Nvidia just spent the summer locking down the one thing money alone can’t buy fast enough: memory. On July 24, 2026, the chipmaker and South Korea’s SK Group unveiled a partnership Nvidia itself is calling worth more than $500 billion, tying Nvidia’s next-generation Vera Rubin chips directly to SK Hynix’s high-bandwidth memory production for years to come. The announcement landed in San Francisco during an AI summit organized by South Korean President Lee Jae Myung, alongside the heads of OpenAI, Anthropic and Broadcom. It came ten days after SK Hynix raised roughly $28 billion in a Nasdaq share sale, and two days after AMD struck its own multi-billion-dollar pact with Anthropic. Three deals, one week, one underlying story: the AI industry’s real bottleneck in 2026 isn’t compute anymore. It’s memory.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Nvidia and SK Group Just Announced
Nvidia and SK Group described the partnership as spanning two things: AI factories, Nvidia’s term for large-scale AI data centers, and next-generation memory supply. The centerpiece is a long-term agreement for SK Hynix to supply Nvidia with high-bandwidth memory, including the HBM4 generation that Nvidia’s upcoming chips depend on. Alongside the memory piece, SK Telecom committed to building a 2-gigawatt AI data center, with the first phase expected online in 2027. Nvidia said that facility will run on its Vera Rubin platform and draw on SK Hynix HBM4 memory, tying the compute and memory halves of the deal to one physical build-out.
Nvidia and SK Hynix also agreed to jointly develop and optimize next-generation AI memory, a research partnership layered on top of the supply agreement. Crucially, the companies signed letters of intent rather than final binding contracts, a distinction that matters more than the press release headline suggests. The announcement also arrived alongside separate news that Naver and Brookfield plan to expand Naver’s own AI data center footprint in South Korea, part of a broader push by Seoul to position the country at the center of global AI infrastructure spending rather than just supplying its components.
Breaking Down the $500 Billion Figure
Headlines are running with “$500 billion,” but the number needs context most coverage skipped. CNBC reported the agreement “could be worth $500 billion over a number of years,” language that signals a multi-year estimate built from projected memory purchases, data center construction and equipment spending, not a single check Nvidia is cutting SK Group this quarter. Neither company has published a year-by-year breakdown of what portion is memory supply contracts, what portion is data center capital spending, or what portion depends on demand materializing as forecast.
That ambiguity isn’t unique to this deal. The AMD-Anthropic pact announced two days earlier used the same “up to” framing for its $5 billion equity commitment and 2-gigawatt compute buildout. Wall Street has started calling this pattern the AI industry’s circular economy: chipmakers, cloud providers and AI labs signing multi-year framework agreements sized in the hundreds of billions, most of them contingent on future capacity, future funding rounds or future demand. That doesn’t make the Nvidia SK Hynix deal meaningless. It means the $500 billion figure describes an ambition and a direction, not a transaction that already happened in full.
Why SK Hynix Holds the Leverage: Inside the HBM4 Bottleneck
The deal reads less like Nvidia doing SK Hynix a favor and more like Nvidia paying for certainty. High-bandwidth memory, the stacked DRAM that sits next to AI accelerators and feeds them data fast enough to keep the chips busy, has become the tightest link in the AI hardware supply chain. SK Hynix, the company that has led HBM production since the first generation shipped for Nvidia’s A100 in 2020, is once again the pacesetter for HBM4, the version Nvidia’s Vera Rubin platform needs.
SK Hynix CEO Kwak Noh-jung has not been shy about how tight that supply really is. Speaking to reporters on July 10, 2026, the day of SK Hynix’s Nasdaq debut, Kwak said: “We forecast that next year will be the worst year in the industry’s history from the supply perspective.” He added: “Our customer demand continues to go up, while our capacity has limitations. We still forecast that customer demand will remain higher than our supply capacity even beyond 2030.” That is a memory company CEO telling investors, on the record, that his factories cannot catch up to demand before the next decade. For Nvidia, locking in a named, multi-year commitment from the company making that admission is close to essential.
