Onpode
Cover art for Nvidia just locked a $500 billion-plus partnership with SK Group across AI factories and memory

Nvidia just locked a $500 billion-plus partnership with SK Group across AI factories and memory

July 25, 2026 · 10 min

Cleo Rios & Mia Drake

Nvidia and SK Group announced a $500 billion-plus AI partnership covering HBM4 memory co-development and a 2-gigawatt SK Telecom AI Factory, but the figure rests on letters of intent — not signed contracts — and a conflated $950 billion headline bundles in Samsung's separate $200 billion Broadcom deal.

On July 24, 2026, at an AI Summit in San Francisco, Nvidia and South Korea's SK Group announced a comprehensive strategic partnership valued at over $500 billion, formalized through letters of intent. The deal spans two major pillars: AI factory construction and next-generation AI memory supply.

0:009:31
Get the next episode on Nvidia

Follow it free — new episodes land in your feed.

Or make your own — any topic, in minutes

More Onpode episodes on Nvidia

About this episode

When Nvidia and SK Group announced a $500 billion-plus AI partnership last week, the headline number landed like a thunderclap. This episode tries to figure out what the deal actually says — and what it quietly doesn't. The $500 billion turns out to be a projected cumulative figure built on letters of intent, not signed contracts. The $950 billion figure that South Korean officials cited bundles in a completely separate Samsung–Broadcom deal. The gap between the headline and actual bound capital is wide. But here's the thing: the soft number may not matter, because one piece of the deal is already structurally real. SK Hynix — the world's largest HBM supplier — is co-developing next-generation HBM4 with Nvidia. Not supplying it. Co-developing it. That means Nvidia has input on the spec, and every other AI company trying to source that memory in 2027 is negotiating with a captive vendor. The episode also sits with the 2027 timeline: a 2-gigawatt AI Factory that requires two not-yet-deployed technologies — Vera Rubin chips and HBM4 — to arrive on schedule. And South Korea's president was physically in the room when the LOI was signed, which means slippage isn't a corporate embarrassment, it's a sovereignty problem. The number may be aspirational. The lock-in is not.

Frequently asked

What is the Nvidia SK Group $500 billion AI deal?

The Nvidia–SK Group $500 billion AI initiative covers SK Telecom building a 2-gigawatt AI Factory on Nvidia's Vera Rubin chips and DSX platform, plus SK Hynix co-developing HBM4 memory with Nvidia. The figure is a projected cumulative total based on letters of intent, not signed contracts, with no specified timeline.

Is the $950 billion South Korea AI deal the same as the Nvidia SK deal?

No. The $950 billion figure bundles two separate deals: SK Hynix's $750 billion in U.S. company supply commitments and Samsung Electronics' unrelated $200 billion deal with Broadcom. South Korean presidential adviser Kim Yong-beom combined them into one headline number; the Nvidia-specific piece is smaller and rests on letters of intent.

What does SK Hynix co-developing HBM4 with Nvidia mean for other AI chip companies?

SK Hynix co-developing HBM4 means Nvidia has direct input on the memory specification, and SK Hynix's entire next-generation roadmap is prioritized around Nvidia's needs. SK Hynix is the world's largest HBM supplier, so competing AI accelerator makers sourcing HBM4 on the spot market in 2027 face a captive vendor effectively controlled by a rival.

Why was South Korea's president at the Nvidia SK Group signing?

South Korean President Lee Jae-myung attended the San Francisco signing in person — not by video or statement — because the partnership represents national economic strategy, not just a corporate vendor agreement. His presence transformed the LOI into a sovereignty commitment: South Korea has staked national prestige on a 2027 deadline tied to technologies not yet deployed at scale.

Can SK Telecom build a 2-gigawatt AI factory by 2027?

The target is less than 18 months from the LOI signing, and both Nvidia's Vera Rubin chips and SK Hynix's HBM4 memory must arrive on schedule simultaneously. A 2-gigawatt facility requires grid allocation, land, permitting, and cooling that typically takes years. The 2027 date appears driven by political optics, not engineering confirmation.

