VLSI KoreaWhy Nobody Can Forecast Semiconductor Prices — The Structural Reasons McKinsey and Morgan Stanley Keep Getting It Wrong
WSTS forecast 2023 at +5.1%; the actual was -8.2%. TrendForce's first DRAM call for a quarter settled four times higher. Why McKinsey, Morgan Stanley and TrendForce are structurally set up to miss, and what to trust instead.
🆓 Free article · Anyone can read this article for free
Start with the conclusion: semiconductor price forecasting is structurally close to unwinnable. To call a price you need supply and demand at the same time. The variables that set supply — yields and wafer allocation — are disclosed by no one. The variables that set demand — the order plans of the largest buyers — are revised by billions of dollars every quarter. That is why McKinsey, the investment banks and the specialist research firms, spending enormous sums, keep getting it wrong. This article verifies, in numbers, how wrong they have been, how their methods actually work, and what is worth trusting instead.
Three numbers first. One: WSTS, the industry's own statistical body, forecast in June 2022 that the 2023 semiconductor market would grow +5.1%. The actual was -8.2% — a wrong-sign error of 13.3 percentage points. Two: TrendForce, the most trusted source on DRAM prices, first forecast fourth-quarter 2025 conventional DRAM contract prices at +8 to +13%. After two revisions inside the quarter, the settled outcome was reported at +45 to +50% — roughly four times its first call. Three: McKinsey put the 2030 market at $1 trillion in 2022, on a price assumption of about +2% a year. In September 2026 it raised that to $2.3 trillion, and said in its own decomposition that the error was price, not volume.
1. Forecast vs actual: the large errors on record
| Episode | Forecast | Actual | Nature of the error |
|---|---|---|---|
| WSTS 2023 market forecast (June 2022) | +5.1% growth, $680bn | -8.2%, $526.8bn (SIA) | Wrong sign, 13.3 points |
| 2017 memory supercycle | WSTS +3.3%; the most bullish forecaster +11% | +21.6%, memory up roughly 60% | Almost the entire miss was memory price |
| DRAMeXchange 2019 DRAM prices (Oct 2018) | -15 to -20% | Gartner estimated -42.1% in-year | Right direction, about half the magnitude |
| TrendForce 4Q25 conventional DRAM contract prices | +8 to +13% (24 Sep 2025) | Reported settled at +45 to +50% | Revised twice inside the quarter; about 4x the first call |
| NVIDIA, May 2023 | Consensus revenue $7.15bn | Guidance $11.0bn; delivered $13.51bn | Consensus 54% low; even guidance was 23% below delivery |
| Morgan Stanley, Winter Looms (15 Sep 2024) | SK hynix target cut 54% to 120,000 won; 2025 HBM supply 250bn Gb vs demand 150bn Gb | Five weeks later SK hynix posted record earnings; HBM sold out through 2025 | The bank admitted its near-term call was wrong; by Sep 2025 its target was 410,000 won on a supercycle thesis |
| McKinsey 2030 market | $1 trillion (2022), ASP +2% a year | Raised to $1.6 trillion (autumn 2025), then $2.3 trillion (Sep 2026); $1 trillion was passed in 2026 | The ASP path, not volume, was the error |
| HBM market size (Micron) | $25bn for 2025 (Sep 2024) | $30bn three months later, then about $35bn; the $100bn year pulled forward two years | Buyer orders revised upward quarter after quarter |
Fairness requires a qualification. Close to year-end, annual totals are often accurate: WSTS in December 2019 forecast -12.8% for 2019 against an actual -12.1%, and Gartner in December 2023 put 2024 at $624bn against a $627.6bn actual. The failures cluster in three places: at 12 to 18 month horizons, at turning points, and at segment and price level rather than total market level. That matters because the numbers banks and consultants sell are usually exactly in those three places. One more trap: a famous hit makes the next miss more dangerous. Morgan Stanley's August 2021 Winter Is Coming call was right on memory, and that reputation amplified the market impact of its 2024 miss. A past hit guarantees nothing.
2. Supply: the numbers the calculation needs are not public
Saleable supply equals wafer starts times die per wafer times yield times product mix, minus inventory change. The most price-sensitive terms in that identity are not disclosed.
- Yields: No major maker discloses them, ever. Every yield figure in circulation is a leak, a rival's claim, or an analyst estimate, and sources disagree.
- Wafer allocation: The defining supply fact of 2024 to 2026 was the conversion of conventional DRAM wafers to HBM. HBM was 5% of DRAM bits but 20% of revenue in 2024, and uses roughly three times the wafer area per bit of DDR5. No supplier publishes its allocation in advance; outsiders infer it from revenue mix after the quarter ends.
- Utilisation: There is no audited, regular series. Analysts proxy it from gross margin and wafer revenue — estimates stacked on estimates.
