The narrative in AI infrastructure over the last two years has been dominated by the enormous and growing demand for compute capacity and its economic consequences, such as the buildout of data centers and the consequent shortages of key resources such as land, water, power, and copper.
But of all these bottlenecks, memory is by far the most significant. The demand for memory is now outpacing the demand for other drivers of compute capacity. The implications of this will ripple through not just the economics of data centers, but the cost of every single consumer and enterprise hardware device.
In this piece, we unpack the market action around memory prices, its ripple effects across the consumer and industrial electronics market, and the supply and demand curve that is emerging around AI. Critically, we explain why the amount of memory being purchased by AI companies like OpenAI seems to be more than what they need, and how the threat of on-device inference might actually be incentivizing an engineered memory shortage.
You must log in or register to comment.
For those who don’t want to read several pages of unnecessary text telling you what you probably already know:
The math, while pretty involved, may tell a straightforward story (if you’re interested in the details of our analysis, see the Appendix). OpenAI has contracted 900K memory wafers per month from Samsung and SK Hynix. Partner commentary seems to indicate that’s a monthly number, so that represents 10.8 million wafers over 12 months. In terms of demand, a fully built-out 10GW Stargate cluster would require ~3 million GB200 Bianca Boards. Each board requires ~50% of a memory wafer in total; split between the HBM3e stacks embedded into its two B200 GPU (~30%) and its 480 GB of LPDDR5X system memory (~20%). That puts total wafer demand for the entire cluster at ~3 million wafers.
Therefore, according to our best estimates, OpenAI likely needs less than 30% of the 10.8 million wafers it’s planning to buy
So this is just putting some numbers to what a lot of people already guessed. The AI companies are not just buying a ton of RAM to build out their data centers. They aren’t buying enough other components to even use that RAM. They’re buying it so that no one else can.
It’s OpenAI in particular trying to screw everyone else. The wafers they contracted from Samsung and SK Hynix are something like 40% of those companies’ production. There isn’t enough production volume for the other AI companies to over order like that.
So this is the strategy of putting 4 houses on your properties in Monopoly and never upgrading them to hotels because that way there are no houses for your opponents to buy
Hopefully this accelerates their crash.
Not if they sell it on the surge.
They may have millions extra, and that just means they’ve now become a shitty version of best buy as they schlepp it at surge pricing to make back the bank.
Oh look, I called it.
deleted by creator
That’s bullshit, this is a good research article. The part you are talking about was like 3 paragraphs to explain the difference between DRAM and NAND because it is relevant to their analysis. They have dozens of citations and explain their calculations in depth, which is that openai bought more than three times the memory they need, and reasoned well about why that might be. Honestly don’t know what you’re talking about.
You are correct. I started skimming before I reached the 2200 word mark on a 4158-word article (not counting the appendix), which does explain just that.
Oligarchs trying to buy up all the digital real estate so they can be the digital landlords of computerland
I guess that’s the point when I’ll just go offline for good and learn a solid, old-fashioned trade.
haven’t read the article just yet, but these companies are known for price fixing and colluding for decades now
deleted by creator
Why sell you a computer, when they can rent you one for $$ per query you do
