Recently, two South Korean tech giants, Samsung and SK Hynix, made headlines by jointly announcing a major move: they plan to raise the prices of server DRAM supplied to cloud giants like Microsoft, Google, and Amazon Web Services by a staggering 60% to 70% in one go! That’s right—not 6% or 7%, but a whopping 60% to 70% increase! This move is expected to generate annual profits of around 150 trillion won for the two companies, nearly two to three times last year’s levels. Currently, global cloud providers are racing to build AI data centers, but production capacity for high-bandwidth memory HBM3E (especially the variant used in Nvidia’s H200) is severely constrained. To meet the surging demand for HBM orders, Samsung and SK Hynix have maxed out their production lines, leading to a critical shortage of standard server DRAM.
The "memory battle" is intensifying. Not only is Nvidia's demand surging, but Google's in-house AI chip TPU is also competing for HBM production capacity. Reports suggest that Google plans to sell 5 million TPUs by 2027, increasing to 7 million by 2028. Meanwhile, as AI services like ChatGPT and Gemini become more widely adopted, the demand for inference servers is soaring, driving up DDR memory prices accordingly. To secure supply, executives from Apple, Microsoft, Google, and Meta have flown to South Korea, setting up camp in hotels in Pangyo and visiting Samsung and SK Hynix headquarters daily to negotiate long-term supply agreements. However, Samsung and SK Hynix have consistently refused long-term contracts, insisting on quarterly agreements instead, as they predict DRAM prices will continue to rise until 2027!
This supply crunch is not limited to memory. KeyBanc's latest assessment report indicates that due to massive procurement by hyperscale cloud service providers, the server CPU production capacity of chip giants AMD and Intel for the entire year has essentially been sold out. To address the extreme supply-demand imbalance and ensure future supply, both companies plan to increase server CPU prices by 10-15%. The core reason for this supply shortage lies in the fact that global hyperscale cloud service providers are entering an intensive architecture upgrade cycle, accelerating the phasing out of older platforms and transitioning to new-generation architectures such as AMD's fifth-generation EPYC "Turin" and Intel Xeon "Granite Rapids." It is estimated that server CPU shipments will grow by 25% this year alone. This upgrade demand presents an opportunity for AMD and Intel, especially for Intel, to expand their data center businesses. KeyBanc predicts that server shipments will increase by 2% quarter-on-quarter in the first quarter of 2026, with full-year growth expected to reach 16-17%.
The outcome was predictable: despite the financial pain, major manufacturers ultimately accepted this round of price hikes. Research institutions point out that customers are investing in AI infrastructure "without batting an eye," as AI commercialization is imminent—no matter how expensive, they have to buy it. This indicates that server prices are expected to enter a sustained upward trend by 2026. On one hand, DRAM prices will continue to rise at least until 2027 due to strategic supply control by manufacturers and HBM production capacity constraints. On the other hand, key components such as CPUs, power supplies, and high-speed interconnects are also facing price increases due to tight supply and demand. Meanwhile, AI data center construction is accelerating, driving continuous growth in demand for complete server procurement. Cost pressures will cascade from upstream chips all the way to the final server price. Starting from the second half of 2026, market prices for complete servers are expected to rise significantly. Companies and institutions with procurement plans are advised to plan and finalize their purchasing strategies as early as possible, aiming to complete deployments by early 2026 when prices are relatively low, thereby avoiding substantial cost increases later. If you're going to buy, you really should do it sooner rather than later.



.png)
