AI task complexity is doubling every seven months, and since 2024 that doubling has compressed to roughly every three months. The original Moore's Law, the one that powered the entire semiconductor revolution, ran on an 18-month cycle. AI capability is improving at six times that rate. That single number is the one most "AI is a bubble" arguments do not survive contact with.

I wanted the bubble case to be true. The contrarian in me, the part that has survived 17 years of enterprise technology by being skeptical of hype, wanted the curve to flatten so we could all catch our breath. The 1999 comparison is emotionally satisfying, the capex numbers look terrifying out of context, and the headline "no clear ROI" is half right. Then I went and looked at the underlying data the analysts I respect were quietly tracking. The data does not support the bubble thesis. It supports something more uncomfortable, which is that the curve is real, the demand behind the capex is real, and the procurement cycle inside most enterprises is now slower than the technology refresh cycle it is trying to evaluate.

The METR Discovery: AI's Own Moore's Law

The organization that changed my mind is called METR, the Model Evaluation and Threat Research group. They have been quietly tracking something remarkable: the rate at which AI systems can handle increasingly complex tasks. Their findings are, frankly, startling.

AI task complexity is doubling every seven months overall. And that rate is accelerating. Since 2024, it has compressed to roughly every three months. To put this in context: the original Moore's Law, the one that powered the entire semiconductor revolution, operated on an 18-month doubling cycle. AI capability is improving at six times that speed.

What does this mean in practical terms? GPT-5 handles tasks of roughly three and a half hours. Claude Opus 4.5 handles five-hour tasks. And Claude Opus 4.6, released weeks ago, sits at fourteen and a half hours. Six months ago, these systems struggled with anything beyond 30-minute tasks. Six months from now, the benchmarks we are using today will be obsolete.

I keep returning to this point: the tools your organization evaluated last quarter are already a generation behind. The procurement cycles most enterprises use for technology acquisition, typically 6 to 12 months, are longer than the entire capability doubling period. You are buying yesterday's technology with tomorrow's budget.

AI capability is improving at 6x the speed of Moore's Law. The tools your team evaluated six months ago are already obsolete.

The Scaling Wall That Never Arrives

I need to address the scaling critics directly, because their argument is the intellectual foundation of the bubble narrative.

The most prominent voice has been Gary Marcus, a cognitive scientist who has predicted the end of AI scaling every year since 2019. In 2019, he said deep learning had hit fundamental limits. In 2020, he said GPT-3 proved scaling was insufficient. In 2021, same argument. In 2022, he said large language models had plateaued. In 2023, GPT-4 proved him wrong again. In 2024, he pivoted to claiming that LLMs specifically had peaked. In 2025 and 2026, the models continued improving.

I have nothing personal against Marcus. He is a serious researcher asking important questions. But his track record on scaling predictions is now 0 for 7 across consecutive years. At some point, the pattern itself becomes data.

The fundamental misunderstanding is about S-curves. Every technology follows an S-curve: slow start, rapid acceleration, eventual plateau. The critics see the plateau of one S-curve and declare the technology dead. What they miss is that AI does not ride a single S-curve. It stacks them. When one paradigm plateaus, a new one ignites underneath it. Transformer architecture plateaus? Chain-of-thought reasoning arrives. Reasoning plateaus? Tool use and agentic architectures emerge. Each new paradigm creates a new S-curve that starts climbing before the previous one fully flattens.

This is not theoretical. We can see it in the benchmark data. Every predicted wall has been broken not by pushing harder on the same approach, but by switching to an entirely new approach that nobody anticipated.

When AI Outperforms PhD Holders

The benchmark that stopped me cold was GPQA, the Graduate-Level Google-Proof Q&A benchmark. It contains questions specifically designed to be so difficult that even domain experts with PhD-level knowledge and full internet access struggle with them.

Human PhD experts score an average of 65% on GPQA. Current AI systems score above 93%. Not on trivia. Not on pattern matching. On the kind of deep reasoning that represents the highest level of human intellectual achievement in specialized domains.

For a hands-on look at how these models compare across dozens of benchmarks, explore the AI Model Benchmarking Dashboard I built, powered by Artificial Analysis data.

But the benchmark numbers are not what genuinely unsettled me. What unsettled me was the discovery at the University of Gothenburg. Researchers gave an AI system access to measurements from dusty plasma experiments, a notoriously complex area of physics. The system did not just analyze the data accurately. It discovered new physical laws that the human researchers had not identified. Novel scientific laws. Found by a machine. With over 99% accuracy.

If you are an executive whose mental model of AI is "a faster search engine" or "a better autocomplete," that mental model is now dangerously outdated. We are in an era where AI is not just processing human knowledge. It is extending it. It is finding patterns in reality that human cognition, with all its brilliance, missed.

AI scored 90% on PhD-level benchmarks where human experts averaged 65%. And then it discovered new physics laws the humans had missed entirely.

The Creative Turing Tests Are Over

This section matters for executives because of a specific objection I hear in almost every boardroom: "AI can do the repetitive stuff, but it cannot touch creative work. That is uniquely human."

That objection is now empirically false.

The Music Turing Test has been passed. In controlled experiments, listeners cannot reliably distinguish between AI-composed music and human-composed music. Not at marginally above chance. At statistical indistinguishability. An AI-generated track has charted on the Billboard Hot 100. AI has won international art competitions, in some cases beating professional artists who had spent decades honing their craft.

