[Feature: Hindsight] The Most Famous Call Was the Most Wrong — A Post-Mortem on 2024 AI-Chip Price Targets

On 15 September 2024, Morgan Stanley double-downgraded SK Hynix. The price target went from 260,000 won to 120,000 won — down 54% — and the rating went from Top Pick to Underweight. The report was titled 'Winter Looms.' It was the single sell-side event that split the chip sector that year, and the stock fell sharply the day it landed. And that call ended up, as of June 2026, against 2,636,000 won — roughly 22 times the target.

The easy conclusion here is "the bear was wrong and the bull was right." That tidy verdict is itself wrong. Run the post-mortem and both the bull and the bear missed at the same spot. The bear had its direction reversed; even the bull who got the direction right priced the supercycle's ceiling tenfold too low. The bear's failure of direction and the bull's failure of scale share one root — anchoring — and that anchor was held by bull and bear together. To see what they dropped anchor on that made both extremes miss, I'll score four forecasts against the actuals and pull them apart.

The Scorecard — Grading Four Calls Against the Actuals

To avoid cherry-picking, you can't put only the most famous bear on the stand. Alongside the two extremes of the 2024 opinion spectrum — the bear (Morgan Stanley) and the bull camp — I add the Nvidia consensus as a benchmark, all on one board. The verdicts are not assigned by an editor. They are computed from the contrast between two numbers: the forecast and the actual.

Table: a post-mortem on 2024 AI-chip forecasts. Verdicts are computed from the contrast between the two numbers, forecast and actual. The bull target is not a precise consensus average but Morgan Stanley's pre-downgrade Top Pick (260K) used as a representative bull proxy, and the Nvidia consensus is marked a 'weak hit' because its forecast date is ambiguous. Sources — Morgan Stanley 'Winter Looms' (2024-09, quadruple-confirmed across KED, Bloomberg, Korea Times), sell-side consensus tallies (TipRanks, StockAnalysis), and each stock's market price (2026-06).

The contrast among three of the calls is sharp at first glance. The bear had its direction reversed (SK Hynix 120K → 2,636K); the bull got the direction right but priced the ceiling ten times too low (260K → 2,636K); only the Nvidia consensus landed within a single-digit error ($177 → $193). The fourth, the Samsung bear, is the in-between case. On 14 November 2024 the stock pierced down through Morgan Stanley's 76,000-won level to a trough of 49,900 won — the direction was right in the short term, and even that price level wasn't pessimistic enough. That short-term direction held for two or three months, but the 'winter' thesis collapsed against 323,000 won in 2026. So it scores not as a clean miss but as a partial.

Fairness requires noting the horizon here. The bull target of 260,000 won was effectively reached at the 12-month mark (autumn 2025) — SK Hynix's 52-week low was 245,000 won. So the bull's ~10× shortfall is an error not against the 12-month target but against the cycle ceiling. Even so, the real question is not hit/miss but this: why did even the strongest bull's frame draw the ceiling ten times too low?

The Bear and the Bull Were Two Directions of the Same Mistake

On the surface, 'Winter Looms' and the supercycle (an endless boom) are opposites. Yet the structure of the failure is identical. Both saw memory as a commodity cycle. The premise: prices revert to the mean, and supply eventually catches up with demand. The bear thought that reversion would come soon; the bull thought it would come a little later — that was the only difference.

Take apart Morgan Stanley's downgrade rationale and the frame shows through plainly. 'Q4 2024 is the peak of the memory-chip cycle, and in 2025 HBM supply doubles (with Samsung newly entering as an HBM supplier to Nvidia), the cyclical DRAM shortage ends, and oversupply erodes margins.' Not one sentence is a faulty inference. In a past memory cycle it would have worked exactly that way. The problem is that this cycle was not a commodity cycle.

What broke the frame was that memory did not move like a commodity cycle. The bull camp anchored this market on size — TrendForce put 2024 DRAM revenue at $90.7 billion and HBM at 20% of that revenue. But the heart of this failure was not size. It was that the commodity cycle's premise — prices revert to the mean and supply eventually catches up with demand — did not hold this time. That demand stayed locked up long before mean-reversion could reach it is a plausible reading, but it is too early to pin it on a single mechanism. All that can be asserted is the outcome — that the premise did not work — and that outcome is stamped directly onto the margins in the next section.

That outcome was stamped onto margins. SK Hynix's Q1 2026 operating margin of 72%, Micron's fiscal Q3 2026 gross margin of 84.9% — both are figures with no precedent in half a century of memory history (each company's 1Q26 and 3QFY26 IR). Samsung Electronics, too, posted Q1 operating profit of 57.2 trillion won, 8.5 times the year before (Samsung Electronics 1Q26 IR). This is not a 'tall supercycle' but an extreme that a mean-reversion model cannot explain. Had this been a commodity cycle, oversupply and margin compression should have arrived by now; the opposite was stamped instead — direct evidence that the mean-reversion premise did not work in this cycle.

So scoring the bear and the bull has to go beyond hit/miss. The two were not wrong in opposite directions; they were wrong in two directions of the same frame.

Why Only Nvidia Hit — the Asset the Anchor Barely Gripped

If the same-frame hypothesis holds, then an asset the frame grips only weakly should be forecast well. Nvidia is both the counterexample and the evidence.

Nvidia's 2024 sell-side consensus averaged about $177 (range $135–$220, roughly 42 analysts, after the 10:1 split adjustment). Against a late-2024 price of about $134, that implied 25–30% upside. The actual: $193.23 in June 2026, with an all-time high of $235.47 in May. The consensus hit the current price to within 9%. Even the strongest bull in memory priced the cycle ceiling ten times too low, yet the Nvidia consensus was nearly exact.

