{"meta":{"version":"2026-09-09.research-1","preparedAt":"2026-09-09","numericalClasses":["reported","derived","model-estimate","assumption-scenario","unavailable"],"publicBoundary":"Public references, independently calculated estimates and explicit scenario assumptions. Commercial price-board histories and private operational methods are not redistributed.","refresh":"This release uses dated, reviewed snapshots. Models recalculate from the bundled validated inputs; this is not a live ingestion feed.","modelPromotion":"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.","researchAgent":{"name":"VLSI ASTRIA","title":"Semiconductor Intelligence Agent","version":"2026.09","attribution":"Research & model estimates by VLSI ASTRIA","scope":"AI-assisted research with reproducible statistical models and explicit scenario assumptions. Publisher facts retain their original source. Experimental forecasts are not independently validated or guaranteed."},"evaluatedAt":"2026-09-09T16:22:21.083Z","productCount":150,"pricedReferenceCount":49,"forecastSeriesCount":4,"densityScenarioCount":8,"relationshipCount":15,"mixBucketCount":29,"foundryCount":8},"view":"methods","count":7,"records":[{"id":"provider-surveys","title":"Survey methodology benchmark","state":"Public disclosure reviewed","question":"How are the commercial price references segmented?","method":"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.","formula":"Product specification × market basis × observation period","improvement":"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.","limitation":"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.","sources":[{"label":"TrendForce research scope","url":"https://www.trendforce.com/research/dram"},{"label":"DRAM price board","url":"https://www.trendforce.com/price/dram/dram_spot"},{"label":"NAND price board","url":"https://www.trendforce.com/price/flash/wafer_spot"}]},{"id":"quote-normalization","title":"Comparable quote aggregation","state":"Implemented · awaiting quote panel","question":"What does a typical comparable quote look like?","method":"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.","formula":"P = median(comparable, deduplicated quotes)","improvement":"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.","limitation":"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.","sources":[{"label":"Source terms","url":"https://www.trendforce.com/about/terms"}]},{"id":"price-forecast","title":"Reproducible price forecast","state":"Experimental · executable","question":"What range follows from the available exact-reference history?","method":"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.","formula":"P(t+h) = P(t) × exp(projected log change); evaluate absolute error against later observations","improvement":"Publish model version, input period, error metrics, baseline comparison, stale-data checks and horizon. Missing-period estimates and future forecasts are labeled separately.","limitation":"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.","sources":[{"label":"Open DDR3 source","url":"https://www.data.go.kr/data/15051125/fileData.do"},{"label":"Official memory release archive","url":"https://www.motir.go.kr/kor/article/ATCL3f49a5a8c"}]},{"id":"density-scenario","title":"Missing-density price scenario","state":"Assumption scenario · executable","question":"How would a missing density price change under explicit assumptions?","method":"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.","formula":"P(target) = P(anchor) × (Gb(target) / Gb(anchor))^elasticity × premium","improvement":"Every result includes its anchor, assumptions and bounds. It can be recalculated as assumptions change and replaced with actual evidence later.","limitation":"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.","sources":[{"label":"Anchor source archive","url":"https://www.motir.go.kr/kor/article/ATCL3f49a5a8c"}]},{"id":"foundry-mix","title":"Process and wafer reconstruction","state":"Reported + derived + scenarios","question":"How much is revenue mix, physical capacity mix or wafer area mix?","method":"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.","formula":"300mm equivalent = count × (diameter / 300)^2; wafer share(i) = [revenue share(i) / ASP(i)] / Σ[revenue share / ASP]","improvement":"Show physical count share and area share separately. Preserve fab subset, period, capacity versus shipments, and node-specific assumptions.","limitation":"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.","sources":[{"label":"TSMC results","url":"https://investor.tsmc.com/english/quarterly-results/2026/q2"},{"label":"UMC reports","url":"https://www.umc.com/en/IR_Reports/annual_reports"}]},{"id":"customer-products","title":"Foundry → customer → product evidence","state":"Source-linked ledger","question":"Which product is linked to a foundry and how strong is that link?","method":"Store one named relationship and product scope per record with the original announcement date, process if disclosed, internal/external relationship, lifecycle status and source.","formula":"Foundry → counterparty → product or component → process → evidence event","improvement":"Preserve chiplet and packaging boundaries, distinguish tape-out from manufacturing, and keep historical announcements from becoming current allocation claims.","limitation":"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.","sources":[{"label":"NVIDIA Blackwell disclosure","url":"https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing"}]},{"id":"supply-demand","title":"Supply–demand model development","state":"Research specification","question":"Can public operating signals improve sparse-product predictions?","method":"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.","formula":"Demand bits = Σ(system shipments × memory content); good bits = wafer starts × gross dies × yield × bits per die","improvement":"Compare incremental forecast error and interval coverage against no-change and time-series baselines, with a final untouched evaluation window.","limitation":"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.","sources":[{"label":"Public provider research scope","url":"https://www.trendforce.com/research/dram"}]}]}