"Memory supply constraints will not be resolved until new fabs are built and reach high utilization rates. Korea has lived through semiconductor winters and harbors deep fears of oversupply, but the earlier semiconductor cycles were entirely a product of the Windows era — a relic of the past. We are now in a fundamentally different AI era."
Samsung Electronics and SK Hynix have recently unveiled mega-investment plans totaling nearly 3,200 trillion won — Samsung Electronics at 2,030 trillion won and SK Hynix at 1,100 trillion won — to build new semiconductor and packaging fabs. Some observers have raised concerns about a semiconductor peak-out, warning that a memory chicken game similar to the one that brought down Japan's semiconductor industry could be unfolding.
Bruce Bateman, chief semiconductor analyst at Omdia Taiwan, flatly dismissed those fears in an interview, calling them a failure to grasp the shift in the IT paradigm. "Even if SK Hynix brings a new fab online, it would add only about 3 percent at most to the total HBM market," Bateman said. "Memory supply constraints will not be resolved until new fabs are built and reach high utilization rates." According to Omdia, HBM is essentially sold out through 2027.
The reason the market appears to be peaking out is that wafer output and revenue are not being viewed separately. According to Omdia, from 2020 to 2026, semiconductor market revenue surged 160 percent and average selling prices jumped 90 percent, while wafer output grew only 33 percent. Rising memory prices, Bateman said, are being mistaken for an expansion in the overall size of the semiconductor market.
Omdia projects the semiconductor market will grow 87.5 percent year on year, from $687 billion in 2024 to $1.6 trillion in 2026. Growth is then expected to moderate, reaching $2.1 trillion by 2028.
"This year represents the peak in the rate of growth, not the peak in market size," Bateman said. A slowdown in the growth rate does not mean the market is shrinking. He projected that HBM prices could begin to fall around 2028 — when new fabs from the three major memory chipmakers come online — and no later than 2030, as HBM becomes more commoditized over the medium to long term.
The real bottleneck driving this super-cycle, however, is power. "Among major US data center projects, between 9 and 12 gigawatts of capacity face the risk of delay or cancellation due to power and water constraints," Bateman said. "The question is whether companies will cancel their Nvidia orders if data center construction stalls, or whether they will accept delivery and warehouse the products." AI accelerators take more than a year to deliver and require upfront payments of $2 billion to $6 billion to place an order.
The most critical factor right now is accurate demand visibility. The reason the three major memory chipmakers have been mentioning long-term agreements, or LTAs, in recent earnings calls is that LTAs allow them to forecast big tech demand over the medium to long term. "Big tech companies like Amazon and Meta have enormous financial firepower — everyone calls them whales — and they are willing to accept 100 percent cash prepayment terms," Bateman said. "But the question that TSMC and Samsung foundry executives ask most persistently in private is: among our big tech customers, who actually holds the cards and who is telling the truth?"
Even if the data center frenzy cools, semiconductor demand will not disappear. "Even if the data center boom loses some steam, AI will naturally permeate on-device agents in mobile devices, low-cost mini PCs and home smart appliances, forming a new hybrid infrastructure ecosystem," Bateman said. The industry is moving beyond the Windows era into the age of agentic operating systems. Nvidia CEO Jensen Huang's recent unveiling of the RTX Spark, a next-generation AI superchip for PCs, reflects an anticipation of the era that comes after data centers, Bateman added.
go@heraldcorp.com
