Everything About AI
NextFin News - South Korea's benchmark KOSPI has fallen 27% from its June peak, yet Goldman Sachs' chief Asia Pacific equity strategist Timothy Moe is keeping his 12,000-point target intact — a call that implies nearly 80% upside from current levels and was already among the most aggressive on Wall Street when he made it three months ago. The market, Moe argues, is making a simple but costly mistake: underestimating how long the artificial-intelligence boom will keep feeding the earnings of the nation's memory-chip giants.
The math behind the call is blunt. Moe applies a 12-month forward price-to-earnings ratio of 7.5 times to Goldman's earnings forecasts. The KOSPI currently trades at 5.3 times forward earnings — roughly half its seven-year average of about 10.6 times. If Korean companies deliver the earnings Goldman expects, Moe says the 12,000 target "isn't as crazy as it might appear." The question for investors is whether the memory supercycle is a structural regime shift that justifies a re-rating, or a cyclical wave that will break before the multiple ever expands.
The Valuation Gap: Why 12,000 Is Not a Fantasy on Paper
Moe's target rests on two legs, and the first is arithmetic rather than optimism. The KOSPI's forward P/E of 5.3x sits about 2.1 standard deviations below its historical norm even after the market's recovery attempts. Goldman's earnings forecasts for the index call for current-year profit growth of roughly 360% — the strongest annual expansion in any Asian market since the recovery from the 1997 Asian financial crisis — followed by another 35% growth in 2027. Apply the 7.5x multiple to those earnings and the index lands near 12,000.
Decomposing the 80% gap makes the bet concrete. The distance from a 5.3x multiple to a 7.5x multiple is about 42% of multiple expansion. The remaining roughly 27% must come from earnings growth being delivered and revised into the forward estimate. In other words, Moe is not asking the market to believe in a bubble multiple; he is asking it to believe in two things only — that Goldman's earnings forecasts land, and that the multiple normalizes to merely 7.5 times, still well below the 10.6-times average. An 80% call built on a below-average multiple is a fundamentally different proposition from an 80% call built on a bubble multiple.
This is where the second-order point matters. Most investors are debating whether memory prices have peaked. Moe is arguing something different: that the earnings base itself has shifted to a permanently higher plane, so even a below-average multiple produces an above-average index level. A market trading at half its historical multiple is not pricing a boom; it is pricing a bust. The entire bet is that the bust narrative is wrong.
"We're still maintaining the target, and the earnings delivery from the companies we track supports it," Moe said in an interview. "The market is underestimating how long this earnings cycle will last."
The concentration of that earnings delivery is extreme. Samsung Electronics and SK Hynix together account for roughly half the KOSPI's weight. When the two memory makers sneeze, the index catches pneumonia — a fact that cut both ways in 2026. The same concentration powered a near-doubling of the KOSPI in the first half of the year, then helped erase 27% of its value from the late-June peak. Both companies briefly entered the trillion-dollar market-capitalization club in May before falling back out as the rout deepened.
The Memory Mechanism: Why This Cycle May Not Behave Like the Last One
Memory chips are the textbook cyclical industry. For three decades the pattern has been reliable: shortage, price spike, capex boom, oversupply, price collapse, capex cut, repeat. Investors who have lived through DRAM cycles are right to be skeptical of any strategist claiming "this time is different." But the structure of the current upcycle differs from its predecessors in at least three ways, and each one pushes the cycle's duration outward.
First, the demand shock is category-creating, not replacement-driven. A single AI server consumes eight to ten times more DRAM than a traditional server, and more than three times the NAND flash storage. Global AI server shipments are projected to reach 1.5 million units in 2026, up roughly 180% year over year, according to industry estimates. This is not consumers swapping phones; it is an infrastructure build-out measured in hundreds of billions of dollars of hyperscaler capital expenditure.
Second, supply has been structurally withdrawn from the commodity end of the market. Samsung, SK Hynix and Micron — which together control more than 95% of global DRAM production, according to IDC — have reallocated manufacturing capacity toward high-bandwidth memory, or HBM, the specialized DRAM that AI accelerators require. HBM consumed 23% of total DRAM wafer output by April 2026, up from 19% in 2025, TrendForce data show. Every wafer devoted to HBM is a wafer not making conventional DDR4 and DDR5, tightening the entire memory market even as AI demand explodes. NAND manufacturers are converting production lines to DRAM to chase the higher margins, which tightens storage in parallel.
