Howard Marks on the AI Bubble, Irrational Exuberance, and Investing Under Radical Uncertainty
Recorded on the day SpaceX completed the largest IPO in history — with Anthropic and OpenAI also filing to go public — Howard Marks argues that the AI moment carries every marker of a classic bubble while remaining genuinely impossible to call. His point is not that AI is overvalued but that it is unvaluable by ordinary methods: the rational response is to treat participation as calibrated speculation, hold a forecast alongside an honest probability that the forecast is wrong, and stay alert to the possibility of loss rather than certain of the upside.
Key ideas
- Exuberance is certain; whether it is irrational is not. Borrowing Greenspan’s phrase, Marks grants that the market shows exuberance — SpaceX pricing at over 100 times sales is the tell — but insists nobody can yet say it is irrational, because nobody can specify what AI will do, for whom, or at what profit. The honest position is uncertainty, not a verdict.
- AI is the least specifiable technological revolution yet. Railroads, radio, the automobile, computers, the internet — each came with a knowable use. AI has ‘unimaginable, unlimitable upside’ and a matching degree of uncertainty, so a value investor ‘can’t put numbers on a pad’. Deciding whether and at what price to participate is, in his South African friends’ phrase, ‘a thumb suck’.
- ‘This time it’s different is never different.’ Every prior innovation drew too much capital, built too much infrastructure, and ended in a money-losing bubble where the capital providers lost out. Marks wrote this year that if the AI boom does not produce a money-losing bubble, ‘it’ll be the first’ — while granting it could happen.
- Deal with the future using two things, not one: a forecast and the probability it is right. The recurring error is confidence. You may forecast an optimistic AI future, but pairing it with ‘I’m highly confident I’m right’ is the mistake; the young investor laying a portfolio’s foundation has no more basis for certainty than anyone else.
- The moat is dying, so the safe ground is shrinking. Disruption — newer than the value-investing tradition — has hollowed out classic moats: the local newspaper, then enterprise software (the ‘SaaS-pocalypse’). Marks now asks what cannot be disrupted by AI, and finds the answer keeps shrinking — even plumbers and masseurs.
Content
Exuberance, and the limits of calling it irrational
Marks opens on the distinction that organises the whole conversation. Exuberance is observable and certain; irrationality is a judgement no one is entitled to yet. He will not declare a bubble, because to do so would require knowing AI’s parameters — ‘what AI will be able to do or when or for whom or how much profit it’ll produce and for whom’ — and he is adamant that no one, including the hyperscalers locked in the arms race, can supply those numbers. This is the discipline of his tradition turned on a problem it cannot ordinarily handle: a value investor refuses both the bull’s certainty and the bear’s.
The least specifiable revolution
Pressed on whether this feels like the dot-com era, Marks says it does — but with a difference that cuts both ways. He has read about technological innovations across 150 years; this ‘may be the greatest, may be the most powerful’, and is ‘in many ways the least specifiable’. The railroad would carry goods coast to coast; radio would carry messages; the uses were legible even when the profits were not. AI’s upside is unbounded and its outcome unknowable, which is precisely why ordinary valuation fails: ‘if somebody will tell me what they think Anthropic net earnings will be in 2036, I’ll bet them that they’re not within 50% of the truth.’ Buying the IPO therefore means accepting that you are ‘closer to speculating — and I don’t say that word pejoratively — than analytical investing’.
‘This time it’s different is never different’
Marks places AI in the lineage of railroads (1860s), radio (1920s), the automobile, computers, and the internet. Each was accompanied by a bubble: excitement, a winner-take-all race, capital flowing ‘in like water’, too much infrastructure built, prices too high, and capital providers who lost money. He cites his own recent memo: if this innovation ‘doesn’t produce a money-losing bubble, it’ll be the first’. He then voices the optimist’s rejoinder — this time it’s different, because the value is incalculable — and dismantles it with the historian’s observation that the line is recited in every bubble: ‘this time it’s different is never different.’ The asymmetry he draws is not a recommendation to abstain but a demand to stay alert: ‘the way people get into trouble is by not being alert to the possibility’ of loss.
Forecast plus the probability you are right
His constructive method is to refuse single-point forecasting. ‘To deal with the future, you need two things, not one. You need a forecast, and you need a judgment regarding the probability that your forecast is right.’ An optimistic AI forecast is permissible; high confidence in it is the error. Notably, Marks tilts bullish on capability — in his last memo he judged AI ‘more likely to be underestimated today than overestimated’ — while staying bearish on the terms: the open questions are how much capital the technology should receive and what a share of an AI company is worth, not whether the technology is real. This is the same explore-the-distribution humility that the wiki’s What Makes a Great Investor theme treats as the core of the craft, and it sits directly against the certainty critiqued in Over-Attribution of Intelligence.
