Samsung Q2 26 – Canary?

A question of demand

• Samsung’s Q2 26 profit has increased more than 19x YoY, but given that this was only just ahead of expectations as opposed to Micron, which blew expectations away, the market is wondering whether or not we are at the top.
• Samsung’s preliminary revenues / operating profit of KRW171tn / KRW89.4tn up more than 2x YoY and 19x YoY respectively.
• However, relative to expectations, these numbers beat expectations by only a small margin, and by much less than Micron did when it reported results a couple of weeks ago.
• In a market that has problems looking past the next quarter, this is a problem, and it is once again asking the question of whether or not we are at the top of the market.
• The answer to this question lies not with the memory makers, the accelerator companies or the data centres but with those who are buying the tokens.
• There are two pools of demand, one which is the free service offered by the LLM-based service providers and the other the paid-for services offered by an increasing number of providers.
• The problem is that the cost of generating tokens is increasing extremely quickly both from the perspective of the capex required to produce the tokens and from the amount of compute that goes into producing each token.  
• For example, the cost of 1GW of Vera Rubin is about 50% greater than that of 1GW of Grace Blackwell and while this produces 3x more tokens per GW, the larger, more complex models that “think” for longer consume a lot more compute for each token produced.
• The net result is that the cost of compute is going through the roof as evidenced by the deals that both Google and Anthropic recently signed with xAI.
• At the beginning of 2026, the going rate for 1GW of H100 or Blackwell was around $10bn / GW, but xAI has been able to charge between $37.7bn / GW and $55bn / GW as a result of the scarcity of compute relative to demand.
• These kinds of prices imply that market power is shifting from LLM-based service providers to the providers of compute, which is why xAI and now Meta have decided to park their in-house AI and sell the compute into the market.
• As a result of this, the market price for tokens has to rise and rise significantly from here if the LLM-based service providers who are buying compute are going to remain in business.
• I estimate that Anthropic is already charging $20bn / GW – $25bn / GW, but in order to make a 40% gross margin on the capacity it is buying from xAI, it will need to increase pricing to around $60bn / GW.
• This means a further doubling of the price that it charges to its enterprise customers, which is going to mean an even tighter control by these companies of their token consumption.
• Tokenmaxxing has quickly given way to token rationing, and the big question is how much the market demand will respond to another doubling of prices.
• For as long as demand continues to greatly outstrip supply, then we do not have a problem, and the memory companies will continue to rise and the compute providers will continue to print cash.
• Meta moving to sell its capacity is not a sign of demand weakening but instead an economically rational response to crazy pricing in the market for compute capacity.
• The cure for high prices is always high prices, and so when there is enough capacity to meet demand, then prices for compute will fall, and the market will normalise.
• This will mean slower growth and lower margins, which will in turn cause valuations to normalise.
• There is no sign of this at the moment, and so, for the moment, the AI trade looks set to continue.
• Hence, while I am getting closer to the point where I should think of selling Samsung, that time is not yet, and in the 9-12 month time frame, there is also an opportunity in Meta (see here).
• Hence, I am thinking of buying Meta as opposed to selling Samsung Electronics, but it looks to me like we are closer to the end of the silicon chip run than the beginning.