Third, the pricing evidence already confirms the squeeze. DRAM spot prices have surged nearly 700% over the past year. NAND contract prices climbed 55% to 60% quarter over quarter in the first quarter of 2026, with a further 70% to 75% jump materializing in the second quarter, according to TrendForce and TechInsights. These are not marginal increases; they are the signature of a market where demand has outrun supply across both major memory categories simultaneously.
The funding behind that demand is equally consequential. Goldman Sachs research estimates U.S. Big Tech capital expenditure will reach roughly $1.2 trillion in 2027, up from about $750 billion expected for 2026. That spending is the demand engine for HBM, enterprise SSDs, and the memory embedded in every AI accelerator. Moe's call is, at its core, a derivative bet on that capex number holding.
The Second Structural Pillar: Korea's Discount and the Tax-Reform Lever
The memory cycle is not the only structural argument in Moe's favor. South Korea's equity market has traded at a persistent discount to Japan's TOPIX, the S&P 500 and Europe's STOXX 600 for years — a "Korea discount" rooted in governance concerns, low shareholder returns and the geopolitical risk premium. More than 70% of KOSPI-listed companies still trade below one times book value, a far higher share than in Japan or the United States.
Seoul is now attacking that discount through the Value Up corporate-governance program and accompanying tax reform. Under the 2026 reform bill, companies with a price-to-book ratio below 0.8 will have inheritance and gift taxes calculated on 80% of net asset value rather than on the depressed market price — a direct incentive for controlling families to stop suppressing share prices, since the tax bill no longer rewards a low quoted valuation. Anti-share-price-suppression measures are being written into the same legislation.
The mechanism here is distinct from the memory cycle but complementary in its effect on the index. Memory earnings lift the numerator of the valuation equation; governance reform and tax incentives support the denominator by narrowing the discount investors demand. If even a fraction of the 70% of sub-book-value companies re-rate toward one times book, the index-level impact compounds with the earnings recovery. This is why an 80% call is not solely a bet on DRAM prices: it is a bet on two independent re-rating channels working in the same direction.
Positioning leaves room for that re-rating to be funded. Goldman Sachs has noted that holdings of South Korean equities within global and emerging-market funds remain underweight, meaning the buyer base for a recovery is not fully invested. Foreign flows were the accelerant on the way up in the first half of 2026 and can be the accelerant on the way back if earnings delivery and capex guidance cooperate.
The Counter-Thesis: Concentration, Debt, and the Cycle That Bites Back
The strongest case against Moe does not question the memory boom. It questions whether a 50%-concentrated index is the right instrument to express it, and whether the funding behind the boom is as durable as the demand.
Concentration risk is the immediate objection. Jonathan Krinsky, chief market technician at BTIG, noted during the June rally that the KOSPI could rise more than 12% over six sessions while market breadth was negative every single day — the signature of an index being carried by a handful of names rather than broad participation. When Samsung and SK Hynix together make up about half the KOSPI, an 80% index call is really a call on two stocks and the memory price cycle that drives them. The other roughly 800 listed companies are passengers, and a rally that does not broaden is fragile by construction.
The funding objection runs deeper. Hyperscalers are increasingly financing their build-out with debt rather than cash. Goldman Sachs projects that in 2027 more than a third of Big Tech AI investment will be debt-funded, with roughly $400 billion of global investment-grade bond issuance from the hyperscalers against about $1.14 trillion in capex. Alphabet already reported its first quarter of negative free cash flow since its 2004 IPO, and Meta has leaned on partnership structures — such as a data-center project in which BlackRock holds an 80% stake funded partly with a $12.5 billion debt financing — to stretch its balance sheet. If financing costs rise or AI revenue fails to monetize at the pace the capex assumes, the $1.2 trillion spending number becomes the first variable to shrink — and with it, the memory demand that underpins Moe's earnings forecasts.
