Mentalité d'Imprévisibilité

First, my verdict: this time, don't go all in on a single path, and don't expect to hedge by spreading across a few directions—diversification itself doesn't resolve true uncertainty. You got stuck last year, and the problem most likely wasn't that you "bet on the wrong direction," but that you treated a "pseudo-era" as a calculable risk and placed heavy bets on it. Real unpredictability means you can't even tell whether it's an era shift at all, let alone calculate probabilities. So this time, the positioning principle is simple: make sure you can survive no matter which path dies, while also ensuring that if one truly succeeds, you capture a big share.

Concretely, break it into three steps.

Step One: Separate the "estimable" from the "non-estimable"

You currently have two types of things on your plate: first, the AI application projects you've invested in yourself, and second, your company's business shifting toward AI. These two have different levels of unpredictability and can't be placed on the same betting table.

The application projects you've invested in—especially early-stage ones—fall under true uncertainty: you can't calculate their survival rate three years out, because the market, technology paths, and regulations could all change, and there's no historical data to reference. For this part, stop trying to calculate "which direction has the better odds"—you can't, and anyone who says you can is lying to you.

The company business transformation is relatively more manageable: your existing customers, cash flow, and team capabilities are known, and you can estimate "how much it will cost to transition and how much old business you'll lose." This part is calculable risk, which can be covered with a margin of safety.

Separate these two, and your decision-making becomes clear.

Step Two: Position with the "barbell strategy," not equal spreading

Taleb's barbell strategy is: one end is extremely conservative (protect the core), the other end is small bets on high payoffs (bet on a few high-ceiling opportunities), and avoid the middle ground. This precisely cures your "stuck last year" problem—the middle ground is those things that "look stable but are actually pseudo-eras," things that seem like upgrades but are really just reskins.

Here's the specific setup:

• Company business: Move only 20%-30% of resources toward AI transformation, and keep the rest defending existing cash flow. Don't do something like "full AI transformation"—that turns calculable risk into an incalculable gamble. For every investment in the transformation, ask one question first: if this technology line is proven to be a pseudo-era next year, how much do I lose? That number must be something your company's annual profit can absorb.
• Personal investment projects: Don't sprinkle money across five or six directions. Pick two to three projects you genuinely understand and that each have the potential to "survive independently," and set each bet at a scale where "losing it all doesn't affect your company's transformation." This is optionality—keeping multiple viable paths open, but none of them life-threatening.
• Cash: Keep a 12-to-18-month cash runway. This isn't conservatism; it's the ticket that lets you keep placing bets when true uncertainty hits. Last year you were stuck for a year, most likely because you didn't have enough cash and were forced to hold on a wrong path.

Step Three: Identify your current points of fragility

Getting trapped by a "pseudo-era" last year means you had a single-point dependency: you bet on "a certain technology path will definitely succeed." Now examine whether your investment portfolio and company transformation plan have similar single-point dependencies. For example:

• Do all projects rely on the same underlying model?
• Is the transformation plan designed around only one application scenario?
• Does the team's core skill set cover only one tech stack?

If the answer is "yes," then decouple: let different projects run on different models, different scenarios, and different customer bases. That way, if any single point collapses, you only lose one piece, not the whole game.

Closing suggestion

One thing you can do today: list all your AI-related investments (personal projects plus company transformation) in a table, and mark three numbers for each—amount already invested, maximum acceptable loss, and exit conditions. Write the exit conditions clearly as "what situation triggers my withdrawal," for example, "if this model doesn't release an iteration in six months, I stop." Once this table is done, you'll know whether you're currently fragile or antifragile. Don't rush to adjust positions; first see clearly where you stand.

First, a direct verdict: your question itself conflates "risk" with "true uncertainty," which is why it feels like no matter how you position it, it's wrong. The 2021 chip shortage was a "risk with calculable probability"—there was historical data, supply chain cycles, and you could estimate roughly how long it would last and how much impact it would have. But the 2026 round is different: DRAM/HBM is being siphoned off by AI data centers, compounded by geopolitical conflicts disrupting logistics, and a large chunk of this is "true uncertainty where even probabilities can't be calculated"—there's no precedent; you don't know how long the suction effect of AI demand will last, or whether geopolitical conflicts will escalate into longer logistics disruptions. So your question "how much to stockpile" is itself misasked; what you should ask is, "How do I position myself so that I can afford to lose even when I can't predict accurately?"

Step 1: Separate the "estimable" from the "non-estimable"

Don't rush to set a stockpile quantity. Take a piece of paper and split two lines:

• The estimable part: Your chip procurement volume over the past three years, supplier lead time fluctuations, inventory turnover days, and the actual number of days your lines were down during the 2021 shortage. These have data and can be calculated. For example, if you were cut off for 6 weeks in 2021, your safety stock should at least cover more than 6 weeks of consumption—that's your safety margin, and it's calculable.
• The non-estimable part: Whether AI demand will further drain capacity, whether geopolitical conflicts will escalate to port blockades, or whether new export controls will emerge. These have no precedent—don't try to calculate probabilities; doing so is just self-deception.

