All just speculation and almost no hard facts. Demand could still drop dramatically if they run out of investors / can't build all the datacenters they are planning. I dont care if manufacturing is "booked" for 100years if no one will be actually buying. Manufacturers know this is not going to continue like this. Thats why no one is building out manufacturing capacity.
> feels like plants outside of Taiwan or South Korea are basically political stunts rather than significant production
Investment seems to be emphasizing "advancements such as process technology upgrades, higher-layer stacking, hybrid bonding, and high-value products such as HBM" over capacity expansion [1].
But capacity is being built out, albeit modestly. (Mostly, it seems, by Micron.)
I have family working on the micron construction. While it is happening, there appears to be a lot of needless spend going on right now and production is perpetually being delayed. My family member was laughing at the fact that at the beginning of them working there they had "First chip by 2025" which has shifted to "2026" and now "2027".
I have family working on the new micron plant. There's definitely a fair bit of make work going on to expand how much money can be pulled from the chips act. Things like picking up and moving equipment and redoing wiring work because of said equipment move.
I can't say for certain exactly the real reasons for hold ups. I suspect it's because it's at least partially because construction wasn't fully planed out before it started.
Some are legitimate efforts but everyone (other than SSD manufacturers) forgets how complex the supply chain is for modern chips. The companies making the ultra-pure chemicals and substrates have already contracted out their future production. Even something local like ultrapure water requires experts who themselves are in demand and are already employed wherever they reside now. Few will be willing to move to a greenfield factory in a country without an established chip industry.
Can't place it in time but there was an interview with one of the big memory manufacturer firms said that they were waiting out to see if it was a temporary trend.
That said, it might've already been a year ago already when CXMT was less of an actor and they probably hadn't had time to reap enough profits from the craze yet to justify spending.
Both sides of this are partly right, and the 2023 memory crash explains a lot of it.
In early 2023 every NAND maker was losing money on each chip after prices collapsed. Samsung cut NAND output to around 50% of capacity by September 2023, and Kioxia pushed its new Kitakami fab back from 2023 to 2024. After that, there were reports of Samsung and SK Hynix looking at converting some NAND lines to DRAM/HBM, because that's where the AI money is.
So it isn't that nobody is building. It's that NAND makers got burned badly two years ago, they're adding capacity cautiously, and some of the floor space that could make flash is making HBM instead. A fab takes years to come online, so any new capacity started now shows up in 2027-2028 at the earliest.
That's also why I'd be skeptical of "second half of next year" for retail. The memory market has always been boom and bust, and the bust usually comes from oversupply. Right now the makers are deliberately avoiding exactly that.
> I dont care if manufacturing is "booked" for 100years if no one will be actually buying.
They can't just not buy what they've booked for this year and next (and 2028+ if they have made agreements further in advance) if their immediate DC plans fall through, they have a contract to buy and there will be financial penalties if they pull out (unless they go out of business or have a nice loophole hidden in said contract).
Even if they could pull out (via loophole or being willing to pay the termination fee), they won't because that would give competitors access to the inventory (which is part the point of block booking future stock) and when they do need the stuff in the near future it will very likely cost more than the current arrangement will be setting them back. It would be much better to take the inventory and sit on it ready for other DC plans, they won't all fall through.
Isn't this is one of the reasons why hyperscales use SPV (special-purpose vehicles) to build AI DC? If they'll decide to slow down they will just let SPVs to go bankrupt and if there is no business there are no penalties. The parent company will stay protected.
When you have a term for the shell company you're going to use to avoid liability you can bet your butt someone using is going to point that out to a judge when they sue you.
The shell company is financed and owned by a group of investors alongside the company(ies) actually making the capital investment. That's the other big reason for using a SPV: you can access outside capital, which lets you leverage expansion faster than your own free cash flow can, and then gets to share in the upside.
If the market for AI disappears, the SPV becomes unprofitable for everyone, so they write it off and let it go bankrupt. Who you gonna sue? It's not like it's one person or one company behind the SPV, and the entity you actually contracted with is the now-nonexistent SPV.
(This, BTW, is the whole purpose behind the legal construct of the corporation: to allow multiple parties to pool financial resources behind a risky venture, limit their downside to the amount of invested capital, and shield them all from legal liability.)
It’s like when the Ellisons were buying Paramount and Paramount wouldn’t accept an offer from the _revocable_ trust.
When they’re giant companies you have to let them be punished for not doing due diligence or be punished for taking an above market rate price at the risk of the deal breaking (the data center shell companies don’t trade as well as the original company.