Grounded in 8 sources
Nvidia, SK Group unveil $500 billion-plus AI data centers initiative, memory partnership - CNA · channelnewsasia.com
Nvidia locks down memory supply from SK Hynix as part of $500 billion AI deal - CNBC · cnbc.com
South Korea's SK Hynix, Samsung Elec sign $950 bln partnership with US big tech - Reuters · reuters.com
South Korea President Lee looking to open new era of AI in with global tech companies - Reuters · reuters.com
Nvidia invests $1 billion in Naver and strikes $500 billion deal with SK Group · thenextweb.com
SK Signs $500 Billion AI Infrastructure LOI With Nvidia - Seoul Economic Daily · en.sedaily.com
Nvidia and SK Group enter $500 billion AI partnership — plan to supercharge AI infrastructure with next-gen memory and massive AI factories | Tom's Hardware · tomshardware.com
SK Group and NVIDIA Expand Strategic Partnership Across AI · globenewswire.com
Read transcript

Cleo Rios: Hey — I have a confession, I genuinely thought this was going to be a boring chip story when I first saw the headline Friday morning.

Mia Drake: Oh no, what changed?

Cleo Rios: The detail that South Korean President Lee Jae-myung was physically in the room — not on a call, not sending a congratulatory statement — in the room in San Francisco when Jensen Huang and SK Group Chairman Chey Tae-won signed the MOU. That's when I went, wait, this is not what they said it is.

Mia Drake: Because a country's president doesn't travel to countersign a vendor agreement.

Cleo Rios: No! And that's the whole thing we're trying to figure out today — the $500 billion Nvidia-SK Group announcement, Jensen Huang calling it 'over $500 billion of business together,' is either the biggest AI infrastructure deal in history or it's geopolitical maneuvering using a supply chain as the cover story. And I genuinely don't know which.

Mia Drake: I'd say it's both, actually — SK Telecom builds a 2-gigawatt AI Factory running on Nvidia's Vera Rubin chips and the DSX platform, SK Hynix locks in as Nvidia's HBM4 supplier, and South Korea gets to say its national conglomerate is the backbone of American AI compute. Those things are all true at the same time.

Cleo Rios: Okay but — SK Hynix is the world's largest HBM supplier. If they're Nvidia's memory department now, every other AI company is buying from a captive vendor. That's not infrastructure, that's a chokehold with a press release.

Mia Drake: And the $500 billion number is doing enormous work without anyone specifying the timeline. Which — that's where I want to start pulling.

Cleo Rios: Okay but — pulling on what? Like, what does pulling even mean when the number itself is a projection over years nobody's specified?

Mia Drake: That's exactly it. Think of it like a friend texting you before a road trip saying 'I'll cover $500 on this.' No route booked, no gas station, just an enthusiastic text. That's a letter of intent. The Nvidia–SK deal was formalized through LOIs — not signed contracts. Jensen Huang says 'over $500 billion of business together,' but business over when? Nobody says.

Cleo Rios: Wait — so the $500 billion is not a contract number.

Mia Drake: It's a projected cumulative figure over an unspecified number of years, based on letters of intent. And then — okay, genuinely wild — South Korean presidential adviser Kim Yong-beom goes out and cites a $950 billion number. Which sounds even bigger! Except that $950 billion is actually just SK Hynix's $750 billion in U.S. company supply deals plus Samsung Electronics' separate $200 billion deal with Broadcom, all bundled into one headline.

Cleo Rios: Stop — the Samsung-Broadcom deal is in the $950 billion? That's not even the same partnership.

Mia Drake: It's a totally separate deal, just conflated into one press-friendly number by Kim Yong-beom. So the Nvidia-specific piece is actually smaller and murkier than the headline suggests — and even that piece rests on LOIs.

Cleo Rios: And nobody's — wait, what about Brookfield? Because I saw the Naver piece, the $1 billion Nvidia investment to expand Naver's data center from 55 megawatts to 200 megawatts, and Brookfield was attached to some number there too.

Mia Drake: Brookfield's commitment to fund up to $9 billion for Naver's expansion is a nonbinding term sheet. Not finalized financing. So you've got an LOI at the top, a conflated $950 billion figure in the middle, and a nonbinding term sheet on the Naver side — the gap between the headline and actual bound capital is the whole story.

Cleo Rios: It's aspirational math dressed up as a deal announcement. I mean — that's not nothing, intentions that big have gravity, they pull real money eventually. But it's not what Friday's headline said it was.

