#HBMShortageBoostsAIChipPrices

The AI hardware story of 2026 can be told as one simple chain. An HBM shortage leads to HBM price increases, HBM price increases push up the cost of building AI chips, and higher AI chip costs are now showing up as higher AI chip prices. Every link in that chain is visible in the hard numbers this year, and it matters to two very different groups of companies, the AI chip vendors who have to buy memory, and the memory manufacturers who sell it.

HBM stands for High Bandwidth Memory. It is a stack of DRAM dies placed vertically on a base die, connected with tiny vertical channels, and packaged extremely close to the GPU die so data can move at enormous speeds. A modern AI accelerator cannot function without it, because large models are limited mainly by how fast memory can feed the compute units. This is also why HBM is so expensive to produce. Each gigabyte of HBM consumes roughly three times the wafer capacity of a gigabyte of DDR5, memory makers earn about three to five times more revenue per wafer from HBM than from conventional DRAM, and new capacity needs 12 to 18 months of lead time before it produces anything.

Supply is concentrated in just three companies, SK Hynix, Samsung and Micron. Together they control more than 95 percent of global DRAM production and they are the only three HBM suppliers in the world. That concentration is the engine of the shortage, because all three have redirected the majority of their advanced capacity toward HBM and server memory, with more than 80 percent of advanced capacity now going to those products. HBM took about 18 percent of total DRAM wafer input in 2025, roughly 22 percent in 2026, and is expected to approach 30 percent in 2027. In bit terms HBM accounts for only 8 to 9 percent of DRAM supply now and about 13 percent in 2027, which sounds modest until you remember how much larger HBM dies are than ordinary DRAM. TrendForce puts the 2026 supply demand gaps at 4.9 percent for DRAM, 5.1 percent for HBM and 4.2 percent for NAND, the widest since 2011.

On top of that, the capacity is already sold. SK Hynix said its entire 2026 memory capacity was sold out before the year began. Micron has sold out its HBM supply for 2026 and its 2027 output, and Samsung pre sold its 2026 HBM capacity as well. By mid 2026, reports indicated all three suppliers had sold through their HBM capacity into 2027.

The price response has been extreme. DRAM contract prices rose 171.8 percent year on year in the third quarter of 2025, then jumped 90 to 95 percent quarter on quarter in the first quarter of 2026, followed by an estimated 58 to 63 percent in the second quarter and 13 to 18 percent in the third quarter, with Jefferies expecting another 30 to 40 percent in the fourth quarter. Server DRAM contract prices rose 60 to 70 percent in early 2026 after a cumulative 64 percent increase in the second half of 2025, and are projected to climb around 270 percent across 2026 as a whole. PC DRAM contract prices more than doubled in a single quarter, mobile DRAM prices rose close to 80 percent, and DDR5 contract pricing moved from about 7 dollars per unit to roughly 19.50 dollars.

HBM itself is where pricing power is strongest. Suppliers planned increases of roughly 20 percent on HBM3E for 2026, and HBM4 is priced far above the previous generation, with stacks quoted between 500 and 600 dollars and premiums of 55 to 70 percent over HBM3E. SK Hynix secured about a 50 percent premium on HBM4 in its NVIDIA deals and has raised HBM4 prices by roughly 70 percent. Bernstein expects HBM4 pricing to climb from about 16.6 dollars per gigabyte in 2026 to around 37 dollars per gigabyte in 2027. TrendForce expects HBM contract prices to rise 70 to 140 percent in 2027 and warns of multiples higher pricing if standard DRAM stays this tight.

All of this lands directly on the cost of an AI chip. HBM now accounts for 30 to 40 percent of the build cost of an AI accelerator, up from under 20 percent two generations earlier. Morgan Stanley's supply chain work on the Vera Rubin rack shows memory costs up 435 percent, lifting memory from 5 to 10 percent of the rack bill of materials to 25 to 30 percent. Printed circuit board costs rose 233 percent, multilayer ceramic capacitor costs 182 percent, and ABF substrate costs 82 percent. The result is a rack price near 7.8 million dollars against roughly 4 million dollars for the current GB300 Blackwell rack. Moving from HBM3E to HBM4 also increases the number of DRAM dies per stack from 12 to 16, a 33 percent rise in memory consumption per accelerator before any growth in unit volumes.

That is exactly why AI chip prices are rising. NVIDIA has told its largest customers to expect increases of more than 15 percent, and around 17 percent on some GB300 and Vera Rubin 200 systems, for hardware shipping in early 2027. A 72 GPU Vera Rubin rack priced near 7 million dollars today could reach roughly 8 million dollars. For a one gigawatt data center, this single round of price increases adds at least 5 billion dollars of cost. Foundry pricing is moving in the same direction, with TSMC raising prices across advanced nodes that make up 74 percent of its wafer business, including 3 to 10 percent on sub 5 nanometer processes, up to 10 percent on mature nodes from 2027, and 2 nanometer wafers near 30,000 dollars against about 20,000 dollars for 3 nanometer, a premium of roughly 50 percent. CoWoS advanced packaging is booked solid, with NVIDIA holding an estimated 60 percent of that capacity.

