I started this article in a Zurich hotel room, scrolling through confident predictions that AI was a bubble about to burst, including a LinkedIn post from a former colleague. Part of me wanted the bubble case to be true. The sceptic in me, the part that has survived a long career in enterprise technology by distrusting hype, wanted the curve to flatten so we could all catch our breath. The comparison with 1999 is emotionally satisfying, the capital expenditure looks alarming out of context, and the complaint that there is "no clear ROI" is half right.
Then I looked at the data that the analysts I respect were tracking. My conclusion is that the capability curve is real, the demand behind much of the capex is real, and the procurement cycle inside most enterprises is now slower than the technology cycle it is trying to evaluate. Valuations may still correct. The technology is unlikely to go away.
What METR measures
METR, an independent evaluation organization, measures how long a task an AI system can complete with fifty percent reliability. Its January 2026 update fitted a doubling time of about 196 days across its full record and about 89 days for models released from 2024 onward. METR notes that these are regression estimates and are sensitive to the mix of tasks in its benchmark.
For comparison, the classic Moore's Law cadence was a doubling roughly every eighteen months. On METR's recent estimate, the length of tasks frontier models can handle is doubling several times faster than that. The frontier has moved from tasks measured in minutes to tasks measured in hours within a few years.
The practical point for executives is timing. Many enterprise procurement cycles for technology run six to twelve months, which is longer than the recent doubling period. A tool evaluated last quarter may already be a generation behind the options available at contract signature.
Why the scaling wall keeps moving
The scaling critics deserve a serious answer, because their argument is the intellectual foundation of the bubble case.
Gary Marcus, the best-known of them, has argued repeatedly since 2019 that deep learning faces fundamental limits. He is a serious researcher raising important questions about reliability and reasoning, and so far capability has kept improving through each predicted limit.
The disagreement often comes down to S-curves. Every technology follows one: a slow start, rapid acceleration and an eventual plateau. It is easy to see the plateau of one curve and declare the technology finished. AI progress has tended to stack curves instead, with a new approach starting to climb before the previous one has fully flattened. As pretraining gains slowed, chain-of-thought reasoning arrived, and as reasoning matured, tool use and agentic architectures followed. Several predicted walls were passed by switching approach rather than by pushing harder on the old one.
Expert-level performance on hard questions
GPQA, the Graduate-Level Google-Proof Q&A benchmark, contains questions designed to be hard even for domain experts with internet access. Experts score about 65% on questions in their own field, and frontier AI systems now score above 90%.
For a hands-on comparison of models across many benchmarks, see the AI Model Benchmarking Dashboard I built on Artificial Analysis data.
The result that unsettled me more came from physics. Researchers gave a machine-learning system measurements from dusty plasma experiments, a notoriously complex area, and it identified physical relationships the researchers had not previously described, with over 99% accuracy in reproducing the data. If your mental model of AI is a faster search engine or a better autocomplete, results like this suggest it is out of date. These systems are starting to extend human knowledge as well as process it.
Creative work is not a safe exception
In many boardrooms I hear the same objection: AI can handle repetitive work, but creative work is uniquely human. The evidence no longer supports that as a planning assumption.
In controlled listening experiments, people often cannot reliably tell AI-composed music from human-composed music, and AI-generated work has won art and photography competitions against professionals. The quality debate is nuanced and ongoing. The strategic point is volume and speed: for many business uses, AI can produce creative output fast enough that the difference between human and machine origin stops mattering operationally.
The companies that do well will treat this as amplification. Strong human creatives working with AI tools can produce work neither would produce alone, while companies that treat creativity as an exclusively human domain risk being outproduced by competitors who integrated AI into their creative workflows earlier.
Forecasts have moved closer
Expert and forecaster expectations for when general AI might arrive have moved sharply earlier in the past five years. Estimates that once placed it several decades out now put meaningful probability within this decade.
Ray Kurzweil's point about exponential change is useful here: exponential progress looks slow until it suddenly looks instantaneous. Imagine filling Lake Michigan by doubling the volume of water each cycle. For most of the process the lake looks almost empty, going from 0.1% full to 0.2% and then 0.4%. Late in the process it is still only half full, and a few cycles later it is full.
The practical implication for enterprise planning is that a three-to-five-year strategy should allow for AI capabilities in 2029 being very different from today's, rather than a little better. Planning for linear improvement in an exponential environment is a common way for organizations to fall behind.
The strongest objection
The strongest version of the bubble case does not deny the capability curve. It says that a real technology can still sit inside a financial bubble. The internet was real in 1999 and valuations still collapsed. Capital expenditure can run ahead of revenue for years, and enterprise evidence of return is thin: most generative AI pilots show no measurable P&L impact, and only a small minority of companies report material EBIT from AI.
I think that objection is largely right about finance and largely wrong about strategy. A correction in AI valuations is possible and would not contradict anything in the capability data. Executives should separate two questions: whether investors are overpaying, which they cannot control, and whether the capability will matter to their business, which they must plan for either way.
The risk that deserves attention
The same people building this technology are worried about it. Dario Amodei, the chief executive of Anthropic, has put the chance of AI going catastrophically wrong at somewhere around 10 to 25 percent. Geoffrey Hinton left Google so he could speak freely about the risks. Yoshua Bengio has compared the current trajectory to driving blindly into fog, and Mustafa Suleyman, a co-founder of DeepMind, has warned about systems that can improve themselves faster than people can understand or control.
I am not a doomer. I have spent several years integrating AI into my own work and into the organizations I have worked with. But the exponential growth that undermines the bubble narrative is the same growth that makes safety urgent. Amodei's figure matters because it comes from someone with close access to frontier research, even if it is not a precise probability. When the person building the car warns that the brakes might fail at speed, you keep driving and make sure the brakes are as good as they can be.
For enterprises, that translates into governance as a strategic priority. With the EU AI Act and ISO/IEC 42001 setting expectations, organizations with mature governance will be better placed than those trying to retrofit it.
Three questions for your next board meeting
- If frontier capability doubles in a matter of months, is your technology evaluation cycle fast enough? A six-to-twelve-month procurement process can leave you deploying a tool that is already a generation behind.
- What is your exposure if the bubble narrative turns out to be wrong for your industry? If you have held back investment on the assumption that the hype will cool, which market positions and capabilities could competitors build that you could not replicate in time?
- Can your governance keep up with capabilities that change within a year? An annual review and a policy document will not govern systems that change every few months.
Monday move
Find out how long your last AI tool evaluation took, from the first request to the signed decision. If it took longer than six months, ask the procurement and IT owners to propose a faster path for low-risk AI tools, with the governance checks that would make that speed safe, before the next evaluation starts.