South Korea’s two memory titans—Samsung Electronics and SK Hynix—are moving in rare synchrony, leaping from manufacturing toward the upper echelons of the AI value chain. Their joint strategic investment in Anthropic, the U.S.-based developer of the Claude generative AI models, signals a deliberate shift: Korean semiconductor firms are no longer content being mere component suppliers. They aim to claim influence in the architecture of generative AI itself. This is not speculative venture capital—it’s a calculated geopolitical and technological positioning.
In 2024, Anthropic secured hundreds of millions in funding backed by Amazon, Google, and notably, Korean conglomerates. Both Samsung and SK Hynix participated. The rationale is clear: Claude’s latest iterations exhibit surging demand for high-bandwidth memory (HBM). According to MLPerf benchmarks, Claude 3.5’s inference workload requires nearly 40% higher HBM3E throughput than its predecessor. With Samsung and SK Hynix collectively controlling over 90% of the global HBM market, investing in Anthropic effectively locks in one of the future’s largest memory consumers.
More profoundly, this move disrupts the traditional chip-system supply chain dichotomy. Historically, memory makers reacted passively to specifications dictated by NVIDIA or AMD. Now, through equity stakes embedded early in model development, Korean firms gain leverage to shape how AI architectures define memory subsystems. I judge this will accelerate the standardization of HBM4 and could catalyze a “memory-centric” AI chip design paradigm—precisely where Samsung and SK Hynix hold deep expertise.
Yet beneath the surface of collaboration lies fierce rivalry. Simultaneously with the Anthropic investment news, SK Hynix’s labor union demanded housing loan benefits matching Samsung’s. On the surface, it’s a labor issue; in reality, it reflects intensifying competition for talent. Engineers skilled in HBM stacking, TSV packaging, and AI workload co-optimization are now among the industry’s scarcest assets. Samsung’s integrated ecosystem and superior employee benefits have long given it an edge in talent acquisition. If SK Hynix fails to close this gap, even winning Anthropic orders may not translate into timely production ramp-up.
Geopolitics compounds this tension. While the U.S. CHIPS and Science Act doesn’t explicitly restrict HBM exports, controls on advanced packaging technologies are tightening. Samsung is building HBM capacity in Austin and Taylor, Texas, while SK Hynix has partnered with Intel to develop CoWoS-like packaging at Intel’s Arizona facilities. Two paths, one goal: secure stable supply to U.S. AI clients while mitigating sanctions risk. Anthropic, as a U.S.-based advocate for “constitutional AI,” offers a politically palatable conduit—allowing Korean firms deeper access to America’s AI ecosystem without triggering national security alarms.
Notably, neither company disclosed its exact stake in Anthropic, but market consensus suggests Samsung holds a larger share—consistent with its lead in HBM3E production. Samsung began volume production of 12-layer HBM3E in Q1 2025, while SK Hynix remains focused on 8-layer variants. This technical gap directly influences client allocation: Anthropic has already prioritized Samsung’s HBM in its internal training clusters. Such dual binding—through both capital and technology—is forging a new moat.
Looking ahead, the deepening U.S.-Korea AI alliance will reshape the global memory landscape. Micron, despite U.S. subsidies, lags by at least a year in HBM yield and scale. Foundries in China Taiwan face constraints in acquiring advanced equipment, limiting near-term competitiveness. For the next three years, the AI memory market will likely remain highly concentrated in these two Korean players. But monopoly carries risk: overreliance on a handful of clients—Anthropic, OpenAI, Microsoft—could erode pricing power, especially as these firms explore in-house compute-in-memory solutions.
South Korea’s semiconductor industry stands at a pivotal juncture. It can either cement its role as the indispensable enabler of AI infrastructure or become a high-end contractor serving American AI ambitions. The answer hinges on whether Samsung and SK Hynix can continue defining the rules of “memory as power” as AI evolves toward multimodal, real-time, and edge deployment. Anthropic is just the opening move. The real test lies ahead.