AI Daily Brief: SK Hynix's $28 billion Nasdaq bet, while GPT-5.6, Gemini and Grok all stay behind glass

The most capable new models in the industry right now have something odd in common: almost nobody can actually use them. GPT-5.6 is out but gated, Gemini 3.5 Pro is delayed again, and Grok 4.5 is confined to two of Elon Musk's own companies. While all three sit behind glass, the money financing the hardware underneath them is moving in full public view, and this week it moved in the biggest foreign listing Wall Street has ever seen.

SK Hynix goes public in New York, chasing the biggest payday of the AI boom

SK Hynix began marketing a $28 billion Nasdaq listing on July 6, offering 17.79 million shares under the ticker SKHY in what's on track to become the largest IPO by a foreign company in US history, behind only SpaceX's $85.7 billion raise last month and ahead of Saudi Aramco and Alibaba's records. Pricing is set for Thursday with trading due to start Friday, and the company has already drawn indications of interest for up to $7 billion from investors including Baillie Gifford and Coatue. The proceeds are earmarked for new fabrication capacity in South Korea and EUV lithography scanners, the machines that make advanced memory production possible at all. A memory chipmaker choosing to raise this much money, this publicly, right now is a wager that the AI industry's appetite for HBM isn't close to peaking, a bet SK Hynix's own stock has already made: it's up roughly 260% year to date.

Anthropic keeps courting the memory makers that just took a stake in it

Samsung and SK Hynix didn't just buy into Anthropic's $65 billion Series H in May, alongside Micron, Blackstone and Temasek; Anthropic is now in early talks with Samsung to manufacture a custom AI chip on its 2-nanometer process, with separate conversations reportedly underway with Microsoft and UK startup Fractile. Nothing is designed yet, chip design and testing haven't started, and Anthropic may not proceed at all. But the direction reads clearly against this week's SK Hynix listing: the AI lab that just took investment from two memory giants is simultaneously trying to become less dependent on the broader chip supply chain they're part of, while those same companies raise fresh capital to expand the very capacity Anthropic would be building around. Everyone in this trade is hedging against everyone else.

Three flagship models, three different kinds of stuck

OpenAI's GPT-5.6 family, Sol, Terra and Luna, formally launched on June 26 but remains restricted to roughly 20 government-vetted organizations under a voluntary preview arrangement with Washington, with no confirmed date for wider release beyond OpenAI's "coming weeks." Google DeepMind, meanwhile, pushed Gemini 3.5 Pro's launch back to July 17, abandoning the existing Gemini 2.5 Pro architecture for a full rebuild rather than shipping something that wasn't yet at flagship quality on coding and long-horizon reasoning. And xAI's Grok 4.5, built on a fresh 1.5-trillion-parameter foundation model roughly three times the size of the public Grok, is running in private beta at SpaceX and Tesla only, with no public API and no independent benchmark to verify Musk's claim that it rivals Opus. The one model actually shipping to the people who already pay for it is, unglamorously, a pricing change: Anthropic's Claude Fable 5 moved off subscription bundles and onto pay-as-you-go pricing on July 7, the same day OpenAI's internal release window for GPT-5.6 was reportedly targeting. Capability is advancing fast across every lab; access to it is advancing much slower, and on three different schedules set by three different sets of gatekeepers.

What it means for India

India spent the week building the physical layer this entire chip story depends on. Prime Minister Modi inaugurated the CG Semi OSAT facility in Sanand, Gujarat on July 5, a packaging and testing plant that's part of the broader India Semiconductor Mission, telling the crowd that India's youth would drive the next wave of AI and robotics on chips actually made in the country. It's a modest facility next to what Samsung, SK Hynix and TSMC operate, and it isn't trying to compete with their leading-edge nodes. What it does is give India a foothold in the assembly and testing stage of a supply chain that a foreign listing like SK Hynix's this week just showed has years of capital still flowing into it. On the compute side, India has onboarded more than 38,000 high-end GPUs at roughly ₹65 an hour, close to a third of the global average cost, through government-backed access programs, an unglamorous but real advantage for startups that can't afford Silicon Valley cloud pricing while they wait for gated frontier models to open up.

Markets and AI money

Indian equities closed July 7 on a mixed note even as the memory trade kept running globally: the Sensex fell 207 points, or 0.26%, to 78,078, while the Nifty 50 dipped 31.65 points, or 0.13%, to 24,398.70, with IT names doing the heavy lifting on the upside, HCL Tech up 3.16%, Tech Mahindra up 2.88% and Infosys up 2.81%, while Adani Ports, Tata Motors and L&T dragged the index down. That IT strength is worth noting given how battered the sector has been on AI-disruption fears in recent weeks; a bounce led by the same companies that were being sold off a few days ago suggests the market isn't fully settled on how to price this trade yet.

The pattern across all of it: capital is moving faster than capability is shipping. SK Hynix can price a $28 billion offering in a week, but OpenAI still can't say when GPT-5.6 reaches the public, and Google needs another nine days just to finish rebuilding Gemini. The market isn't waiting for any of them to be ready.