SK Hynix’s $28 Billion Nasdaq Debut Adds Another Layer
The timing of the Nvidia deal is hard to separate from what SK Hynix did two weeks earlier. On July 9, 2026, SK Hynix priced a Nasdaq listing of 17.79 million new American depositary shares, raising roughly $28 billion to $29 billion, one of the largest new share sales anywhere in the world this year. Trading began July 10 under the temporary ticker SKHYV, before switching to the regular ticker SKHY on July 13. The underlying common shares received an additional listing on Korea’s KOSPI exchange on July 29.
The listing was widely read as SK Hynix tapping U.S. capital markets directly while the global AI boom is still funding memory expansion at a scale the company couldn’t finance through Seoul alone. SK Group chairman Chey Tae-won framed the moment in broader terms. In comments reported by Chosun Biz, he said: “With the spread of AI data centers, AI factories, and AI PCs, memory demand will keep growing.” Two weeks later, Nvidia showed up with a $500 billion validation of that thesis. Whether that sequencing was coordinated or coincidental, SK Hynix walked into its most important customer negotiation of the year fresh off a record capital raise and a wave of investor attention.
Vera Rubin: The Chips This Memory Deal Is Built For
Vera Rubin is Nvidia’s next rack-scale AI platform, and it’s the reason HBM4 supply matters so much right now. Nvidia says Vera Rubin is already in full production, with partner systems shipping in the second half of 2026 and production shipments starting this fall. The platform pairs a Rubin GPU with a Vera CPU. Nvidia’s technical materials list the Rubin GPU at roughly 336 billion transistors, up to 50 petaflops of inference throughput, 35 petaflops of training throughput, and HBM4 memory delivering 22 terabytes per second of bandwidth.
The Vera CPU side uses 88 Arm-based Olympus cores across 176 threads, with 1.2 terabytes per second of memory bandwidth and a 1.8 terabytes per second NVLink-C2C link back to the GPU, according to Nvidia’s developer documentation. Every one of those numbers depends on HBM4 shipping in volume, on schedule, from a supplier Nvidia trusts. That is the real function of the SK Group deal. It functions less like a memory purchase order and more like an insurance policy on Nvidia’s own product roadmap.
SK Telecom’s 2-Gigawatt Data Center and the 2027 Timeline
On the data center side of the deal, SK Telecom committed to a 2-gigawatt AI facility, with the first phase expected online in 2027. Two gigawatts is a useful yardstick for how large AI infrastructure commitments have gotten. It’s roughly the continuous output of two full-sized nuclear reactors, dedicated to a single company’s AI computing footprint. Nvidia positioned the facility as serving sovereign, physical, agentic and enterprise AI services, corporate shorthand for everything from national government AI projects to robotics workloads to enterprise chatbots.
The regional framing matters too. Nvidia explicitly tied the buildout to AI demand across Asia-Pacific, with South Korea positioned as the anchor market. That lines up with the diplomatic backdrop: the deal was announced at a summit that President Lee Jae Myung organized specifically to bring AI infrastructure leaders to San Francisco, alongside separate news that Naver is expanding its own AI data center capacity with Brookfield. Seoul is treating 2026 as the year South Korea stops being just a component supplier to the AI industry and starts operating as a computing hub in its own right.
Nvidia’s 2026 Partnership Sprint, By the Numbers
The SK Group deal is the largest of several Nvidia-adjacent mega-partnerships announced in 2026, but it isn’t the only one reshaping how AI infrastructure gets built and financed. The table below lines up the year’s biggest disclosed announcements by value and scope.
