Hard Drives on Backorder for Two Years as AI Data Centers Trigger HDD Shortage

Do you mean relative to six months ago, or now? Because a lot of the prices have already more than doubled.

(I’m upset because the computer I’ve been planning to build, which three months ago would have come to around ₹90,000, is now up to ₹1,20,000 and climbing week by week, half due to price increases on the same part, half due to forced substitutions on RAM since the cheaper 32GB 6400MT/s DDR5 sticks are completely unavailable. And looking into laptops, for the first time ever I’m seeing manufacturet SODIMM or SSD upgrades being cheaper than aftermarket.)

Per the op:

> and the ongoing DRAM shortage is proof of this, with memory kits costing more than double what they did just a few months ago.

> While enterprise-grade QLC SSDs would entirely power this pivot, Sandisk has already raised NAND prices by 50%, according to another DigiTimes report, after initially warning of a 10% increase two months ago.

So you're basically saying prices may return to normal in two years, and that's somehow a good thing compared to them not being inflated in the first place?

GPUs? No way. The datacenter cards don't even have video output ports, and I think the chips destined for AI / ML training also have everything video/render related removed from the silicon, makes for more yield.

And the other way around, using (cheap) consumer GPUs in servers, I think at least NVDA tries to prevent that with driver-based DRM, so there won't be any flooding coming from there either.

Enterprise SSDs are not expensive only because they have better flash chips, but they have much more of them.

A top of the line write oriented SSD comes with 4-7x more capacity than what it says on the tin, but that extra capacity is used for cell replacement rather than capacity itself.

Mixed use comes with 2-4x overprovisioning, and read oriented is around 2x IIRC.

I think it's the higher density that makes it better for cold storage, which generally has infrequent access, and more reads than writes.

Hence the QLC's endurance being "sufficient for cold storage".

You have to actually examine the real bits on the drive. Resellers don't want to take the time to actually zero a drive, they usually just nuke the partition table.

You also need to physically examine the drive. Corroded fingerprints on the PCB, wear on the port contacts, scratches from mounting rails, etc.

That's how it found out that the last "new" drive I bought on Amazon was actually a used Backblaze drive. It contained terabytes of customer data, and a shit ton of cleartext files. SMART, of course, reported it was a brand new drive with zero hours. Cleartext logs on the drive showed many thousands of hours of runtime.

Physical examination is the only reliable method.

I wouldn't shop there at all. It's a literal scam market. Allegedly.

But why ? They are the good ones. TEMU and Shein bad. /s

- Air conditioners

- generators

- Racks/Cages

- fire supression

- Concrete, Steel, power lines

etc

Downstream of that, AI is effectively also responsible for the current generation of game consoles never declining in price.

Because game consoles are fixed platforms that continue to be manufactured over 5+ years, normally the most expensive parts in the system (the CPU and GPU) would gradually get cheaper to manufacture [and in turn, cheaper to buy] over the course of the console's lifetime—which was often passed onto the consumer in the form of the console's MSRP gradually decreasing. Either the process node for the console's silicon design would stay fixed, and demand for this process node would gradually decrease as larger fab customers move on to newer nodes, decreasing the (effectively auction-based) pricing for fab time on the older node; or the console manufacturer['s silicon vendor] would put their silicon design through a process-shrink, and so, while still paying top dollar for use of the fab's newest node, would be getting more chips per wafer out of that, and again could charge less.

But instead, what we've seen since the start of the AI boom is that there's no longer any price-reduced timeslots to be sold to manufacture the low-BOM parts for these price-sensitive console-maker customers. Instead, both Nvidia and AMD are now getting such high value out of even the older nodes, that 1. the fabs know they can squeeze them, charging full price for those slots as well; and 2. both Nvidia and AMD, in their roles as silicon vendors to the console makers, haven't been able to justify using (very much of) the time they're paying so much for to fab low-BOM-cost parts to fulfill their pre-existing outstanding purchase orders, when they could instead be fulfilling much-higher-margin new POs from the hyperscalers.

Thus every console of the ninth generation (PS5, Xbox Series, Switch) still selling for their launch price (with no help from process shrinks); thus none of these consoles having been able to be produced in excess of demand to the point that supply-and-demand ever drove retail prices below MSRP; and thus the only tenth-gen console so far, the Switch 2, taking ~4 years longer than anticipated to release†.

---

† Nintendo were very likely waiting for Nvidia to run off enough of the Tegra T239 with a sufficiently low passed-on fab cost, for Nintendo to both 1. be able to build a backstock and non-run-dry pipeline of Switch 2s, and 2. be able to be positive-margin on charging the same price as the Switch 1 for them. They waited four years, and neither thing ever happened; so they eventually just gave up and priced the Switch 2 higher, and also built out an entirely novel D2C + online-marketplace-partner distribution pipeline so they could ration the tiny initial supply of units they had been able to build with the chips Nvidia had supplied them so far.

Though, that being said, Nintendo actually got a double whammy here. They were also waiting for fast NAND to come down in price, so that they could have physical game cards manufactured for a trivial BOM price while still enabling the "direct GPU disk-streamed assets" pipeline that games of the last generation had begun relying on. Obviously, as today's article points out, that hasn't been happening either! Thus game key cards; thus SD Express cards only beginning to trickle out, with no sizes above 128GiB available at the Switch 2's launch time; and thus those SD Express cards being ridiculously priced for their capacity compared to equivalent transfer-speed + die-size NAND (as seen in e.g. low-profile/flush-mount flash drives.)

* Trust that what we are watching or reading is human-sourced and real

* Time for real creators' unique voices to remain unique

* Hope for society's future

Every tech startup circa 2023 doesn't do anything without smearing AI all over their front page.