SK Hynix asks U.S. investors how much they still believe in memory

Good morning. The market walked into Thursday with oil anxiety in one hand and AI capex math in the other.

The macro tape was not subtle: reports of fresh U.S.-Iran tension pushed crude higher and stocks lower, with the S&P 500 and Nasdaq slipping in early trading. But underneath the risk-off headline, the tech trade kept doing what it has done all year: separating AI winners by supply chain layer. GPUs are crowded. Power is scarce. Debt buyers are getting pickier. And memory, somehow, is still the place where the seller may have more leverage than the buyer.

That is why today’s lead is not another model demo. It is a share offering.

SK Hynix is bringing one of the year’s cleaner AI infrastructure questions to Nasdaq: if everyone says the AI trade is crowded, will investors still fight for exposure to the part of the stack with the tightest supply?

The company is expected to start trading U.S. depositary shares on Friday in an offering that could value the U.S.-listed shares at up to $28 billion. Reuters reported that demand was more than seven times the number of shares available at the top of the range, with 16.8 million ADSs projected at $158 to $168 each. Ten ADSs represent one common share, and SK Hynix is not selling new shares or receiving proceeds.

That last detail matters. This is not a company using the AI window to fund a factory spree. It is a liquidity event and a price-discovery test. Investors already know SK Hynix is central to high-bandwidth memory, the specialized memory used around AI accelerators. What they are being asked now is whether the recent selloff changed the cycle thesis or just made the entry point less theatrical.

The stock has been one of the market’s purest AI-memory winners, up more than 680% over the previous 12 months, according to Reuters, even after falling almost a quarter over the prior two weeks. On LSEG figures cited by Reuters, it was still trading at about 5.5 times forward earnings, below Micron’s 6.66 times. In other words: the move is huge, the multiple is not, and that is exactly why the offering has market attention.

Bloomberg framed the bull case around pricing power, not vibes. On the July 8, he said hyperscale customers were trying to lock in pricing because they feared paying more three months from now. His cleanest line: “the best pricing power right now is memory.”

That is the hinge. Nvidia can be debated on custom chips, China restrictions, cloud concentration, and valuation. Memory is simpler and less forgiving: if AI clusters need it, and customers are reserving it ahead of price increases, the supplier has room to breathe.

There are still risks. China-related rules can move quickly, and the same AI exposure that made SK Hynix a market darling also makes it vulnerable when investors sell semis in one basket. But Thursday’s demand signal says the memory trade is not yet exhausted. It has merely become more selective.

OpenAI had two very different pieces of news land in the same tape, and together they say a lot about what the company is becoming.

First, Bank of America extended OpenAI a new $520 million credit line. PYMNTS, citing Bloomberg reporting, said the facility has not been drawn and that BofA had previously passed on OpenAI’s earlier borrowing round. The reversal matters less for the cash itself than for the relationship. A lender that skipped the old deal now wants a place near the company before a possible future IPO.

Second, OpenAI is releasing GPT-5.6 after a staggered rollout shaped by Washington. Axios reported that the company will make the model more broadly available Thursday after the White House gave its blessing, following a limited preview to government, academic, and industry partners. The same report said the Trump administration had delayed a June release under a June 15 order restricting powerful models, but later determined that GPT-5.6 did not cross the threshold requiring continued limits.

That combination is the story. OpenAI is no longer just a product company shipping frontier models into the wild. It is a capital-markets entity with bank relationships and a regulated strategic asset whose release timing can become a government question.

The details are worth separating. On the finance side, the new line sits alongside a much larger credit stack. PYMNTS said Morgan Stanley led a $4 billion credit line in October 2025 that was expanded to $10 billion earlier this year. OpenAI paying a premium versus investment-grade borrowers is unsurprising; the company has enormous strategic value, but it is still burning through infrastructure needs at a pace that makes lenders underwrite growth, not comfort.

On the product side, GPT-5.6 is being positioned as OpenAI’s strongest model yet for software engineering, data analysis, and cybersecurity assessments. That last category is exactly where the policy pressure lives. The government wants U.S. companies to keep leading, but it also wants to avoid handing advanced cyber tools to adversaries by default.

The process may work for a single marquee launch, but it is hard to imagine every frontier model release becoming a bespoke White House calendar event. If the rulebook is not standardized, the biggest companies will learn to lobby the process while smaller labs wait outside the door.

Apple’s Broadcom announcement is not the flashiest AI story of the week, which is exactly why it works.

The company said it will increase spending with Broadcom under a new multiyear commitment expected to exceed $30 billion, leading to more than 15 billion U.S.-made chips. The deal covers custom silicon components and advanced wireless connectivity technologies, with Broadcom producing radio-frequency components, including FBAR filters, at its Fort Collins, Colorado facility.

Apple also said the agreement is its largest American Manufacturing Program commitment to date and will help Broadcom expand and modernize Fort Collins through a $1.5 billion capital expenditure investment. The company tied the deal to its broader pledge to invest $600 billion in the U.S. economy over four years.

That makes the transaction useful in three directions at once.

For Apple, it gives Tim Cook a concrete domestic-manufacturing headline at a moment when tariff politics still matter. For Broadcom, it reduces concern that Apple will in-house every valuable chip function over time. And for Washington, it is an easy photo-op supply-chain win: advanced components, a Colorado facility, and a familiar consumer brand.

Bloomberg Tech’s Mark Gurman added the product context: Broadcom is already central to Apple’s RF filter stack, and the expanded relationship may also touch custom silicon used around Apple’s AI server ambitions. That does not turn Apple into Nvidia overnight. But it does put Broadcom closer to Apple’s next wave of connectivity and private-cloud hardware.

There is a labor footnote. Bloomberg Tech also highlighted a projected U.S. chip-sector gap of 157,000 skilled workers by 2030, citing McKinsey, SEMI, and the National Science Foundation. Only about 3% of U.S. engineering students go into the chip industry. Deals can be announced quickly; fabs, technicians, manufacturing engineers, and yield learning arrive more slowly.

Today’s issue is really about who gets to tax the AI boom.

SK Hynix taxes it through memory scarcity. OpenAI taxes it through model access and eventually capital markets. Broadcom taxes it by being the supplier Apple can point to when domestic manufacturing matters. Amazon’s lenders, Meta’s power planners, and SambaNova’s investors are all trying to answer the same question from different angles: when AI spending keeps rising, where does the margin actually settle?

The market does not have a single answer yet. But it is getting much less patient with vague ones.