AI at the End of 2025: The Line Between Progress and Overreach

The Line Between Progress and Overreach

By Mark Evans MBA Founder 360 Strategy, Aka Rogue Entrepreneur

Previous article referenced: “The Truth About the AI Bubble: A Perez–Azhar Playbook for Business Leaders” — https://www.360strategy.co.uk/post/the-truth-about-the-ai-bubble-a-perez-azhar-playbook-for-business-leaders

Howard Marks chose his moment carefully. On the 9th December 2025 he published a memo titled Is It a Bubble? for Oaktree Capital’s clients and the wider investment world (Marks, 2025). The topic is the AI boom and whether we are drifting into bubble territory.

Howard Marks is not a passing commentator. He co founded Oaktree Capital Management and built his name by navigating credit cycles with a cool head. His memos are written for serious investors who want judgement rather than slogans. Warren Buffett has previously said they are the first thing he reads when they land in his inbox, which gives a sense of how widely they are respected.

So when he turns his attention to AI and starts asking hard questions about bubbles, it is worth pausing. His memo has also pushed me to go back to my own article from earlier in 2025 (link above) and test where I was right, where I was early and where I missed important signals.

What Howard Marks brings into focus

Marks starts by separating two things.

One is the behaviour of companies building AI. Chips, data centres, models, tools. The other is how investors and lenders are behaving around them. He is explicit that he cannot judge the detailed industrial logic of every AI player, but he can analyse prices, capital structures and psychology (Marks, 2025).

On the builder side, the story is clear enough. Useful products. Strong demand. Rapid model improvement. Real productivity gains in structured work.

On the financial side, the picture is more troubling. Valuations that seem to price in a very smooth future. Long dated borrowing used to buy hardware that may feel old in two years. Circular deals that record revenue and commitments on both sides without proving long term economic value. A tone in parts of the market that has moved from curiosity to near certainty

.

His point is not that AI has failed. It is that the gap between technological reality and financial expectations is widening. That gap is where fragility tends to appear.

Inflection bubbles, mean reversion and why AI feels different

One of the most interesting parts of the memo is where Marks picks up a distinction from Byrne Hobart and Tobias Huber’s book Boom: Bubbles and the End of Stagnation (Hobart and Huber, 2024; Marks, 2025).

They describe two broad types of bubble:

• Mean reversion bubbles, where a financial fad rises and falls without changing much. For example, subprime mortgage securities in the mid 2000s. The product is temporary. Prices overshoot, then fall back.

  
  

• Inflection bubbles, where speculation forms around a genuine technological shift. Railways, electrification, the early internet. In these cases the bubble bursts, investors lose money, but the underlying infrastructure reshapes the world that follows (Hobart and Huber, 2024; Perez, 2002).

  
  

Ben Thompson at Stratechery picked this up as the “good” versus “bad” bubble idea in his piece The Benefits of Bubbles, arguing that inflection bubbles front load the investment that society would never make if it relied on cautious planning alone (Thompson, 2025). Derek Thompson made a similar point in AI Could Be the Railroad of the 21st Century. Brace Yourself, where he draws explicit parallels between today’s AI build out and nineteenth century railroad manias (Thompson, 2025; Marks, 2025).

Marks does not romanticise any of this. He accepts that inflection bubbles often accelerate useful change, but he also reminds his readers that they still leave a lot of wreckage behind for investors. The railway network transformed Britain. Many railway shareholders were wiped out on the way. The same was true of early telecom and internet overbuilds (Perez, 2002; Marks, 2025).

If AI is in a bubble at all, the implication is that it sits closer to the inflection side than the mean reversion side. That is not a comfort. It simply means the technology is likely to outlive the overpricing.

Lottery ticket thinking and distorted judgement

The other term from the memo that matters for business leaders is lottery ticket thinking.

Marks gives an example. Imagine backing a start up chip maker on the basis that if it has even a one in a thousand chance of becoming a trillion dollar company, then paying a hundred million dollars for a stake looks “cheap” in expected value terms (Marks, 2025). On a spreadsheet, that can be made to look rational.

The problem is that once this logic takes hold, it changes behaviour. Investors start to accept very low probabilities of extreme outcomes as a reason to keep pouring in capital. They behave as if a small chance at a huge win justifies almost any price. That is precisely how lottery tickets are sold.

This kind of thinking is a long way from the world most of your clients live in. SMEs cannot fund their roadmap on hopes of one in a thousand payoffs. They have to deal in cash, margin, staff and solvency. Yet the same mindset can leak into corporate planning in subtler ways. For example:

• Betting on a single AI project because it might “transform the business”, with no clear path in between.

• Signing multi year vendor deals on the assumption that being early must be better than being thoughtful.

• Treating AI adoption as a rare event window rather than a sequence of testable steps.

  
  

Marks is not just talking about venture capital. He is describing a pattern of belief that can creep into any decision making when people are dazzled by potential (Marks, 2025).

Revisiting my own view from early 2025

In my earlier article I argued that AI looked like the early stage of a familiar pattern. Speculative overbuild. Uneven returns. A painful correction at some point. Then, over time, a new layer of infrastructure that the economy comes to rely on. That argument was heavily influenced by Carlota Perez and her work on technological revolutions and financial capital (Perez, 2002).

Perez describes two main phases in each big technology wave. An installation phase, where financial capital pours into new technologies and overbuilds capacity. Then a deployment phase, where that capacity is finally used well and becomes the backbone of the wider economy (Perez, 2002; Perez, 2002 cited in Technological Revolutions and Financial Capital, 2002).

I said earlier this year that AI was still in installation. I stand by that. Where my original piece now looks too light is in two areas.