Mia Drake: But here's where the soft number stops mattering — because even if the $500 billion evaporates, the SK Hynix piece is already real. They're co-developing HBM4 with Nvidia. Not supplying it. Co-developing it. SK Hynix's next-gen roadmap is literally being written around Nvidia's needs.

Cleo Rios: Wait — co-development meaning Nvidia has input on the actual spec?

Mia Drake: That's what it means. And SK Hynix is the world's largest HBM supplier. So picture this — a semiconductor engineer at, say, a company building custom AI accelerators, not Nvidia, sitting in a procurement meeting in mid-2027, trying to source HBM4 on the spot market. And the world's largest HBM supplier has its entire next-gen roadmap prioritized for one customer.

Cleo Rios: That engineer is just — gone. They're done. They're calling Samsung and getting a much worse conversation.

Mia Drake: And this deal didn't happen because things were fine — it happened because there was already a global HBM shortage. Nvidia didn't partner with SK Hynix out of generosity. They were solving a bottleneck. Only now that bottleneck still exists, it's just... redirected.

Cleo Rios: Nvidia solved its own chokehold by becoming the chokehold for everyone else. That's — no but that's actually the whole story, right there.

Mia Drake: Does supply grow at least? Like, I mean — okay, actually, does the shortage shrink even if it's concentrated?

Cleo Rios: Concentration isn't growth. You haven't built more water, you've just moved the reservoir inside Nvidia's fence. And the part that makes this genuinely terrifying comes later — South Korea has staked national prestige on a 2027 deadline that requires two technologies that haven't deployed yet to show up on schedule.

Mia Drake: Oh, that's the part I want to get into.

Cleo Rios: Two technologies not yet deployed at scale — Vera Rubin and HBM4 — and the first SK Telecom AI Factory has to be online in 2027. That's less than eighteen months from when they signed the LOI.

Mia Drake: And that's not a corporate target someone can quietly revise in a Q3 earnings call, because Lee Jae-myung stood at that signing. This is national economic strategy now. The political cost of slippage isn't embarrassment for SK Telecom — it's credibility for South Korea.

Cleo Rios: Wait — a 2-gigawatt facility. Not two hundred megawatts. Two gigawatts. That's the grid allocation of a small country. You don't permit that, acquire the land, solve the cooling, and connect to the grid in eighteen months. That's — I mean, that timeline exists because a president needed a ribbon-cutting date, not because the engineering said yes.

Mia Drake: Which means either the physical buildout was already underway before July 24th — in which case why frame it as a new announcement — or 2027 slips and everyone scrambles to call a partial deployment a success.

Cleo Rios: And that's before you get to Vera Rubin and HBM4 both having to arrive on schedule. Both of them. At the same time.

Mia Drake: That's the structural risk nobody's naming cleanly. And zoom out one more level — Samsung's $200 billion deal with Broadcom is part of the same wave. So between SK Hynix's $750 billion in U.S. supply commitments and Samsung's Broadcom deal, South Korea has bet nearly a trillion dollars that the U.S. tech stack is where the future lives. That's not investment — that's surrendering your optionality on who defines AI infrastructure.

Cleo Rios: No exit ramp.

Mia Drake: None. So the LOI caveat matters legally, but it doesn't matter strategically anymore. The dependency is real whether or not the $500 billion ever fully closes — because the roadmap integration is already done.

Cleo Rios: So the calibrated take is: the number is soft, the timeline is a political artifact, but the lock-in is genuinely structural. South Korea doesn't get to hedge this. The bet is placed.

Mia Drake: And the asymmetry is the thing, right — when Vera Rubin is late, Jensen Huang issues a roadmap update. When South Korea is late, it's a national reckoning. Those are not the same category of consequence.

Cleo Rios: I mean — that's it, isn't it. I started this whole conversation thinking, oh, a president showed up to a boardroom, that's a weird flex. And now I'm sitting here realizing Lee Jae-myung wasn't there to add prestige. He was there because South Korea already can't afford for this to be just a letter of intent. The fine print stopped mattering the second he walked in the room.

Mia Drake: Chey Tae-won and Jensen Huang signed a document. Lee Jae-myung made it a sovereignty statement. That's a different instrument entirely.

Cleo Rios: So — not a boring chip story.

Mia Drake: Not even slightly. Thanks for going down this with me — genuinely did not expect to end on sovereign risk.

Nvidia just locked a $500 billion-plus partnership with SK Group across AI factories and memory · Onpode