Silicon Analysts

Returns 18 AI accelerators (H100/H200/B100/B200/GB200/GB300/Rubin, MI300X/MI355X/MI455X, Gaudi 3, TPU v5p/v6e, Trainium 2/3, Maia 100, MTIA v2) with structured fields: chip, vendor, processNode, dieSizeMm2, memoryType, memoryCapacityGb, memoryBandwidthTbS, fp8TflopsSparse, bf16TflopsDense, packageType, estMfgCostUsd, estSellPriceUsd, chipGrossMarginPct, costBreakdown.{logicDie, hbm, packaging, testAssembly}, interconnect. USE THIS for: comparing manufacturing cost or sell price across vendors; looking up published specs of a current accelerator (includes early-ramp 2026 parts like Rubin and MI455X, flagged via provenance.confidence_tier). DO NOT USE for: chips not in the catalog (use get_market_pulse for market news/forecasts); custom chip cost modeling (use calculate_chip_cost); HBM market dynamics (use get_hbm_market_data). Filters: vendor (enum), chip (substring match), fields (projection list). Returns empty array if filters match nothing — does not error. Each chip record carries provenance.last_updated; data refreshes monthly.

Pure-function chip cost estimator. Given die dimensions (mm), process node, and optional packaging/HBM parameters, returns: estimatedChipCost (USD), dieArea (mm²), grossDiesPerWafer, frontendYield (%), totalYield (%), and a costBreakdown {waferCostPerGoodDie, packagingAndTestCost, hbmCost, marginCost}. USE THIS for: hypothetical chip cost modeling, sensitivity analysis, fabless tapeout decisions. DO NOT USE for: published cost of an existing accelerator (use get_accelerator_costs); wafer pricing only (use get_wafer_pricing). Required: dieWidth, dieHeight (1–33 mm reticle limit). Errors with INVALID_PARAMS if outside bounds. processNode defaults to tsmc-n5; valid nodes via get_wafer_pricing. Estimates are directional ±15–20%. Optional energy adder: pass energyRegion (texas|ohio|arizona|china|korea|taiwan|germany) to get a conditional `energy` block — regional manufacturing-electricity cost per die (SA estimate; wafer price already embeds foundry energy, so treat it as a scenario delta). energyFacilityOverhead=false drops the ~1.75× facility multiplier. Optional substrate scenario: pass substrate=panel-310x310 to model CoPoS panel-level assembly — applies the midpoint of Yole's realistic 20–30% panel cost-savings band to the packaging cost ONLY (silicon GDPW unchanged; panels are back-end). Conditional `substrateScenario` block + SA-scenario meta note. TSMC CoPoS: pilot ~June 2026, mass production 2028–29 — a forward-looking scenario, not a quote.

Heuristic chip manufacturing LEAD TIME estimator (MANUFACTURING CYCLE TIME). Given total mask layers (or a processNode to default them), foundry utilization % (optional — defaults from live foundry-allocation data), and packagingType, returns min/max bands: fabDays, fabWeeks, packagingWeeks, totalWeeks, plus effectiveDpml (days per mask layer), the operating-curve weight, a resolved-inputs echo, assumptions, methodology, and public-source citations. USE THIS for: "how long to manufacture this chip" — wafer-fab cycle time + packaging assembly/test time for hypothetical chips; cycle-time sensitivity to fab utilization or packaging class (conventional vs flip-chip vs CoWoS). DO NOT USE for: booking windows / allocation lead time — how long until a booked-out foundry STARTS wafers, publicly 52–156+ weeks at N3-class nodes and CoWoS (use get_foundry_allocation); chip cost (use calculate_chip_cost / get_accelerator_costs). Provide maskLayers (integer 10–200) or processNode (tsmc-n3 | tsmc-n5 | tsmc-n7 | tsmc-n16 | tsmc-28 | samsung-3nm | samsung-5nm | samsung-7nm | samsung-14nm | intel-7 | intel-16 | gf-12lp | gf-fdx | umc-22-28 | umc-40 | smic-28). packagingType accepts coarse classes (conventional | flip-chip | cowos, default flip-chip) or any platform packaging id (fc-bga, wirebond-bga, cowos-l, copos, ...). Utilization ≤80% settles at the best-case band; ≥95% converges to the worst-case bound (FabTime operating-curve shape). Heuristic from public DPML benchmarks — directional, confidence LOW, not a foundry quote. Cite as "Silicon Analysts — Lead Time Estimator".

Returns 10 HBM market sub-tables: accelerators, specs, marketShare, spotPrices (RETIRED 2026-07-28 — frozen), leadingIndicators, qualificationFeed, revenueForecast, supplierRevenue, validationChecks, bitDemand. Optional `table` parameter narrows to a single sub-table; omitting returns all 10. USE THIS for: HBM3/3e/4 generation specs, SK Hynix/Samsung/Micron share of HBM revenue (marketShare: Q2 2026, Counterpoint Research, the latest quarter; the quarterly series and IDC full-year actuals are get_market_dataset dataset='hbm-market-share'; not share of bit shipments), derived HBM bit demand by SKU class and customer type (bitDemand, EB ranges, monthly). spotPrices is a RETIRED series: no public HBM spot market exists in any generation — HBM sells via annual/multi-year LTAs. Its rows are frozen estimates served for the record with per-row retired marking; for current, sourced HBM pricing use get_market_dataset with dataset='hbm-pricing'. bitDemand is NOT a workload split — it is a SKU-class/customer-type cut. Dominant HBM SKUs are dual-use, so a training-vs-inference HBM attribution would be dishonest; no public source publishes one. DO NOT USE for: HBM price history or current HBM pricing (use get_market_dataset dataset='hbm-pricing'); per-accelerator HBM cost in a specific chip (use get_accelerator_costs.costBreakdown.hbmCostUsd); HBM cost in a hypothetical chip cost calc (use calculate_chip_cost with hbmStacks/hbmCost). Returns INTERNAL_ERROR if the upstream Supabase HBM tables are unreachable. Only two sub-tables refresh on a schedule: qualificationFeed (Mon/Thu) and bitDemand (monthly, 1st). accelerators, specs and validationChecks are hand-curated reference tables, the editorial tables are revised by hand, and spotPrices is frozen (retired 2026-07-28). meta.table_freshness dates each sub-table; meta.data_freshness.stale flags an overdue refresh.

