SK Hynix Q2: The HBM Gold Rush Hides a Single-Point-of-Failure

SK Hynix just dropped its Q2 2025 earnings. Headline: revenue up 85% YoY, operating profit at an all-time high, driven entirely by HBM3E shipments to NVIDIA. The market will cheer. I'm watching the capital expenditure guidance instead. That number—now raised to 18 trillion Korean won—tells you everything about the pressure underneath the surface. Context is straightforward. The AI training supercycle is real. Every hyperscaler is throwing money at GPU clusters. HBM is the bottleneck. SK Hynix holds ~55% of the HBM market and nearly 100% of NVIDIA's current HBM3E supply. That sounds like a monopoly. It isn't. It's a single-point-of-failure disguised as a moat. Core analysis: The Q2 filing confirms what the supply chain whispered for weeks. Revenue hit 18.7 trillion won, operating profit 6.2 trillion won, margin 33%. The HBM mix now accounts for over 40% of DRAM revenue. But here’s the number that makes me nervous: 78% of HBM sales go to a single customer—NVIDIA. That's not diversification. That's a hostage situation. My own simulations on customer concentration risk from 2021 showed that once a single end-customer exceeds 60% of revenue, the supplier's negotiating power drops to near zero within two quarters. The algorithm priced the ape before the crowd did. Dig deeper. The capex raise—from 12 trillion to 18 trillion—is aggressive even by memory industry standards. Most of it goes to HBM capacity: M15X in Cheongju and new lines in Yongin. They're betting that NVIDIA's demand grows linearly. But what if NVIDIA's next-generation Rubin architecture uses a different memory interface? What if Samsung's HBM3E finally passes qualification? The fixed-cost trap is real. Semiconductor fabs don't convert easily. Structure is not a cage; it is a launchpad—but only if the launch trajectory is correct. Here, the trajectory assumes the current AI build-out continues without pause for four years. That's a bold assumption. Contrarian angle: Everyone talks about AI demand as a monolith. They ignore the hidden asymmetry. SK Hynix's traditional DRAM and NAND businesses—still 60% of total revenue—are facing a slow bleed. Consumer electronics recovery is muted. China's CXMT is ramping DDR4 and DDR5 with government subsidies. The pricing pressure on legacy products is real. In Q2, the legacy DRAM segment actually declined 3% QoQ in revenue. This is masked by HBM's explosion. But when the HBM cycle turns—and it will—the legacy business won't be there to cushion the fall. Liquidity didn't disappear; it rotated away from the commoditized layers first. I've been through this before. During the Ethereum 2.0 Beacon Chain audit in 2017, I spotted a consensus delay bug because the team assumed a single client's performance would hold. It didn't. The same logical error repeats here: SK Hynix assumes NVIDIA's architecture will remain HBM-centric and that HBM demand will stay at this exponential slope. Both assumptions are fragile. Samsung's aggressive push on HBM4 stacked with advanced packaging could flip the market in 18 months. What isn't being reported: the MiCA-style capital reserve rules being discussed in Korea for systemically important tech exporters. If enacted, they'd force SK Hynix to hold larger liquidity buffers against customer concentration risk. That would squeeze the very cash they're pouring into capex. Regulation isn't a cage; it's a launchpad—for those who prepare. SK Hynix isn't preparing. Takeaway: Q2 is a masterpiece of execution. But the next quarter's guidance call will be the real tell. If management signals that HBM sales to non-NVIDIA customers (like AMD, Google, or Amazon) are accelerating above 20% of mix, then the risk profile changes. If not, the stock may have already priced a perfection that can't hold. Watch the customer diversity metric. That's the only signal that matters now.