Micron's 87% Margins, SK Hynix Debut & AMD's Packaging Crunch

**Jamie Cole** (0:00)
AI Hardware & Chips Daily News I'm Jamie Cole, thanks for joining me. Today, memory margins hit software scale levels. How long does it last?
Micron's data center business just posted 87% gross margins. That's not a software company, that's a memory chip maker. And that single number is the clearest signal yet that the AI hardware cycle has done something memory has never quite managed before, turned a historically cyclical, commoditized business into something that looks, at least for now, like a structural bottleneck with real pricing power. The details behind that margin are worth sitting with. Micron's Compute and Data Center unit pulled in $25 billion in quarterly revenue.
Customers have already handed over $22 billion in locked deposits. That's not spot market behavior, that's hyperscalers prepaying to secure supply they can't afford to miss. The important distinction is that this isn't a demand spike. It's a capacity constraint locking in for years.
SK Hynix made that case directly to public markets this week, debuting on Masdak at roughly a $1 trillion valuation. The stock has run seven times in the past year. The company leads in high bandwidth memory, the stacked DRAM architecture that sits directly on AI accelerators and solves the data movement bottleneck that otherwise dominates GPU latency and power consumption. A $1 trillion debut says the market believes HBM demand isn't peaking. Whether that confidence holds is the real question. The risk is visible in the valuation itself. SK Hynix is trading at around 70 times forward earnings. Memory has a well-documented boonbust history. The bull case requires AI capital expenditure to stay elevated through at least 2027, which is when new fabrication capacity begins to come online. Any deceleration in hyperscalar spending before then compresses those margins fast. The margin story connects directly to the packaging story, and that's where AMD finds itself in a genuinely difficult position. Advanced packaging, specifically TSMC's co-double OS process, has quietly become the real bottleneck in AI chip production. It's not foundry node availability anymore, it's the backend. AMD holds roughly 11% of total co-double OS capacity, Nvidia holds around 60%.
The gap matters, because AMD now has to split its constrained allocation between two products simultaneously, Venice EPYC server CPUs and Mi400 AI GPUs. These are different markets with different demand curves, and AMD can't dynamically reallocate between them. Capacity is locked years in advance. If the demand shifts between CPU and GPU faster than AMD's procurement locked in, the result is either stranded supply or lost revenue. That's the packaging paradox.
AMD is competitive on silicon. It's constrained on packaging. Here's the thing. On the geopolitical side, the ASML situation escalated from rumour to active US government inquiry. Commerce Secretary Lutnick pressed ASML executives directly over alleged covert EUV machine exports to China. ASML pushed back with compliance documentation, asserting no machines reached China. What remains unresolved is whether the concern involves complete machines, components or technical knowledge transfers. That distinction matters for how the Wassenaer arrangement gets enforced, and how Dutch-US export control alignment holds together going forward. China isn't waiting for that resolution. The response is now multi-track and funded. Lin Man's team at the Shanghai Optics Institute demonstrated a solid state laser EEV light source that sidesteps ASML's carbon dioxide laser patents entirely. It's a real technical result. Mass production is still years away, likely beyond 2028
Separately, three Chinese AI chip startups, Suan Miao, TranStreams and Dongfang Suanshin, raised capital specifically to pursue three-dimensional GPU stacking as a way around US export controls.
Thermal management and yield at scale remain unproven.
The timeline is real but uncertain. The signal to track is where the memory contract structures hold. If 22 billion in customer deposits is the new normal, suppliers have fundamentally repriced their leverage in the AI supply chain. If AI CapEx slows before new HBM capacity arrives, those locked contracts become the floor, not the ceiling. Watch AMD's next earnings for any commentary on Koro's allocation pressure.
And watch where the commerce moves from questioning ASML to actual enforcement action.
The memory margin story is real. The question this briefing keeps returning to is the same one the market is pricing, how structural is structural? Thanks for listening. This podcast was built using AI technology. A YesOui production.