- Inventory: Listed makers report one aggregate balance-sheet line, a quarter late, with no split by product or by supply-chain stage. Distributor, module-house and customer-warehouse inventory is not reported at all.
- Equipment lead times: EUV lead times are reported at 12 to 18 months, so this year's leading-edge supply was largely decided 18 to 36 months ago. Near-term supply is extrapolable only if you know starts, mix and yields — which the points above say you do not.
The 2024 Samsung HBM qualification saga shows the opacity live. Reuters reported, from anonymous sources, that Samsung HBM failed NVIDIA tests on heat and power. Samsung denied it. Jensen Huang said the chips had not failed; the work was not done yet. A later pass report was in turn denied by Samsung. For about a year, the single most consequential supply fact in AI memory was unknowable from public sources, with the principals contradicting each other. No outside model can input a variable like that correctly.
3. Demand: harder than supply, and the order data itself lies
Supply at least has physical substance: wafers made last month extrapolate a few months forward. Demand, as the supply chain sees it, is orders — and orders equal true consumption plus inventory adjustment plus double ordering plus strategic stockpiling. That is the bullwhip effect, formalised by Lee, Padmanabhan and Whang in 1997.
- Average chip lead times rose from 13.9 weeks in fall 2020 to a record 26.6 weeks in March 2022. End demand did not double; every buyer ordered earlier and larger to hold a place in the queue. Forecasters then read that inflated order book as demand.
- Analyses put precautionary over-ordering at 10 to 20% above need; in autos in 2022, orders implied about 120 million vehicles against about 83 million in forecast sales.
- By the second half of 2022, consumer chips had flipped to glut while autos stayed short. A single end date for the shortage was a category error: dozens of node and product markets cleared at different times.
- Hyperscaler capex, which now sets demand, is revised intra-year. Alphabet's 2025 plan went from $75bn in February to $85bn in July to $91 to $93bn in October, up 22 to 24% within the year. Meta raised its range as well. A January price forecast embeds a number the buyer itself will revise by October, because boards approve capex quarterly against a moving cloud backlog.
4. Even the published price is mostly not a price
- The spot market is under 10% of the market, by TrendForce's own statement in 2020. Spot and contract prices have moved in opposite directions at the same time. Spot is the residual market, so it overstates both crashes and spikes.
- Most volume moves on B2B contracts. Published contract prices are surveyed settlement ranges for a reference part, collected after negotiation — a part that is a shrinking share of value as server DDR5, LPDDR5X and HBM dominate.
- HBM has no public price at all. It is a custom, qualification-gated component sold inside accelerator packages on annual agreements bundled with volume, generation mix and packaging allocation. Every published HBM price is either supplier revenue divided by estimated bits, or a leaked single transaction. One reported spot print for an HBM3E stack was about four to five times the long-term-agreement level — on the same day. The question of which one is the price of HBM does not survive inspection.
- A published forecast range becomes a negotiating anchor for both sides, so the forecast can move the outcome it predicts. The press then quotes the latest range without its revision history, converting a moving survey into a false point estimate.
5. Methodology map: what they actually see
| Institution | Real method | Data genuinely held | What must be estimated | Structural conflict |
|---|---|---|---|---|
| WSTS / SIA | Bottom-up member billings, with a twice-yearly consensus negotiated by member marketing experts | Member shipments — the best coincident data in the industry | Customers' true consumption; members see orders, not consumption | The forecasters are the suppliers. No member is paid to call a downturn early |
| Gartner / IDC / Omdia | Top-down GDP and end-equipment models cross-checked bottom-up | Unit forecasts for PCs, phones, servers, autos | ASP — the residual that turns units into revenue, and exactly where the error lives | The same forecast is sold to suppliers and buyers; boldness risks credibility, late revision costs accuracy |
| TrendForce / DRAMeXchange | Daily spot collection plus contract-negotiation tracking | Spot prints and negotiation colour from both sides | The contract price itself, until it settles — hence intra-quarter revisions | It sells data to participants in the market whose price range it publishes |
| Sell-side banks | Channel checks plus company models | Colour weeks ahead of published data | Everything else, plus a valuation overlay: a target price is a forecast times a chosen multiple, so target revisions overstate forecast revisions | |
| McKinsey / BCG / Bain | Top-down scenario work: demand pools times semiconductor content times an assumed ASP path | Structural content trends | The ASP path — in 2022, about +2% a year, and the whole error in 2026's own decomposition | A bigger 2030 market justifies bigger client investment programmes and bigger engagements |
All five types drink from the same three wells: WSTS billings, company guidance, and channel colour. Five agreeing forecasts are therefore not five observations; they are one observation cited five times. The actual uncertainty for 2023 looked like the full spread of forecasts, from -3.6% to -22%, not the clustered headlines. Finally, a stark fact: no peer-reviewed study was found that independently audits the long-run accuracy of WSTS, Gartner, TrendForce or bank semiconductor forecasts. The most-cited numbers in an industry approaching a trillion dollars have never had the kind of published accuracy audit that exists for macro or earnings forecasts. The only systematic scorer is a small specialist tracker, Semiconductor Intelligence — and its annual winner rotates, with small firms repeatedly beating large ones. Spend does not buy turning-point accuracy.