| Deal | Announced | Headline Value | What It Covers | Key Milestone |
|---|---|---|---|---|
| Nvidia + SK Group | July 24, 2026 | $500B+ over multiple years | AI data centers plus HBM4 memory supply | First SK Telecom facility live in 2027 |
| AMD + Anthropic | July 22, 2026 | Up to $5B equity, up to 2GW compute | Instinct MI450-series GPUs, Helios racks | First 1GW live in H1 2027 |
| OpenAI + Broadcom | June 24, 2026 | Custom inference chip (“Jalapeño”) | In-house ASIC, about 50% cheaper inference vs. GPUs | Taped out in nine months |
| Micron (U.S. expansion) | July 10, 2026 | $250B through 2035 | Domestic HBM/DRAM fab capacity | New York facility groundbreaking |
| Meta + Microsoft (leases) | 2026, ongoing | $850B in combined data center leases | AI data center real estate | Multi-year build-out underway |
The Competitive Squeeze: Samsung and Micron Race to Catch Up
SK Hynix’s position doesn’t leave Samsung Electronics or Micron standing still. Both companies are pushing to expand their own HBM4 output and win a larger share of Nvidia’s supply chain, though neither has published the kind of named, dollar-denominated Nvidia agreement that SK Hynix just landed. Micron used the same mid-July window to raise its own ambitions, lifting its U.S. investment target to $250 billion through 2035, breaking ground on a new facility in New York, and lining up supply deals with Ford and Meta.
Samsung, historically the world’s largest memory maker by volume, has spent much of 2025 and 2026 working to close a yield gap with SK Hynix on advanced HBM. Industry reporting describes all three major suppliers as working through Nvidia’s qualification process for HBM4, without a confirmed public timeline for Samsung reaching the scale of commitment SK Hynix now has in writing. For Nvidia, spreading orders across three suppliers is the obvious hedge against exactly the kind of single-vendor shortage SK Hynix’s own CEO has been warning about.
Historical Context: From GPU Scarcity to Memory Scarcity
This is the third distinct hardware shortage the AI industry has lived through since ChatGPT’s 2022 debut, and each one has moved a layer further down the supply chain. The first shortage, in 2023 and 2024, was raw GPU scarcity. Nvidia couldn’t make H100 and A100 chips fast enough, and cloud providers waited months for allocations. The second, through 2025, was a power and data center capacity shortage, as hyperscalers realized they had chips to install but not enough electricity or shell space to run them, a constraint tech-insider.org has tracked closely in South Korea’s own $880 billion chip plan hitting its own power wall.
The current shortage is memory. HBM production requires specialized packaging and testing capacity that takes years to build, not months, which is exactly why SK Hynix’s CEO is comfortable forecasting a shortage lasting into 2027 and demand outstripping supply beyond 2030. Each shortage has pushed AI infrastructure spending further upstream, from finished chips to power grids to the raw memory dies themselves. The Nvidia SK Hynix deal is best read as Nvidia trying to get ahead of a fourth shortage before it happens, by owning a piece of the memory roadmap directly instead of buying whatever capacity is left on the open market.
AMD-Anthropic vs. Nvidia-SK Group: A Week of AI Mega-Deals
Two days before the Nvidia-SK Group announcement, AMD and Anthropic unveiled their own strategic partnership at AMD’s Advancing AI 2026 conference. The contrast between the two deals says a lot about how differently companies are placing their bets this year.
| Category | AMD + Anthropic | Nvidia + SK Group |
|---|---|---|
| Date announced | July 22, 2026 | July 24, 2026 |
| Headline value | Up to $5 billion equity investment | $500 billion-plus, multi-year |
| What’s being secured | Compute demand (chip adoption) | Memory supply (HBM4) |
| Core hardware | Instinct MI450-series, Helios racks | Vera Rubin GPUs, SK Hynix HBM4 |
| Capacity commitment | Up to 2 gigawatts | 2 gigawatts (SK Telecom facility) |
| First deployment | First 1GW in H1 2027 | First facility in 2027 |
| Contract status | Strategic partnership announced | Letters of intent signed |
The structural difference matters. AMD’s deal ties an equity investment directly to a customer adopting its chips, essentially paying Anthropic to commit. Nvidia’s deal instead locks in a supplier, paying for guaranteed memory rather than guaranteed demand. Both bets are proxies for the same underlying anxiety: that whichever company lacks a locked-in position by the end of 2026 will be the one paying spot-market prices once the shortage SK Hynix’s CEO is describing actually arrives. For more on the AMD side of that same week, see our coverage of AMD Advancing AI’s Helios bookings and gigawatt-scale commitments.