Disruption and the death of the moat
Marks marks the internet’s arrival (~30 years ago) as the moment a new force entered value investing: disruption. The moat — the protective barrier a cautious investor prizes — has proved perishable. His worked example is the local newspaper, once an near-perfect moat (irreplaceable classifieds, trivial price, daily repurchase, no viable competitor), destroyed by digital communication. More recently, enterprise software looked impregnable until coding models triggered what he calls the ‘SaaS-pocalypse’ around the start of February, when the market briefly concluded the whole software industry might be obsoleted because ‘AI writes its own software’. The unsettling conclusion: he can no longer name an industry safe from AI disruption — not plumbers, not masseurs — so ‘the probabilities that can be assigned to the future are much broader today than ever’. This connects to Value Investing and to the no-moat economics argued from the technology side by Gary Marcus on AI's Overstated Intelligence, LLM Economics, and the Limits of Scaling.
The investment spectrum: hyperscalers to lottery tickets
Asked how one actually invests around the unknowable, Marks lays out a spectrum and tells the listener to choose a point and a portfolio weight. At the lower-uncertainty end sit the hyperscalers (Amazon, Google, Meta, Microsoft) — established moats and ‘enormous operating cash flow’, but diluted AI upside because other businesses hold back the growth rate. In the middle sit the AI one-product companies (Anthropic, OpenAI, Nvidia) — harder to specify, but with a high probability of still mattering in five to ten years even if not number one. At the far end sit startups — possibly no revenue, possibly no defined product, but the ground-floor lottery ticket where one winner pays ‘an incalculable amount’, and most buyers lose everything. He is sharp on the danger of the moment: the market is ‘rebranding lottery tickets as certain safe investments’, citing SpaceX (losses up 700% year over year) and the ‘profits don’t matter, the top line’s growing’ justification he distrusts. On Risk Posture terms, the discipline is to know where on the spectrum you are standing and size accordingly.
Profitability, ultimately
Marks traces the recurring retreat of the valuation anchor: from earnings, to sales, to — in 1999-2000 — ‘per eyeball’ and ‘per click’. His conviction is that ‘ultimately, it always comes down to value’, and that profitability will matter again when ‘exuberance is replaced by sobriety’. He invokes Buffett on the internet (2000): it would add to efficiency, ‘but that’s not the same as adding to profitability’. The question he leaves open is distributional — if AI is a labour-saving device and the providers fight a price war, the benefit may accrue to the customer (the shipping, retail, or warehouse company) rather than to the AI provider, so ‘who gets the money’ is unsettled.
Private credit and the tide going out
On the fears around private credit, Marks judges them ‘overblown’. Direct lending for mid-size buyouts has existed under different names since the late 1970s; most loans will pay. The real unease, he argues, is illiquidity surprise: retail investors entered non-traded vehicles (BDCs) that always limited redemptions to ~5% per quarter, did not read the prospectus, and now — in a turn from complacency to alarm — discover they cannot exit at will. He frames both moods as mistakes. The deeper point is cyclical: the 17 years since March 2009 have been ‘salad days’, and ‘it’s only when the tide goes out that we find out who’s been swimming naked’ — of ~700 direct-lending managers, only about 3% predate the global financial crisis, so most have never been tested by a harsh environment. This is the Mastering the Market Cycle thesis applied to the present.
Advice for a young investor
Closing, Marks gives career advice that doubles as a statement of temperament: investing has no reliable physical rules, so ‘if you’re the kind of person who wants to be successful every time, don’t become an investor — become a dentist’. He cites Taleb’s dentistry analogy from Fooled by Randomness, and Buffett’s attribution of a 60-70-year record to roughly twelve great investments. The field rewards those who can deal with ambiguity and live with a batting average ‘far from a thousand’, and who never reach the point of saying ‘I’ve got this figured out’.
Related
- Howard Marks — guest; the wiki’s anchor voice on risk and cycles
- Scott Galloway — host
- Ed Elson — host
- Value Investing — the tradition AI’s unvaluability strains, and the moat-erosion problem
- Risk Posture — calibrating aggressiveness along the speculation spectrum
- What Makes a Great Investor — the forecast-plus-probability humility
- The Most Important Thing — Marks’s book on second-level thinking and risk
- Mastering the Market Cycle — the tide-going-out, salad-days argument
- Gary Marcus on AI's Overstated Intelligence, LLM Economics, and the Limits of Scaling — the no-moat / unprofitability case argued from the technology side
- Over-Attribution of Intelligence — the certainty Marks warns against, named as a concept