AI Demand Fuels Enterprise SSD Growth; 3Q25 NAND Flash Prices Likely to Rise Further

[TheLostSwede]

TrendForce's latest investigations reveal that continued AI investments by major North American CSPs are expected to drive a significant increase in enterprise SSD demand in the third quarter of 2025. The enterprise SSD market is likely to shift toward undersupply with finished product inventory levels remaining low, supporting a potential price increase of up to 10% QoQ.

TrendForce notes that earlier this year, suppliers adopted a conservative production strategy as they gradually brought the NAND Flash market back into supply-demand balance. However, the introduction of new U.S. reciprocal tariff policies in early April disrupted the market's momentum in Q2 and introduced volatility into price trends. Although some PC manufacturers accelerated shipments in Q2, it failed to substantially boost overall demand for NAND Flash products. Meanwhile, persistent weakness in the retail market has prompted suppliers to tighten capacity controls even further.




Nonetheless, demand for storage has recently picked up. This surge is partly driven by the shipment of high-end AI servers such as NVIDIA's GB200, and partly by increased HDD orders since early this year—signaling a broader trend of enterprise infrastructure expansion. Both SSDs and HDDs are expected to benefit from rising enterprise capital expenditures with steady server deployments by CSPs, ushering in a new wave of order growth.

View at TechPowerUp Main Site | Source

 

[TumbleGeorge]

Is it true that training and operating an LLM is not only very energy-intensive, but also degrades hardware faster than any other activity, thus creating a lot of electronic waste?

 

[lambda]

Is it true that training and operating an LLM is not only very energy-intensive, but also degrades hardware faster than any other activity, thus creating a lot of electronic waste?

Don't think so, the hardware for hyperscalers is built differently from say our consumer grade stuff. They are built to run at a specific load for a specific period of time.

The update/upgrade cycles are probably shorter given the AI hype and hardware advancements.

All that said, the advancements in terms of hardware are not trickling down to consumer grade electronics and hardware afaik. Hardware is getting more expensive in general.

 

[bonehead123]

Sooo..... to summarize (without all the techno/marketing-babble): Well, DUH !

Supply & demand.....

Demand & supply......

Supply & demand.........

yada yada yada, there's only so many ways to say it, but that about sums it up..

 

[Rover4444]

How does AI require so much flash? Even when training uncompressed models it should only be a couple of terabytes...

 

[darakian]

How does AI require so much flash? Even when training uncompressed models it should only be a couple of terabytes...

I would guess that a few different things are happening.
1. Data which otherwise would have been deleted is now being kept.
2. Companies are experimenting with different mixes of data to make more applicable (for them) models and doing a poor job of deduping data
3. More individual workers are working on these data sets locally.

Upshot is that whenever the bubble pops we should see some nice deals on consumer ssds at least.

 

[TumbleGeorge]

Upshot is that whenever the bubble pops we should see some nice deals on consumer ssds at least.

Make it!

 

[TheinsanegamerN]

Is it true that training and operating an LLM is not only very energy-intensive, but also degrades hardware faster than any other activity, thus creating a lot of electronic waste?

No. for NAND, wear comes from writing data, more specifically, overwriting data. When you "delete" data on a SSD, it isnt actually deleted because removing the charge produces even more wear on the cell, so the charge remains until it is overwritten. If a SSD has a 100TBW rating, it doesnt matter to the SSD if that 100TB is movies, video games, your dirty bedroom content, games, or AI training data, it writes all the same and will wear out at the same rate.

For the processors heat cycling is what damages them, if they are kept at constant load from working on things it wont matter unless the workload causes overheating, which these LLM companies should weed out int he pre alpha stage.

 

[FreedomEclipse]

Better grab what I need before the 50% price hike.

 

[Kwadratowicz]

Where the heck are cheap sata ssd?
So they will sell more stuff, and that's why they will rise the price?
I hate the world. It looks like manufacturers are all together and are trying to contain consumers in mud hole.

 

[Bwaze]

Upshot is that whenever the bubble pops we should see some nice deals on consumer ssds at least.

I wouldn't count on that. When cryptomadness bubbles popped in 2018 and 2021, GPU prices didn't fall. In fact, after spring 2018 crypto collapse we have gotten RTX 20x0 in autumn with terrible price / performance, and the prices remained high for the whole generation.

Similar story occured in 2021, cards like RTX 3080 launched at $700 and were sold for more than $1500, but after crypto crash it took till the end of the generation for cards to fall to MRSP, and the next generation launched with terrible price hike. And due to server AI demand, we haven't come back since!

I think once the companies take a ride on a gravy train, they're very reluctant to go back to normal business, catering to consumers, they would rather try to find some new magical hen that lays golden eggs. I can imagine those plans go down well with investors that like to hear about possibilities of skyrocketing revenue.

I bet if AI bubble would burst, or at least shifted away from Nvidia, they would rather try to restart cryptomining hype on home GPUs than go back to deal with gamers that at this point hold quite a big grudge.

Same thing with SSD makers. I bet they would try to diversify into many possible server uses, even look at things like Chia blockchain mining before considering home consumers that have very little range on what they're prepared to spend.