** originally published here; also on Linkedin

For a couple of months now my timeline keeps serving me videos of people predicting the end of the AI party - and not just random doom accounts, these are people with serious credentials. A former Wall Street fund manager explaining that a three trillion dollar AI bubble is about to burst. A billionaire investor, one who genuinely called 2000 and 2008, telling almost nine million viewers that his personal advice on US tech stocks is to "sell them all". Well produced, convincingly delivered. And yet some questions kept nagging at me while watching: should I really be worried? What about my own modest investment portfolio? And, if the end is really this close and this evident, then shouldn't I be able to find it in the numbers myself?

So I went looking. I've been running a long conversation about the AI investment cycle for a few months now anyway - one of those slow burning discussions I keep coming back to, started under Opus 4.6, continued under Opus 4.8 and today I decided to put GPT 5.6 to the test with the earlier results as this seemed like the moment to put it under pressure with one of the latest frontier top guns. But even that was not enough so after GPT had its merry way with the knowledge I handed the whole history to Fable (yes, the same model I wrote about here last month, still consuming twice the tokens) and asked it to attack the reasoning instead of summarising it: where does this break? And after that, to check what remained against actual sources. It came back with four weak points in my own analysis, and I have to admit that most of them held up.

My own blind spot first

The one that hurt: I had assumed that the cheap Chinese and open models are already eating the enterprise market from below, because that is what all the usage charts in my feed suggest. In the actual enterprise spending numbers they sit at around one percent, and the share of open models even declined last year. So my own analysis contained exactly the kind of convenient gap I was about to accuse the doom videos of. (Which is exactly what the exercise was for, but it stings a bit anyway.)

What the doom videos get right

Then I went back to the videos, this time with the transcripts and the numbers side by side. I was expecting to catch them on the facts and that is not quite what happened. To my surprise, the fund manager builds his case on credit: private credit now funds a serious share of the AI build-out, and those funds are showing withdrawals and gating. That is not hysteria - my own digging points at the same soft spot. AI-linked companies sold roughly $244 billion in bonds in the first half of this year, more than double all of 2025. Analysts project the combined free cash flow of the big cloud builders to cross below zero around this quarter - the era of AI capex paid from own pocket is ending in front of us. Chip vendors are financing their own customers at a relative scale that, by some counts, already exceeds what Lucent did at the top in 2000. On the uncomfortable numbers, the doom side and I largely agree.

The billionaire surprised me even more. His central argument is that the great bubbles always form around the most important ideas, precisely because everyone can see they are real - railroads, the internet. Amazon fell 92% in the dot-com crash and then went on to inherit the retail world. If anything, that is the most useful frame in this whole discussion: the difference between a great technology and a great investment. On that point he and I are simply on the same side.

Where we part ways

So where do we part ways? First, on the facts that seem to be overlooked, forgotten for the sake of the argument. Nvidia trades at roughly 21 times forward earnings after another record quarter (the billionaire quotes 35 to 40 times for the broad market; the S&P's plain trailing multiple is closer to 29, the cyclically adjusted one closer to 41 - either way well short of Japan's 65 in 1989, by his own comparison). The two AI-driven corrections we already had this year, one in software in February and one in semiconductors in June and July, were both absorbed - everyone who called those the beginning of the end was right for about three weeks and then not anymore. Cloud revenues are still accelerating, at least for now. There has, so far, not been a single major AI-linked default. And the "Chinese model just as good at a third of the price" from one of the videos actually tripled its price compared to its predecessor.

Second, and this matters more to me, on the shape of the claims. Asked when the collapse comes, the billionaire's honest answer in the interview is careful: it "would be compatible with history" for the peak to be very soon - and elsewhere, that it could be weeks, months or years. The title above that same interview reads: "I'm SELLING everything. The crash is already here!" Somewhere between the recording and the upload, a careful old bear was turned into a crystal ball. The fund manager, meanwhile, closes his case with "credit always ends the party. Always." It sounds like analysis, but a rule without a date, a threshold or a condition under which it would be wrong is not something you can be wrong about. I'm aware that a lot of video and article titles have a measure of clickbait in it, obviously the producers want us to read/watch their content. But I find it annoying when people make high stake claims without the actual content of their product backing it up, it wastes my time(!) and it is misleading. I wrote here a few weeks ago about conviction doing the work that evidence should be doing - this is the same mechanism, at macro scale and with better production values. Though in fairness, part of the conviction turns out to be added in the edit.

And while I was writing this, the algorithm escalated and served me the third layer: the anonymous, well-produced channels that recycle the credentialed bears into something darker. The one it picked for me cites the billionaire by name, quotes Cisco's peak price-earnings ratio of 201 correctly, and never finds room for what Nvidia actually trades at today (just over thirty times trailing earnings, if you want the honest comparison). Its numbers stop somewhere in 2024: "over $300 billion a year" of AI capex, while this year's guidance is more than double that, and an insider-selling wave from February 2024 presented as the smart money leaving "near the highs" - two and a half years ago, with the indices far higher since. Even here a real mechanism sits buried inside (the circular money flow between investors, AI labs and chip vendors is real, and it's on my list below), but by this layer there is no forecast left at all. It is a finished story with culprits, and a helpful link to the next crisis video at the end. Funny enough, the people closest to the source are the most careful with their claims - the certainty only really starts a few steps down the chain.

A watchlist instead of a prediction

So what the exercise left me with is not a counter-prediction but a list of things to watch - you could say I took their mechanisms seriously and gave them thresholds. Does the free cash flow of the hyperscalers recover, or does it keep sinking while the building continues? Whether the withdrawals and gating in private credit spread further, or settle down (demand in the bond markets has been thinning since February, so this one I check first). Whether anyone shortens the depreciation schedule on their GPUs, or takes a first real impairment. And whether those cheap open models ever get past the low single digits of enterprise spend after all. For each of these I've written down a threshold, a source and a date to review it. Not very exciting, I know, but at least it is checkable.

More relevantly: there are a couple of key dates on that list which beat all the others. The most quoted numbers in this entire debate - the revenues and margins of the frontier labs themselves - have never been audited. What circulates are run rates announced in the middle of fundraising rounds, and leaked figures that don't reconcile with each other, sometimes not even close. Both of the big labs have now reportedly filed confidentially for an IPO. Which means there is an end date to this whole debate: the day those prospectuses become public, most of the numbers people are now throwing at each other become checkable for the first time. Hours of doom content, and that date is not mentioned once.

So, should I be worried?

And, since I raised the question myself earlier: What about my own modest portfolio? I'll admit I did change the mix somewhat - less dependent on the US stock market, a bit more Europe, and away from the most heated and maybe overcommitted tech funds. Do I think the markets will crash? No, I don't. Do I believe there might be some downward corrections along the way? Yes, there probably will be - and that is the reason for the change, nothing more dramatic than that. Which sounds, I realise, like a modest version of the billionaire's advice. The difference is that my version comes with a list attached, and with conditions under which I'll change it back.

So no prediction from me, but a question instead, and I mean it for both camps. If you are convinced there is a bubble: what evidence would talk you out of it? If you are convinced there isn't, same question. I have my list. And if you have signals on yours that I'm missing, or if you think this whole watchlist approach misses the point, I'd genuinely like to hear it.

Sources

The videos discussed in this post

Judge for yourself whether the titles match the content — that is rather the point of the post.

Financing and credit

Valuations and market data

All figures were last verified against these sources on July 23, 2026. If you spot a number that no longer holds, that is exactly the kind of comment I am hoping for.