SK Hynix files $28B US IPO as HBM becomes AI's bottleneck

SK Hynix brought in $10 billion in revenue from High Bandwidth Memory, the specialized silicon that lets GPUs move data fast enough to actually run frontier models, in just the first six months of 2026. Against that backdrop, the company's $28 billion US IPO filing this week is not a routine capital raise. It is a significant disclosure moment for a dominant component in the AI infrastructure stack getting audited by public market disclosure.

HBM is the memory architecture that sits directly on top of GPU processors. It determines how much data can flow between the GPU's compute cores and its cache per nanosecond. Without sufficient HBM, a model cannot be served at scale, you hit the memory wall, and inference latency explodes. NVIDIA's H100 and H200 GPUs ship with HBM stacks built by SK Hynix as the primary supplier, though Samsung also supplies HBM for NVIDIA's H200, B200, and other AI accelerators. AMD's MI300 series uses SK Hynix HBM too. The moment frontier AI shifted from training to inference in 2025–2026, HBM demand became the binding constraint. Amazon, Alphabet, Meta, and Microsoft are collectively spending $650 billion on AI infrastructure in 2026 according to public guidance. All of that money flows through data centers built around GPUs that cannot function without HBM. SK Hynix supplies approximately 50-55 percent of the world's HBM as of July 2026. The company is a critical vendor in the AI buildout, controlling a significant portion of a supply-constrained component.

The $10 billion revenue figure released with the IPO filing, covering only H1 2026, rewrites how the infrastructure market understands supply constraints. This is not HBM as a margin line item in a larger chip business. This is HBM as a business that rivals the entire annual revenue of most semiconductor specialists. For comparison, Broadcom's Semiconductor Solutions segment generated $15 billion in Q2 FY2026. SK Hynix's HBM business alone is running at a $20 billion annual pace, and the company has disclosed capacity expansion plans extending through 2028. That timeline matters because it means every AI model deployment at scale, every LLM serving millions of concurrent users, every multimodal inference endpoint, every enterprise RAG (retrieval augmented generation) system running on private cloud, depends on memory shipments SK Hynix has already committed to deliver or has not yet committed to produce.

The IPO filing requires SK Hynix to disclose utilization rates, capacity expansion capex, and customer concentration for the first time in public filings. Investors will see that the company supplies HBM to four customers: Nvidia (which consumes HBM built into reference designs and sells them to hyperscalers), AMD, Intel, and Chinese state-backed semiconductor suppliers serving domestic AI infrastructure. There is no fifth player. No Asian competitor has achieved parity. No US or European vendor is in the queue. That concentration creates both leverage and liability. SK Hynix can price HBM at whatever the market will bear, and the market right now is willing to bear a lot. But if geopolitical tension escalates around semiconductor exports from South Korea, or if SK Hynix's own capex roadmap slips, the entire global AI inference infrastructure stalls. Public market scrutiny will force the company to quantify that risk explicitly.

What to watch: SK Hynix's Q3 2026 earnings guidance and capacity disclosures. If the company projects flat or declining HBM utilization, it signals that the AI buildout is overheating relative to actual deployment demand, and pricing will compress hard. If utilization remains above 95 percent through 2027, it means memory bandwidth remains the constraint, and SK Hynix has pricing power indefinitely. Second, watch whether AMD or Intel accelerate their own HBM development or secure long-term supply agreements outside SK Hynix's roadmap. If they do, it signals competitive pressure. If they don't, it confirms SK Hynix's dominant position in the market is structural, not temporary. Finally, track whether US policymakers attempt to incentivize HBM production domestically, via IRA grants or allied manufacturing subsidies, before the 2027 election cycle. If they do, it means the Biden and Trump administrations both agree that AI memory bandwidth is now a national security asset. That is the real story inside the IPO filing: HBM is not a component anymore. It is infrastructure.