Research and estimates by VLSI ASTRIA: AI-assisted public-source research, reproducible statistical models and explicit assumptions. Original publisher facts remain separately attributed. Forecasts are experimental.
A candidate can be described as improved only after chronological out-of-sample validation against a declared baseline, with errors, horizon, sample size and interval coverage published. No superiority over commercial providers is claimed.
Public disclosure reviewedSurvey methodology benchmark
How are the commercial price references segmented?
TrendForce / DRAMeXchange separate spot, contract, module, mobile memory and SSD references, with product-specific publication schedules. Public service descriptions cover supply, demand, capacity and price outlooks.
Product specification × market basis × observation period
Product records preserve the stated specification, object, size, unit, basis and publication date. The next schema step is structured speed, organization, grade and package fields; the quote engine already checks exact specification signatures before aggregation.
The reviewed public pages do not disclose a reproducible panel, sample size, weights or full forecasting algorithm. No claim of reproducing their proprietary methodology or outperforming their forecasts.
Implemented · awaiting quote panelComparable quote aggregation
What does a typical comparable quote look like?
Deduplicate within an exact specification, period, currency, unit and price basis; calculate a median and dispersion with source counts. Quotes remain offers or survey references unless transaction evidence exists.
P = median(comparable, deduplicated quotes)
The engine returns a deduplicated observation count, median, low/high values and a reason when a cell is incompatible. Source-panel concentration and a full rejection ledger are development targets.
No independent dealer panel is currently connected. A median of asking prices is not a transaction price. Public commercial-board visibility does not establish a reusable price-history feed.
Experimental · executableReproducible price forecast
What range follows from the available exact-reference history?
Compare no-change, recent log drift and damped log trend in chronological walk-forward evaluation. Keep each product and native frequency separate and return empirical error bands with sample counts.
P(t+h) = P(t) × exp(projected log change); evaluate absolute error against later observations
Publish model version, input period, error metrics, baseline comparison, stale-data checks and horizon. Missing-period estimates and future forecasts are labeled separately.
Small samples and regime changes can dominate results. Latest-vintage evaluation is retrospective, not an as-published trading simulation. Error bands are empirical and do not guarantee future coverage.
Assumption scenario · executableMissing-density price scenario
How would a missing density price change under explicit assumptions?
Anchor to a dated public reference in the same generation and object class. Change density with a user-selected elasticity and premium range. Keep the output as a generic sensitivity scenario, separate from manufacturer products.
P(target) = P(anchor) × (Gb(target) / Gb(anchor))^elasticity × premium
Every result includes its anchor, assumptions and bounds. It can be recalculated as assumptions change and replaced with actual evidence later.
The elasticity and premium are illustrative, not fitted. The range is not a statistical prediction interval. Does not infer a speed-bin, vendor SKU, LPDDR, GDDR, HBM stack, or finished-module market price from commodity DRAM.
Reported + derived + scenariosProcess and wafer reconstruction
How much is revenue mix, physical capacity mix or wafer area mix?
Keep issuer revenue shares on their original denominators. Convert disclosed physical wafer counts by diameter to a common area basis. When only revenue is known, allow relative wafer-ASP assumptions to produce a scenario.
300mm equivalent = count × (diameter / 300)^2; wafer share(i) = [revenue share(i) / ASP(i)] / Σ[revenue share / ASP]
Show physical count share and area share separately. Preserve fab subset, period, capacity versus shipments, and node-specific assumptions.
Revenue alone cannot identify wafer counts. Capacity is not production; utilization, yield, outsourcing and time-varying ASP remain separate inputs. Individual share bounds need not sum to 100% because they describe different scenarios.
Source-linked ledgerFoundry → customer → product evidence
Which product is linked to a foundry and how strong is that link?
Store one named relationship and product scope per record with the original announcement date, process if disclosed, internal/external relationship, lifecycle status and source.
Foundry → counterparty → product or component → process → evidence event
Preserve chiplet and packaging boundaries, distinguish tape-out from manufacturing, and keep historical announcements from becoming current allocation claims.
The ledger is incomplete. A named customer is not a customer share; absence from public sources is not proof of no relationship. Unknown allocation remains unknown unless a disclosed denominator or explicit scenario inputs support it.
Research specificationSupply–demand model development
Can public operating signals improve sparse-product predictions?
Candidate inputs are issuer bit-shipment growth, inventory days, utilization, disclosed capacity, end-market shipments and product memory content. Test a regularized log-price model using only information published before each origin.
Demand bits = Σ(system shipments × memory content); good bits = wafer starts × gross dies × yield × bits per die
Compare incremental forecast error and interval coverage against no-change and time-series baselines, with a final untouched evaluation window.
Not yet fitted or deployed. Physical yield, dies per wafer and allocations require evidence or explicit ranges; public customs values do not identify exact product ASP or bit shipments. No accuracy improvement is claimed before evaluation.
This release uses dated, reviewed snapshots. Models recalculate from the bundled validated inputs; this is not a live ingestion feed. Public references, independently calculated estimates and explicit scenario assumptions. Commercial price-board histories and private operational methods are not redistributed.
Market