TSMC just posted a record quarter, and the fight over what comes next is not about yield or price. Nvidia, Apple, and Broadcom are reportedly competing for the same production slots at the same foundry, and CNBC’s reporting calls the contest for capacity “fierce.” That word matters more than the earnings number. A record quarter paired with a capacity brawl means demand has already outrun what the world’s most advanced foundry can physically produce. The question worth asking is not whether companies want more AI chips. It is what happens once the chips themselves become the thing everyone is rationing.
On Bluesky, a user writing as headfallsoff.com raised a milder version of that question: whether you can avoid AI because, as the post puts it, it is “in all software now.” The post frames this as an argument that might be true, which casts AI adoption as a decision engineers and product teams still get to make, team by team, app by app.
Put that framing next to TSMC’s earnings call and the disagreement becomes obvious. One side describes a preference. The other describes a production line running at its limit while Nvidia, Apple, and Broadcom bid for whatever is left. A preference can be reversed. A production constraint cannot, not by anyone holding only an opinion about software architecture.
That is the moment the obvious reading gives way to a truer one. The debate over whether to put AI into software only makes sense if the supply of AI hardware is elastic enough to let every team opt in or out on its own schedule. It is not. Albert Rutherford’s The Systems Thinker explains why a system built around one deepening dependency behaves this way.
Once a structure leans hard enough on a single input, the reinforcing loop around that input dominates everything meant to balance it, and the cost of pulling back rises faster the longer you wait. Rutherford writes about oil fields, not chips, but the mechanism is the one TSMC’s capacity fight is showing in real time: “The larger the stock of oil they have to begin with, the longer the reinforcing loop will dominate the balancing loop.”
Replace oil with advanced-node wafers and the logic holds. Every company that built its roadmap on access to TSMC’s most advanced nodes has already deepened its dependence past the point where backing out is a live option. Nvidia cannot simply need less capacity. Apple cannot un-design silicon that assumes it. The loop reinforces itself, and stepping away gets more expensive every quarter the record continues.
So the headfallsoff.com question answers itself, just not the way either side of a software debate expects. Teams are not embedding AI into every application because it is fashionable or because leadership read the same memo. They are doing it because the physical capacity to build AI hardware sits inside one foundry, rationed among a handful of buyers with the leverage to secure it, and growing tighter relative to demand with every record quarter.
Once the hardware supply narrows that far, software follows the hardware, not the other way around. The companies that locked in TSMC capacity early get to decide what AI looks like in every application downstream; everyone else inherits that decision. Systems rarely fail from one bad choice. They lock in through a thousand small dependencies, and by the time anyone notices, leaving costs more than staying ever did.
Leave a Reply