Market and Investor Reaction
Public reaction to the Nvidia-SK Group announcement has been more measured than the headline number might suggest, partly because so much detail (exact spending splits, equity components, binding contract terms) remains undisclosed. CNBC’s coverage noted that Wall Street’s AI-chip enthusiasm had already been broadening beyond Nvidia alone well before this deal. Reporting from May 8, 2026 tracked investor attention shifting toward Intel, AMD and Micron as the AI trade widened past a single company.
That backdrop shapes how investors are reading this deal: less as a surprise that sends any single stock sharply in one direction, and more as confirmation of a trend that has been building all year. The AI chip trade is increasingly a memory and infrastructure trade, with Nvidia at the center but SK Hynix, Samsung, Micron and AMD all repricing around the same underlying scarcity story.
What It Means for Gamers and PC Buyers
None of this is happening in a vacuum for anyone buying a graphics card or a prebuilt gaming PC in 2026. HBM isn’t the memory inside a gaming rig (that’s GDDR and standard DRAM), but HBM, GDDR and consumer DRAM all draw on the same fabs, the same wafer capacity, and often the same packaging lines at SK Hynix, Samsung and Micron. When a memory maker’s most profitable, best-guaranteed revenue comes from HBM contracts with Nvidia, new capacity tends to get prioritized there first.
tech-insider.org has already tracked the downstream effect on consumer hardware pricing this year, from SSD prices climbing as NAND supply tightens to broader RAM shortages hitting gaming PC builders directly. A $500 billion memory commitment from the industry’s largest chip buyer doesn’t reverse that pressure. It signals that the companies setting memory allocation priorities for the next several years are optimizing for AI data centers first, with whatever capacity is left over trickling down to consumer DRAM and graphics memory.
Risks: Letters of Intent Are Not Signed Contracts
The biggest caveat in the entire announcement is also the easiest one to miss. Nvidia and SK Group didn’t sign a completed, binding supply agreement. They signed letters of intent to formalize one. That’s a meaningful legal distinction. Letters of intent signal serious commercial direction, but they typically leave pricing, volume commitments, exclusivity terms and penalty clauses to be negotiated later, and they can be renegotiated if market conditions change.
Neither Reuters, CNBC nor Nvidia’s own release disclosed a specific dollar breakdown, an equity stake, or a binding delivery schedule beyond the 2027 target for SK Telecom’s first data center phase. Given how many “up to” and multi-year qualifiers appear across this deal and the AMD-Anthropic announcement from the same week, the honest read is that both companies have committed to a direction and a scale, with the hardest negotiating (who bears the cost if AI demand growth slows, what happens to pricing if the shortage eases before 2030) still ahead of them.
Five Predictions for the Next 12 Months
Based on the pattern of deals announced across 2026, here’s how this is likely to play out over the next year.
- Samsung and Micron will each announce their own named HBM4 supply agreements within two to three quarters rather than cede the narrative entirely to SK Hynix.
- The letters of intent will convert into definitive agreements with disclosed volume and pricing terms sometime in 2027, likely alongside SK Telecom’s first data center milestone.
- Consumer DRAM, GDDR and NAND prices stay elevated through at least the first half of 2027, since none of the new 2026 capacity is earmarked for retail channels first.
- At least one more gigawatt-scale mega-partnership lands before the end of 2026, following the same template as AMD-Anthropic and Nvidia-SK Group: multi-year commitments framed around a headline dollar figure that describes potential rather than a completed transaction.
