DIGITIMES Biz Focus

As the AI computing race accelerates and semiconductor manufacturing capacity is increasingly redirected toward advanced memory, the memory industry is shifting from price-driven competition to a battle over resource allocation and supply security. The rapid expansion of high-bandwidth memory (HBM) production continues to consume advanced fabrication capacity, constraining the supply of conventional memory products and creating persistent shortages across the market. This concentration of high-end production has widened the supply-demand gap, driving the average selling prices of certain premium memory products to levels eight to ten times higher than previous norms

Rugged edge computing brand – Cincoze will showcase its latest embedded computing solutions at COMPUTEX Taipei 2026, from June 2–5 (Hall 2, Booth Q0401a). Cincoze will display its full range of edge computing solutions around the theme of "Smarter, Faster, More Reliable - Powered by Edge AI" encompassing the full spectrum of edge application environments in four dedicated zones, including DIN-Rail computers, industrial panel PCs and monitors, rugged embedded computers, and GPU embedded computers

Marine debris management is entering a new phase of data-driven applications. Converting years of accumulated coastal imagery into actionable tools for surveying, identification, and monitoring has emerged as a critical challenge in the digitalization of ocean governance.Under the guidance of Taiwan's Ocean Affairs Council (OAC), the National Academy of Marine Research (NAMR) is hosting the "2026 International Marine Debris Image Recognition AI Challenge." Featuring a dataset of over 20,000 real-world marine debris images, the competition invites AI, data science, computer vision, and marine science teams from Taiwan and abroad to participate.The competition is supported by Amazon Web Services (AWS) as the AI technology partner, with model evaluation and competition operations managed through the Industrial Technology Research Institute's (ITRI) AIdea AI Co-Creation Platform. Registration is now open.NAMR sets the challenge: bringing AI to the frontlines of ocean governanceMarine debris has long been a fundamental issue in coastal environmental governance - and one of the most difficult to address in the field. Coastal debris is diverse in type, scattered in distribution, and frequently degraded by sun exposure, seawater erosion, sand burial, and physical damage, making manual surveys and image interpretation highly labor- and time-intensive.To accelerate digital transformation, NAMR has established MDImageNet, an AI-Ready marine debris image dataset covering the ICC19+1, NAMR26+1, and NAMR33+1 marine debris category schemes. The competition draws on NAMR's existing marine debris image dataset, comprising over 20,000 images annotated with YOLO-format bounding boxes. The dataset covers common coastal waste categories including plastic litter, fishing-related debris, and other anthropogenic waste.A "post-mapping" strategy is adopted: participants first train models using the original class labels provided, then map predictions into 20 official recognition categories during inference, with final scoring based on 19+1 primary marine debris target classes.The competition design confronts teams with the real-world constraints of field data - cluttered backgrounds, diverse object classes, and significant appearance variations. By requiring quantitatively evaluated object detection models, the challenge goes beyond open data sharing: it validates whether AI can be transformed into deployable tools built upon existing survey infrastructure.The competition comprises preliminary and final rounds, evaluated primarily on mean Average Precision (mAP) at an IoU threshold of 0.5. A "Best Lightweight Optimization Award" is also offered, using NetScore to balance detection accuracy, model size, and computational complexity-encouraging models suitable for practical deployment in coastal patrols, UAV-based image analysis, and long-term environmental monitoring.AWS cloud resources and AIdea platform power hands-on AI developmentThe competition integrates resources from AWS and ITRI's AIdea AI Co-Creation Platform to support teams throughout model development, training, testing, and evaluation. AWS provides the cloud development environment, offering Amazon SageMaker AI and computing resources for machine learning workflows. Technical workshops are also scheduled to familiarize participants with cloud-based AI development tools and model training pipelines.ITRI's AIdea Platform handles competition execution, dataset support, and automated scoring. The standardized cloud environment and evaluation framework ensures all teams operate under identical conditions, maintaining fairness and reproducibility.The competition is open to high school students, college students, and professionals, with teams of two to five members. Cross-institutional and interdisciplinary collaboration is encouraged, combining expertise in AI, computer vision, data science, and marine science to produce evaluated model outcomes for marine debris recognition.Registration for the "2026 International Marine Debris Image Recognition AI Challenge" is open until August 10. The total prize pool is NT$300,000, with two additional Best Lightweight Optimization Awards.NAMR aims to bring together the AI community, research institutions, academia, and industry to convert marine debris imagery into deployable environmental monitoring models-building Taiwan's AI application experience and marine data foundation for ocean governance.For competition details and registration, visit the official website (link).

