llms.blog
SK hynix

SK hynix commits $39 billion to two new fabs as AI memory demand reshapes its roadmap

SK hynix's board has approved 54 trillion won (about $39 billion) to build two new memory fabrication plants, the company announced on August 7, 2026, turning the first two arms of a much larger investment plan into funded construction schedules. The larger project, Y2, is the second fab at the Yongin Semiconductor Cluster. At 35.2 trillion won it is a DRAM plant of roughly 1.13 million square meters, with ground breaking set for July 2027 and its first cleanroom opening in June 2029 to produce

axc
axc
2 min
SK hynix commits $39 billion to two new fabs as AI memory demand reshapes its roadmap

SK hynix's board has approved 54 trillion won (about $39 billion) to build two new memory fabrication plants, the company announced on August 7, 2026, turning the first two arms of a much larger investment plan into funded construction schedules.

The larger project, Y2, is the second fab at the Yongin Semiconductor Cluster. At 35.2 trillion won it is a DRAM plant of roughly 1.13 million square meters, with ground breaking set for July 2027 and its first cleanroom opening in June 2029 to produce high-bandwidth memory and next-generation DRAM.

The second, M17, is a new NAND fab in Cheongju at 19.1 trillion won. It breaks ground earlier, in February 2027, and opens its first cleanroom in December 2028. Cheongju already houses three SK hynix fabs, so M17 plugs into existing power and water infrastructure for the fastest possible timeline.

Infographic of the two fabs: Y2 DRAM at 35.2 trillion won and M17 NAND at 19.1 trillion won

A 1,100 trillion won master plan

Both sit inside the 1,100 trillion won framework SK hynix laid out in June 2026, split across Yongin (600 trillion won), Cheongju (100 trillion won), and a future southwestern hub (400 trillion won). The approvals move Yongin and Cheongju from announced intentions to construction.

Betting on structural demand

The company is betting that memory demand is structural rather than a passing supercycle. Research firm Omdia projects DRAM and NAND demand both growing at a 19 percent compound annual rate from 2025 through 2030. SK hynix says supplying the required volume at the moment customers need it, not just leading on technology, is now the competitive advantage.

The demand is specific to AI. HBM sits at the center of accelerator memory, while enterprise SSDs and key-value cache storage for inference workloads are pulling NAND into AI infrastructure in a way that barely existed two years ago.

Notably, SK hynix is decoupling the building shells from the equipment inside them. Fabs go up on the master schedule, but cleanroom fit-out and tool installation happen in step with customer orders, letting the company commit construction capital now while metering the far larger equipment spend against actual demand.

Sources

SK hynix Approves Two New Fabs as AI Memory Demand Reshapes Its Buildout - Unite.AI: https://www.unite.ai/sk-hynix-approves-two-new-fabs-as-ai-memory-demand-reshapes-its-buildout/

SK hynix fab facility investment announcement (Aug 7, 2026): https://news.skhynix.com/en/fab-facility-investment-2026/

SK hynixSemiconductorsAI Infrastructure
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Embedding Model Fine-Tuning in Production: Hard Negative Mining, Synthetic Data Pipelines, and Contrastive Distillation

    Production retrieval-augmented generation (RAG) and semantic search architectures frequently suffer from domain mismatch when relying on general-purpose embedding models. Off-the-shelf bi-encoders trained on broad web corpora often experience a 15% to 30% degradation in retrieval metrics such as NDCG@10 and MRR@10 when deployed on specialized enterprise corpora, including proprietary codebases, internal API schemas, clinical trials, and technical documentation. While downstream cross-encoder re

    1 min

  • The Gumbel-Softmax Trick: How Continuous Relaxations Enable Differentiable Discrete Sampling

    The Gumbel-Softmax Trick: How Continuous Relaxations Enable Differentiable Discrete Sampling In modern deep learning, end-to-end training depends on reverse-mode automatic differentiation. When an architecture operates on continuous tensors, computing gradients via the chain rule is straightforward. However, many foundational artificial intelligence problems involve discrete choices: selecting tokens from a fixed vocabulary, routing tokens to expert networks in a Mixture-of-Experts (MoE) archit

    1 min

  • Token Healing and Partial Token Alignment in Production LLM Serving: Architecture, Prefix Trie Rollback, and Serving Trade-Offs

    Modern large language models operate on discrete subword tokens generated by greedy compression algorithms like Byte-Pair Encoding (BPE), WordPiece, or Unigram. While subword tokenization enables high compression rates and fixed vocabulary sizes, it introduces a structural defect at the interface between raw user text and autoregressive inference: the partial token problem, commonly known as the prompt boundary problem. When a user prompt terminates mid-token or at a punctuation boundary that c

    1 min