Getty Images

The world's two dominant memory chipmakers have contractually locked in the high-bandwidth memory supply for the two most powerful AI accelerator platforms through the end of the decade — and companies not covered by those agreements now face years without a path to the same chips. South Korean President Lee Jae Myung convened a high-stakes AI Summit at The Midway in San Francisco on Friday, yielding $950 billion in semiconductor and AI infrastructure commitments between South Korea's two largest chipmakers and America's most powerful technology companies. (Exchange rate as of July 25, 2026; any KRW-to-USD conversions are approximate.)

Samsung Electronics signed a sweeping five-year memorandum of understanding with Broadcom worth over $200 billion, covering advanced high-bandwidth memory supply, cutting-edge foundry manufacturing, and next-generation packaging. Separately, SK Group committed to $750 billion in long-term memory chip supply partnerships with Nvidia and a broader coalition of US technology firms — a figure that dwarfs any prior semiconductor supply agreement on record.

The deals, structured as forward purchase contracts rather than equity investments, are designed to lock in the HBM supply chain that underpins the world's most powerful AI accelerators through 2030. That matters for AI companies not in the room: with all three major HBM suppliers already sold out through 2026 and meaningful new capacity not arriving until 2027 or 2028, any organization that didn't secure a long-term supply agreement Friday faces a multi-year hardware constraint that money alone cannot solve.

What "Forward Purchase Contracts" Actually Means — and Why the Distinction Matters

Presidential adviser Kim Yong-beom drew a distinction at a post-summit briefing that is easy to miss but crucial to understanding the agreements' economic logic.

"These are advance contracts — advance orders for memory chips on that scale. This is not an investment," Kim told reporters. "The agreements are commitments that say: 'If you produce the memory chips, we will purchase them in those expanded volumes.'"

The structure serves two different purposes. For Samsung and SK Hynix, forward purchase commitments provide the demand certainty needed to justify the billions in capital expenditure required to build new HBM capacity — investments that cannot be responsibly committed without guaranteed offtake. Building a new HBM-capable fabrication facility takes three or more years and costs upward of $10 billion; no company commits that capital on the basis of spot-market signals alone. "Our companies have won long-term supply contracts, which will allow them to plan production with far greater stability," Kim said.

For Nvidia and Broadcom, the structure converts an unpredictable and increasingly expensive spot market into a managed, multi-year supply pipeline. Global contract DRAM prices rose 90 to 95 percent quarter-over-quarter in the first quarter of 2026 — TrendForce's February 2026 forecast, revised sharply upward from a prior estimate of 55 to 60 percent, was then confirmed when the actual realized increase came in at 93 to 98 percent, the largest quarterly memory price spike ever recorded. NAND flash prices rose more than 50 percent in the same period. Industry analysts warn that supply will remain tight for at least 12 to 18 months, with meaningful new HBM capacity not arriving until 2027 or 2028 at the earliest.

Samsung and Broadcom: A $200 Billion Stack From Silicon to Packaging

The headline deal on the memory and foundry side came from Samsung. The company and Broadcom signed an MOU expected to be valued at more than $200 billion across memory chips, foundry services, and advanced packaging over five years through 2030.

On the memory front, Samsung will supply Broadcom with its most advanced high-bandwidth memory products — HBM4 and its successor HBM4E — for integration into Broadcom's next-generation AI accelerators. On the foundry side, Samsung's manufacturing arm will produce Broadcom's AI and networking chips on its 2-nanometer and below process technologies, including wireless broadband communication products. The agreement extends to advanced packaging: Samsung will apply its 2.3D and 2.5D integration capabilities to stack memory and logic into unified modules, enabling higher-performance, more power-efficient AI silicon.

The MOU was signed at the summit with Samsung Foundry President Jinman Han and Broadcom CEO Hock Tan both present.

"AI is driving unprecedented demand for tightly integrated semiconductor technologies spanning memory, logic and advanced packaging," said Young Hyun Jun, Vice Chairman and CEO of Samsung's Device Solutions Division. "By expanding our collaboration with Broadcom across these critical technologies, we look forward to delivering greater value to customers while advancing the AI infrastructure of the future."

