SK Group’s expanded cooperation framework with Nvidia and Samsung Electronics’ new memorandum with Broadcom put more than $700 billion of planned AI supply-chain activity around the same constraint: obtaining enough high-bandwidth memory and advanced packaging to turn cutting-edge processors into deployable AI servers.
The headline figures are large, but Bernstein said in a July 27 report that they should not be treated as a fresh wave of chip purchase orders. Neither arrangement discloses binding purchase volumes, prices, product mix or delivery schedules. Instead, the commitments place memory supply, chip manufacturing, packaging and data-center construction within longer-term strategic partnerships that could shape which companies can deliver AI systems on time through the end of the decade.
AI server makers increasingly need coordinated access to several scarce components. GPUs and custom AI accelerators provide the computing power, but HBM supplies the high-speed memory needed to keep those chips fed with data. Advanced packaging then connects processors and stacks of HBM closely enough to support the bandwidth required by large AI workloads. A shortage, yield problem or qualification delay in any one of these stages can hold up a completed server shipment.
SK Group links memory development to AI data centers
SK Group and Nvidia disclosed an AI supply-chain cooperation framework valued at more than $500 billion. The arrangement combines SK hynix’s work on next-generation memory with SK Telecom’s plans to build AI computing infrastructure, rather than describing a single contract for Nvidia hardware.
According to company disclosures, SK Telecom’s planned AI factory could reach up to 2 gigawatts of capacity, with its first site targeted to begin operating in 2027. That timetable points to a multi-year buildout requiring power infrastructure, computing systems and a reliable stream of memory-equipped servers.
The cooperation also references Nvidia’s DSX infrastructure, the company’s Vera Rubin computing platform and SK hynix’s HBM4 development. HBM4 is expected to be a new generation of stacked memory designed for increasingly demanding accelerators, though the companies did not provide commercial shipment volumes or a detailed delivery calendar in the material cited by Bernstein.
The structure matters for SK Group because it connects two parts of the conglomerate’s AI strategy: SK hynix’s role as a memory supplier and SK Telecom’s attempt to operate large-scale AI infrastructure. That model would give the group a route to participate in both supplying components and deploying the systems that consume them.
Samsung and Broadcom broaden the manufacturing scope
Samsung Electronics and Broadcom have signed an MOU covering planned cooperation worth more than $200 billion over five years through 2030. Samsung described the arrangement as spanning HBM supply, support for Broadcom’s next-generation AI accelerator and foundry manufacturing tied to its 2-nanometer process technology.
The proposed work also includes 2.3D and 2.5D integration, packaging approaches that combine multiple chips and memory components in dense configurations. Such designs are increasingly central to AI hardware because performance depends on moving data rapidly among the compute chip, HBM stacks and other components.
Samsung framed the agreement as an integrated “memory + foundry + packaging” offer. That places it against manufacturing models led by Taiwan Semiconductor Manufacturing Co., which has built a strong position in advanced chip production and CoWoS packaging, a family of technologies widely used to integrate high-performance logic chips with HBM.
Bernstein compared Samsung’s Cube-S and Cube-E/R approaches with TSMC’s CoWoS-S/L/R platforms, focusing on their role in joining logic processors to stacked memory through high-bandwidth interconnects. The competition is less about a single chip design than the ability to deliver a qualified package at scale, with acceptable yields and dependable lead times.
Capacity has become the commercial pressure point
Bernstein said market attention has moved beyond the availability of compute chips alone. HBM production capacity, packaging lines, substrates and the time required to qualify each component have become equally material constraints on AI server delivery.
This places Samsung’s proposed Broadcom cooperation under a demanding execution test. The company would need to align HBM output, leading-edge foundry capacity, packaging, substrate supply, manufacturing yields and logistics. A problem in one layer could reduce the value of capacity available in another, particularly where an accelerator cannot ship until its memory and packaging have passed qualification.
The near-term competitive impact on TSMC may be limited, Bernstein said, because demand for advanced-node manufacturing capacity remains fully booked. Samsung’s opportunity therefore rests less on immediately displacing established foundry orders and more on proving it can offer customers a workable alternative across several tightly linked production stages.
Memory market forecasts put the commitments in context
Bernstein cited consensus expectations for annual global memory-industry revenue of about $0.9 trillion in 2026 and roughly $1.3 trillion in both 2027 and 2028. TrendForce separately projected the global memory market at about $1.28 trillion in 2027, with Samsung, SK hynix, Micron, KIOXIA and Chinese suppliers among the companies serving the market.
Those estimates explain why HBM has become strategically important beyond its direct sales value. AI systems require relatively specialized, high-performance memory, and HBM capacity competes for engineering attention and production resources with other memory products. Suppliers that can expand output while maintaining yields could gain leverage in negotiations with accelerator designers and system builders.
Bernstein maintained Outperform ratings on Samsung Electronics, SK hynix, Micron, Nvidia and Broadcom, while retaining an Underperform rating on KIOXIA. Its report listed target prices of 440,000 won for Samsung common shares, 3.30 million won for SK hynix, $315 for Nvidia and $550 for Broadcom.
The SK-Nvidia and Samsung-Broadcom arrangements show that AI infrastructure planning is moving toward integrated supply commitments, where access to HBM and advanced packaging can determine whether a processor roadmap reaches a functioning data center. The first planned SK Telecom AI factory is not targeted for operation until 2027, underscoring how far ahead companies are now trying to secure the manufacturing chain behind their AI ambitions.
Explore how institutional flows, tariffs, and liquidity shape AI-focused assets—start with our latest market outlook on macro crypto trends.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