Consumer hardware shows the pass through even more clearly. GeForce RTX 50 series cards rose up to 30 percent in South Korea in August 2026, pushing the RTX 5090 past 5,100 dollars, while Chinese retail listings for the RTX 5070 Ti jumped 21.8 percent inside a single 24 hour window and the RTX 5070 rose 16.2 percent. AMD raised GPU and GDDR7 kit prices by about 10 percent for its board partners from July 2026, and 2GB GDDR7 modules now trade in the low 20 dollar range. Qualcomm told partners that Snapdragon chip prices will rise by double digits from September 1, 2026, after its handset chip revenue fell 20 percent year on year. Chinese AI chipmakers raised prices by 20 to 50 percent in only two months, with Huawei's Ascend 950DT quoted above 250,000 yuan, roughly 37,255 dollars.

The knock on effects reach ordinary devices as well. Gartner expects combined DRAM and solid state drive prices to rise 130 percent by the end of 2026, which translates into 17 percent higher PC prices and 13 percent higher smartphone prices, alongside a 10.4 percent fall in PC shipments and an 8.4 percent fall in smartphone shipments. Lenovo, Dell, HP, Acer and ASUS have warned customers about 15 to 20 percent increases, and at HP the memory share of PC build cost rose from 15 to 18 percent to 35 percent in one quarter. A 64GB DDR5 server module moved from 873 dollars in the first quarter to roughly 1,586 dollars by the fourth quarter, a gain above 80 percent. A 16 gigabit DDR5 chip went from about 6.84 dollars to about 27.20 dollars, up 298 percent, while a 32GB DDR5 kit that once sold below 90 dollars now trades near 529 dollars. Amazon's Echo Dot rose 60 percent to 79.99 dollars and the Kindle 37 percent to 149.99 dollars.

The clearest winners of this chain are the memory makers. Global DRAM industry revenue reached 97 billion dollars in the first quarter of 2026, up 81 percent quarter on quarter, then climbed another 59.5 percent to 154.73 billion dollars in the second quarter. Samsung's DRAM revenue rose 93.4 percent in the first quarter to 37.32 billion dollars and then 63.4 percent to 60.98 billion dollars in the second, taking a 39.4 percent share. SK Hynix grew 62.5 percent and then 37.9 percent to 38.59 billion dollars, while Micron grew 81.6 percent and then 65.5 percent to 36 billion dollars. Samsung's second quarter operating profit reached 89.4 trillion won, up about 1,810 percent year on year. Micron's gross margin hit 74.4 percent, and SK Hynix's operating margin has run near 72 to 77 percent. Micron's share price gained more than 120 percent year to date, SK Hynix around 146 percent and Samsung about 121 percent, with SK Hynix passing a one trillion dollar market value and Micron crossing 1,000 dollars per share. In HBM specifically, SK Hynix held 50 percent of revenue in the second quarter of 2026, Samsung 33 percent and Micron 18 percent. The HBM market itself grew from about 4 billion dollars in 2023 to roughly 34.6 billion dollars in 2025 and could reach 54.6 billion dollars in 2026, a further 58 percent increase.

For AI chip companies the picture is more mixed. Demand remains enormous, with NVIDIA data center revenue at 89 billion dollars in the quarter ended July 2026, up 117 percent year on year, and global AI server shipments expected to grow around 31 percent in 2026. But memory is now the largest cost input that chip designers cannot control, and the shortage is even reshaping product designs. NVIDIA has been evaluating lower memory configurations for Rubin Ultra, including 8 high HBM4, which would cut capacity to about 192 gigabytes per accelerator instead of 288 gigabytes, purely because DRAM supply is expected to be tightest in 2027. Custom AI accelerator sales are projected to grow 45 percent in 2026, faster than GPU shipment growth of 16 percent, partly because ASIC designers enjoy more flexibility on memory content. Memory is also expected to absorb 68 percent of major cloud providers' capital expenditure in 2027, and cloud capex itself is projected to rise 98 percent in 2026 and another 50 percent in 2027.

Looking ahead, the pressure does not fade quickly. TrendForce expects HBM bit shipments to grow 50 to 60 percent in 2027 and still fall short of demand. SK Hynix's chief executive has called 2027 the worst year for the shortage, with constraints possibly lasting to 2030, and meaningful new supply is not expected before 2029 or 2030. That is the real difference between this cycle and earlier memory cycles, it is a structural reallocation of the world's most advanced manufacturing capacity toward AI memory rather than a temporary supply accident. HBM shortage, HBM price increase, AI chip cost increase, AI chip price increase, that chain is now visible in data center budgets, graphics cards, laptops, smartphones and quarterly earnings at the same time, which is precisely why this news matters so much to both AI chip companies and HBM manufacturers, one side paying for it and one side profiting from it, with the balance between them set to define the economics of AI hardware well into 2027.
#Gate广场中秋团圆局