I am not celebrating this. I am reporting it. And I am reporting it because the strategic implications are enormous. If your competitive moat is "we have creative people," that moat is evaporating. Not because AI is more creative than humans. The quality debate is nuanced and ongoing. But because AI can produce creative output at a volume and speed that makes the distinction between human and machine creativity operationally irrelevant for most business applications.

The companies that will thrive are the ones that understand this is not about replacement. It is about amplification. The best human creatives working with AI tools will produce output that neither could achieve alone. But the companies that assume creativity is an exclusively human domain, that refuse to integrate AI into their creative workflows, will find themselves outproduced by competitors who made that integration 18 months ago.

The Timeline Compression Nobody Expected

In 2020, the median prediction among AI researchers for when we would achieve Artificial General Intelligence was 2070. Fifty years away. Comfortably distant. Something for the next generation to worry about.

By 2025, that estimate had collapsed. Surveyed researchers now assign a 25% probability to AGI arriving by 2027 and a 50% probability by 2031. That is not a revision. That is a paradigm shift in expectations. The goalpost did not move. It teleported.

Ray Kurzweil, who made specific AGI predictions in 1999 that were widely mocked, is now looking remarkably prescient. His core thesis, that exponential progress appears slow until it suddenly appears instantaneous, keeps proving itself. The Lake Michigan analogy is useful here: if you are filling Lake Michigan by doubling the volume of water every cycle, the lake looks essentially empty for the vast majority of the process. It goes from 0.1% full to 0.2% full to 0.4% full. At 93% of the way through the process, the lake is still only half full. Then in the final few cycles, it fills completely.

We might be at that inflection point right now. And the executives who are betting that AGI is "decades away" might be making the same error as someone looking at a half-empty Lake Michigan and concluding they have plenty of time.

The practical implication for enterprise planning: your three-to-five-year strategic plan needs to account for the possibility that AI capabilities in 2029 will be categorically different from what exists today. Not incrementally better. Categorically different. Planning for linear improvement in an exponential environment is how organizations become irrelevant.

AGI estimates collapsed from 2070 to possibly 2027. Planning for linear improvement in an exponential environment is how organizations become irrelevant.

The 25% Question: Why Even the Optimists Are Worried

I could have ended this article at the previous section. The data makes the case. AI is not a bubble. The scaling is real, the capabilities are accelerating, and the timeline is compressing.

But I would be doing you a disservice if I did not address what keeps me up at night. Occasionally literally, in Zurich hotel rooms.

Dario Amodei, the CEO of Anthropic, one of the three companies at the frontier of AI development, assigns a 10 to 25% probability that AI will lead to catastrophic outcomes for humanity. Not "disruption." Not "job displacement." Catastrophic outcomes. One in four odds, from someone building the technology.

Geoffrey Hinton, widely considered the godfather of deep learning, left Google specifically to speak freely about the risks. Yoshua Bengio, another founding figure, describes our current trajectory as "blindly driving into fog." Mustafa Suleyman, co-founder of DeepMind, warns about recursive self-improvement: the point at which AI systems can improve themselves faster than humans can understand or control the improvements.

I am not a doomer. I have spent three years enthusiastically integrating AI into every aspect of my work and the organizations I've worked with. But I am a realist. And the reality is that the same exponential capability growth that makes the bubble narrative wrong also makes the safety concerns urgent.

The 25% figure from Amodei deserves executive attention not because it is a precise probability, but because it comes from someone with complete access to the frontier research. When the person building the car tells you there is a one-in-four chance the brakes might fail at highway speed, you do not stop driving. But you do make absolutely sure the brakes are the best they can possibly be.

For enterprises, this translates into a practical mandate: governance is not optional. The organizations that treat AI safety and ethics as a compliance checkbox rather than a strategic priority are not just being irresponsible. They are being strategically naive. When the regulatory environment catches up, and the EU AI Act and ISO/IEC 42001 suggest it is catching up fast, the organizations with mature governance frameworks will have a decisive advantage over those scrambling to retrofit compliance.

What This Means for Your Next Board Meeting

I started this article in a Zurich hotel room, scrolling through confident predictions of AI's demise. I am ending it with a much simpler observation.

The bubble narrative is comforting. It suggests that the disruption is temporary, that the investment frenzy will collapse, and that we can return to familiar patterns. I understand the appeal. I felt it myself, staring at my former colleague's LinkedIn post.

But comfort is not strategy. The data tells a different story. AI capability is doubling every three months. The scaling walls keep falling. The benchmarks keep being surpassed. The timelines keep compressing. And the people building this technology are simultaneously amazed by what it can do and genuinely worried about where it is going.

Three questions for your next board meeting:

  1. If AI capability doubles every three months, is your organization's technology evaluation cycle fast enough to stay current? Most enterprises operate on 6-to-12-month procurement cycles. That means by the time you deploy a solution, it is already one to four generations behind.
  1. What is your exposure if the bubble narrative is wrong? If you have been underinvesting in AI based on the assumption that the hype will cool, what happens when it does not? What market position do you lose? What capabilities do your competitors build that you cannot replicate in time?
  1. Do you have a governance framework that can scale with exponential capability growth? Static compliance is dead. If your AI governance consists of an annual review and a policy document, you are governing 2024's AI with 2019's tools.

The question is no longer whether AI is a bubble. The evidence has answered that. The question is whether your organization is moving fast enough to capture the value, and carefully enough to manage the risks, of a technology that is accelerating faster than any in human history.

That LinkedIn post from my former colleague is still up. I thought about commenting. I decided not to. In three years, the data will have made my argument far more convincingly than any comment thread ever could.