: “As long as you have a supply of capital and a race dynamic you’re guaranteed overcapacity,” the same way it went with canals, railroads and electrification.
• The commitments are huge and locked in. Anthropic’s IPO filing shows compute contracts it can’t cancel, and it doesn’t give a total. Prof G Markets added them up and got to more than $800 billion.
• Buyers feel it as a cost, not an investment. In a Kyla Scanlon skit, the buyer puts it plainly: “An investment is something that I buy once and it makes me money. A cost is something that I have to pay for every month.” Her read on the real risk is “not that the technology isn’t useful, but perhaps we got a little over our skis with all the build out.”

The case that it isn’t:

• A correction isn’t a crash. Scott Galloway, on Prof G Markets: “I agree there’s going to be a correction here.” What he doesn’t buy is that it’s “so clearly a bubble waiting to pop viciously.”
• Companies are getting results, slowly. Azeem Azhar, in the same Klein debate, calls it a productive boom. He hears the same thing from the CEOs he talks to: “It’s harder than we thought. We’re going to keep doing it and we’re starting to get results.”
• The economics are improving. In the first half of 2025, Anthropic spent twice its revenue on compute. A year later, compute was 55% of revenue, according to the same filing.

Scanlon’s other video, on Anthropic’s IPO, lands on the most honest answer of the bunch: “Nobody knows what’s happening.”

Let’s play out both outcomes

Say it is a bubble and it pops. Funding will dry up, some providers will fold or get bought, and the market will consolidate. The tools and workflows you’ve built will survive that, because the real asset is the work that went into getting your team comfortable using the tools and thinking through the best way to execute on work with AI as a core enabler of how you work (much like the internet and mobile are taken for granted as enablers of how we work today). By leaning in now, your people will build the muscle memory and the institutional knowledge of how your work should run in these new ways, and that’s what lets you move those tools to whichever vendor survives, or to whatever kind of tool wins out.

Say it isn’t a bubble and usage continues to grow significantly. In this case, the labs will keep building, and they’ll increasingly be in a position to raise prices, because they’ll have the demand to do so. They’ll also need to, because they have commitments to fill. Anthropic alone has committed to spend at least $114.5 billion with Google through 2033 regardless of actual usage. We’re watching that pricing reach our clients. One client we work with moved to a pooled, credit-based plan because the licenses they’d handed out weren’t getting used regularly, and it was the easiest way to spread usage between heavy users and more basic users. Anthropic is moving Enterprise customers to metered billing at renewal, and Microsoft bills its newest tooling, Copilot Cowork, by usage. The most capable models cost extra on every major platform. Some aren’t for sale at all: Anthropic keeps its Mythos model for vetted partners, and OpenAI does the same with its most capable cyber models.

What carries you through both of these scenarios is the same discipline:

• Your people know how to use these capabilities well. For example, they know which work needs which model.
• The best way to execute on recurring, meaningful pieces of work is packaged so it runs the same way every time, on whatever model and vendor you’re on. The process of creating that packaging is what makes these implementations portable, which is what insulates you from the bubble bursting.

Every platform has its own name for these tools (a skill, a project, an agent). Whatever the platform, we define adoption as the point where specific, meaningful tasks in someone’s week have graduated to shared tools that the team itself maintains. Seat counts and usage dashboards don’t tell you that, but once you’ve reached this point, you’ll be able to ride the waves of future change in this space much more easily.

What that looks like at a mid-size firm

Even today, many organizations we start working with have handed people these tools with few guardrails and little guidance. Worse, once they’re on an enterprise plan, it becomes a wild west of managing usage and spend.

One firm we worked with kept raising its monthly usage cap. When we looked at the analytics, the tasks its teams ran every week (the weekly report, proposal prep and account research) were all running on models that didn’t need that level of horsepower. A lower-tier model would have done each of them, and people were sending the tool digging through connectors for data it should have been handed.

So we pinned everyone to the lower-tier models as a starting point. The cap came down 25% almost immediately, because a lot of people didn’t need the top tier for how they were actually using the tools in their day-to-day.

Then we trained them. It was an overview of the models and what each is best used for, built on real examples of the tasks people do. Everyone gets access to the basics from day one, but to use the more powerful (and more expensive) tools, you need to show that you’re ready to handle that. Ethan Mollick argues that companies should “be biased towards allowing people to use AI,” and he’s right, which is why the base and mid-level models stay open to everyone.

Last, we built shared tools for the weekly report, the proposal prep and the account research. Each one packages the best version of that task and runs it on a more cost-effective model. They saved money, and they also made the proposals better, faster, and more consistent. Some of the data behind them didn’t need a connector at all. A scheduled export did the job. With Salesforce charging for each action its AI agents take, getting savvy about how you access your firm’s key information also belongs on the list for managing costs. And, as before, these are the right moves regardless of what happens with the AI bubble.

That’s the job: get the most out of every dollar you’re already spending, both in tokens and in the tools you’ve already rolled out across the organization, AI or not.

The case for waiting

The strongest objection comes from Aswath Damodaran at NYU Stern. Talking about Apple holding back while its rivals spent hundreds of billions, he said it will be “a classic case study 10 years from now as to whether it’s better to wait out the uncertainty and then decide what kind of factory to build.” That could be read as a critique of what I’ve proposed above. If nobody can figure out where this is all going, why spend now?

It’s a valid question for the companies deciding whether to build the underlying technology. Very few of us are working in firms that are in that position. An organization focused on putting AI to work is in a different one, because organizational change is a process you can’t speed-run. In our work, it takes about a quarter for people to build the habits for effective use. It takes around another quarter before a team starts meaningfully redesigning parts of the work, and that only happens once people are already thinking and working in these new ways. You don’t get to start competing in your space differently, with significantly better operating margins, without going through adoption and redesign first.

So, start where you are today. If you’re early, get people using the tools, and create a structured path for them to become savvier users. If you’re further along, move your highest-volume work into shared tools.

Let everyone else argue about the bubble. A market correction doesn’t refund the time it takes to get good at looking at your work in fundamentally different ways, and the organizations that started this quarter will be two quarters ahead of the ones that waited.