The scope of the deal carries competitive significance for Samsung's position in HBM. The company lost ground to SK Hynix during the HBM3E generation — SK Hynix holds approximately 56 percent of the global HBM market by revenue as of early 2026, according to IDC data in the company's SEC filing, with Samsung holding a substantially smaller share. A long-term, multi-dimensional commitment from Broadcom — one of the most important fabless chip companies — represents a significant validation of Samsung's HBM4 roadmap and the maturity of its 2-nanometer process node.

SK Hynix and Nvidia: $500 Billion and HBM Memory Through the Decade

If the Samsung-Broadcom deal reshaped perceptions of Samsung's competitive trajectory, the SK Group-Nvidia agreement redefined what a semiconductor supply commitment even looks like.

SK Group and Nvidia announced a $500-billion-plus comprehensive partnership spanning AI factory construction and next-generation memory supply, signing letters of intent at the summit. The broader SK Group commitment — encompassing Nvidia and other US technology partners — totals $750 billion, according to Korea Herald reporting on Kim Yong-beom's briefing.

The Nvidia component is structured around two interlocking pillars. On the memory side, SK Hynix will establish a long-term partnership to secure and jointly develop next-generation HBM for Nvidia's AI accelerator roadmap. "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," said Raj Mirpuri, Nvidia's enterprise vice president.

On the infrastructure side, SK Telecom will build a 2-gigawatt AI factory powered by Nvidia's Vera Rubin DSX platform and SK Hynix's HBM4 memory. The first facility is expected to come online in 2027, with Nvidia providing priority access to its Vera Rubin systems. Nvidia's Vera Rubin DSX, which entered full production in May 2026, delivers 10x agent throughput compared with the prior Grace Blackwell platform and represents Nvidia's reference architecture for next-generation AI factories.

Nvidia separately announced a $1 billion investment in Naver to expand GPU supplies and support the Korean internet company's plans for a $10 billion global AI factory, with infrastructure fund Brookfield providing up to an additional $9 billion, according to Korea Herald reporting on the summit briefing.

Jensen Huang, speaking at the summit, described South Korea as indispensable to global AI — noting that without HBM memory invented in Korea, Nvidia would not have been able to build the AI supercomputers that power today's models, and calling the current moment "the golden age for Korea."

SK Group Chairman Chey Tae-won noted that AI customers are now seeking "far more memory than previously anticipated," citing discussions with Nvidia, Anthropic, OpenAI, and Broadcom as evidence of a demand surge that is reshaping what supply commitments need to look like, Reuters reported.

"I wouldn't be surprised if, the next time I see [Huang], he'll tell me those figures were too low and say, 'I want more,'" Chey said.

How High-Bandwidth Memory Works — and Why Only Three Companies Can Make It

High-bandwidth memory is not a standard commodity. The technology stacks multiple DRAM dies vertically and connects them via microscopic through-silicon vias — essentially drilling thin electrical wires through stacked silicon chips — and then mounts the resulting assembly on a silicon interposer directly adjacent to the GPU or AI accelerator it serves. The result is a dramatically wider data path than conventional memory can achieve.

Where standard GDDR memory uses a 32-bit-wide connection per channel, a single HBM stack uses a 1,024-bit or 2,048-bit interface — a fire hose versus a garden hose. HBM4, whose JEDEC standard was published in April 2025, doubles the interface width from 1,024 bits to 2,048 bits, delivering roughly doubled bandwidth compared with its predecessor HBM3E — up to 2 terabytes per second per stack from the prior 1.2 to 1.3 TB/s — while improving power efficiency by more than 40 percent. SK Hynix has achieved per-pin speeds above 10 gigabits per second in mass production, exceeding the JEDEC baseline of 8 Gbps at Nvidia's request.

Building new HBM capacity takes years and costs billions — and the window between committing capital and delivering chips is long enough that demand forecasts must be locked in well in advance. Samsung's next new HBM-capable facility in Pyeongtaek is not expected to be operational until 2028; SK Hynix's new M15X facility in South Korea is slated for utilization by mid-2027; and SK Hynix's US advanced packaging facility in Indiana is not expected to reach mass production until the second half of 2028.