Six months into 2026, the story isn't whether the electronics supply chain has stabilized. It has, broadly. The real story is that it has split into two distinct markets moving in opposite directions, and most procurement teams are still planning as if there's just one. The AI-driven leading edge remains capacity-constrained and pricing-positive, while the mature-node segment is loosening into a genuine buyer's market. Knowing which side of that line each line item on your BOM sits on is the single most useful thing you can do heading into the second half of the year.The Two-Speed Market, BrieflyTSMC is reportedly preparing to raise prices 3–10% on its sub-5nm offerings, and with Nvidia and Apple having already locked in large blocks of capacity through year-end, second-tier buyers are increasingly competing for allocation that may not exist in H2. Memory tells a similar story. Combined output from Samsung, SK Hynix, and Micron is expected to grow sharply by 2030, led by a projected rise in HBM production, but that new capacity doesn't meaningfully arrive until 2027. In the meantime, DRAM and HBM remain the tightest categories in the entire component ecosystem.Mature-node wafer pricing has actually returned to pre-pandemic levels, down 5–8% year-over-year, as Chinese fab capacity comes online and automotive/industrial utilization climbs into the 80–85% range. Consumer electronics demand is stabilizing too, which is easing pressure on mature-node semiconductors and passives.What This Actually Means for Your Procurement StrategyThis is where most H2 outlooks stop short. Here's what to actually do with this picture, category by category.1. Segment your BOM by risk profile, not just by part number.Advanced logic tied to AI-adjacent applications faces a fundamentally different supply reality than commodity discrete or mature-node passives. Treat these as two separate procurement strategies, not one blended approach. A BOM review that groups parts by constrained, stable, or loosening rather than by function or supplier will surface where your actual exposure sits, and it's often not where teams assume.2. Use the mature-node buyer's market now, not later.If you were forced into single-sourcing during the 2021–2023 shortage years, H2 2026 is the window to qualify second sources for those mature-node components while pricing and availability both favor you. This window won't stay open indefinitely. As automotive and industrial utilization keeps climbing toward capacity, the leverage shifts back to suppliers.3. Be strategic, not reactive, on memory buys.Memory is not one uniform story. Leading-edge densities tied to HBM demand carry a real price premium, but previous-generation DDR4 and LPDDR4 may still offer value while pricing is structurally supported mainly at the high end. If your designs can tolerate a prior-generation module, this is the year to lock it in rather than wait and hope for relief that isn't coming until 2027.4. Watch for new procurement categories forming in real time.Optical networking components for AI data centres are moving from niche to mainstream as AI clusters push toward much higher bandwidth per rack. If your roadmap touches high-bandwidth AI infrastructure at all, get ahead of this now. Procurement categories that don't exist yet in your sourcing playbook have a way of becoming urgent overnight once a design win locks them in.5. Rebuild safety stock around true risk, not blanket buffers.Broadly increasing inventory across the board is expensive and imprecise. The more effective move is targeting buffers at the small percentage of components, often a single connector, capacitor, or legacy memory module, that actually drive downtime risk if they disappear. Combine that with forecasting discipline: suppliers increasingly prioritize allocation and pricing based on how credible and consistent your rolling forecasts are, which means forecast accuracy is now a negotiating asset, not just a planning exercise.6. Don't sleep on non-obvious demand drivers.Rising defense spending across Asia-Pacific is quietly adding a new, non-cyclical source of demand for industrial and mature-node electronics, one that doesn't show up in most consumer-electronics-driven forecasts. If your end markets touch defense, aerospace, or industrial automation, factor this in as upside demand pressure, not background noise.The Bottom Line for the Rest of 2026H2 2026 doesn't call for a single supply chain strategy. It calls for two, running in parallel. Where you're constrained, the priorities are allocation planning, forecast credibility, and locking in what capacity you can. Where you're not, the priority is using the current leverage to diversify, qualify alternates, and rebuild resilience before the window closes. Teams that treat this as one undifferentiated “stabilizing market” will miss both opportunities.If you're navigating either side of this, securing allocation on constrained parts or taking advantage of loosening mature-node availability, that's exactly the kind of sourcing challenge our team works through with customers every day. Reach out to your Fusion Worldwide representative and let's talk through your BOM.(Article Sponsored by Howard Tan, Director of Purchasing, China Fusion Worldwide)