- South Korea keeps leaning into AI diplomacy, using SK Hynix’s Nvidia relationship and Samsung’s own ambitions as leverage in trade and investment talks with Washington, building on the summit that produced this announcement.
What Industry Voices Are Saying
Four on-the-record statements from the companies involved capture the stakes better than any outside analyst summary.
“The expansion will include a co-develop opportunity for us on the next-generation SK Hynix AI memory, and this will help us secure a stable supply of HBM memory.”
Raj Mirpuri, Enterprise Vice President, Nvidia — via CNBC
“We forecast that next year will be the worst year in the industry’s history from the supply perspective.”
Kwak Noh-jung, CEO, SK Hynix — via Yahoo Finance
“With the spread of AI data centers, AI factories, and AI PCs, memory demand will keep growing.”
Chey Tae-won, Chairman, SK Group — via Chosun Biz
Read together, the statements describe an industry where the company selling memory is telling the world it can’t make enough of it, and the company buying that memory is responding by committing to a nine-figure, multi-year deal to make sure it’s first in line anyway.
Frequently Asked Questions
What exactly did Nvidia and SK Group announce?
On July 24, 2026, Nvidia and SK Group announced a partnership Nvidia values at more than $500 billion over multiple years, covering AI data centers and a long-term SK Hynix memory supply agreement, including HBM4 for Nvidia’s Vera Rubin chips.
Is the $500 billion a signed, binding contract?
No. Nvidia and SK Group signed letters of intent, not a completed supply contract. CNBC described the figure as an estimate of what the partnership “could be worth… over a number of years,” not a lump-sum payment.
What is HBM4 and why does it matter for this deal?
HBM4 is the newest generation of high-bandwidth memory, the stacked DRAM that sits beside AI chips to feed them data fast enough to keep them running at full speed. Nvidia’s Vera Rubin platform is built around HBM4, and SK Hynix is currently the industry’s leading HBM supplier.
How does this affect gaming GPU and RAM prices?
Not directly, since HBM is a different product from the GDDR and DRAM used in consumer PCs. Indirectly, all three memory types share fabrication and packaging capacity, so prioritizing HBM4 for AI customers can leave less capacity for consumer memory, adding to pricing pressure already seen in 2026.
What is Nvidia’s Vera Rubin platform?
Vera Rubin is Nvidia’s next rack-scale AI computing platform, pairing a Rubin GPU (about 336 billion transistors, up to 50 petaflops of inference performance) with a Vera CPU (88 Arm-based cores). Nvidia says it is in full production, with partner systems shipping in the second half of 2026.
How does this compare to the AMD-Anthropic deal?
AMD’s deal, announced two days earlier, commits up to $5 billion in equity investment tied to Anthropic adopting up to 2 gigawatts of AMD Instinct MI450-series GPUs. Nvidia’s deal instead secures memory supply from SK Hynix rather than demand from a customer, making the two deals complementary rather than directly competing.
When does the first SK Telecom data center come online?
Nvidia and SK Group say the first phase of the 2-gigawatt SK Telecom AI facility is expected online in 2027, running on Nvidia’s Vera Rubin platform and SK Hynix HBM4 memory.
Are Samsung and Micron part of this deal?
No. This specific agreement is between Nvidia and SK Group. Samsung and Micron are separately working to expand their own HBM4 output to compete for a larger share of Nvidia’s future memory orders, and Micron announced its own $250 billion U.S. expansion plan the same week.
Related Coverage
- AMD Advancing AI: Helios Hits $5.25M, 12GW Booked [2026]
- South Korea’s $880B Chip Plan Hits Power Wall [2026]
- RAM Shortage 2026: AI Chips Hit 5 Gaming Platforms’ Prices
- AMD MI350P vs Nvidia H200: 40% Faster, $30K Price [2026]
- SSD Prices Double as NAND Shortage Hits Gaming PCs [2026]
- OpenAI’s Jalapeño Chip Undercuts Nvidia by 50% [2026]
- More AI Chip Coverage
