Contents
Four Conditions Must Hold for the ADR to Reach $245
Where Inference Growth Will Drive Memory Spending
How Technology Leadership Becomes Profit—and How It Can Be Diluted
What Do Five-Year LTAs Actually Guarantee?
How Much of the Profit Growth Comes From Pricing, Operations, and the Share Count?
From Accounting Profit to Distributable Cash
Who Actually Receives the “Nearly 42% Three-Year Return”?
How Much Operating Improvement Could the ADR Premium Absorb?
Reverse-Engineering the $245 Target Into Adjustable Assumptions
Where the Strongest Disconfirming Evidence Could Emerge
How Future Public Disclosures Should Change the Thesis
This content is compiled from publicly available information and research reports. It does not constitute investment advice or represent any individual's views. For educational reference only; please read critically.
JPMorgan’s $245 price target for SK hynix’s ADR rests on a longer memory upcycle, more dependable cash flow, and stronger shareholder returns. The key question is how much of the upside comes from operations, how much from buybacks, and how much simply reflects the premium US investors are willing to pay.
Four Conditions Must Hold for the ADR to Reach $245
On September 10, 2026, JPMorgan initiated coverage of SK hynix’s American depositary receipt with an Overweight rating and a June 2027 price target of $245. Against its September 8 reference price of $185.55, this implies approximately 32.0% upside. The timing matters: $185.55 is a price on a specified date, while $245 is a forward valuation contingent on future assumptions. Neither should be treated as a live market quote.
The valuation first requires continued growth in AI demand and that this demand translate into memory purchases. SK hynix must then convert those purchases into revenue and profit, and ultimately turn those profits into distributable cash. Finally, US investors must remain willing to pay a premium for the same underlying company. These four layers are connected but not interchangeable. Strong HBM demand does not automatically imply equally strong cash flow, while better cash flow does not guarantee that the ADR premium will persist.
JPMorgan expects the AI-driven memory upcycle to last more than five years and forecasts an EPS CAGR of approximately 34% over the next two years. Its valuation applies roughly 7x earnings to the average of 2026 and 2027 EPS for the Korean common shares, followed by a 20% premium for the ADR. A P/E multiple measures how much investors are willing to pay for one unit of annual earnings. Averaging two years reduces the impact of fluctuations in any single year, but it does not eliminate cyclicality. If both years fall within an unusually profitable phase, the average may still exceed sustainable earnings.
The central issue is therefore the durability of memory-sector profits. Traditional memory investing involves a familiar contradiction: when prices surge, profits are strong and P/E multiples appear low; capacity then expands, inventories rise, prices fall, and the earnings that once looked cheap quickly disappear. A low multiple can represent undervalued cash-generating capacity only if customer demand, supply constraints, and contractual structures collectively extend the period of elevated profitability.
Long-term agreements and shareholder returns form the bridge in this rerating thesis. LTAs are intended to reduce order volatility, while shareholder returns reduce uncertainty over whether cash will remain trapped on the balance sheet. If both work as intended, investors may accept more stable earnings forecasts and a higher valuation. If LTAs merely secure customer relationships while pricing remains market-based, or if profit growth is absorbed by capacity expansion and working capital, the basis for a higher valuation weakens.
Where Inference Growth Will Drive Memory Spending
SK hynix’s AI exposure extends beyond its reputation as an HBM supplier. High-bandwidth memory sits close to compute chips and supplies data for high-throughput workloads. Server DRAM provides larger-scale working memory, while enterprise SSDs store persistent data, model files, and information that can be moved out of higher-cost memory. Each addresses a different trade-off among speed, capacity, cost, and persistence.
Training repeatedly processes large datasets and model parameters. Inference continuously generates outputs in response to user inputs. As context windows lengthen and concurrent requests increase, systems must manage more intermediate state. AI agents may also break a single user task into multiple model calls, tool queries, and feedback-driven revisions. Demand can therefore grow through both a larger user base and greater compute intensity and data retention per task.
A useful conceptual relationship is that total memory demand depends broadly on the number of concurrent tasks, the data that must remain resident for each task, and the redundancy reserved for performance and reliability. If concurrent workloads rise 50% and data per task increases 20%, all else equal, the demand multiplier is 1.5 × 1.2 = 1.8. This illustrates the mechanism; it is not a shipment forecast for any product. If compression and scheduling reduce unit demand by 25%, the net multiplier under the same assumptions falls to 1.35.
This is also why token volume cannot be equated directly with HBM purchases. A token is a unit for measuring information processed by a model. Better caching, batching, quantization, and memory scheduling may allow the same hardware to process more tokens. Alternatively, lower service costs may attract more users and encourage more complex applications, consuming the resources saved through efficiency improvements. The first path reduces hardware requirements per unit of activity; the second expands total usage. The net effect depends on actual workloads and budgets.
