This Is Not the Dot-Com Bubble. It Is Also Not a Free Lunch.

Every few years, someone asks whether the current technology wave is a bubble. The answer is almost always the same, and almost always wrong in the same way: yes and no, with the emphasis in the wrong place.

In a recent conversation on The Pragmatic Engineer, Martin Fowler and Kent Beck — two people who have watched more technology cycles than most — put AI adoption in historical context. Beck recalled holding the Intel 4004 and thinking “that chip is a computer?” Fowler traced the resistance patterns from object-oriented programming, through the internet, through Agile, and into the present. Their conclusion is not the one the hype merchants or the skeptics want to hear.

It is the mature one. And it requires holding two truths at the same time.

Truth One: The Technology Is Real

The dot-com comparison gets invoked reflexively whenever infrastructure spending outpaces visible revenue. Hyperscaler capex crossed $300 billion in 2025. Goldman Sachs warned back in 2024 that roughly $1 trillion would be spent on AI infrastructure “with little to show for it so far.” MIT’s NANDA initiative reported that 95% of enterprise generative AI pilots fail to move the P&L. These are real numbers. They look bubble-adjacent.

But that is where the analogy breaks.

Pets.com filed for IPO in February 2000 and liquidated in November of the same year. Its revenue model was a hope. Most of the dot-com era’s headline losses were companies with no economic engine under them. The infrastructure layer — the fiber, the servers, the nascent ad networks — survived and powered the next fifteen years of growth. The financial bubble was real. The technology was also real. Both things were true.

AI in 2026 is not Pets.com. OpenAI and Anthropic are posting billions in revenue. Frontier developers are six times more productive than laggards in the same companies. Ramp, a $32B fintech, has pushed internal AI adoption above 99% — and the work we did on that case showed the lift is real when the organization reshapes around the tools. Fortune 500 adoption of AI coding tools is already ahead of where cloud computing stood at a comparable point in its curve.

The technology is delivering value today, to the companies that know how to receive it. This is the part the bubble narrative misses.

Truth Two: The Industry Around It Is Replaying Old Mistakes

Here is where Fowler and Beck earn their stripes. Fowler calls it the “Agile industrial complex” — the layer of consultants, certifications, frameworks, and conference circuits that grew around Agile and eventually buried its original intent under process theater. He sees the same pattern forming around AI.

He is right. The snake oil is already thick.

Vendors are selling “AI transformation” roadmaps that are PowerPoint with a GPT wrapper. Consultancies are billing for prompt-engineering workshops the way they used to bill for SAFe rollouts. Boards are being told that their organization is “behind on AI” by people whose business model depends on that being true. The middle layer — the layer between the technology and its productive use — is being industrialized into a machine for extracting fees from fear.

This is not new. Beck’s observation cuts deepest: “people don’t want faster, cheaper, better” when organizational incentives point elsewhere. Agile died in most organizations not because the ideas were wrong but because the incentives around headcount, promotion, and procurement were pointed at the opposite outcome. AI will die the same death in the same organizations for the same reasons, unless leaders are honest about what they are actually measuring.

The telltale sign is already visible: companies tracking pull request volume, code line counts, and “AI usage percentage” as if those were outcomes. They are not outcomes. They are activity. A team shipping more pull requests of worse code is not more productive. It is just louder.

What Prior Cycles Actually Teach

Every cycle Fowler and Beck have lived through — microprocessors, object-oriented programming, the internet, cloud, mobile, Agile — followed roughly the same arc. A real capability emerges. The people closest to the work recognize it immediately. An industrial complex forms around it and starts selling certainty to people who feel behind. A washout eventually clears the consultants who added no value. The substrate survives. The practitioners who focused on the substrate survive with it.

Three disciplines separated the survivors from the victims in every one of those cycles:

They measured outcomes, not activity. In the Agile era, the teams that thrived tracked deployment frequency, change failure rate, and time to recover. The teams that died tracked story points and velocity. In the AI era, the equivalents are cycle time on real work, defect rates, and customer outcomes. PR count is the new story point. Resist it.

They invested in the boring layer while others chased demos. In the dot-com era, the companies that survived were the ones that kept their database architecture sane while their competitors spent on Super Bowl ads. In the AI era, the boring layer is testing, observability, evaluation harnesses, and governance. Kent Beck is not wrong that TDD is more relevant now than ever — tests are the contract layer between humans and agents that do not remember what they wrote yesterday.

They bought the technology, not the vendor pitch. The companies that got cloud right in 2010 were the ones that learned AWS primitives, not the ones that bought “private cloud” from their existing data center vendor. The companies that will get AI right are the ones learning how foundation models, evaluation, and agent orchestration actually work — not the ones buying “AI transformation” from the firm that sold them digital transformation, cloud transformation, and agile transformation in the previous three cycles.

What Is Genuinely Different This Time

Mature analysis requires acknowledging what prior cycles did not have.

The adoption velocity is real. ChatGPT reached 100 million users in two months. No prior general-purpose technology came close. That is not marketing; that is a measurable fact. The speed compresses the window in which organizations can afford to be wrong.

The open-source commodity layer is unprecedented. Llama, DeepSeek, Qwen, and their successors have pushed inference costs down faster than any prior infrastructure curve. The vendor lock-in that defined cloud’s middle years is structurally weaker here. This is underrated.

The regulatory layer arrived early. The EU AI Act is fully effective in 2026. No prior cycle had compliance obligations this early in its maturity curve. Governance is not a late-cycle bolt-on this time; it is table stakes from day one. The companies that treat it as a blocker will lose to the companies that treat it as a design constraint.

And finally, the geopolitical dimension is new. AI is the first major tech cycle that is simultaneously a commercial race and a state-level contest, with export controls, sovereign compute, and national strategies built in from the start. That changes the arithmetic of dependency in ways the internet era never had to reckon with.

The Mature Position

It is not “AI will change everything.” It is not “AI is a bubble.” It is both of these statements held together, without the emotional satisfaction of picking a side.

The technology is real. The fees being extracted around the technology are often not. Value accrues to the organizations that invest in the substrate — testing, governance, measurement, workflow redesign — while their competitors invest in the theater. Leaders who make AI adoption a procurement decision will lose to leaders who make it an operating decision.

Fowler has a quiet heuristic that belongs in every boardroom conversation about AI: watch for when you are producing negative value and stop. The same discipline applies at the organizational level. If the AI program is generating activity without outcomes, it is producing negative value, and the right move is to pause, measure, and redesign — not to buy more tools.

The cycles always rhyme. The winners are always the ones who remembered the last verse.

---

This analysis synthesizes Cycles of Disruption in the Tech Industry from The Pragmatic Engineer featuring Martin Fowler and Kent Beck (April 2026), Goldman Sachs’s Gen AI: Too Much Spend, Too Little Benefit? (June 2024), MIT NANDA’s State of AI in Business 2025, and the Agile Manifesto (February 2001).

Victorino Group helps enterprise leaders separate AI signal from noise and build the boring layer that makes adoption compound. Let’s talk.