Chey Tae-won runs SK Group, the conglomerate behind SK hynix — the company that supplies most of the high-bandwidth memory feeding the world’s AI accelerators. At Sunday’s press briefing at the Korea Chamber of Commerce and Industry’s Jeju Forum, he put numbers on 2027 that should stop every AI capex model mid-spreadsheet: customers are asking SK hynix for 60 to 100 percent more AI memory next year than this year. With AI now eating more than half of total semiconductor consumption, he pegged overall demand growth at a minimum of 50 to 60 percent. And supply’s answer, in his words: no company has meaningful new capacity coming online next year.
Then came the part that turns a supply-chain story into a diplomacy story. Foreign governments, Chey said, have begun intervening on behalf of their domestic industries to lock down memory supply — treating chip access as a matter of “economic security.” His prediction: governments will start pressuring other governments soon. When the man whose company held 58 percent of global HBM revenue in the first quarter — per Counterpoint Research, with Micron and Samsung at 21 percent each — says allocation is going political, that’s not analysis. That’s a report from inside the room.
SK is building as fast as concrete allows. The first clean room at its Yongin mega-cluster was pulled forward to February 2027 from May. An extra 21.6 trillion won — about $14.5 billion — was committed in March, and the Cheongju M15X fab is being converted into a dedicated HBM base. But here’s the twist worth sitting with: the same man warning of shortage also called current memory prices “abnormal” and said they need to come down. Sellers don’t talk down their own prices — unless they’ve concluded that prices this high invite exactly the political intervention he’s describing.
When supply can’t move, allocation goes political
The compute scramble was already visible at the top of the stack — labs are renting each other’s datacenters by the $10 billion, and countries are standing up national AI factories to guarantee themselves a seat at the table. Chey’s warning says the real bottleneck sits a layer lower, in the memory stacked next to every GPU.
Our take: The GPU crunch was always solvable with money — TSMC could print more. Memory is different: a fab takes years, and every 2027 wafer is already spoken for. When supply physically can’t respond to price, the market stops allocating and politics starts. That’s what “economic security” means in practice: your datacenter buildout — or your country’s national AI program — now waits on a Korean fab schedule and on who lobbies hardest for a place in line. Read the 60–100 percent number for what it is: hyperscalers revealing their true 2027 plans to their supplier before any earnings call. And Chey calling his own prices “abnormal” is the tell of a man managing political blowback, not maximizing margin.
What to watch
- Prepayment deals. Hyperscalers locking 2027 HBM supply with cash up front — third-quarter capex disclosures will show who paid to jump the queue.
- Yongin’s February date. The clean-room opening has already been pulled forward once. Any slip ripples through every 2027 AI roadmap — and the buildout is already fighting physics and neighbors on other fronts.
- Samsung and Micron’s answer. Twenty-one percent each is a lot of incentive. Watch for capacity announcements engineered to make Chey’s “no new capacity” line obsolete.
- The first government-brokered allocation. A subsidy, a trade-talk line item, an export priority — the first state-arranged memory deal confirms the economic-security era has arrived.