Korea Cycle Data

Samsung and SK hynix hire thousands of engineers a year, and the questions they ask are a decent window into what they care about operationally. Yield, tool uptime and data come up again and again. This post maps ten recurring interview questions to Samsung's published evaluation criteria and to reported 2026 interview accounts, then shows how a strong answer is built. KEY TAKEAWAYS 1. Samsung's careers site splits the final round into three interviews: executive (character and fit), job competency (knowledge of major and motivation) and creativity (original ideas and reasoning). 2. SK hynix's 2026 first-half job interviews reportedly put 1–2 candidates in front of 3–4 panelists, drilling into the written application. 3. Eight of the ten questions here test technical concepts or hands-on experience. Under an illustrative uptime model, the classic PM-vs-BM question has a numeric answer: 94.3% to 99.2%. Author's classification, based on Samsung's published c...

Samsung is putting more than 50,000 NVIDIA GPUs into its chip manufacturing flow. The US Bureau of Labor Statistics still projects semiconductor technician jobs to grow 8% over the next decade. Both are true, and the gap between them explains what AI actually replaces in a fab: tasks, not jobs. KEY TAKEAWAYS 1. Wafer transport is already fully automated: about 1,850 overhead hoist vehicles ran in Samsung's Pyeongtaek P1 fab as of 2022. The AI question is about judgment, not muscle. 2. AI is taking the front end of judgment. SK hynix-backed Gauss Labs reports about 29% lower process variation from virtual metrology in deposition. 3. Headcount is still rising: BLS projects +2,600 US process technicians (+8%) for 2025–2035, and SK hynix grew from 32,390 to 34,549 employees in 2025. 2027 is McKinsey's projected demand peak. Values are read from the published chart and are approximate. What Samsung actually announced On October 31, 2025, Samsung announced an "AI Me...

Chip fabs ban makeup, and the reason is more specific than "keep it clean." A 1984 IBM study counted roughly 3 billion particles of 0.5 µm or larger from a single coat of mascara, and found 15 elements in common cosmetics, including sodium and potassium. One is a physical threat to wiring, the other an electrical threat to transistors. KEY TAKEAWAYS 1. One mascara application released about 3 billion particles ≥0.5 µm; a full face, about 5.1 billion (Phillips et al., IES 1984). An ISO 3 cleanroom allows 35 such particles per cubic meter of air. 2. The industry roadmap treats a particle half the size of the circuit half pitch as a killer defect. By 2016, IRDS noted that killer size was already approaching 10 nm. 3. A single particle usually kills a single die. The expensive event is a contamination excursion: under illustrative assumptions, a yield drop from 91% to 64% across 20 lots is about $1.35 million. Different bases (one application vs. one cubic meter of air)...

Open a Korean chipmaker's job posting for the first time and the job titles are where most people stall: process technology, process design, equipment technology, manufacturing technology, device, PI. They sound alike. The work differs a lot. Worse, the same word means different things at different companies. This piece uses only the job descriptions Samsung Electronics and SK hynix have published to separate the three core engineering roles, and then asks why the distinction matters to investors. Three points up front • Samsung's DS memory business lists nine job families; process design, process technology and equipment technology correspond to what are usually called device, process and equipment roles. • The Korean word "seolbi" means equipment at Samsung but fab infrastructure (buildings, power, gases) at SK hynix. SK hynix's equipment engineers sit inside "manufacturing technology". • The US Semiconductor Industry Association estimates ...

A long memory upcycle has brought a wave of hiring news from Korea's chipmakers. One number makes it concrete: SK hynix hired 739 new graduates in 2023 and 3,201 in 2025. This piece maps the path into Korea's semiconductor industry for 2027 using only published hiring data and the government's workforce outlook: how many are hired, where they come from, and which roles each academic major leads to. Three points up front • SK hynix's new-graduate hiring went from 739 in 2023 to 3,201 in 2025, roughly 4.3 times in two years. • An industry outlook cited by the government puts Korea's chip workforce at 177,000 in 2021 and the requirement at 304,000 by 2031, an increase of 127,000. • In SK hynix's published list of majors by job family, electrical and electronic engineering appears in eight of nine families. What is confirmed and what is not Confirmed: SK hynix new hires of 739, 942 and 3,201 for 2023 to 2025 (Financial News, July 30, 2026). The fu...

