Korean memory suppliers lock in AI giants with long-term HBM deals amid supply squeeze

Korea’s high-bandwidth memory market faces a surge in prices and supply constraints as major suppliers secure multi-year deals with AI firms, intensifying the shortage and elevating costs across the tech industry.

Artificial intelligence groups buying memory years in advance are pushing Korea’s high-bandwidth memory market into a more rigid and expensive phase, with July export data showing the average HBM unit price above $70 for the first time and spot prices for top-end parts far above contract levels. Seoul Economic Daily reported that a 36GB HBM3E stack was changing hands at about $2,100 on the spot market, while TrendForce said Samsung had reportedly committed about 70% of its memory capacity through 2031 to long-term agreements with customers including Nvidia, Microsoft and Google. (en.sedaily.com)

The underlying trade figures point less to a short-lived spike than to a supply squeeze. According to the Korea International Trade Association data cited by Seoul Economic Daily, the average export price per HBM unit reached $76.13 at the end of July, up 9.5% from June, even as shipment volume fell 27.3% to 132.47 million units and export value slipped to $10.09 billion. TrendForce added that exports of DRAM used in AI chips, including HBM and LPDDR, fell 13.2% between May and July, yet export value still rose 18.5% to $13.55 billion and the average unit price climbed 36.6% to $22.90. In other words, fewer chips are leaving Korea, but each unit is worth markedly more. (en.sedaily.com)

Part of the explanation is technical. Ahn Ki-hyun of the Korea Semiconductor Industry Association told Seoul Economic Daily that “HBM has a characteristic where yields decline with each upgrade to a newer version”, while HBM4 is priced at roughly twice the previous generation. Samsung’s own interim report helps explain why that matters: the company said it began mass production and shipment of HBM4 in February 2026, using 1c DRAM and a 4nm base die, with capacities ranging from 24GB to 48GB and speeds of 11.7Gbps to 13Gbps. The same filing says Samsung’s 36GB HBM3E 12-high part, first developed in February 2024, achieved that capacity by stacking 12 layers while keeping the height of an eight-layer HBM3 package. More layers and tighter yields mean more conventional DRAM is consumed by each premium module. (en.sedaily.com)

The contracting model itself is also changing. French consumer-tech site Frandroid said multi-year supply agreements of this kind were historically more often limited to one or two years, making five-year visibility unusually long for the memory market. TrendForce said the rise of long-term agreements is tightening the free market by locking more output into named customers, and added that inventories are being reshaped around future commitments rather than unwanted stock: Samsung’s semiconductor inventories rose 32% in the first half of 2026 to 38.06 trillion won, while SK hynix’s total inventories rose 26% to 17.99 trillion won. TrendForce also said Nvidia’s long-term purchase commitments rose to $279 billion in the second quarter of its 2027 fiscal year from $119 billion in the first, with the latest commitments described as primarily related to memory procurement. (frandroid.com)

Manufacturers are trying to respond, but the options are slow and capital-intensive. Seoul Economic Daily reported that Samsung executives are weighing whether to convert the S5 foundry line at Pyeongtaek into memory production equipment as early as 2027, while SK hynix is reviewing a possible joint venture plant in Japan and deeper co-operation with Kioxia. Wccftech’s English-language account of the Korean report put the local pricing gap even more starkly: roughly 2.87 million won for a spot-market 36GB HBM3E module against 500,000 to 700,000 won under long-term agreements, with a 16-layer HBM4 part at about 4.8 million won. Those numbers help show why late buyers face a severe penalty. (en.sedaily.com)

SK hynix’s own market outlook suggests suppliers do not expect this imbalance to disappear quickly. The company said it was strengthening packaging ties with TSMC, building the Cheongju M15X fab and setting up a dedicated HBM organisation and a Global AI Research Center to prepare for rising demand. In the same note, it cited UBS as expecting SK hynix to take about 70% of the HBM4 market for Nvidia’s Rubin platform in 2026, and said focused investment in HBM was tightening the broader DRAM balance by pulling resources away from general-purpose memory. Even in that bullish assessment, the company acknowledged that some researchers expect HBM prices to correct after 2026 as competition and production expand. (news.skhynix.com)

That strain is already spreading beyond AI accelerators. Samsung said in its interim report that higher component costs could weigh on PC and mobile demand in the second half, and that its System LSI business was dealing with pricing pressure from device makers because of rising memory costs. Frandroid made the consumer version of the same argument more bluntly: if about 70% of premium memory output is reserved for AI customers, less remains for smartphones, PCs, consoles and other mainstream electronics. The immediate winners are the suppliers with scarce HBM output; the likely losers are smaller AI firms and consumer-device makers that lack guaranteed access. (images.samsung.com)

What matters now is not simply that spot prices are high, but that the market is still being repriced through contracts and capacity plans rather than clearing in the usual way. Seoul Economic Daily said even some large technology groups were not receiving all the volumes they had contracted, and TrendForce said the divergence between contract and spot prices reflected supply tied up in long-term deals rather than a market already back in balance. As long as HBM4 ramps continue to absorb more DRAM, and as long as the largest buyers keep securing years of output ahead of time, memory is likely to remain one of the clearest choke points in the AI supply chain. (en.sedaily.com)