Samsung, SK Hynix Fall Below 10 Days of Memory Supply

Samsung Electronics and SK Hynix, which together hold most of the world's DRAM manufacturing capacity, saw their combined finished-memory inventory fall below 10 days of supply during the third quarter, Seoul Economic Daily reported on Sept. 7, citing a note from KB Securities. Kim Dong-won, the firm's head of research, was blunter about what comes next than the inventory number alone suggests: the memory market in 2027 will see "the tightest supply conditions in history," he wrote.

Ten days of stock is not itself a crisis; DRAM inventories have run leaner before. What makes the figure notable is the demand behind it. KB Securities revised its estimate for global AI infrastructure investment in 2027 to $1.3 trillion, up 60 percent from the prior year's spending, and it expects memory to absorb 57 percent of that total, up from 40 percent in 2026 and just 14 percent in 2025. TrendForce, working from separate estimates, puts memory's 2027 share even higher, at 68 percent. The two firms disagree by 11 percentage points on a number in the trillions of dollars, which is itself the more honest way to read a forecast like this: nobody actually knows the exact share, only that it is rising fast enough that two credible estimates land more than ten points apart.

The chip everyone assumes is the profitable one, and the wafer that actually is

The assumed story is that high-bandwidth memory, the stacked DRAM used inside AI accelerators, is where all the pricing power sits, since HBM is the component every GPU maker is fighting over. TrendForce's own numbers complicate that story. HBM per-wafer revenue was overtaken by DDR5 64-gigabyte RDIMM modules, the conventional server memory used outside AI accelerators, in the first quarter of 2026, according to a TrendForce press release from June 2. Conventional DRAM prices have climbed on a market starved of the capacity that HBM production is absorbing, to the point that a technology Samsung and SK Hynix were selling as commodity memory two years ago now generates more revenue per wafer than the specialized product it is supposedly subordinate to.

That inversion is why TrendForce expects HBM contract prices to "surge multiples higher" in 2027: suppliers have to restore the gap or lose the incentive to keep prioritizing HBM's more complex, lower-yielding production process over ordinary DRAM that is paying better per wafer. The firm's wafer-allocation data shows the top three suppliers putting about 18 percent of total DRAM wafer input toward HBM at the end of 2025, rising to roughly 22 percent by the end of 2026 and an estimated 30 percent by the end of 2027.

Year	HBM share of DRAM wafer input	Memory share of AI infra spend
End 2025	~18%	14%
End 2026	~22%	40%
End 2027 (est.)	~30%	57% (KB) / 68% (TrendForce)

What the capacity per chip is actually buying

The reason HBM keeps eating a larger share of wafer input despite the pricing inversion is that each new accelerator generation needs more of it. TrendForce's figures show HBM capacity per AI chip rising from a range of 96 to 192 gigabytes in 2026 to 216 to 288 gigabytes, with Nvidia's upcoming Rubin Ultra platform, expected in 2027, targeting 384 gigabytes per chip. That trajectory tracks the same memory pressure documented elsewhere on the accelerator side: DeepSeek's own open-weight model needs 475 gigabytes of GPU memory before it can run at all, and the industry's response on the software side has been aggressive quantization, like the technique that squeezed a 27-billion-parameter Qwen model down to 8.4 gigabytes. Software is compressing models precisely because the hardware supplying the memory to run them uncompressed is the scarcest, most contested link in the entire AI supply chain right now.

What the stock market already priced in

Samsung and SK Hynix shares are down 38 percent from their peaks over the past three months, even as the underlying supply story has tightened rather than loosened, per KB Securities' note. That decline has pushed both companies' price-to-earnings ratios down to roughly 3, based on next year's expected results, a level that in a normal cycle would signal a company in decline rather than one facing the tightest supply conditions its own analysts have described. The market is not disputing KB Securities' inventory math; it is betting that a memory supercycle this steep reverses faster than the wafer-allocation numbers above suggest, the same skepticism that shows up whenever a commodity input to AI infrastructure gets priced for a demand curve nobody can fully verify yet.

That same wariness about betting five years out is visible elsewhere in how AI infrastructure gets financed. Nvidia has increasingly taken direct equity stakes in the companies that buy its chips rather than relying on open-market demand alone, the same instinct toward guaranteed exposure over a market bet that shows up in Samsung and SK Hynix hedging their own multi-year commitments even while collecting record pricing today.

What would change this read

The number to watch is whether TrendForce's 68 percent 2027 memory-share estimate or KB Securities' 57 percent turns out closer to right, since an 11-point gap on $1.3 trillion in projected spending is worth roughly $143 billion, larger than most single companies' annual capital budgets. If the lower KB estimate holds, some of the urgency behind the sub-10-day inventory figure eases on its own as conventional DRAM supply catches up faster than the more aggressive forecast assumes. If TrendForce's number is closer, expect the DDR5-over-HBM revenue inversion documented above to reverse hard in 2027, because suppliers cannot sustain putting 30 percent of wafer input toward HBM while a competing product pays better per wafer indefinitely.