Samsung's HBM4 Squeeze, OpenAI Revenue Shock & DRAM Margin Fade

YesOui (0:00)

AI Hardware & Chips, Daily News.

And Jamie Call, thanks for joining me.

Today Samsung's HBM4 Gambit can supply scarcity by foundry wins.

Lisa Su flew to Seoul on October 7th, and that tells you almost everything about where the power sits in the AI hardware supply chain right now. AMD's chief executive was there to discuss production of the Instinct MI455X accelerator, and, more pointedly, whether Samsung could become AMD's foundry partner for 2-manometer logic. The meeting happened because Samsung has something AMD needs badly, HBM4 memory.

SK Hynix has locked most of its HBM4 output to Nvidia. That leaves AMD exposed, and Samsung is using that exposure as a negotiating lever to pull AMD toward its foundry business. It's a classic supply chain pressure play, and it's the central thread running through everything happening in the chip industry this week. The context for that leverage matters. Samsung just posted operating profit of $107.4 trillion won for the third quarter of 2026 Memory profits surged ninefold year over year. By any measure, that's a dominant performance on the memory side. The important distinction is what sits underneath it. Samsung's foundry division reportedly lost around $2 trillion won in the same quarter. So the company is running a high-margin memory business to subsidize a foundry operation that still can't compete at scale with TSMC. The HBM4 gambit is Santum's attempt to use one strength to fix the other. Whether that trade actually works is a different question. Here's the problem AMD has to weigh. Samsung's SF2 2-manameter process is reportedly yielding around 60%.

The industry standard for attracting flagship customers is closer to 70 That gap isn't trivial. For high-volume, high-stakes production like Instinct AI accelerators, a 10-point yield deficit translates directly into cost and supply reliability risk. The C for entry point being discussed is consumer Ryzen parts, not flagship AI chips. That's telling. AMD isn't going to migrate its most critical workloads to a process that hasn't proven itself yet, regardless of how tight the HBM4 situation gets. The signal here is that Samsung's leverage is real but time-limited. If yield improves and HBM4 scarcity persists into 2027, the negotiating window stays open. If yield stalls, AMD has every reason to stay with TSMC, even at a premium. Here's the thing. Underneath the supply tightness story, there's a pricing signal worth watching. TrendForce is forecasting DRAM contract price growth of 10-15% in the face quarter. That sounds reasonable, until you compare it to the 60% surge in the second quarter. The deceleration is sharper than many expected. And it suggests the premium memory margin window is narrowing faster than the supply timeline would imply. Supply tightness is projected to extend to 2028, but pricing power doesn't necessarily follow supply curves when capacity additions are stacking up across the industry. Then there's the demand side shock that hit on October 9th. OpenAI disclosed annualized revenue of $50 billion.

The prior figure circulating was $68 billion. That's not a rounding error. The revision triggered drops of 3-8% across Nvidia, AMD, Broadcom, Intel, CoreWeave, Oracle and Super Micro Computer. In a single session, the key implication isn't the number itself. It's what investors are now pricing in. The possibility that AI infrastructure spending is running ahead of the revenue base that's supposed to justify it. Samsung reporting record memory profits in the same week that AI hardware stocks sold off on demand doubts is not a contradiction. It's a lag. The question is, how long that lag holds before the two curves converge?

Two shorter developments worth flagging. ASML is applying a flat 10% increase to EUV and EUV replacement parts for Samsung and SK Hynix starting January. This kind of across-the-board maintenance hike is unusual. It raises fab operating costs at exactly the point when memory margins are compressing, and TSMC reports earnings shortly. The numbers to watch are just revenue. It's pricing guidance, margin commentary, and any update on fab diversification progress in the US, Japan, and Germany. TSMC still holds over 70% of foundry revenue share, and its 2-manometer capability remains effectively unreplicable near-term. Any signal on pricing power from Taipei sets the tone for the entire sector. The real test over the next quarter is whether Samsung's yield trajectory improves fast enough to make the AMD foundry conversation a genuine decision rather than a diplomatic meeting. Watch the MI455X production timeline. Watch whether AMD signals any diversification away from TSMC in its next earnings call.

And watch DRM pricing in January when the ASML cost hike lands alongside what may be the first clear signal of whether memory margins are stabilizing or sliding.

Supply scarcity can buy leverage. It can't buy yield.

That's the trade-off Samsung is running right now. And it's the one that matters most.