The difference lies in the structure of the asset. Nvidia is close to a differentiated quasi-monopoly in the AI-accelerator market — this post-mortem won't fix a precise share figure, but the world's #1 market cap reflects that market position indirectly. For a supplier whose product is differentiated and whose switching costs are high, substitute supply doesn't easily slip in, so the commodity-cycle intuition — 'supply catches up and prices revert' — grips weakly to begin with. The three memory makers, by contrast, are an oligopoly, but their product is relatively commoditized and they carry a history of the capacity race. So the build-out reflex — 'soon someone adds capacity and oversupply arrives' — fired strongly, and this time wrongly, even though the three makers' pricing power itself is, if anything, on the strong side. The contrast — a differentiated quasi-monopoly (Nvidia) versus an oligopoly exposed to commoditization and the capacity race (memory) — is the crux: the more firmly the frame gripped an asset, the more badly it missed.

One caveat has to be flagged. The Nvidia consensus's forecast date is indeterminate — late 2024 or an early-2025 snapshot. If it was struck later, on more information than the memory calls, then its better fit may owe to the timing of measurement rather than to asset structure. That is why it's a 'weak hit,' not assertable as firmly as the memory calls. Even so, the observation points one way — on the asset the anchor gripped weakly, the forecast landed; where it gripped strongly, both extremes missed together.

Citation Volume Was Not a Signal of Accuracy

Here one common assumption breaks. "The most famous report is the most trustworthy."

The single event that defined the sector in 2024 was Morgan Stanley's bear call — the most-cited sell-side report of the year, transcribed at once by KED, Bloomberg, and Korea Times. And it was the call that missed the most. The bull camp's target got the direction right (even if not the scale), and the Nvidia consensus was nearly exact. You can't claim a correlation from a single sample. But at least in this post-mortem, fame was not a signal of accuracy.

The reason is structural. The more marquee the call (the one that draws the most attention), the more it is a contrarian bet running against consensus, and a contrarian bet, when right, is right big — and when wrong, wrong big. 'Winter Looms' double-downgraded a stock that was posting record results: contrarian by definition. Fame measures the boldness of the bet, not its accuracy. Weight your reading by 'the more famous the report, the more trustworthy,' and the data points the other way.

A Forecast Is Not a Point but a Trajectory

The last clue of the post-mortem was left by Morgan Stanley itself. Track how the issuer revised its position and far more information comes out than from a point-estimate target.

Morgan Stanley retreated within five weeks of the downgrade. When SK Hynix posted record third-quarter results, it conceded 'in the short term our outlook was wrong' and nudged the target from 120,000 won to 130,000 won (2024-10-24). In hindsight, the bear was already being neutralized from the very year of the forecast. SK Hynix's 2024 revenue was 66.19 trillion won (+102% YoY), it swung from loss to profit, and HBM revenue jumped 250%. And in 2026, Morgan Stanley reversed completely with 'Memory, Double Up,' raising the SK Hynix target to 1,100,000 won.

And yet even that post-capitulation target (1.1M) fell short of the actual (2.6M). The reversal came a step late, and its magnitude fell short too. That it still missed the scale even after switching wholesale to a structural-shift frame deflates even the hope that a better frame would have nailed the ceiling — a frame can explain the direction and the existence of a ceiling, but it cannot, within the horizon, pin down the size of an unprecedented tail. The pattern — that in a market where mean-reversion intuition runs strong the issuer always climbs up a step behind — is, for that reason, all the sharper.

So a forecast of this kind deceives you if you read it as a single-point snapshot. The speed and direction of the revisions — 120K → 130K → 1.1M — were a more honest signal than the lone point estimate of 1.1M.

So What Should We Read Differently

A post-mortem is not a place to write down lessons but to calibrate the next forecast. I'll land on falsifiable propositions, only as far as the data supports.

First, change the unit of reading. Read not a single point-estimate target but the issuer's speed and direction of revision as the signal — the Morgan Stanley trajectory is the evidence. In the next cycle too, more than the absolute level of a target, how fast and in which direction an issuer bends its position will be the information that moves first.

This calibration has a scope, though. On an asset where the commodity-cycle frame grips strongly — somewhere like memory, where the product is commoditized and exposed to the capacity race — both bear and bull systematically show a bias toward setting the mean-reversion clock too short. So in a bull market it is reasonable to read the consensus with an upward correction. Conversely, for a supplier like Nvidia, differentiated with high switching costs, the consensus already fit well, so the over-assertion that 'the sell-side is always wrong' is not backed by the data.

And one more — the symmetry trap. To declare 'the cycle is dead' in the face of unprecedented margins may itself repeat 2024's anchoring failure with only the sign flipped. So this post-mortem's frame hypothesis is provisional too — the next-generation cycle, with a 2027 peak now being floated, is its test bed. There, does the most famous marquee call again miss the most? Do the issuers again reverse a step late? If both signals repeat, the frame hypothesis is reinforced; if they break, I revise it.

That is as far as the propositions go. One last question remains — who bore the cost of the missed forecast? The investor who dumped SK Hynix following the marquee bear in September 2024 missed the entire surge that followed, and the investor who sold, content at the 260,000-won bull target, missed everything above it. Morgan Stanley did respond — it admitted it was 'wrong in the short term,' and in 2026 it reversed completely with 'Memory, Double Up.' But that response was absorbed into the issuer's reputation, while the loss stayed with the side that moved on the call. A sell-side target is not a fiduciary promise, and there is no invoice to be settled. When a forecast moves the market and still misses, the books of the side that bears the cost and the side that made the call are never reconciled in the end.

The next time some target splits the sector, ask once more: what did this call drop anchor on — a commodity cycle, or a structural shift?