Samsung Warns AI Memory Chip Shortage Will Last Through 2028

Samsung Electronics has delivered one of the most consequential supply forecasts of the current AI buildout era: the global shortage of advanced memory chips will not only persist but deepen in 2027 and remain tight through 2028. The warning, delivered by Jaejune Kim, executive vice president of Samsung's memory business, during the company's second-quarter earnings call, pushes the industry's expected recovery horizon out by another year and signals that the AI infrastructure boom is creating structural scarcity in one of computing's most foundational components121516.

What Samsung Actually Said

The core of Samsung's message was blunt. The supply shortage in 2027, Kim told analysts, is expected to be worse than this year, and the tightness is projected to continue into 20281519. The company framed unmet demand from the current year as likely to carry over into the next, compounding an already strained allocation picture rather than allowing it to clear10.

The context for the warning was a quarter of almost comical financial magnitude. Samsung's semiconductor division reported an operating profit increase of more than 250-fold year over year, driven by data center demand and the company's deliberate prioritization of high-value AI server products such as high-bandwidth memory (HBM), server DRAM, and enterprise solid-state drives131518. The Register's framing captures the irony precisely: a record pile of cash for Samsung, and years of elevated prices for everyone buying chips16.

Samsung's explanation for the shortage's durability rests on physics and lead times. Kim noted that the interval between starting construction of a new fab and actually producing wafers exceeds three years, meaning that even the aggressive capacity investments now underway across the industry cannot meaningfully relieve supply before the late 2020s16. This is a supply-side argument, not a demand-side one — the industry is spending heavily, but memory manufacturing simply cannot be scaled quickly.

The Long-Term Contracts Behind the Forecast

The most strategically significant detail in Samsung's disclosure is not the 2028 warning itself but what the company is doing about it. Samsung has signed long-term supply agreements with the world's five largest data center operators and is reportedly close to finalizing similar deals with five more major customers1417.

These contracts run for at least five years, and Samsung expects them to account for roughly 60% to 70% of its long-term production capacity1315. Critically, the agreements include upfront payments and floor pricing — minimum price provisions designed to reduce earnings volatility across the cycle1417. In effect, the hyperscalers are paying now to guarantee allocation later, and Samsung is converting a speculative boom into contracted, bankable revenue.

This structure matters beyond Samsung's balance sheet. It changes the economics of the memory industry. Historically, DRAM has been a brutally cyclical commodity in which suppliers over-expanded into booms and suffered ruinous pricing in busts. Multi-year take-or-pay arrangements with the world's biggest cloud companies — covering a majority of future output — are a de facto industrial policy shift, turning spot-market chaos into something closer to infrastructure contracting. If Samsung's model holds, memory pricing may stay structurally elevated even if AI demand cools, because the floor pricing has already been locked in.

Where the Reporting Diverges: The Market Isn't Buying It

Here the coverage splits into two camps, and the tension is the real story.

One group of reports, including IBTimes and the Economic Times' data center desk, treats Samsung's forecast as evidence that AI infrastructure spending is more durable than skeptics believe, noting that the guidance counters investor fears of a hyperscaler spending slowdown1213. The Free Press Journal similarly framed the results as easing concerns that heavy AI infrastructure outlays by major tech companies might slow20.

The other camp points out that the market has already voted, and it voted against. Bitget's analysis notes that despite explosive second-quarter profits at both Samsung and SK Hynix, the two companies' stock prices were hammered with sharp declines following their results11. The reason is what that outlet calls a cycle divergence of more than a year: management insists shortages will extend to 2028 and is signing long-term contracts accordingly, while investors worry about slowing cloud capital expenditure and rising memory production capacity in China — pricing in an early peak to the cycle11.

The Register's angle reinforces the bearish read from a different direction, arguing the shortage persists largely because memory makers themselves are choosing to keep capacity tight, prioritizing AI server components and leaving mainstream PC, smartphone, and consumer device memory short — a shortage that pads Samsung's profits at the expense of everyone downstream16. Traders Union's report similarly acknowledges that investors are openly questioning whether record profits can be sustained14.

My reading: the market skepticism deserves to be taken seriously, but it is fighting a structural headwind. Samsung has already converted 60-70% of its long-term capacity into contracted volume with floor pricing — meaning much of the revenue visibility the skeptics doubt has been legally locked in by customers with enormous balance sheets. It is hard to talk about a demand collapse in 2027 when five hyperscalers have pre-paid for supply stretching to the early 2030s. The bears are effectively betting that the contracts' floor pricing becomes a liability and that China's capacity ramp arrives faster than Samsung's three-year fab lead time suggests is possible. One of those bets requires a demand break; the other requires an execution miracle.

Why Memory Is the AI Bottleneck That Actually Bites

The AI chips conversation usually orbits GPUs and accelerators, but Samsung's warning is a reminder that the binding constraint in AI infrastructure is increasingly memory. As Techbooky notes, memory is not optional — AI servers require enormous DRAM and HBM allocations, and every wafer of HBM capacity the industry adds displaces conventional DRAM production18. HBM, a vertically stacked form of DRAM positioned close to processors, consumes disproportionate fab capacity per bit produced, which is exactly why the AI boom has tightened the entire mainstream memory market rather than just a premium niche8.

The downstream consequences are already visible. Samsung's memory unit has been prioritizing AI server demand, and the resulting shortfall of conventional DRAM has been pushing up prices for laptops, phones, storage, and gaming devices throughout 20261618. Apple said earlier this year that rising memory prices had begun pressuring its profitability, echoing warnings that PC and smartphone makers would bear the brunt of the shortage as capacity diverts toward HBM for AI servers9. Samsung's own mobile division has now been caught on the wrong side of the trend — TechSpot highlights that the phone business posted its first loss even as the chip business booms, partly because the component costs Samsung's chip unit is commanding are the same costs squeezing its handset division1720.

The Bottom Line

Samsung's 2028 forecast is best understood not as a prediction about AI demand lasting forever, but as a statement about how long it takes to build memory fabs. Three-year construction lead times mean the supply response to 2025-2026 demand cannot arrive before roughly 2028 regardless of what anyone spends today16. Whether demand remains strong enough to absorb that eventual supply wave is the open question — and Samsung itself conceded that beyond 2029 visibility is limited10.

For buyers of AI infrastructure, the message is that scarcity pricing is now a multi-year condition, and locking supply early is the only rational strategy — which is precisely what the top ten data center operators have done. For everyone else — PC makers, phone vendors, and consumers — the forecast implies another stretch of elevated prices and constrained product configurations stretching well past 20271016.

The disagreement between Samsung's management and the stock market will be settled by capital expenditure data in the second half of this year11. Until then, the burden of proof sits oddly: the skeptics must explain away billions in pre-paid, floor-priced contracts, while Samsung must hope that the AI data center buildout it has bet 70% of its future capacity on keeps its footing for at least two more years. One of those positions is backed by signed paperwork. The other is backed by market sentiment. In the memory business, history suggests paperwork tends to win.