More than 30,000 Samsung Electronics workers have taken to the streets of South Korea demanding a greater share of the company’s AI-driven profits, marking one of the largest organized labor actions in the tech giant’s history as its semiconductor division posts record earnings fueled by surging global demand for high-bandwidth memory and AI accelerators.
The Profit Paradox: AI Boom vs. Stagnant Wages
Samsung’s semiconductor business reported operating profits of 10.4 trillion KRW ($7.6 billion) in Q1 2026, a 218% year-over-year increase driven primarily by HBM3E sales to NVIDIA and AMD for AI training clusters. Yet despite this windfall, base salaries for production-line engineers and fab technicians have risen just 3.2% annually over the past three years, according to data from the Korea Employers Federation. Workers argue that although Samsung allocates billions to AI R&D and shareholder buybacks, the human labor enabling its chip dominance remains structurally undercompensated.
This disconnect reflects a broader trend in the AI supply chain: extreme profitability concentrated at the design and IP layers, while manufacturing labor sees diminishing returns. Unlike TSMC, which ties 15% of executive bonuses to workforce productivity metrics, Samsung’s compensation model still weights heavily toward tenure and title rather than output in high-stress, cleanroom environments.
Under the Hood: Why Memory Workers Hold Leverage
The striking workers are primarily from Samsung’s Pyeongtaek and Hwaseong campuses, which produce over 60% of the world’s HBM3E memory — a specialized DRAM stacked with TSV (Through-Silicon Vias) and thermocompression bonding to achieve 1.2TB/s bandwidth per stack. These modules are not commoditized; each HBM4 stack requires 28-day fabrication cycles involving 1,400+ process steps, with yield sensitivity to nanometer-scale particulate contamination.
Replacing this skilled labor isn’t trivial. While Samsung has invested in AI-guided defect detection and EUV lithography automation, the final 30% of HBM assembly — including die stacking, TSV formation, and thermal interface material application — still relies on human intervention for micron-level alignment. A single misaligned die can render an $800 HBM3E stack useless, creating a bottleneck that even advanced robotics struggle to fully eliminate.
“You can automate lithography, but you can’t automate the tactile feedback needed to sense when a die is seating correctly in a 2.5D package. That’s still human intuition, and it’s undervalued.”
Ecosystem Bridging: The AI Profit Concentration Problem
Samsung’s labor action highlights a growing fault line in the AI economy: value capture is increasingly detached from physical production. NVIDIA reports a 78% gross margin on its H100 GPUs, while Samsung’s HBM3E — which constitutes 40% of the H100’s bill of materials — carries a gross margin of just 32%. This disparity mirrors the smartphone era, where Qualcomm’s Snapdragon SoCs earned 50%+ margins while contract manufacturers operated at single-digit profitability.
The implications extend beyond wages. If Samsung concedes to profit-sharing demands, it could trigger a wage-price spiral across the Asian semiconductor supply chain, pressuring TSMC, SK Hynix, and Micron to revisit labor contracts. Conversely, a hardline stance risks accelerating unionization efforts at Samsung’s Austin and Xi’an fabs, potentially disrupting just-in-time delivery to AI hyperscalers.
From an open-source perspective, the unrest underscores tensions in proprietary hardware ecosystems. Unlike RISC-V, where IP licensing models encourage collaborative innovation, Samsung’s HBM designs remain locked behind NDAs and patent thickets, limiting third-party verification of manufacturing claims. This opacity fuels worker skepticism when executives cite “market conditions” to justify wage stagnation.
What This Means for the AI Hardware Race
Samsung’s ability to scale HBM4 production — critical for next-gen AI accelerators like AMD’s MI400 and Intel’s Falcon Shores — hinges on resolving this labor conflict. HBM4 targets 1.6TB/s per stack using 12-layer stacking and advanced MR-MUF (Mass Reflow Molded Underfill), processes that demand even greater precision than HBM3E. Any disruption in yield or throughput could delay AI chip rollouts by hyperscalers, indirectly benefiting competitors like Micron, which has begun piloting HBM4 with a profit-sharing pilot program for its Boise fab workers.
the Samsung strikes reveal a fundamental tension in the AI era: as software-defined value explodes, the physical labor enabling it risks becoming invisible — unless workers refuse to be.