The HBM production constraint also radiates outward. Every HBM wafer produced displaces roughly two to three conventional DRAM wafers because HBM requires substantially more process steps and clean-room time. Micron CEO Sanjay Mehrotra stated on the company's fiscal Q1 2026 earnings call that "the gap between the demand and supply for all of DRAM, including HBM, is really the highest that we have ever seen." The three companies that control more than 90 percent of global DRAM — Samsung, SK Hynix, and Micron — simultaneously warned of multi-year shortages as their order books extended into 2027.

Who Gets Left Out: The Supply Lock-In's Hidden Consequence

The deals announced Friday benefit their signatories precisely because HBM supply cannot be conjured on demand. That same scarcity is what makes the lock-in consequential for the rest of the industry.

HBM capacity from all three major suppliers — SK Hynix, Samsung, and Micron — was fully sold out through 2026 before Friday's summit. Industry analysts have characterized the situation as not just a cyclical shortage driven by mismatched supply and demand, but a potentially permanent, strategic reallocation of the world's silicon wafer capacity.

The consequences are already visible. Gaming GPU production has been cut 30 to 40 percent in the first half of 2026 as memory manufacturers prioritize higher-margin HBM over the GDDR7 used in consumer graphics cards. Consumer DRAM prices have tracked the same surge: a 32GB DDR5 kit that sold for $100 to $200 in October 2025 had risen above $350 by early 2026, and DDR4 kits that sold for $60 to $90 in the same period were trading at $150 to $180.

AI startups and mid-tier cloud providers without long-term HBM supply agreements face compute rationing or sharply higher GPU costs through 2027 as Nvidia prioritizes the hyperscalers that hold supply agreements.

The antitrust dimension has also surfaced. On June 25, 2026, 17 plaintiffs filed a federal antitrust class action against SK Hynix, Samsung, and Micron in the US District Court for the Northern District of California, alleging the three companies coordinated to curtail conventional DDR3 and DDR4 production using the HBM transition as cover. The allegations are unproven and the companies have not been adjudicated liable.

How Does HBM4 Actually Improve AI Performance?

HBM4's doubled interface width is the answer to a specific bottleneck: AI models — particularly the large-scale language and multimodal models that drive today's frontier applications — require continuous, high-volume data transfer between memory and compute. A GPU running at peak capacity can process data faster than conventional memory can supply it. HBM solves this by co-locating memory physically adjacent to the compute die on a silicon interposer, eliminating the distance data must travel and widening the data path dramatically.

HBM4's 2,048-bit interface delivers 32 independent channels — double the 16 channels of HBM3E — enabling greater access parallelism for the kind of massive tensor operations AI training and inference require. For the SK Telecom AI factory that will use Vera Rubin and HBM4, this architectural step-change is what makes 2-gigawatt-scale AI infrastructure economically viable: more memory bandwidth per watt means more AI computation per dollar of electricity.

Beyond Chips: Anthropic, Robotics, and What South Korea Is Actually Building

The summit's agreements extended well beyond silicon. Anthropic CEO Dario Amodei confirmed at the summit that Anthropic has signed supply deals with both Samsung Electronics and SK Hynix. In May 2026, Anthropic raised funding at a post-money valuation of $965 billion, naming both Korean chipmakers as critical partners in its memory and logic supply chain.

Amodei told President Lee during their one-on-one meeting that South Korea is "the best place at present to build an AI data center in a short period of time," and signaled openness to expanding Anthropic's footprint, contingent on the pace of regulatory approvals, according to Kim Yong-beom's post-summit briefing. On the data center front, Kim said South Korean companies and global tech giants agreed to pursue partnerships in large-scale AI data centers with a combined power capacity of approximately 5 gigawatts and roughly 2 million GPUs. Samsung SDS and Anthropic additionally signed a strategic partnership to jointly develop new AI markets and train AI engineers.

In physical AI, Hyundai Motor Group and Nvidia agreed to jointly develop a standardized robot reference platform for universities and startups. Hyundai also unveiled an "autonomous vehicle foundry" plan with Waymo.