6. The most credible source is the buyer — with two corrections
The strongest demand signals come from NVIDIA, Apple and the hyperscalers, for three reasons. They are the demand: NVIDIA's order book is a large share of AI accelerator demand, and Apple has accounted for roughly a quarter of TSMC revenue. Their long-term agreements, prepayments and packaging reservations hand them the supplier's forward book as a by-product — when SK hynix and Micron say HBM is sold out, or Micron says price and volume for its entire 2026 HBM supply are agreed, that is harder information than any bit model. And their internal supply-chain organisations and data infrastructure are larger than most research firms: in May 2023, analysts modelled NVIDIA at $7.15bn for the quarter while NVIDIA guided $11.0bn, a number only the holder of the order book could see.
Related reading
- Market Bulletin: $64 Billion and Counting — TSMC and ASML Say Fab…
- Morgan Stanley's "memory valuation trap" warning — the 3–4x P/E trap…
- Micron's $54.23B Record Quarter: HBM Sold Out Through 2027, DRAM Prices…
- Why Semiconductor Stocks Are Falling: SOX -5.4% and 7 Reasons (Lam,…
📩 Following this story?
Get the follow-up analysis — earnings, reports and HBM supply — in your inbox daily at 10:30 KST. Subscribe free →
Two corrections keep this honest. Buyers know their own demand, not the market's: NVIDIA cannot see its rivals' or China's orders. And buyer and supplier statements are negotiating instruments — talking up scarcity raises the price of the next negotiation. NVIDIA itself delivered 23% more than its own $11.0bn guide, so even the best-informed forecaster does not publish its true central estimate. Read buyer statements as the best demand data available and as interested-party statements, at the same time.
7. A practical framework: trust commitments, discount opinions
Trust signals that are commitments or prints: hyperscaler capex revisions and cloud backlog; long-term agreements and sold-out books with price and volume attached; TSMC monthly revenue and WSTS monthly billings, read for their inflection rather than their level; lead times paired with inventory days and cancellation colour; equipment bookings such as ASML's, a costly signal about supply in two years' time rather than price next quarter; and contract ranges after negotiations have settled — accurate precisely because, by then, they are reports rather than forecasts.
Discount point price forecasts two or more quarters out for DRAM, NAND or HBM, where errors of two to four times at turning points are the historical norm from the best specialist source available; long-range market-size headlines quoted without their scenario range; target prices, which multiply a forecast by a multiple and a mood; spot-price extrapolations from a market that is under a tenth of volume; and precision itself — a supply table stated to the billionth gigabit failed in 2024 on model choice, not arithmetic. In this industry, an extra significant figure is evidence against a forecast, not for it.
A minimal protocol for readers: date and archive every forecast you use, and score it next quarter. Build your view from commitments — capex revisions, accelerator orders, packaging and HBM allocations, wafer starts — and treat price as the residual that clears the last step, with an interval rather than a point. Watch the three dials that turn first: lead times, inventory days, and intra-quarter forecast revisions; when the forecasters start revising, the direction and speed of revision is the forecast. And always ask who is selling what to whom.
Frequently asked questions
Q. Should TrendForce numbers be ignored entirely?
A. No. They are the best public source — but read the revision history, not the first range. In the 4Q25 case, the direction and speed of revision was the forecast.
Q. Why is WSTS accurate by December?
A. Because its system is bottom-up billings. As the year completes, the forecast converges to a sum of actuals — accurate exactly when it stops being a forecast.
Q. Are bank reports useless?
A. Discard the target price; keep the channel checks, the supply tables and the guidance-revision tracking — after checking the model underneath. Modelling a qualification-gated, contract-sold product like HBM as commodity bits produces precise-looking wrong answers.
Q. What is the single best leading indicator?
A. Hyperscaler capex guidance revisions. In the current cycle, the demand shock happens there first; every supply-side number is revised after it.
Sources: SIA and WSTS official releases; TrendForce press releases (2023 to 2026); Gartner releases; McKinsey semiconductor insights (2022; September 2026 revision); Micron earnings materials; SK hynix third-quarter 2024 results; Semiconductor Intelligence annual forecast scoring. Yield figures and customer-level packaging allocations are not disclosed by the companies concerned; where they appear they are press or analyst estimates and are identified as such in the text.
📩 Missed today's analysis? Get tomorrow's in your inbox
Daily Silicon at 10:30 KST plus earnings and analyst-report breakdowns, first. Samsung Electronics, SK hynix, TSMC and HBM supply, read by practicing engineers. Join 300+ subscribers.
Subscribe free →
Free · No spam · Unsubscribe anytime