Step 2: Assess your current position

You said "you've been burned before and your production lines have stopped," which means you're likely fragile right now—if supply is cut off, your lines stop and your cash flow breaks. The essence of fragility is "single dependency + no buffer." Your current position is: betting everything on "the supply chain won't have a major event," and if something happens, you shatter.

Step 3: Use a barbell strategy to set your position, not "how much to stockpile"

"Stockpiling to prevent supply disruptions" and "keeping cash flow from blowing up" aren't an either/or—you're using a flawed single strategy. The right approach is a barbell:

• One end: Conservative, deterministic safety margin. Use the 2021 shortage duration as a baseline, add 30% redundancy, and calculate your "minimum survival inventory." This part must be stockpiled—even if it means tying up cash, you have to protect it—because the cost of stopping your lines far exceeds inventory costs. This is your floor, not a "what if," but a known risk that "something similar will definitely happen."
• The other end: Small bets on high-payoff optionality. Don't put all your chips into stockpiling. Use a small amount of capital to build "alternative options": for example, sign framework agreements with two or more non-mainstream suppliers, lock in backup logistics routes in advance (even if pricier), or convert some non-core production lines into quickly switchable flexible lines. These aren't stockpiling—they're buying "options"—low cost in normal times, but if supply is cut off, you have multiple paths to take.

Step 4: Identify and dismantle single points of dependency

What's your most fragile point right now? Likely "main supplier + single logistics channel." The way to decouple: at least two supply sources for core components, even if the second costs 10% more—it's worth it because that's your insurance. For logistics, don't rely on just one shipping line; even if it costs more to use air freight or a detour route, keep a backup channel.

Closing action

Go do one thing today: calculate the loss for every day of the 2021 shortage, including line stoppage losses, penalty fees, and customer attrition, then divide that by the capital cost of the inventory you plan to stockpile. If the former is far greater than the latter, your stockpile isn't enough yet; if they're about equal, you're stockpiling reasonably; if the latter is far greater, you're over-stockpiling. This number will directly tell you where your position should be—no guessing needed.

First, split your question into two halves: one is "whether predictions are reliable," and the other is "whether unpredictable thinking will become an excuse for not planning." For these two, I'll give a direct judgment on the former, but the latter needs to be unpacked.

Predictions are indeed unreliable, but "not predicting" and "not planning" are two different things.

Knight, back in 1921, split "not knowing" into two types: what can be calculated as probability is risk, and what can't even have a probability calculated is true uncertainty. The "nine out of ten predictions are wrong" you see—the vast majority of errors fall into the latter category. Things like which AI player will succeed, which tech route will win, or what the market will look like next year—these have no historical data to rely on, and you can't even estimate probabilities. When experts force a number, it's essentially packaging "true uncertainty" in the guise of "risk." So you're right; this kind of prediction isn't worth paying attention to.

But "not looking at predictions" doesn't equal "not planning." Quite the opposite: unpredictable thinking forces you to shift the focus of planning from "guessing which path will win" to "surviving on any path." That's smarter planning, not no planning. The object of planning changes: before, you planned for the "optimal path"; now, you plan for "survival capability in the worst case." You don't need to know which AI company will come out on top; you need to know—if every path you bet on fails, can you afford the loss?

So, will it become an excuse for not planning? Yes, but that's a misuse. The entire value of this framework lies in forcing you to answer three questions, and once you answer them, you'll know what to do:

First, segment. Break down the decision you need to make: which part is risk with calculable probabilities (like existing customer churn rate, cost curves of mature technologies), and which part is truly unpredictable (like next year's policies, paradigm shifts). For what's calculable, estimate with data; for what's incalculable, clearly admit you can't calculate it—don't force a gamble.

Second, assess positioning. Are you currently fragile, robust, or antifragile? Fragile means "breaks at a touch"—like cash flow only lasting three months, all revenue tied to a single client, or a tech stack locked into one route. Robust means "can withstand but doesn't get stronger"—you have buffers but don't leverage shocks. Antifragile means "benefits from being hit"—limited downside, open upside. First, find the fragile points; those are the most critical.

Third, adjust positioning. Use three tools, and don't mix them: margin of safety (cash runway, time buffers)—applies to the part you can't afford to lose; barbell strategy (one end extremely conservative to preserve life, one end small bets with high payoffs)—applies to the part you want to gamble on; optionality (keep multiple viable paths, include exit clauses in contracts, identify single points of dependency and decouple them)—applies to the part you haven't seen clearly yet.

Specific to your scenario: which AI tech route wins is truly unpredictable—don't predict it, don't go all in on a single route. But "how long your company's cash flow can last" is calculable—calculate it clearly and leave a sufficient margin of safety. For tech selection, use a barbell—use mature solutions for core business to preserve life, and allocate a small percentage of budget to one or two high-odds new directions; if you lose it all, it won't hurt your core. Include exit clauses in contracts, and don't rely on a single point in your supply chain.

Finally, here's an action you can implement immediately: today, write down the decision you're agonizing over, list "calculable probabilities" on the left and "incalculable probabilities" on the right, then ask only one question about the right side—if this fails completely, what's my worst outcome, and can I bear it right now? The part you can't bear is where you need to immediately add margin of safety and decouple. Once you complete this action, you'll find that "unpredictability" not only doesn't make you lie flat, but it makes your moves steadier.