SK Hynix Posts 72% Margin Record; Meta Tracks Keystrokes for AI Agents

SK Hynix posts a 72% margin on AI memory demand as Meta starts harvesting employee keystrokes for AI agent training.

🗄️ SK Hynix Posts All-Time-High 72% Operating Margin on HBM Demand

Decoded: South Korean memory chipmaker SK Hynix reported Q1 2026 revenue of 52.58 trillion won ($35.55 billion) — the first quarter in company history to exceed 50 trillion won — and an operating margin of 72%, an all-time record. Operating profit rose fivefold year-over-year and nearly doubled sequentially from Q4 2025. SK Hynix is the world's leading supplier of high-bandwidth memory (HBM) chips required by AI accelerators, and attributed the record to rising memory prices and accelerating AI infrastructure investment. "Despite the fact that first quarter is typically a seasonal downturn, strong demand persisted due to expanded investments in AI infrastructure," the company said in its earnings release. An SK Hynix executive added that "customers are prioritizing procurement over price" — a signal of critical allocation pressure at the HBM layer. Results were broadly in line with analyst estimates; revenue came in slightly below the 53.55 trillion won consensus. (CNBC, April 23, 2026)

Why it matters: A 72% operating margin in a traditionally weak seasonal quarter confirms that AI infrastructure buildout is overriding cyclical memory patterns at the supply layer that matters most. SK Hynix holds the largest share of global HBM production — the memory stack required in every Nvidia H200 and Blackwell GPU. Its Q1 actuals quantify the pricing power that a structural demand-supply imbalance generates: fivefold year-over-year profit growth even before seasonally stronger Q2 demand arrives. For Micron (MU), the results validate the HBM pricing environment ahead of its own earnings. For Nvidia (NVDA), the "procurement over price" commentary signals continued constraints on Blackwell shipment timelines through mid-2026 as customers accept higher memory costs to secure allocation.

🤖 Meta Deploys Keystroke and Screen-Capture Software on Employee Computers to Train AI Agents

Decoded: Meta is installing software called Model Capability Initiative (MCI) on U.S.-based employees' computers to capture mouse movements, clicks, keystrokes, and periodic screen snapshots for use in AI model training, Reuters reported April 21, citing internal memos. The initiative is part of the company's Agent Transformation Accelerator (ATA) program, announced by CTO Andrew Bosworth. The goal is to gather behavioral data where AI models underperform humans — tasks like selecting from dropdown menus and using keyboard shortcuts. Bosworth wrote that the vision is for AI agents to "automatically see where they can contribute" to work tasks while employees focus on directing, reviewing, and improving them. Meta said the data will not be used for employee performance reviews. The tool runs on work-related apps and websites. (Reuters, April 21, 2026)

Why it matters: Meta's MCI converts its workforce into a continuous AI training dataset at scale. The company employs roughly 76,000 people globally; real behavioral telemetry from workers performing live software tasks generates computer-use training data at a volume and authenticity that cannot be replicated synthetically or through contractor labeling. The timing compounds the signal: the MCI rollout arrives alongside Meta's planned 10% workforce reduction, creating a dynamic in which the company simultaneously cuts headcount while using remaining employees to accelerate the AI agent capabilities intended to replace them. For investors, MCI is evidence that Meta SuperIntelligence Labs is treating enterprise-grade computer-use agents as a near-term product priority — the same category where Anthropic Claude and OpenAI Operator are already commercially deployed. If proprietary behavioral training data produces agents that outperform publicly available alternatives, the capex-efficiency case for META strengthens considerably.

Stay decoded. See you tomorrow.

— The Get AI Decoded Team