Talk about memory prices or capital spending long enough and the word "fab" comes up every time. What actually happens inside one over a day is less widely known. This piece follows a semiconductor fab's day using only what companies have published in their newsrooms and press tours, plus the international cleanroom standard: how people get in, how wafers move, and why the factory never stops. Three points up front • An ISO Class 3 cleanroom allows at most 35 particles of 0.5 micrometres or larger per cubic metre, about one millionth of typical city air. • According to Samsung Semiconductor's newsroom, the overhead hoist transport (OHT) vehicles in a single fab travel about 100 km a day. • A wafer spends roughly 40 to 100 days inside a fab depending on the process, which is why fabs run around the clock on rotating shifts. What is confirmed and what is not Confirmed: ISO 14644-1:2015 sets the Class 3 limit at 35 particles (≥0.5µm) per m³. The old US...

If you planned to replace a laptop this winter, the question is probably whether to buy now or wait for 2027. Korea's consumer price index offers a blunt starting point: in September 2026 the "computers" item was 27.4% higher than a year earlier. This piece answers the timing question with numbers rather than instinct. A laptop's price ultimately follows the price of the memory inside it, so the path of memory contract prices tells you most of what you need to know. Three points up front • Korea's computer CPI rose 27.4% year on year in September, more than nine times the 2.9% headline rate. • Memory's share of PC cost roughly doubled in a year: 16% to 23% by Gartner's estimate, and 35% by HP's own account. • Forecasts for when memory prices turn range from mid-2027 to 2028, and NAND (SSDs) is likely to ease before DRAM. What is confirmed and what is not Confirmed: September computer CPI +27.4% and headline CPI +2.9% (Statistics Kore...

Link two DGX Spark 64GB boxes with a single cable and you get twice the memory and twice the memory bandwidth. NVIDIA says you get up to 1.7x the performance. The missing 0.3x is not a marketing rounding error. It is the cost of moving data between two machines, and it is the clearest way to understand what actually limits local AI hardware. KEY TAKEAWAYS 1. Two clustered 64GB units double capacity (64 GB to 128 GB) and memory bandwidth (273 GB/s to 546 GB/s combined). The 1.7x performance figure is the maximum from NVIDIA's own Qwen 3.8 27B test, with no test conditions disclosed. 2. The bottleneck moves to the wire. Each unit's memory runs at 273 GB/s; the ConnectX-7 link between them is 200 Gb/s, or about 25 GB/s, roughly one-eleventh as fast. 3. There are two ways to buy 128 GB: two 64GB units at $9,998 in list prices, or one 128GB unit at a press-reported $6,950. One buys a single memory pool; the other buys double the bandwidth plus a communication tax. What NVIDI...

A purchase order for generators is not a running data center. On October 1, Enerflex said it will build roughly 450 MW of behind-the-meter gas generation for a North American data center developer, with deliveries in 2027 and 2028. That is real demand for power equipment. It is not 450 MW of AI compute, and it is not online yet. This post breaks the deal into the steps between an equipment order and GPUs drawing power, and uses public filings to show where the time goes. KEY TAKEAWAYS 1. Enerflex was hired to design, engineer, fabricate and assemble about 450 MW of gas-fired generating units. Deliveries are scheduled to start in 2027 and finish in 2028. No start-up date, customer name or contract value was disclosed. 2. Nameplate generation is not IT load. With an assumed 15% reserve and a PUE of 1.3, 450 MW supports roughly 294 MW of servers. That is an illustrative calculation, not a company figure. 3. Large turbines are scarce. GE Vernova reported 116 GW of gas equipment in b...

The AI capex debate is usually about how many GPUs get bought. The quieter question is how many hours those GPUs actually serve traffic. On October 1, Nebius bought a startup whose entire pitch is shrinking the dead time between "a request arrives" and "the model is ready to answer." That dead time is the cold start, and it is a bigger cost line than it looks. KEY TAKEAWAYS 1. Loading a model is slow relative to serving it. In an OSDI 2024 experiment, loading LLaMA-2-70B onto 8 GPUs took 84 seconds; generating one token typically takes under 0.1 seconds. 2. Slow loading forces operators to keep GPUs warm for peak demand. In a simple illustrative model, scaling with demand serves the same day with 32% fewer GPU-hours. 3. Nebius spent about $5.7bn on capex in Q2 2026, roughly 10x its $0.58bn quarterly revenue. Inferize is its third inference-stack addition this year, aimed at getting more work out of that hardware. What Nebius actually announced According to ...