Samsung Sold Out Every Gigabit of Its 2026 HBM4 Memory Chips

Samsung Electronics Co. has sold out its entire 2026 supply of HBM4 memory chips, with customer demand concentrated on the product, according to Chief Financial Officer Park Soon-cheol on the company's April 30 earnings call. Samsung's HBM4 revenue surpassed $1 billion four months after mass production began in February, and TrendForce raised its 2026 shipment outlook from 3.5 billion to about 4 billion gigabits. The shortage highlights that memory, not just GPUs, is a critical bottleneck for AI hardware, with Nvidia Corp.'s Rubin GPU and Advanced Micro Devices Inc.'s Instinct MI455X requiring up to 288GB and 432GB of HBM4 respectively.

Samsung's 2026 HBM4 supply is already spoken for, and that tells you where the real bottleneck in AI hardware now sits: not just in GPUs, but in the memory feeding them. Samsung's own finance chief put the shortage on the record during the company's April 30 earnings call. Park Soon-cheol, Samsung Electronics' chief financial officer and vice president, said customer demand had concentrated on HBM4 and that the sales volume was already completely sold out, according to Seoul Economic Daily. That's the line buyers should sit with. You can't build a Rubin server or an MI400 rack from a forecast. The sales pace backs him up. Yonhap reported on June 23 that Samsung's HBM4 revenue had crossed $1 billion four months after mass production and shipments began in February, with industry sources expecting the figure to pass $1.2 billion by the end of June. TrendForce also said Samsung's 2026 HBM4 shipment outlook had been raised from 3.5 billion gigabits to about 4 billion. Short supply is not a theory here. It's the operating condition. Samsung has needed this turn. SK Hynix led the HBM race through the HBM3E cycle. Nvidia's hunger for memory made that lead extremely valuable. But Samsung moved early on HBM4: sixth-generation 10-nanometer-class DRAM, a 4-nanometer logic base die. Its February announcement put the chip's steady transfer speed at 11.7 gigabits per second, with capability up to 13Gbps. That speed matters. The next AI accelerators are memory-hungry in a way older data center chips weren't. Nvidia says its Rubin GPU carries up to 288GB of HBM4 and delivers up to 22 terabytes per second of memory bandwidth. AMD's Instinct MI455X goes higher still on capacity: 432GB of HBM4, up to 23.3 terabytes per second, according to AMD's own product data. That's a lot of stacked memory. The race has moved from qualification to allocation Nvidia said on May 31 that Vera Rubin was ramping into full production, with Taiwan's major server makers and supply-chain partners building systems at scale. Once that happens, memory qualification becomes less interesting than allocation. The supplier with real volume gets the order. The one still ramping gets a later slot. TrendForce's June 23 HBM market bulletin described the split clearly: Samsung completed HBM4 validation first and began shipments in the second quarter, while SK Hynix faced interface synchronization delays that pushed meaningful mass production into the third quarter. TrendForce said SK Hynix also redirected some capacity toward DDR5 and LPDDR5, where pricing had become more attractive. That is the part investors sometimes miss. HBM is the glamour product, but wafer economics still matter. TrendForce said tight conventional DRAM supply had pushed some server memory returns above HBM returns in early 2026, which gives suppliers a reason to be careful about how fast they move capacity. Frankly, if you are waiting for a quick flood of cheap HBM4, you're waiting for the wrong industry. Micron isn't standing still either. In March, the company said it had started volume shipments of a 36GB 12-layer HBM4 product designed for Nvidia Vera Rubin, with more than 2.8 terabytes per second of bandwidth and more than 20% better power efficiency than its HBM3E. In July, Micron raised its planned U.S. fab and technology investments to more than $250 billion through 2035. That tells you the shortage is not a one-quarter problem. Prices now have room to run Buyers care about delivery, but they should care about price just as much. TrendForce said negotiations had already shifted toward 2027 HBM4 supply agreements, and its June pricing work pointed to stronger supplier power as DRAM remains tight. MarketWatch reported this week that UBS analyst Timothy Arcuri expects stronger-than-expected pricing for HBM4 and HBM4E, with average selling prices rising sharply in 2027. The market has noticed. Barron's reported on Thursday that Micron shares were up more than 200% for the year, even after a pullback from their late-June peak. Samsung has had its own moment: Yonhap reported that its combined market capitalization, including preferred shares, passed 2,000 trillion won on May 29 after the company began shipping HBM4E samples to major customers. You don't need to dress this up. AI companies can talk about model roadmaps, inference demand and rack-scale systems, but the buildout still comes down to physical parts. HBM4 is one of them, and Samsung has already sold the year's supply. Also read: Trump Administration Turns AI Loose on China's Tariff Evasion Networks https://startupfortune.com/trump-administration-turns-ai-loose-on-chinas-tariff-evasion-networks/ • IBM Stakes $240 Million on Renting Out AI Inference, Not Training It https://startupfortune.com/ibm-stakes-240-million-on-renting-out-ai-inference-not-training-it/ • DeepSeek Ships V4 Pro, Rivaling Claude and GPT-5 for a Fraction of the Cost https://startupfortune.com/deepseek-ships-v4-pro-rivaling-claude-and-gpt-5-for-a-fraction-of-the-cost/