As the four U.S. big tech corporations (Alphabet, Amazon, Meta, Microsoft) again raised this year's artificial intelligence (AI) infrastructure investment budgets, expectations are strengthening that the memory supercycle (boom period) will be prolonged. These corporations are cited as the largest customer group for AI Server Memory and storage such as high bandwidth memory (HBM), server DRAM, and enterprise solid-state drives (eSSD) for AI servers.
On the 3rd, compiling the latest earnings announcements from Alphabet, Amazon, Meta, and Microsoft (MS), the combined capital expenditure (CapEx) outlook the four big tech corporations presented for this year is $720 billion to $745 billion (about 1,041.12 trillion to 1,077.27 trillion won). Compared with the previous combined outlook of $695 billion to $725 billion (about 1,004.97 trillion to 1,048.35 trillion won), the lower end rose by $25 billion (about 36.15 trillion won) and the upper end by $20 billion (about 28.92 trillion won).
A semiconductor industry official said, "There was analysis that facility expansion could slow because monetizing AI services may not keep pace with rising investment costs, but as the four big tech corporations instead increased their budgets, concerns over an 'AI bubble theory' or 'memory peak-out (post-peak)' are easing."
◇ Big tech's four corporations move to expand AI data centers… "Responding to growing AI demand"
Alphabet, Google's parent, raised its annual facility investment outlook from $180 billion to $190 billion (about 260.28 trillion to 274.74 trillion won) to $195 billion to $205 billion (about 281.97 trillion to 296.43 trillion won). Alphabet Chief Financial Officer Anat Ashkenazi said, "The main reason for the upward revision is that we pulled forward capacity supply to respond to increasing demand."
Amazon expanded its annual investment plan from $200 billion (about 289.2 trillion won) to $220 billion (about 318.12 trillion won). Amazon CEO Andy Jassy said, "Even if we invest $220 billion, we will not secure enough capacity to meet all of this year's demand." Meta adjusted its outlook range from $125 billion to $145 billion (about 180.75 trillion to 209.67 trillion won) to $130 billion to $145 billion (about 187.98 trillion to 209.67 trillion won).
MS said its annual investment decreased from $190 billion (about 274.74 trillion won) to $175 billion (about 253.05 trillion won) due to a change in lease accounting, but that there was no change in the actual plan. It expected fiscal 2027 facility investment to increase from the prior year, and expenditure in July–September to exceed $50 billion (about 72.3 trillion won).
◇ About 24.6 trillion won in the second quarter alone… solid AI demand confirmed
The four big tech corporations' capex-like expenditure executed in the second quarter this year totaled $170.08 billion (about 245.94 trillion won). That was up 30.2% from $130.65 billion (about 188.92 trillion won) in the first quarter. MS said, "Customer demand continues to outstrip available capacity." About two-thirds of MS's facility investment executed during this period went into assets with short useful life, such as central processing units (CPUs) and graphics processing units (GPUs).
The market interprets the increase in facility investment by the four big tech corporations as confirmation of actual growth in AI and cloud demand. In the second quarter, MS's Azure and other cloud services revenue rose 43% from a year earlier. Amazon Web Services (AWS) grew 37%, and Google Cloud grew 82%. Meta's advertising revenue also increased 27% thanks to improvements in AI-based recommendation and ad systems.
A researcher at a market research firm said, "The results are far from the claim that data centers are being built excessively without demand," adding, "As cloud revenue, backlog, and usage all increase together, the spare capacity at AI data centers secured by the four big tech corporations is being rapidly depleted."
◇ "In two years, hyperscalers' memory expenditure up tenfold… will account for 73% of next year's facility investment"
The four U.S. big tech corporations are cited as the players that directly purchase memory semiconductors or place orders for GPU accelerators and servers/storage, thereby determining actual intake of HBM, DRAM, and NAND flash. According to global investment bank UBS, the top 11 hyperscalers (large data center operators), which include them, account for about 90% of demand for AI Server Memory HBM and double data rate (DDR) DRAM. Their share is also about 80% for general-purpose server DDR DRAM and NAND for servers and storage.
UBS estimated their memory expenditure would rise from $72.8 billion (about 105.27 trillion won) last year to $335.657 billion (about 485.36 trillion won) this year, and to $761.317 billion (about 1,100.86 trillion won) next year. It also projected the share of memory in total facility investment to climb from 14% to 38% and 73% over the same period.
UBS presented the following itemized expenditure estimates for the top 11 hyperscalers next year: ▲HBM $144.384 billion (about 208.78 trillion won) ▲DDR DRAM $448.862 billion (about 649.05 trillion won) ▲NAND flash $168.071 billion (about 243.03 trillion won). This means demand is spreading not only for HBM but across products going into servers and storage.
Previously, Nomura, a Japan-based financial firm, projected that memory demand derived from data centers would increase from $454 billion (about 656.48 trillion won) in 2026 to $1.398 trillion (about 2,021.51 trillion won) in 2030.
Memory manufacturers also do not expect the upcycle to end quickly. Samsung Electronics said in its second-quarter earnings, "As demand for server DRAM, eSSD, and HBM accelerates, supply shortages are expected to continue." SK hynix likewise said, "Major big tech customers are expanding infrastructure investments due to rising AI service usage and insufficient computing capacity, and appear to be steadily increasing memory purchases based on growth in revenue and profit generated from AI services," adding, "In fact, major customers are still requesting even more memory supply."
With memory semiconductor shortages continuing, contract structures are also changing. SK hynix has concluded long-term agreement (LTA) negotiations with about 10 customers, including key customers. By including mechanisms to respond to price fluctuations and performance assurances such as deposits, it has increased medium- to long-term demand visibility. Samsung Electronics also said, "Nearly all customers are requesting multi-year supply contracts," adding that it has signed with the world's five largest data center corporations and is in the final stages of negotiations with five other major customers.
A semiconductor industry official said, "As big tech's investment expansion is boosting demand not only for HBM but also for server DRAM and eSSD, this upcycle could last longer than in the past."