The constraint is utilization, and Samsung's 8nm-and-below lines have now reached maximum capacity. DRAM wafer starts alone could not fill the schedule; the base die in an HBM4 stack, fabricated on Samsung's 4nm process, changes the math. Every HBM4 unit sold routes a logic die through the foundry, and Samsung is the only major memory maker that owns the logic lines to make it.

HBM4 mass production began in February, pairing 1c DRAM with a 4nm base die. By the second half, HBM4 is projected to exceed 60 percent of total HBM revenue, with Q3 HBM4 revenue more than tripling quarter-over-quarter. Kiwoom Securities forecasts the foundry and System LSI operating loss shrinking from 6.74 trillion won in 2025 to 3.92 trillion won in 2026, a 41.8 percent reduction. Quarterly losses that topped 2 trillion won in each of the first-half 2025 quarters are expected to land at 777 billion won in Q3 2026.

Structurally, memory revenue now carries foundry-linked volume. Samsung's turnkey contracts bundle memory, foundry, and advanced packaging under one roof. The July MOU with Broadcom, worth more than $200 billion through 2030, includes HBM4 and HBM4E supply for custom AI accelerators. For a buyer evaluating a custom accelerator platform, that means the memory stack, the logic die, and the package come from a single vendor with a single yield accountability chain.

Die Brief's read: the 41.8 percent loss reduction is a utilization story, not a pricing story. Kiwoom itself shifted its investment thesis away from commodity DRAM price increases, capping Q4 DRAM price growth at 3 percent quarter-over-quarter, toward HBM volume and foundry share. What the document does not prove is that the foundry will reach breakeven. The 777 billion won Q3 figure is a narrowing, not a crossing. For the buyer, the constraint shifts from securing HBM4 supply to pricing that reflects a foundry still operating at a loss.

What remains unmeasured: actual yield on the 4nm base die at volume, the split between internal HBM4 consumption and external turnkey revenue in Q3, and whether the Broadcom MOU converts to booked orders before 2027 capacity expansion. HBM4E samples went out in May; no customer qualification timeline is stated. Watch the Q3 earnings call for the foundry revenue line broken out from memory, and for Exynos 2800 wafer starts as a second logic-volume driver.

The 41.8 percent loss reduction is a utilization story, not a pricing story, and the document does not prove the foundry will cross into profitability. The buyer's real constraint is that HBM4 pricing will carry the cost of a foundry still running at a 777 billion won quarterly loss.

HBM4 base-die volume on 4nm fills foundry lines that DRAM alone could not, cutting the operating loss from 6.74T to 3.92T won. Kiwoom's own Q4 DRAM price cap of 3% QoQ confirms the driver is volume, not commodity pricing. Breakeven is not demonstrated; 777B won in Q3 is still negative. The Broadcom MOU and turnkey structure give Samsung a single-vendor yield chain, but the buyer inherits the foundry's margin deficit in the HBM4 price.