Samsung Plans 110 Trillion Won AI Chip Spend

Samsung Electronics plans to invest more than 110 trillion won (about $73bn) in facilities and research in 2026, the largest annual outlay in its history, in a push to lead the market for AI memory chips.

Samsung disclosed the figure on 19 March 2026 in its 2026 Corporate Value Enhancement plan, a regulatory filing reported to its board the day before. Yonhap reported that the plan is 21.7% above the 90.4 trillion won Samsung invested in 2025. Reuters put the total at $73.24bn and noted that 52.7 trillion won of the 2025 figure was capital expenditure.

The filing is short on detail but clear on intent. Samsung describes itself as the only chipmaker able to offer a “one-stop solution” across memory, foundry and advanced packaging, and says it wants a “strong and sustainable leadership position” in high-value memory, including high-bandwidth memory (HBM). The same document says it will pursue “meaningful M&As” in advanced robotics, medical technology, automotive electronics and heating and cooling (HVAC).

Samsung’s 110 trillion won covers fabs at home and in Texas

The money is split between facilities and research and development; the filing does not break it down further. Yonhap reported that Samsung is improving efficiency at its P4 plant on the Pyeongtaek campus, installing key equipment for the planned P5 line, building new fabs in the Yongin chip cluster south of Seoul and aiming to start operations at its new foundry in Taylor, Texas, by the end of 2026.

HBM is at the centre of the plan. It is the stacked DRAM placed beside AI accelerators to feed them data, and in 2025 it was the part of the memory market where Samsung trailed. Bloomberg reported that the increase was aimed at regaining ground in AI memory, where SK Hynix had pulled ahead, and that Samsung’s total outlay exceeded the roughly $50bn TSMC set aside for capital spending in 2026. Our earlier report on Nvidia asking Samsung, SK Hynix and Micron to make HBM base dies at TSMC shows how closely memory makers now depend on their largest customer’s roadmap.

Samsung is spending alongside the cloud giants, not instead of them

Samsung’s plan landed weeks after the largest US platforms set out their own budgets. In its fourth-quarter 2025 results on 4 February, Alphabet said it expected 2026 capital expenditure of $175bn to $185bn, after $91.4bn in 2025. Meta’s results on 28 January guided 2026 capital expenditure, including finance-lease principal payments, at $115bn to $135bn, against $72.2bn in 2025.

Those budgets pay for data centres and the servers in them, and every AI server needs memory. That is the link to Samsung: the hyperscalers’ spending is the demand that the Korean memory makers are building capacity to meet. For context on how those data-centre budgets are being spent, see our analysis of who is winning the AI infrastructure race after Nvidia GTC.

Policy also shapes the field. US export controls limit what equipment Chinese rivals can buy, and the proposed MATCH Act would tighten those limits further. Chinese DRAM makers are still advancing: in October, Samsung and SK Hynix shares fell after CXMT claimed a denser DRAM design.

Samsung’s HBM bet depends on yields and customer qualification

The size of the number says less than it seems. A capital budget is a plan, not output, and Samsung’s own filing warns that its forward-looking items “may be subject to change depending on future market conditions”. New memory fabs take years to move from cleanroom construction to wafer output, so money committed in 2026 mainly buys capacity for 2027 and 2028. The more telling test is whether customers qualify Samsung’s HBM for their accelerators, because HBM is sold against specific products and a supplier that misses a qualification window can sit out a whole product cycle.

The plan also reads as a reaction to a weaker position, not a sign of an existing lead. In 2025 Samsung lost ground to SK Hynix in HBM, the fastest-growing and most profitable part of the memory market. Spending more is the obvious lever for a firm with Samsung’s balance sheet, but money cannot alone fix the packaging and stacking yields that decide how many usable HBM stacks come off each wafer. That is why the filing pairs the investment pledge with a goal of “improved investment efficiency” rather than just more capacity.

For UK readers the story matters mainly through prices and supply. Britain has no memory fabs, so UK data-centre builders and device makers buy DRAM and HBM on the world market. When the largest suppliers say their HBM output is sold out a year ahead, smaller buyers usually face longer lead times and higher contract prices, whichever company wins the market share contest.

Update, 2 November 2026: In its third-quarter 2026 results on 29 October, Samsung reported revenue of 195.3 trillion won and operating profit of 107.4 trillion won, and said it now expects full-year 2026 capital expenditure of about 84.4 trillion won, 76.8 trillion won of it in the chip division. On the earnings call, Samsung said HBM4 sales more than tripled from the previous quarter to 60% of its HBM sales, that its HBM market share had reached the level of its overall DRAM share, and that its 2027 HBM supply is already sold out. See our report on Samsung’s record quarterly guidance.

Correction, 12 October 2026: this article originally said Meta committed $65 billion to AI infrastructure in 2026 and that Samsung’s plan nearly doubled its previous spending. Meta guided 2026 capex at $115–135 billion, and Samsung’s plan is about 22% above its 2025 level.