Good discussion about this one recent Vergcast. They did a whole like 2 hour episode just on dram. AI demand is a real thing, but the bigger problem has been consolidation in the DRAM industry when the market was down (20 companies to three), the amount of time it takes to build new wafer capacity, and those firms reluctance to build that new capacity given the recent history of dram bust cycles killing companies. A wafer factory is a huge building, already hard to build, that also needs to be one of the most sterile places known to man. No east feat and a big investment the three surfing companies aren’t going to make just because pc gamers are mad about high prices. Add onto that the geopolitics of where ram is and can be safely built (Taiwan, SC, other SEA countries, etc) with the skilled labor to run them, lead times on the machines that etch the wafers (I think it’s just one company somewhere in Europe iirc) and you’ve got a mess.
Basically demand was low, prices went down, companies went bust. Demand started to pick up but the three (really two) surviving companies were reluctant to invest in capacity for fear of another bust cycle. Now demand is high, capacity is being built, but there is going to be a long lead time on that easing prices. And if demand continues to increase, capacity may not ever catch up enough to ease prices.
I know it’s fashionable to blame everything we can on AI, and it’s true AI is a factor here in driving demand, and it’s far less fun to say “well actually dram supply is a complicated problem spanning economics, supply chain realities, and geopolitics with a lot of nuance”. But the latter is true.
And it would be useful to pay more attention to this. The silicone supply chain is crazy important and crazy fragile. Instead of just reflexively going “AI bad” to this kind of event, I’d be useful to be talking about how we can diversify supply chains and silicone capacity.
Think dram prices are high now? China just encircled Taiwan in a show of force and the US expects them to make a move to take the island by 2027, the whatever anniversary of Taiwan’s independence. Not just the huge impact that would have due to Taiwan, given TSMC manufacturers most of our advanced chips of all kinds, but all the regional instability that would follow. It’s a big complicated world and there are problems beyond ai.
The ending with AI bubble may be similar to what was with GPUs after the crypto mining. The only issue is that greed most likely will force companies to not lower prices too much.
Also, as far as I know, TSMC have kill switches for their production lines, what means that they will not let occupy their machines by force. Which is a good thing for controlling the technology and a bad thing for the rest of us if someone would actually try to invade Taiwan.
Good discussion about this one recent Vergcast. They did a whole like 2 hour episode just on dram. AI demand is a real thing, but the bigger problem has been consolidation in the DRAM industry when the market was down (20 companies to three), the amount of time it takes to build new wafer capacity, and those firms reluctance to build that new capacity given the recent history of dram bust cycles killing companies. A wafer factory is a huge building, already hard to build, that also needs to be one of the most sterile places known to man. No east feat and a big investment the three surfing companies aren’t going to make just because pc gamers are mad about high prices. Add onto that the geopolitics of where ram is and can be safely built (Taiwan, SC, other SEA countries, etc) with the skilled labor to run them, lead times on the machines that etch the wafers (I think it’s just one company somewhere in Europe iirc) and you’ve got a mess.
Basically demand was low, prices went down, companies went bust. Demand started to pick up but the three (really two) surviving companies were reluctant to invest in capacity for fear of another bust cycle. Now demand is high, capacity is being built, but there is going to be a long lead time on that easing prices. And if demand continues to increase, capacity may not ever catch up enough to ease prices.
I know it’s fashionable to blame everything we can on AI, and it’s true AI is a factor here in driving demand, and it’s far less fun to say “well actually dram supply is a complicated problem spanning economics, supply chain realities, and geopolitics with a lot of nuance”. But the latter is true.
And it would be useful to pay more attention to this. The silicone supply chain is crazy important and crazy fragile. Instead of just reflexively going “AI bad” to this kind of event, I’d be useful to be talking about how we can diversify supply chains and silicone capacity.
Think dram prices are high now? China just encircled Taiwan in a show of force and the US expects them to make a move to take the island by 2027, the whatever anniversary of Taiwan’s independence. Not just the huge impact that would have due to Taiwan, given TSMC manufacturers most of our advanced chips of all kinds, but all the regional instability that would follow. It’s a big complicated world and there are problems beyond ai.
The ending with AI bubble may be similar to what was with GPUs after the crypto mining. The only issue is that greed most likely will force companies to not lower prices too much.
Also, as far as I know, TSMC have kill switches for their production lines, what means that they will not let occupy their machines by force. Which is a good thing for controlling the technology and a bad thing for the rest of us if someone would actually try to invade Taiwan.