Ben Thompson just published an argument that quietly dismantles the entire "AI bubble" narrative. Not by defending the hype, but by showing the economics have already shifted in ways that make the bubble framework irrelevant.
His central claim: we're not in a bubble anymore because agents fundamentally changed the demand equation. And if he's right, a lot of the scepticism around AI investment is looking at the wrong metrics entirely.
The Demand Problem That Disappeared
Here's the traditional bubble logic: AI requires enormous compute. Compute is expensive. For the investment to pay off, you need mass adoption - billions of users generating queries that justify the infrastructure spend. If adoption stalls, the bubble bursts.
Thompson's counter-argument is elegant. In his Stratechery piece, he points out that agents don't require billions of users to justify massive compute demand. A single company deploying an agentic system to handle customer service, logistics optimisation, or financial analysis can generate more compute demand than thousands of casual ChatGPT users.
In simpler terms: you don't need everyone using AI. You need a few organisations using AI intensively. That's a completely different adoption curve - and it's already happening.
Integration Over Models
The second shift Thompson identifies is more subtle but possibly more important. The value isn't in the model alone anymore. It's in how deeply the model integrates into existing workflows, systems, and decision-making processes.
This is why the "who has the best model?" race feels less urgent than it did a year ago. A slightly less capable model that integrates seamlessly with your CRM, your inventory system, and your support queue is worth more than a more powerful model that requires you to rebuild everything around it.
For business owners, this is the practical takeaway: the question isn't "which model is best?" but "which system fits our architecture?" That's a procurement decision, not a gamble on future capability.
What This Means for Builders
If Thompson's right - and the pattern I'm seeing suggests he is - then the opportunity shifts dramatically. The next wave of valuable AI companies won't be the ones with the most impressive models. They'll be the ones solving integration problems.
How do you connect agentic systems to legacy databases? How do you ensure compliance when an agent is making decisions autonomously? How do you audit AI actions across distributed workflows? These are unglamorous, deeply technical problems. They're also where the actual business value lives.
For developers watching this space, the implication is clear: the infrastructure layer is wide open. Model providers are competing on benchmarks. The real opportunity is building the plumbing that makes agents useful in production environments.
Why the Bubble Narrative Persists
So why does everyone still talk about an AI bubble if the economics have already shifted? Thompson suggests it's partly a lag in how we measure technology adoption. We're still looking for consumer-scale numbers - billions of users, viral growth, network effects - when the actual value is happening at the enterprise level, quietly, in ways that don't generate headlines.
A logistics company deploying agents to optimise routing doesn't tweet about it. A financial firm using AI to process loan applications faster doesn't issue press releases. But the compute demand is real, the ROI is measurable, and the adoption is accelerating.
The bubble popped. We just didn't notice because the crash never came. Instead, the market matured into something more boring and more profitable: infrastructure that works.
The Uncomfortable Truth
Here's what makes this argument uncomfortable: if Thompson's right, then a lot of the scepticism around AI investment was mistimed. The bubble wasn't in the technology - it was in the assumption that AI needed consumer-scale adoption to justify its costs.
Agents broke that assumption. And now we're in a different game entirely - one where fewer users, deeper integration, and enterprise deployment create sustainable demand for compute at scale.
For those of us building in this space, the lesson is simple: stop waiting for the bubble to burst. It already did. What's left is the hard, practical work of making AI systems that integrate, comply, and deliver measurable value.
That's not a bubble. That's an industry.