What happened
SK Hynix, the South Korean chipmaker that supplies most of the world's high-bandwidth memory — the specialized chips stacked directly onto Nvidia's AI accelerators — began trading on Nasdaq on July 10. The American depositary receipts priced at $149, raised roughly $26.5 billion, and opened around $170, up more than 14% on the first day.
It is the largest ADR offering ever and the biggest US listing by a foreign company in history, surpassing Alibaba's 2014 debut. Proceeds go toward new chip plants in South Korea, built to feed AI infrastructure demand the company cannot currently keep up with.
The most oversubscribed part of the AI economy isn't the models. It's the memory they sit in — and the market just priced years of demand into one morning.
The detail almost everyone will miss
Put this next to what happened twenty-four hours earlier. On July 9, OpenAI launched GPT-5.6 and cut the price of capable AI work — the cheapest tier now runs at a fifth of the flagship's cost. On July 10, the scarcest physical input to all of that work raised record money and still jumped double digits on debut.
Those are not two stories. They are one story read from opposite ends of the supply chain. Model providers are cutting your price per token as strategy — to win volume, to keep you from leaving. Underneath them, the memory that determines how many accelerators ship at all is spoken for, and its maker is raising $26.5 billion to build capacity it won't have online for years.
Your token price is set by strategy. The capacity behind it is set by physics — and physics just told you demand is outrunning supply.
Why this matters if you run a business
If AI is becoming a real line item in your operation, the cheap-inference era you're enjoying is a pricing decision, not a law of nature. It is being subsidized by providers fighting for share, on top of hardware that is getting more expensive and harder to book.
When capacity tightens, it shows up for you as rate limits, waitlists, and quiet price-floor resets — not as a press release.
One caution in the other direction: a first-day pop is sentiment, not a supply forecast. This listing is evidence that sophisticated money expects AI infrastructure demand to stay brutal for years. It is not proof your API bill rises next quarter. Treat it as a weather report, not a storm.
What to do about it
Nothing here requires panic. It requires three habits, started now:
- Treat today's model prices as a window, not a floor. Lock in the unit economics of your AI-dependent workflows while the price war is running, and re-check them quarterly instead of assuming the curve only bends down.
- Keep your switching costs low. A capacity crunch at one provider shouldn't become your outage. If a workflow matters, know today which second model clears your quality bar and how long the swap takes.
- Meter your spend per task, not per month. You cannot manage a cost line you don't measure. Cost-per-invoice-processed or cost-per-report-drafted is the number that tells you early when the weather changes.
The companies that get hurt by an infrastructure squeeze won't be the ones spending the most on AI — they'll be the ones who never measured what it costs and built as if cheap was permanent.