"A Global Initiative": South Korea's Calculated Bet

The summit was the culmination of a strategy months in the making. Last month, President Lee unveiled a domestic industrial plan backed by corporate commitments totaling at least $880 billion, anchored by Samsung and SK Group, aimed at expanding chip production, building AI data centers, and strengthening leadership in semiconductors and physical AI. The San Francisco commitments are the international complement — the moment when government strategy, chaebol capital, and American technology demand converged into binding commercial relationships.

Presidential policy chief Kim Yong-beom framed the significance directly: "This visit was intended to demonstrate at home and abroad that the three megaprojects announced last month constitute a major global initiative jointly developed by Korea, with its powerful semiconductor industry, and big tech companies seeking strategic investment partnerships," Kim told reporters at a post-summit briefing.

Nvidia also announced a joint AI research lab with the Korea Advanced Institute of Science and Technology (KAIST) — the first joint AI lab between a Korean university and any global technology company.

Huang and other US executives, speaking informally at a post-summit dinner, expressed confidence that investment in Korea could produce a top-tier frontier AI model within a short period, and urged South Korea to develop its own domestic AI model rather than relying solely on those built elsewhere.

After the summit, the assembled US tech executives and Korean industrial leaders gathered for dinner at a nearby restaurant, where President Lee raised a toast and called out "We are—" with the assembled executives responding: "Family," according to the Korea Herald's coverage of Kim Yong-beom's account of the evening.

Frequently Asked Questions

What is a forward purchase contract in the context of the Samsung-Broadcom and SK Hynix-Nvidia deals?

A forward purchase contract is a long-term supply agreement in which a buyer commits to purchase specific volumes of a product at agreed prices over a defined future period — in this case, through 2030. Presidential adviser Kim Yong-beom explicitly distinguished these from equity investments: they are advance orders for memory chips, not stakes in the companies. For chipmakers, the agreements justify the multi-billion-dollar capital expenditure needed to build new HBM fabrication facilities, because construction requires years and committed demand. For buyers like Nvidia and Broadcom, they convert an unpredictable spot market into a managed supply pipeline — critical when memory prices have risen 93 to 98 percent in a single quarter.

What is HBM4 and how does it differ from previous generations?

HBM4 is the sixth generation of high-bandwidth memory, standardized by JEDEC in April 2025. It doubles the interface width from 1,024 to 2,048 bits, enabling roughly twice the bandwidth of its predecessor HBM3E — up to 2 terabytes per second per stack, versus 1.2 to 1.3 TB/s for HBM3E. Manufacturers including SK Hynix have pushed per-pin speeds above 10 gigabits per second in production, above the JEDEC baseline of 8 Gbps. The practical effect for AI workloads is greater data throughput per watt, allowing AI accelerators to sustain higher compute utilization.

Will AI companies and cloud providers that aren't covered by these deals be able to get the same HBM?

In the short term, no. All three major HBM suppliers — SK Hynix, Samsung, and Micron — had their 2026 production fully sold out before Friday's summit. Meaningful new HBM capacity is not expected until 2027 or 2028 from new facilities currently under construction. AI startups and mid-tier cloud providers without long-term supply agreements face compute rationing or substantially higher hardware costs through the end of 2027, because long-term forward contracts with hyperscalers take priority over spot-market buyers. The 3-to-1 wafer displacement ratio — each HBM wafer produced displaces roughly two to three conventional DRAM wafers — also means HBM ramp puts direct pressure on consumer and enterprise memory supply chains.

What does the $950 billion total actually represent, and is it real money changing hands?

The $950 billion represents the projected value of committed forward purchases over the next five years — not a single cash transfer, stock purchase, or investment. Samsung's $200 billion covers the value of HBM supply, foundry manufacturing, and advanced packaging Broadcom has committed to purchase from Samsung through 2030. SK Group's $750 billion (including SK Hynix's $500-billion-plus Nvidia partnership) represents committed memory chip and AI infrastructure purchases. These are binding commercial commitments with real consequences: they obligate Samsung and SK Hynix to produce at agreed volumes and obligate buyers to purchase that output, giving both sides the certainty needed to justify multi-year capital plans.