The Forecast Is a Path, Not a Number: What 44 Revision Chains Reveal About Semiconductor Consensus

In June 2022, WSTS projected that the global semiconductor market would reach $680 billion in 2023, growing 5.1% [1]. Eighteen months later, its final word on the same year was $520 billion [2].

Both numbers are correct. Both were published by the same organization about the same twelve months. Between them sit two more prints — $557B in November 2022, $515B in June 2023 — and the sequence is the actual object of interest. A single quote from any one of those four dates would have been cited, at the time, as "the WSTS forecast."

Silicon Analysts archives third-party semiconductor forecasts under the date they were originally published, so a revision becomes a new row rather than an overwrite. The archive currently holds 228 observations across 144 series, reaching back to January 2018. Forty-four of those series have been restated at least once. This is what those 44 chains show.

Forecasters move as a bloc

Group the chains by the year they forecast, and the revisions are close to unanimous.

Every chain targeting 2023 was cut. Every chain targeting 2024, 2025 and 2026 was raised. The single mixed year is CY2022 — the year the cycle turned, when forecasters were revising in both directions at once because they disagreed about when the peak had happened.

The practical consequence is about independence. Citing WSTS and Gartner and SEMI on the same target year reads like triangulation across three sources. The archive suggests it is closer to one signal quoted three times: they revise in the same direction, in the same quarters, in response to the same reported data.

The revisions converge

Among the 24 chains with three or more prints, the median first step is 8.2% and the median final step 3.5%. Sixteen of the 24 shrink from beginning to end.

Gartner's CY2023 revenue-growth chain is the pattern in miniature. In July 2022 it projected a 2.5% decline [3]. By November 2022 that was 3.6%. In April 2023 it jumped to 11.2% — the large late correction. By December 2023, with most of the year observable, it settled at 10.9% [4]. Four prints, a total move of 8.4 percentage points, and almost all of it in one step.

SEMI's equipment forecasts show the same shape at larger amplitude. Its December 2020 forecast for CY2022 equipment sales was $76.1 billion [5]. By July 2022 the number was $117.5B; by December 2022 it settled at $108.5 billion [6] — a 42.6% move from first print to last, most of it arriving in the middle of the target year itself.

This is the honest reading of a two-year-out forecast: it is a placeholder that will be replaced. The forecast that matters is the one published closest to the period, and by then it is barely a forecast.

The largest revision in the archive is happening right now

Convergence describes the ordinary case. The archive's widest move is what happens when the ordinary case breaks.

In June 2025, WSTS put the 2026 semiconductor market at $760.7 billion, growing 8.5% [11]. Twelve months later, in the Spring 2026 release, the same organization put the same year at $1.51 trillion, growing 89.9% [12] — a 98.7% revision to a single target year, and the largest in this archive by any measure.

The revision is almost entirely one segment.

Memory accounts for roughly $589 billion of the $750 billion revision — about 79% of the total move — as the Spring 2026 release puts memory growth at 249.5% year over year on AI infrastructure and HBM demand. Logic, the other AI-exposed segment, contributes most of the rest. Analog, discretes, sensors and optoelectronics were barely touched.

This is the case that makes the argument concrete. Anyone who carried the June-2025 figure into a 2026 model was not slightly off; they were modelling half the market. The number was not wrong when it was published — it was the considered forecast of the industry's own statistics body, and it was superseded by a demand shock that had not yet happened. The only way to see that is to keep both prints.

The record stops before the year does

This is the finding with no obvious workaround.

How long before the target period ended did each forecaster publish its last word? Across the 43 chains targeting a calendar year:

The one remaining chain targets a fiscal year (Micron FY2026) and stopped 165 days before that period ended, so the pattern holds across all 44.

Not one chain in the archive contains a revision dated after its target period ended. Forecasters restate right up to the edge of the year and then stop.

That leaves a specific gap. The last published number silently becomes "what they forecast for 2023" — but it was published in November 2023, with the year not yet complete, and it was never compared by its author to what the year actually produced. The organization that would be best placed to say "we opened at $680B and the year came in at X" is precisely the organization that stops publishing before X exists.

None of this is misconduct. A forecast house sells forward-looking work, and re-publishing a closed year has no commercial audience. But it means the public record of any target year is permanently missing its final entry.

Guidance moves harder than research

Splitting the chains by who published them produces a counterintuitive result.

Company guidance revises roughly twice as hard as third-party research, despite management having incomparably better information about their own spending.

The mechanism is structural, not epistemic. Guidance is a commitment restated on a quarterly clock, so it tracks the operating plan in near-real time. Meta's CY2023 capex guidance stepped $36.5B → $31.5B → $31.5B → $28.5B → $28B across five consecutive earnings releases [7] [8]. Its CY2024 guidance ran the same cadence in the opposite direction, $32.5B → $39B. Alphabet's CY2025 capex went $75B → $85B → $92B over three quarters [10]. A research firm publishing twice a year cannot produce that resolution, and is not trying to.

The widest single move in the archive belongs to neither camp cleanly: TrendForce's forecast for 2024 HBM bit-supply growth went from 105% in August 2023 [9] to 260% by March 2024 — a 155-percentage-point revision inside seven months, as HBM capacity plans were rewritten around accelerator demand.

What this changes about using forecasts

Three things follow for anyone building a model on published forecasts.

Date every forecast you cite. "WSTS projects $680B for 2023" is not a claim about 2023; it is a claim about what WSTS believed in June 2022. Without the as_of date the sentence cannot be evaluated, and a model that inherited that number kept it long after its author had moved 23.5% away from it.

Discount early prints explicitly. If the median first step is 8.2% and the median final step 3.5%, a forecast published two years out should not enter a model at the same weight as one published two months out.

Do not treat multiple houses as independent draws. In this archive they move together, in the same direction, within the same year.

Method and limits

Every row in the archive carries the originator, the metric, the target period, the value, the unit, the original publication date, a source URL, and a verbatim quote containing the number. Rows are collected from primary sources — press releases, IR statements, SEC filings — and where the original page has been replaced, from a dated web archive capture. Nothing is inferred; if a forecaster published a range and no midpoint, the range is what is stored and displayed.

Two limits are worth stating plainly.

This archive records forecasts; it does not grade them. Nothing here says any forecaster was right or wrong, because the archive holds no realized outcomes to compare against. Every figure above describes revision behavior only. Silicon Analysts grades its own projections, which are frozen monthly into write-once rows and scored against realized values once the target period matures; that is a separate, deliberately separate, surface.

And the archive is incomplete by construction. It holds what was found and verifiable — 228 observations, not the industry's full published output. A forecaster absent from a given year is absent from the archive, not from the record. Coverage is deepest where organizations publish on a predictable calendar: WSTS and SEMI (11 chains each), Gartner (7), Meta (5).

The full archive, including the 100 series that have been published only once and are not yet revision chains, is at Forecast Vintages [13] — readable without an account, and queryable through the API and MCP tool with a free key.