I softened the speed and brutality of the hardware cycle

Chips and supporting kit are ageing even faster than I expected. Payback windows are short. Renewal cycles are quick. If you buy the wrong thing, or buy at the wrong time, there is not much room to grow out of the mistake. You feel the hit in a very short period.

When I wrote the first article I recognised this in theory. I did not give it enough weight in practice.

I underestimated how fast debt would move in

I also expected debt to show up later. Instead, we have already reached a point where significant long dated borrowing is funding assets whose competitive edge may be short lived. Marks is clear about the risk that creates, because he has watched similar structures unravel before (Marks, 2025).

On those two points, his memo has forced me to sharpen my view.

Where I still hold my ground

Despite all that, some of my original conclusions feel stronger, not weaker.

AI continues to create real value when it is paired with the right complements. Clean data. Thoughtful workflow design. Adequate training. Clear governance. The unglamorous architecture of change.

Moreover, nothing in Marks’ memo contradicts that. If anything, his focus on investor exuberance reinforces the need for business leaders to stay anchored in their own operational reality rather than market narratives (Marks, 2025; Perez, 2002).

The lesson from Perez, Hobart and Huber, and now Marks, is that bubbles can fund useful infrastructure, but they do not guarantee sensible adoption. That part is still on leaders.

Where Marks’ lens needs adjusting for operators

Marks writes for people who allocate capital across companies and markets. Boards and founders are trying to keep a specific organisation solvent, competitive and coherent.

Three differences matter.

• Most businesses are not issuing bonds to build data centres. Their risk sits in vendor lock in, dependency on a small number of platforms, weak internal controls and unclear accountability.

• A market level debate about “good” bubbles versus “bad” ones does not help the CFO who has to decide whether to sign a three year AI contract in January.

• Even if AI valuations correct sharply, the tools will still be there. The day to day questions around workflow, skills and governance will remain.

  
  

So, while Marks, Hobart, Huber and Perez are useful guides to the wider terrain, you still have to translate their language into decisions about specific projects and trade offs inside your own business.

What late 2025 actually feels like

On the ground, it feels like this.

The technology keeps improving. Each quarter brings better models and more capable tools.

The financial structures around it feel stretched. Capex plans are ambitious. Debt is rising. Some deals already look like they were designed to keep numbers moving rather than to create long term value.

Expectations in many boardrooms are ahead of capability. People talk in big outcomes. The hard questions about data readiness, skills and process sometimes show up later, if they show up at all.

And in the background, you can sense the psychological shift that Marks describes. Less curiosity. More assumption. Less “what if this works” and more “of course this will change everything” (Marks, 2025).

The questions business leaders should be asking now

The bubble question is interesting, but it is not the most important one if you run a business. More useful questions are closer to home.

• Can your AI investment pay for itself in a timeframe that reflects how quickly the underlying technology will move on

• Are you choosing use cases because they relieve real pressure in your organisation, or because they sound impressive

• Do you have a clear boundary between experiments and production so that pilots cannot quietly become permanent obligations

• Is your spend on data, skills and workflow at least in line with your spend on licences and infrastructure

• How concentrated is your exposure to a handful of vendors, and what is your plan if they change pricing or strategy in ways you do not like

  
  

If you cannot answer those cleanly, the main risk is not Marks’ bubble. It is your own internal set up.

My position as the year ends

I am not calling an AI bubble. I am not dismissing the idea either.

What Marks, Hobart, Huber and Perez have done is give us a clearer vocabulary. Installation versus deployment. Inflection bubbles versus mean reversion bubbles. Lottery ticket thinking. These are not abstract terms. They describe real patterns that we can already see in the AI market if we choose to look (Perez, 2002; Hobart and Huber, 2024; Marks, 2025).

The technology itself remains powerful. The economics around it are noisy. Investor psychology is becoming louder. That combination is familiar from previous turning points.

For leaders, the task is simple to describe and hard to do. Stay involved. Stay sceptical. Build things that matter. Invest in the complements, not just the shiny front end. Leave yourself room to adapt if the wider market has mispriced the future.

Most of the damage in past cycles has come not from the technology, but from the stories people told themselves while the music was still playing.

We finish 2025 with more tension and more ambiguity than we had at the start of the year. That feels honest. The hype curve and the reality curve have started to pull apart. Marks has done us a favour by naming that gap and reminding everyone that conviction without doubt is rarely a safe way to manage risk (Marks, 2025).

References

Howard Marks (investor) (2025) Howard Marks (investor). Wikipedia. Available at: https://en.wikipedia.org/wiki/Howard_Marks_%28investor%29 (Accessed: 10th December 2025).

Hobart, B. and Huber, T. (2024) Boom: Bubbles and the End of Stagnation. San Francisco: Stripe Press. Amazon+1

Marks, H. (2025) Is It a Bubble? Memos from Howard Marks, 9 December. Oaktree Capital Management. Available at: https://www.oaktreecapital.com/insights/memo/is-it-a-bubble (Accessed: 10th December 2025). Oaktree Capital

Perez, C. (2002) Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages. Cheltenham: Edward Elgar. Elgar Online+2carlotaperez.org+2

Thompson, B. (2025) ‘The Benefits of Bubbles’, Stratechery, 5 November. Available at: https://stratechery.com/2025/the-benefits-of-bubbles/ (Accessed: 10th December 2025). Stratechery by Ben Thompson+1

Thompson, D. (2025) ‘AI Could Be the Railroad of the 21st Century. Brace Yourself’, Derek Thompson [Substack], 4 November. Available at: https://www.derekthompson.org/p/artificial-intelligence-could-be (Accessed: 10th December 2025). derekthompson.org+1