SK Hynix reported operating profit of 8.4 trillion won ($6.1 billion) for Q2 2026, a 557% year-over-year surge driven by HBM3E demand for NVIDIA AI accelerators. But shares slid 4% after the company guided 37 trillion won ($27 billion) in 2026 capital expenditure for HBM4 lines, a new Indiana fab, and advanced packaging. The tension between record earnings and ballooning capex reflects a market starting to question whether the AI memory cycle can sustain current investment levels. SSD prices have risen 220% in 12 months as HBM production consumes fab capacity otherwise used for NAND and conventional DRAM.
The Numbers #
SK Hynix reported a 557% year-over-year profit increase on July 29, powered by the insatiable demand for high-bandwidth memory used in AI training and inference. The South Korean memory giant posted operating profit of 8.4 trillion won ($6.1 billion) for the June quarter, crushing analyst estimates.
But the headline number hides a tension that the market noticed immediately. Shares slid 4% on the earnings release despite the blowout results. The reason: capital expenditure guidance ballooned to 37 trillion won ($27 billion) for 2026, with most of it going to HBM4 production lines and new fab construction.
The market is asking a question that no one has a clear answer to: how much memory does the AI industry actually need, and how long will this cycle last?
The HBM4 Arms Race #
SK Hynix is the dominant supplier of HBM3E memory to NVIDIA, and it's racing to be first to HBM4 qualification. Samsung is closing the gap with its own HBM3E and HBM4 roadmaps, and Micron has re-entered the HBM race after years on the sidelines.
The $27 billion in expansion costs covers:
- <<<BOLD>>>New HBM4 production lines<<<BOLDEND>>> at the M15X facility in Cheongju, South Korea
- <<<BOLD>>>A new fab in Indiana<<<BOLDEND>>> announced earlier this year, targeting 2028 production
- <<<BOLD>>>Advanced packaging capacity<<<BOLDEND>>> needed to stack 16 and eventually 24 DRAM dies per HBM module
- <<<BOLD>>>Research and development<<<BOLDEND>>> for HBM4e and beyond, including hybrid bonding
HBM4, expected to enter mass production in the first half of 2027, will use 2048-bit interfaces (double HBM3E) and deliver over 2 TB/s of bandwidth per stack. Each module requires roughly twice the DRAM die area of HBM3E, meaning capacity expansion is not just expensive — it's exponentially so.
Why It Matters #
The AI memory shortage has cascading effects across the entire technology industry. SSD prices have risen 220% over the past 12 months, according to Tom's Hardware, because HBM production consumes fab capacity that would otherwise go to NAND and conventional DRAM.
- Consumer DRAM prices are up 80% year-over-year
- Enterprise SSD pricing has doubled
- Graphics card memory is becoming a cost driver
- The DIY PC market is being priced out of high-capacity builds
The AI selloff that began in late July has hit chip stocks particularly hard, with SK Hynix, Samsung, and TSMC all declining 10-15% from their peaks. Bloomberg reported on July 29 that "AI Investment Boom Faces Reality Check from Markets and Regulators," suggesting the market is beginning to question whether the infrastructure buildout can sustain current valuations.
The Supply-Demand Gap #
Despite the capex increase, industry analysts project that HBM supply will remain tight through 2027. NVIDIA alone is expected to consume 70% of HBM4 output in 2027, leaving little for everyone else.
The wild card is whether AI training demand continues at its current trajectory or moderates as inference efficiency improves. If model scaling laws hold, demand for HBM will keep rising. If the industry pivots to smaller, more efficient models — as the success of models like DeepSeek and Qwen have shown is possible — the memory demand curve might flatten.
The Takeaway #
SK Hynix's 557% profit jump is the sound of a company that caught the right wave at the right time. But $27 billion in expansion costs against an uncertain demand horizon is a bet that the AI buildout has years of runway left.
The market is starting to hedge. SK Hynix's post-earnings stock drop says more about what investors fear than what the company actually delivered.
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained #
Inference Running a trained model to make predictions on new data.
NVIDIA The dominant provider of AI hardware.
Scaling Laws Mathematical relationships showing how AI model performance improves predictably with more data, compute, and parameters.
Training The process of teaching an AI model by exposing it to data and adjusting its parameters to minimize errors.