AI Newsround – Memory, OpenAI & Oracle

SK Hynix and Samsung make the first cut.

• It appears that Micron has lost out on supplying HBM4 to Nvidia for Vera Rubin, but I am not convinced this means very much, as I suspect it will be brought on board later in the year when shipments scale up.
• There are numerous reports suggesting that Nvidia has selected Samsung and SK Hynix to supply HBM4 for Vera Rubin, and given that they are relatively independent of one another, I am inclined to think that there is some truth to them.
• However, this does not mean that Micron will not be a player in this generation, as, typically, Nvidia will source from two suppliers in the beginning and then expand to others as volumes ramp up.
• This is good news for Samsung, as I was expecting it to be the third supplier brought on later, but once again, it has fared better than I expected.
• I am pretty sure that Micron will be brought on board soon because Nvidia is currently sole-sourcing from Korea, which introduces geopolitical risk that is easy to avoid.
• Furthermore, I do not expect that Micron will suffer a loss of revenues, as everyone is fully booked out for 2026, and so any capacity that was earmarked for Nvidia will easily be sold elsewhere.
• Consequently, this looks like a storm in a teacup, as there no sign of demand weakening.
• The only take-home message is the outperformance against expectations from Samsung.  
• Hence, I remain very comfortable holding Samsung, and if the war-based volatility continues, I might consider buying some more.

Business model wobbles.

• The cancellation of capacity expansion at Abilene, Texas, is not a big deal in the grand scheme of things, but I think it points to the fragility of the AI compute business model, further supporting my view that something needs to change to make it more viable.
• OpenAI and Oracle have decided not to proceed with an 800MW expansion, which would have taken OpenAI’s compute capacity in Abilene, Texas, from 1.2GW to 2GW.
• This does not mean that OpenAI is slowing down its expansion, as it will simply put the capacity elsewhere, but it does raise significant questions.
• There is some thought that reliability played a part in Oracle and OpenAI deciding to go elsewhere, but I think that this is unlikely.
• This is because Crusoe, the site’s operator, has managed to line up Meta as a candidate to take over the space that OpenAI has left, which would not have been this easy if there were reliability issues.
• I suspect that this was more about money, financing and the business model, where RFM Research has concluded (see here) that the business model of compute is barely economic, which makes it highly sensitive to fluctuations in financial assumptions.
• Hence, any small change to the conditions or timing of payments or interest rates could have a profound effect on the financial outcome of the installation.
• For example, RFM Research has concluded that the 5-year IRR for a 1GW data centre is just 1%, but if the capital cost can be reduced by 20% or the revenue increased by a similar amount, then the 5-year IRR increases to 8%.
• I suspect it was a reduction of this nature in the financial estimates that caused the disagreement between Crusoe, OpenAI and Oracle rather than a technical or outright demand issue.
• Hence, I don’t see this as a sign that the bubble is bursting, but it reiterates the fragility of the compute business model, underlying that it needs to change.