Tiered storage adds another dimension. Some data previously held in expensive memory may migrate to higher-capacity, lower-cost tiers, creating incremental demand for enterprise SSDs. That does not mean HBM, DRAM, and SSD demand will expand at the same rate. Simply adding capacity growth across every tier could double-count the same data as it moves through the hierarchy. SK hynix’s revenue will also depend on the architectures customers adopt and the company’s share at each tier.
JPMorgan forecasts global HBM bit demand growth of 76% in 2026, 63% in 2027, and 44% in 2028. Bits measure memory capacity, so these projections do not represent revenue growth. Revenue also depends on price per unit of capacity and the generational product mix. A higher share of advanced products can lift average selling prices even if prices for comparable products remain unchanged. Conversely, falling prices may prevent strong capacity growth from producing equivalent revenue growth.
SK hynix operates across DRAM, NAND, and enterprise storage, positioning it to capture demand at multiple tiers. Under JPMorgan’s segment forecasts, DRAM—including HBM—accounts for approximately 76% of revenue in 2026 and 80% in 2028. DRAM therefore remains the primary earnings driver, while NAND and enterprise SSDs provide additional sources of growth and risk. Understanding this mix is essential when assessing which portion of earnings will be affected by a particular change in server technology.
Ultimately, the AI-demand thesis should reconcile three accounts: application usage must grow, customers’ AI revenue must support their budgets, and memory vendors must deliver products and collect cash. If hardware orders rise while utilization, revenue, and cash collections fail to keep pace, the evidence for durable demand remains incomplete.
How Technology Leadership Becomes Profit—and How It Can Be Diluted
HBM competitiveness depends on multiple links in the production chain. Memory dies must meet performance and power requirements; stacking and packaging must achieve stable yields; products must pass customer qualification; and suppliers must deliver on time and at scale. Leadership at one stage demonstrates technical potential. Sustainable commercial advantage requires the entire chain to operate repeatedly and reliably.
The importance of customer qualification is particularly easy to underestimate. High-end computing systems have long development cycles, and memory failures or supply shortfalls can disrupt delivery of the entire platform. Customers therefore value engineering collaboration, consistent quality, and volume-production capabilities. Suppliers already embedded in core platforms typically gain earlier visibility into next-generation requirements and can adjust R&D; and capacity accordingly. This collaboration can reinforce leadership, but sustaining it requires continued investment.
JPMorgan views SK hynix’s HBM execution, allocation of advanced DRAM resources, and relationships with major compute-chip customers as competitive advantages. Their economic value lies in helping the company win high-specification orders, maintain a favorable product mix, and exercise greater pricing power during periods of constrained supply. Investors should focus on whether these advantages translate into revenue and cash flow, not merely on the sequence of product announcements.
Leadership should not be mistaken for permanent exclusivity. Customers need supply security and generally prefer to avoid dependence on a single vendor, giving them an incentive to cultivate second sources. Once a competitor clears key qualification thresholds, it may begin winning orders. Even without rapidly capturing a leading share, incremental supply can change the dynamics of the next pricing negotiation. A leading supplier can simultaneously command high market share and strong margins, but neither is guaranteed to remain unchanged indefinitely.
On the supply side, HBM and commodity DRAM compete for resources. Redirecting wafers, equipment, and engineering capacity toward HBM may constrain conventional DRAM supply and indirectly support its pricing. HBM demand can therefore lift high-end revenue directly while also affecting the profitability of other products through resource allocation. This is one mechanism by which AI can amplify the impact on the broader memory industry.
Nominal capacity, however, is not the same as saleable bit output. Effective supply also depends on capacity produced per wafer, process maturity, product mix, and yield. Even with unchanged wafer starts, technology migration and yield improvements can increase effective supply. NAND output is additionally affected by changes in layer counts. Looking only at new fabs or wafer input risks overlooking supply growth driven by productivity.
High profitability therefore contains the seeds of its own correction. Technological complexity and long expansion cycles delay new supply and help industry leaders sustain returns. At the same time, strong profits increase the incentive to expand and encourage customers to seek alternatives. Long-term analysis must consider both sides: how long bottlenecks will persist and whether supply growth will exceed incremental demand once those constraints ease.
For SK hynix, more meaningful competitive indicators include whether next-generation products qualify on schedule, volume-production yields remain stable, orders convert into deliveries, and high-end growth improves companywide profitability. Any market-share figure should also specify whether it measures bits or revenue. Because advanced products command higher prices, revenue share may materially exceed capacity share, and the two measures capture different dimensions of competitive strength.