There is also the valuation trap. The KOSPI's forward P/E looks cheap only if forward earnings actually arrive. The gap between the index's trailing P/E and its forward P/E is unusually wide, which means analysts are assuming exceptionally rapid earnings growth with little room for disappointment. If AI demand slows, the multiple and the earnings estimate can contract simultaneously — the classic double compression that turns a "cheap" market expensive in hindsight.
Finally, China remains the wildcard. A low-cost AI model breakthrough could reduce the amount of Western compute infrastructure required per unit of AI capability, pressuring the capex assumptions underneath the entire memory trade. Chinese producers CXMT and YMTC hold an estimated 5% to 10% of DRAM capacity, mostly in lower-end segments, and lack leading-edge HBM capability today — but the memory industry's history is written in the blood of incumbents who assumed their technological lead was permanent.
Cyclical or Structural: The Call That Decides the Trade
Here is the judgment the market has to make, and it is worth stating plainly: the memory upcycle is cyclical in its price path but structural in its demand floor. These are not the same thing, and conflating them is how investors lose money on both sides.
The price path is cyclical. DRAM and NAND prices will not rise in a straight line forever; at some point capacity will catch up, inventory will rebuild, and spot prices will give back some of their gains. That is the nature of the industry, and no amount of AI enthusiasm suspends it. Evidence for the cyclical leg: three decades of boom-bust history, the current 700% spot-price surge that will inevitably invite capacity responses, and the fact that HBM still represents less than a quarter of DRAM wafer output — meaning the commodity end of the market can still swing violently.
But the demand floor is structural. AI inference and training require a category of memory that did not exist at scale five years ago, and the compute-to-memory ratio of an AI workload is an order of magnitude higher than a conventional workload. This is a regime change in the memory intensity of computing, comparable to the shift from feature phones to smartphones in the 2000s. Even if prices mean-revert from their peaks, the baseline level of memory demand per unit of compute has shifted up permanently. Samsung, SK Hynix and Micron executives have stopped describing the shortage as cyclical; they now talk about forward visibility, committed AI capacity and multi-year capital planning.
The practical implication for Moe's call is that the earnings base does not need memory prices to stay at their peaks. It needs them to stay well above the pre-AI trough. That is a materially easier hurdle, and it is what makes a 7.5x forward multiple defensible rather than heroic. The market is pricing the cyclical peak-to-trough swing; Moe is betting the structural floor sits far above the trough the market fears.
What to Watch: The Signals That Confirm or Break the Thesis
The thesis stands or falls on three observable signals over the next two quarters. First, HBM's share of DRAM wafer output: if it keeps climbing past 23% while conventional memory supply stays tight, the structural capacity-withdrawal argument strengthens. Second, hyperscaler capex guidance for 2027: any material downward revision from the $1.2 trillion trajectory is the single most direct falsifier of the demand premise. Third, the KOSPI's forward P/E: a move back above 7x without earnings downgrades would signal that the multiple expansion Moe needs is actually beginning.
The falsifying signal is specific: if U.S. Big Tech capex guidance for 2027 is cut by more than 15% from current estimates, or if DRAM spot prices fall more than 30% from their 2026 peaks within two consecutive quarters, the structural-demand thesis is wrong and the 12,000 target becomes a cyclical mirage.
Scenarios split cleanly by time horizon:
- Short term (sentiment and liquidity): volatility dominates. Margin loans in Korea hit a record 38.5 trillion won in late June, more than double the level a year earlier, and forced selling can amplify any AI-capex headline. In this window the index can overshoot in both directions regardless of fundamentals.
- Medium term (earnings delivery): this is where Moe's call is won or lost. If the roughly 360% current-year earnings growth forecast lands and the 35% 2027 growth follows, the index re-rates toward the target even at a sub-average multiple. A miss on either number breaks the arithmetic.
- Long term (structural resolution): if AI memory intensity proves permanent, Korea's valuation discount narrows and 12,000 becomes a waypoint on a longer re-rating; if it proves cyclical, the discount widens again and the target becomes a multi-year ceiling rather than a destination.
The bottom line: Moe's 80% call is not a bet that memory prices keep climbing. It is a bet that the market has mislabeled a structural shift as a cyclical spike — and that a 5.3x forward multiple is the price of that mislabeling. If he is right about the duration of the earnings cycle, 12,000 is not aggressive. It is late.
Explore more exclusive insights at nextfin.ai.
