Why Are Hyperscalers Still Over-Buying GPUs? Blame Apple (kind of).

At this point it’s become so screamingly obvious that the hyperscalers are drowning in GPU silicon that they can’t plug in that even the corporate press is noticing. Credit to Ed Zitron (@edzitron on X) for being the loudest voice in the room on this subject early on. But the question becomes why? Why did these super-smart people order so much of the most expensive silicon ever made that it’s piling up in warehouses and dark data center space? Microsoft, Amazon, and Google know the data center construction business quite well. They knew there would be power issues. But they bet big and bought anyway. There’s a rational reason (still for very speculative values of “rational”). And the company behind that reason is Apple, even though Apple is the one big tech firm that didn’t play the “buy infinite GPUs game.” Sound crazy? No, it makes perfect sense.

A Few Very Brief Paragraphs on Semi Fab Economics:

Bleeding-edge semiconductor fabs are bleedingly-expensive to build. TSMC’s 2N (the best of the best, right now) and 3N (best of the best, two years ago) fabs cost around $20 billion each. Because construction costs are so extreme, downtime for any reason (gotta pay off that investment!) costs an average of around $2 million per hour. Financially, it is survival-level important that TSMC keep these fabs booked solid and running as close to 24/7/365 as humanly possible.

There are two reasons a fab might stop running:

1. Equipment malfunction. It happens, because these fabs involve sticking their toes over the edge of human understanding of physics and materials science. The engineers that build and calibrate them write off a decent percentage of success to straight-up witchcraft. Even so, at $2 million per hour if something breaks you can imagine the screaming.

2. Not enough orders booked. This is where our focus is.

TSMC cannot have downtime on these fabs due to lack of orders. Fortunately, right now there is practically infinite demand and production runs are sold in advance. Way, way, in advance. This is not always the case, which will become important later. But right now, if you’re getting a GPU built on N2 technology it’s because you placed the order about three years ago. For N3 (the previous generation) it’s a little bit looser: you placed the order about two or two-and-a-half years ago. TSMC might allow small adjustments over this period, but for practical purposes orders placed are “take-or-pay.” Even so, this alone doesn’t explain the absolute flood of GPUs. Again, Microsoft, Amazon, and Google (and probably Oracle, SpaceXAITeslaSolarCity, etc.) knew they were not going to be able to power them and they all bought anyway. The neoclouds just signed while trying not to drool on the documents.

And Now On to the Why:

Zitron speculated that Jensen Huang is twisting their arms so hard that they’re continuously buying stuff they don’t need. This doesn’t sit with me, and doesn’t explain why other vendors and internal products like Amazon’s Tranium GPUs are overproduced as well. Here’s a far more realistic reason as to why they did it: Even years in advance you couldn’t just roll up to TSMC with a pile of cash and place an order on N2 when it was brand new or N3 when it was the hotness, because prior to the AI madness everyone knew good and well that the first 9-12 months of production on any new TSMC process are the property of Apple. This goes back many years, to when Apple moved from Samsung (a competitor) to TSMC for chip manufacturing. Apple’s CPUs are known for completely annihilating their competitors in the markets they target. It’s not even close. Comparisons are not so much benchmarking as they are humiliation rituals. Part of this is because Apple’s silicon design teams are just that good. The other part is because Apple placed a monstrously huge financial bet on TSMC’s ability to perform the witchcraft necessary to make new processes work (an area where Intel struggled horribly for many years and nearly bankrupted the company). Apple completely bought out the first year or so of production on every new TSMC process. This greatly reduced TSMC’s financial risk of building these insanely expensive new fabs, and it helped Apple maintain their absolute performance hegemony. A classic business win-win.

This relationship, however, put the AI companies in a bind. They wanted to take a big chunk of bleeding-edge production capacity from Apple, not just on N2 but on N3 as well. It’s not easy to disrupt a very long-standing and mutually-beneficial business partnership like that. Apple was TSMC’s most important customer, hands down. To push them out you need huge, huge, huge orders, at much higher profit margins. And, again, these production slots are take-or-pay. Even if you somehow manage to walk away and survive the lawsuit, being blackballed by TSMC is death in the high-end fabless silicon business. Game over. So what almost certainly happened is that back in 2022-ish Jensen Huang went to his potential customers, explained the facts of life to them (not difficult, because none of this is a secret), and they got together and placed one of the biggest and most irrevocable bets in business history. They all got together and broke Apple’s purchasing monopoly on the best chips, and they did it by risking a historically-huge pile of cash.

Unfortunately, it looks like this bet is not paying off. It relied on two things:

1. The leading AI research companies (Anthropic and OpenAI) hitting much higher revenue figures and demonstrating a reasonable path to profitability, and

2. The patience of venture capitalists and the bond markets funding them until (1) happens, and it’s not happening fast enough (if ever).

So here we are, and the multi-trillion-dollar questions are:

1. How deep are the commitments from Nvidia and the hyperscalers to TSMC? TSMC didn’t break their arrangement with Apple for one or two or probably even five years of business.

2. Can these companies actually afford to keep their commitments? TSMC is sending teams to check out data center construction progress and try to figure out what’s real, which is something they have never done and shows a certain lack of faith.

Where Do We Go From Here?

Here are my speculations:

Microsoft (for now), Amazon, and Google can take the financial hit, but it will be an ugly hit. Oracle cannot. The neoclouds are toast.

A good number of these GPU servers never find homes (see my article here). They require very specialized power, very specialized cooling, etc. You can’t just put them anywhere. By the time potential data center space exists, they’ll be obsolete.

In the long run, the biggest loser is TSMC. Apple is known to hold grudges against vendors that stab them in the back. Apple will probably maintain a second-sourcing relationship with Intel, as long as Intel doesn’t shoot itself in the foot too badly (Intel is looking better now but I still think they’re still iffy a few years out). But without Apple’s massive up-front financial commitments, a downturn in the famously cyclical semiconductor business will disrupt or straight-up block TSMC’s ability to build new bleeding-edge fabs, and we’re probably heading for one of those soon. Expect TSMC to grovel to get Apple back to their previous arrangement, and expect Apple to go along with it but give the pricing knife a few extra-hard twists along the way. It’s unlikely that Apple will build its own fabs, even though they have the money to. The key problem is finding the people with the witchcraft skills to run them, even at any price. They won’t take that route unless they believe they have no other option.

Nvidia will take their profits, but their revenue and share price come back down to Earth. They burned a lot of bridges while surfing this wave (weird metaphor mix, but whatever), and we’ll see how much that bites them.

AMD goes back to status quo. They didn’t manage to cash in hard like Nvidia, but they didn’t make any new enemies either. Their status quo is “much better than Intel, but still stuck with x86 in a world moving towards ARM.”

Intel also remains status quo: the company that sells CPUs to people who can’t get enough allocation from AMD, and while trying to claw its way back from the most epic fall from grace in the history of tech. Their biggest burden remains an absolutely cursed board of directors (who should be forced to attend meetings wearing clown suits and full makeup).

The biggest winner is probably Apple. They bet big on local inference, which seems to be the most important AI hardware market going forward. It looks like their first truly built-for-it systems (based on the M7 Ultra in late 2027 or early 2028) will be hitting right when costs start coming back down to normal. And as a bonus they’ll probably have much better pricing terms with their most expensive supplier.

Microsoft has other problems, possibly fatal (yes, really), but that’s a subject for another post.

This article was originally published on my Substack on October 23, 2026.