SK Hynix expects DRAM demand to grow mid-20% in 2027

SK Hynix expects DRAM demand to grow by the mid-20% range in 2027. NAND demand is forecast to climb in the high-teens range over the same year.

The driver is artificial intelligence. Data centers built for AI workloads need enormous amounts of memory, and the industry is struggling to make enough of it.

The forecast and the shortage behind it

A quick primer helps. DRAM is the short-term working memory a computer uses while it runs tasks. NAND is the storage that keeps data after the power goes off, the stuff inside SSDs and phones.

Both are now in short supply, and SK Hynix’s leadership is not sugarcoating it. CEO Kwak Noh-jung has said 2027 will bring the worst supply situation the memory industry has faced.

He went further. Kwak projected that demand will keep outrunning supply even beyond 2030, despite the expansions chipmakers have planned.

The company’s mid-20% DRAM estimate sits at the upper end of a wider set of forecasts. Industry projections put 2027 DRAM bit demand growth somewhere between 19% and 25%, with much of the range clustered in the low-to-mid 20s.

Bit demand, for the uninitiated, measures total memory capacity required rather than the number of chips sold. As chips get denser, each one holds more bits, so this is the cleaner way to track real appetite.

Rival Micron has expressed similar expectations, which matters. When competitors who fight over the same customers reach the same conclusion, the forecast carries more weight than a single company’s optimism.

NAND estimates run slightly cooler. Projections place its 2027 demand growth in the high-teens to low-20% range, and supply conditions could ease somewhat later in the year.

Why supply can’t simply catch up

The answer comes down to three stubborn bottlenecks.

The first is equipment. The tools needed to expand memory production come with long lead times, so new capacity cannot materialize on short notice.

The second is contracts. Major memory manufacturers have reportedly locked in long-term agreements covering approximately 70-80% of their planned 2027 DRAM and NAND output.

For high-bandwidth memory, the commitment is even more complete. Nearly 100% of HBM production is already spoken for.

The third bottleneck is HBM itself. High-bandwidth memory stacks DRAM layers together so data can move to AI processors much faster. It is the premium memory that sits alongside the accelerators powering large AI models.

The catch is that HBM eats more wafer capacity per chip than conventional memory. Every wafer devoted to HBM is a wafer not producing standard DRAM for PCs, phones, and servers.

HBM is expected to consume nearly 30% of industry DRAM wafer capacity in 2027. That is up from around 20% currently.

A ten-point shift in wafer allocation pulls a meaningful chunk of output away from the conventional memory market at the exact moment that market is also growing.

Who pays for the squeeze

For memory makers, projected demand growth paired with constrained capacity points toward greater pricing power for companies like SK Hynix and Micron in the coming years.

Pre-committed contracts add a layer of visibility that chipmakers rarely enjoy. When most of next year’s output is already sold, revenue planning gets considerably less stressful.

The other side of that trade is less cheerful. Hyperscalers, the giant cloud operators building AI data centers, are likely to face rising memory costs.

PC manufacturers are exposed too. They depend on conventional DRAM, which is precisely the product getting crowded out as wafer capacity tilts toward HBM.

What to watch next

Capacity expansion is the big one. If equipment deliveries accelerate or new fabs come online faster than expected, the shortage could soften. If they slip, the squeeze tightens further.

HBM allocation is another lever. The share of wafers flowing to high-bandwidth memory directly shapes how much conventional DRAM reaches the broader market, so any change in that mix will ripple through pricing.

NAND deserves its own attention. With projections hinting at possible easing later in 2027, storage could diverge from DRAM and behave more like a traditional cycle.

For now, the message from SK Hynix is unusually blunt. Demand is rising faster than factories can respond, the contracts are largely signed, and the company’s CEO expects the imbalance to persist well into the next decade.