Pulling out or not, if they don't need them, they can just sell the contract on to a retail supplier. Even if the booked devices are not suitable for retail, if they aren't actually produced yet, there's the possibility of renegotiating, even paying a penalty to change the contract.
Exactly. Thats how serverpartdeals and others get their stock basically. Massive overbuying from hyperscalers and HotSpares that get sent back to the manufacturer and recertified.
Unlikely. That will be costly guarded information: neither side will want competitors or other potential partners knowing exactly what was agreed as it could affect future negotiations to their detriment.
They won't be allowed to fold. If the big US AI companies are permitted to be in trouble that would give China a huge advantage and that will not be permitted to happen. Even ignoring the US-China competition, those companies folding would have such an effect on the economy it would not be allowed to happen (much like the banks tax payers propped up in 2008 and other financial panics).
>Thats why no one is building out manufacturing capacity
But they are.
And there is no putting this genie back into the bottle.
The more companies and users use AI, the more data is generated, the more storage is required.
Byte-per-byte, my data generation at least 1000x'd. Code, projects, user storage for each of it, images, videos, songs, voiceovers, games, sessions, worktrees, throw-off sideprojects, work data, inference done on data, trained model weights... insanity.
And with coming years, more and more people will do that. The upcoming capacity isn't even close to meeting basic extrapolated demands, but we still have the "economic disbelief" that it will scale like that.
On another note, Apple somehow still offers the cheapest cloud pricing out of all providers.
It's more than a year since future shortage was inevitable, but eventually the small HDD player Toshiba has announced: "Toshiba to double hard-disk drive supply to fill AI chip memory gap" (02.10.2026)
>Users will stop using AI to generate images, videos or text
This is actually the worst customer segment for AI as they won't pay for anything but it's using a lot of resources. Most model providers found out pretty early on the target market is businesses. Practically the same way as social media is "free" to end users and funded by companies.
> Byte-per-byte, my data generation at least 1000x'd
Call me crazy, but I don't think AI will create more data than putting video recording cameras into everybody's pockets. Your code projects can't take up more gigabytes than my dog videos.
A lot of the generated data is slop and often not worth storing long-term. People will realize this and introduce more aggressive deletion mechanisms - how much and how quickly depends on the pricing of storage.
This has happened in the past: Email providers (e.g. Microsoft) offered "infinite" storage, i.e. growing over time, but eventually crawled back and applied a hard limit.
You can't just claim scaling will go on like this forever. Extrapolating the number of Huang's affairs makes it look like he will have dozens of them by next year.
But we are at what now - lets say year 4 - of a new, incredibly innovative technological wave that is radically changing how we approach, use and work with technology.
It is incredibly early, we are at the beginning stages of a massive buildout, and to treat it as just a "hype" thing that will pass or a "bubble" is nothing but a massive misapplication of pattern recognition, as in "oh dotocom was a buble, crypto was a bubble, this surely is a bubble too".
If you asked people in 2000's how much storage they need, they'd say 40GB is surely enough.
If you said to someone "Amazon will make nearly 200B a year renting servers", people would think you were mad.
We have had the chance to be alive and witness how one insane technological revolution (internet) progressed and evolved, changed how humans interact with the world forever and built dozens of different industries. At year 4, the scales of today were unimaginable to anyone but the wildest hackers.
If you're referring to the exponential growth: Moore's law also ended all of a sudden. No one knows where this is going, and anyone predicting the future is just speculating.
This seems to misunderstood what a "bubble" is. A bubble does not nean that the underlying product is all fake/hype/useless/etc. In fact, some of the most famous economic bubbles in history formed because the underlying technology was so obviously revolutionary.
Railroad Mania in the UK began in the first years of a gigantic effort to build railroad routes across the country. 20 years later, there were indeed thousands of miles of railroads built throughout Britain and had reshaped society just as had been predicted.
But the initial wave of enthusiasm and investments to capture value from speculative railroad route contracts still led to many people losing a lot of money after the bubble popped and contracts for routes were cancelled left and right. The railroad build out continued afterwards, but not in the exact way investors had been betting on.
I find my 1TB MacBook disk to be already at its limits with a single project. Optimized AI models as well as image assets with texts machine translated to 30 languages are taking a lot of disk space.
It isn't just me, I understand GitHub usage is way up year over year.
That said, with growing production capacity, it looks like 2028 could be a good year to buy new GPU hardware, memory, and storage.
The Frequency is actually increasing at every event which means grid load is reduced for a short while every minute. Those UPS installations basically briefly switch to Battery every Minute to test Capacity/internal resistance and so on. This so called "load dump" is visible in grid frequency since a lot of these systems work synchronous thanks to NTP
reply