I'll give this a shot, as I spend my days managing a portfolio exposed to sector rotations on the mid-length timeframe.

______________________________________

It is not ideal to think in terms of amorphous projections where you sum up every probability. Try and consider at least some path dependence and realize that there is a risk distribution.

What you are specifically looking at is the scenario where there is a) an AI bubble in the public markets (remember, much of AI of it is marked numbers in the VC world) and b) the possible scenario where the grey swan gets big enough and the risk of violent unwind is high enough to where you see arguable value in shifting to defensive positioning in a somewhat diversified portfolio.

I'll assume you are weighted heavy to QQQ/NASDAQ in the portfolio. If you have a truly diversified portfolio with rules for systematic rebalancing, you never manually shift to defensive (obviously what I and most advisors would recommend).

The first note is that an "AI bubble" in public markets, assuming there is one, is very much tied to the overall state of high market concentration in the trillion dollar tech names. To see a violent AI bubble unwind, it will almost necessarily involve the "concentration bubble"/Nifty 50 2.0 unwinding as well. That means the AI Bubble narrative in public markets is tied to overall market liquidity just as much as the AI specific factors. The passive complex will put a large continual bid in for NVDA/GOOGL/AMZN/etc even in the case of an AI downturn, which should neutralize the violent unwind scenario unless both impulses turn negative (active money moving out of AI, and passive money pulling out of broad based ETFs entirely).

Of course the relationship above is reflexive, which is to say that NVDA gapping 20% down on earnings may well spur wider outflows (as you personally alluded to when your response would be to "go full defensive in my portfolio"), and vice versa. So what you want to be watching is specifically for signs of this contagion spreading.

If liquidity/flows strongly turns negative, the concentration bubble unwinds, NVDA and co tank, and that likely triggers a sell-off in the AI speculative names and qualifies as an "AI Bubble Burst". Therefore you want to be watching general economic weakness like recession signals, layoffs, etc. Bond yields blowing out will also likely kill the AI bubble, as it's capital intensive. Even with the fortress balance sheets of the hyperscalers, the long end blowing out (it's starting to press into dangerous territory currently) would likely kill a hypothetical "AI Bubble". So watch bond yields (and bond volatility / MOVE index).

More specific areas I would be watching is Hyperscaler Capex projections in earnings calls (watch it like a hawk), and some aspects of the trade war such as global digital services taxes in particular (because profit shocks for hyperscalers will likely result in reduced capex to preserve their earnings targets). I would also be watching competition for NVIDIA in the more grey swan areas, like Huwawei revealing a chip with twice the performance per watt as NVDA chips for half the price, and a surprise software stack that kicks ass (so the low single digit outcome areas regarding Chinese competition particularly, which wouldn't actually be bad for AI, but would likely result in a shock move of money leaving US related AI names on public markets, which would be a repeat of the February to April move we saw earlier this year in response to DeepSeek. Markets have memory.)

So let's sketch out some risks over the next few years. Now we get to the part that usually adds little to no value (making predictions about where markets will go). Hint: nobody can do this broadly. Sometimes one may have specific value-adding strong convictions picked out of the universe of possibilities, but those are the exceptions.

I'll even put a % chance guess since that's what you're looking for, but it adds literally no value.

Bonds blowing out: low/moderate likelihood. 15-20% over the next few years for a major bond event imo, and I'd carry that directly to a 20% chance of an AI unwind.

Heavy economic downturn: moderately likely. 25% or so.

And since the scenarios overlap somewhat (stagflation would involve both, and it's a leading wider risk currently), let's move the risk to 30%. But remember that this is a risk scenario that will involve market cap weighted passive indexes as well, so we're really talking about market downturn risk rather than just the AI Bubble popping concurrently, since the latter is less meaningful in comparison even when we're focusing on portfolios.

Now on to the AI specific scenarios, which would be AI falling out of favor with consumers. I personally put this under 5%, so I have strong conviction this won't be the case. I'm mostly disregarding this risk. Most would probably put it closer to 20%, and some would be 50/50 if they think it's a fad. You will find a wide array of predictions here.

How about tech industry specific shocks like global digital services taxes causing a sharp drop in AI investment from the huge balance sheets driving it all? I think it's fairly unlikely as well, maybe 5-10% or so. I have high certainty that Google and Apple in particular are going to have great cashflow for the next few years. There are some side-lines to watch here though. Ex: the US government itself is strongly fending off digital services taxes in trade negotiations and using a big stick, but if the tariffs are confirmed to be nullified (recent court ruling), and tariffs are unwound, the world may use the opportunity to tax the multinational American tech companies while the US is stunned and weak, so to speak, especially if some of the other lines are playing out like bond yields getting dangerous, which puts pressure on global govs to raise revenue. That's an example of the sort of thinking that is required to really position a portfolio for things like esoteric crashes in specific sectors and have positive expectancy. You have to take into account all of these self-reinforcing feedback loops and get creative/contrarian to some degree, and find out where the market may be underpriced/offer value hedges and hidden synergistic correlations. Or you can pay up for generic tail protection/put options, and just pay the tax to cover the risk. In your case, perhaps that would be a good option? Just pay up for a put spread on NVIDIA that has a 10 to 1 payout ratio, spend 0.5% of your portfolio, and hey, you get a 5% payout in the case of an AI unwind (you don't need to go overboard, a 5% cash payout is a huge gift in a market crash while most are in panic).

On NVIDIA in particular. I do think there's a moderate risk that NVIDIA is in an outright bubble currently. I could easily see a scenario where guidance comes in weak, margins come down for some reason (even something like a string of bad tape-outs at TSMC... could be anything), and NVIDIA re-rates lower. That would be painful for the entire market, and this is the risk I would be paying up to hedge (I'm currently hedging this a bit). I think there's a good 1/3 chance that NVIDIA tanks sometime over the next few years, but half of these scenarios would just be contained to NVIDIA and more to do with margins coming down to earth a bit (wow, this is quite nice just to be able to conjure up BS % numbers like that).

____________________________________

So in closing, just worry about having a properly diversified portfolio that self-rebalances according to proven rules. Ask yourself if there was an AI unwind, how would my portfolio look, and does this aligns with my financial goals and risk tolerance? Also ask yourself what would happen if it further picked up steam, since upside risk/opportunity cost is important as well. As previously mentioned, you are concerned with a scenario that closely correlates to index concentration as well, so also ask yourself how it looks if the current market leaders perform badly, and other sectors/market factors drive performance going forwards.

Thank you so much, what a great comment. I'm trying to get to that diversified portfolio in 2026.