Rihard Jarc-Assistant

We are committed to pushing the model frontier across cost efficiency, capability, and speed. Starting today, we are reducing prices for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% , and offering a faster option for GPT-5.6 Sol in the API. Luna and Terra’s lower prices are reflected in how usage is counted in Codex and ChatGPT Work, so your usage goes further.

“Nvidia’s CMX is equipped with 576 SSDs, providing a total storage capacity of 9,600 TB. Industry estimates suggest that NAND demand for CMX will surge from 35 million TB this year to more than 100 million TB next year.” Nvidia’s Rubin CMX will quite literally soak up NAND supply like a sponge absorbs water. To put 100 million TB into perspective, it is roughly equivalent to adding another Apple-sized source of demand to the NAND market.

Hey guys, I like this post. I have one question though. I spoke with some people in the industry, and historically in semiconductors and memory there have been a lot of LTAs. One of the biggest double-edged swords here is that, to get companies to commit to production, demand tends to get slightly exaggerated upfront. Even though these contracts are binding and customers are obligated to take the full volume, we’ve seen multiple cases where projects get canceled or scaled down. The customer still takes delivery, which then leads to inventory build-up on their side. That inventory doesn’t just sit there. It often gets pushed into the market at lower prices, effectively competing with the original supplier’s product. So the dynamic I’m pointing out is: Customer signs an LTA → demand softens or a project gets canceled → inventory builds → excess supply leaks into the market at lower prices. The real risk for suppliers here is that this can distort pricing for months, sometimes even years, depending on the size of the overhang. At the same time, it can lead to underutilized production capacity, which is very costly given the fixed cost structure. Curious how you guys think about this risk and how you’ve addressed it in your article.

I was going to avoid discussing LTAs, but I suppose I have to address them after all. Ultimately, I believe both memory companies and investors need to determine whether LTAs are genuinely capable of capping earnings upside while protecting the downside. Samsung Electronics and SK hynix need to provide the same level of detail about their LTAs as Micron has. Otherwise, they will inevitably trade at a discount to Micron, which has been much more explicit about its LTA arrangements. This is especially true now, when the market is pulling forward concerns over the supply expected to arrive in 2028. I hope SK hynix will provide greater clarity on this issue during its earnings call on the 29th. Now let us hear what the LTA bears have to say. To those who chastise us for having learned nothing from analog semiconductors, I would offer a different answer. The reason LTAs in analog semiconductors unraveled so easily was that the industry’s upcycle proved far too short. That upcycle was driven by COVID—an event no one had anticipated and for which no one had prepared. Precisely because it was so unexpected, everyone became overly optimistic about what was ultimately a fortunate but unsustainable upcycle. Consequently, the parties entering into LTAs failed to design the agreements with sufficient rigor—particularly the risks and penalties customers would bear if they terminated them. When the downcycle arrived, those risks were small enough that customers were willing to accept them and break their contracts. This time, however, I would argue that memory could be different. There is a broad consensus that the supply shortage will persist through 2027. I believe this prolonged period of tight supply will encourage memory manufacturers to temper their ambitions and adopt a longer-term perspective. This should be particularly true when demand is driven by AI, which is both more durable and more predictable. Under these conditions, I expect memory manufacturers to take a more calculated approach to structuring LTAs. Micron is already demonstrating this through agreements that deliberately cap its upside. To be completely honest, I do not have a definitive answer either. I cannot say with certainty whether these LTAs are genuinely non-cancellable or whether memory stocks can ultimately be rerated to more than 10 times earnings. What I have noticed recently, however, is the emergence of factors suggesting that these LTAs may be more robust than many assume. For example, rather than terminating their memory LTAs when the commitments become burdensome, neocloud companies appear to be exploring financial hedging options to manage the associated risks.

In the basics, this is of course a time-will-tell argument... it's very hard to look five years into the future. But I have my personal arguments why I think a much higher gross margin than the level they had before is totally possible. Because in my opinion people are pattern-matching HBM to old commodity DRAM, and that's a huge miss. HBM isn't just sold off a spot price anymore, it's really co-designed. HBM4 carries a lot of customer-specific base dies, and it's being designed into the accelerator roadmap for years ahead. So that brings product specialization, which almost equals higher margins if you take it really simple. The more expertise you need, the more expensive it costs. The more different variants and packages you want, the more that expertise is worth. That's just plain simple. And then on the other side you have the EPIC Center, where SK Hynix and Micron signed on as founding partners in Applied Materials' facility to co-develop cutting-edge materials, process and packaging for next-gen memory, with a lot of engineers working side by side. And in my opinion, you don't co-develop your process tech in a joint lab like that for something that's becoming a commodity. That's the integration depth. Every generation the memory maker gets deeper into the customer's design process and the equipment maker's process, into hybrid bonding, taller stacks, the thermal challenges. Everything gets harder to build, not easier to commoditize in my opinion. These 3 thesis foundations I use and the timeline beyond 2030 is just far away man. Would be guess work. Gun to my head i would say it will work trough the half of next decade also like this and beyond because it is literally a postion change for the big 3 not only a supply story.

"Even if the memory shortage persists throughout 2027, research firms, sell-side analysts, and other prominent industry observers broadly agree that the supply-demand imbalance will begin to ease in 2028." Yes, that is correct. I actually have an interview transcript (I cannot disclose due to sensitivity) where the senior employee of one semiconductor company in Taiwan states: "The supply and demand for DRAM will reach a balance in H2 2028 to stabilise the price." He explains it along the lines of: 2026: The current demand gap is hovering around 4% to 5%. Original 2025 projections of an 8% to 9% shortage proved exaggerated because major manufacturers like Samsung and SK Hynix successfully transitioned from 1b to 1c process technologies. This node upgrade boosted bit output and filled much of the anticipated shortfall. 2027: The supply gap is projected to shrink to under 3% by the second half of the year as Micron and Samsung ramp up entirely new production facilities. 2028: The market will face a minor 2% to 3% deficit in the first half of the year (H1) before finally achieving a complete supply-demand equilibrium in the second half (H2), driven by the massive volume of new capacity deployed between late 2027 and 2028. $MU $EWY $DRAM