Returns curated supply-chain headlines with trend direction (up/down/neutral), source attribution, and impact analysis. Categories: logic, memory, packaging, connectivity, power, geopolitics. Defaults to all categories, all trends, no limit. USE THIS for: "what's happening in HBM this quarter?", "any geopolitical moves affecting TSMC?", recent supply/demand inflections. DO NOT USE for: structured pricing data (use get_wafer_pricing, get_hbm_market_data); published cost of a specific chip (use get_accelerator_costs). Per-item dates are formatted strings (e.g., "Jan 2026") — not ISO 8601. Cache: 5 minutes server-side. Returns empty array if all items filtered out.

Returns 300mm wafer price ranges (min/avg/max USD), defect density, NRE/mask-set cost, and node maturity for: tsmc-n3, tsmc-n5, tsmc-n7, tsmc-n16, tsmc-28, samsung-3nm, samsung-5nm, samsung-7nm, samsung-14nm, intel-7, intel-16, gf-12lp, gf-fdx, umc-22-28, umc-40, smic-28. Optional `node` filter narrows to one. ALWAYS read `citation` before using a price in a cost model: it names the corroborating sources and carries the caveat that decides whether the number is usable. Some sellers report a foundry segment operating loss, so their quote is a positioning price rather than a cost-recovering one; `citation` says so explicitly, and `defectDensity`/`nreCost` are null where no public basis exists rather than being estimated. USE THIS for: looking up wafer cost for cost modeling, comparing foundries at the same node. DO NOT USE for: per-chip cost (use get_accelerator_costs or calculate_chip_cost); packaging-related cost (use get_packaging_costs). Returns INVALID_PARAMS if node is not in the valid set. Each record carries the source attribution string. Refreshes monthly.

Returns two sub-arrays: `packaging` (per-tech cost benchmark + capability matrix for CoWoS-S/L, EMIB, SoIC, InFO-PoP, FC-BGA, FC-CSP, etc.) and `hbmSpecs` (HBM2 through HBM4 cost per stack + bandwidth/capacity). Optional `type` filter narrows packaging array to one technology. USE THIS for: packaging cost lookup, comparing CoWoS variants, getting HBM stack pricing for cost modeling. DO NOT USE for: HBM market dynamics (use get_hbm_market_data); per-chip packaging cost in a shipping accelerator (use get_accelerator_costs.costBreakdown.packagingCostUsd). Returns INVALID_PARAMS for unknown type. Refreshes monthly.

Sourced HBM qualification tracker: which memory vendor (SK Hynix, Samsung, Micron) passed which AI-accelerator customer's qualification (NVIDIA Vera Rubin/GB300/B300/H200, AMD MI350/MI325X, Broadcom), by generation (HBM3/HBM3E/HBM4) and stack height. Returns `matrix` (current status per vendor×customer×generation, each row dated + source URL + confidence) and `timelines` (per-relationship status-change history back to 2022, e.g. sampling → in_qualification → qualified → volume_shipping). Refreshed Mon/Thu 09:00 UTC; a status change auto-publishes once corroborated (a primary source, or 2+ independent sources that agree) and is not human-reviewed before publication. meta.data_freshness.stale reports whether the refresh is current; superseded timeline events carry superseded: true and the published correction. USE THIS for: "who supplies HBM4 for Vera Rubin?", "did Samsung pass NVIDIA qualification?", "Micron HBM4 status", qualification timeline/history questions, HBM supply-eligibility analysis. DO NOT USE for: HBM pricing/market share (use get_hbm_market_data); per-chip HBM cost (use get_accelerator_costs). Filters: vendor (enum), customer (substring), generation (enum), include_timelines (boolean). Anonymous callers may receive timelines truncated to the latest event per relationship — full history with a free API key (https://siliconanalysts.com/developers). Cite as "Silicon Analysts — HBM Qualification Tracker".

Foundry & advanced-packaging ALLOCATION — the current-state snapshot per node/tech (TSMC/Samsung/Intel/... × N2/N3/CoWoS-L/SoIC/...) plus optional time-series HISTORY. Current fields: allocation_status (fully_booked → available), lead_time_weeks_min/max + trend, utilization, price_trend, geo_risk, customers, capacity_current/target, customer_shares, allocation_note. With include_history=true, returns the tracked series from capacity_signals: lead_time / booking-window, pct_locked (%-capacity-locked), customer_allocation (publicly-reported per-customer share), cowos_capacity, foundry_utilization — each point dated (as_of) with provenance. No competitor publishes allocation as a structured, queryable feed. USE THIS for: "who has CoWoS allocation and how much?", "what's the booking lead time for N2?", "how locked is 2026 CoWoS capacity?", allocation/lead-time trend over time. DO NOT USE for: per-chip cost (use get_accelerator_costs / calculate_chip_cost); HBM market share/pricing (use get_hbm_market_data); HBM qual status (use get_hbm_qualification). Filters: foundry, node, category, customer, history_metric, include_history (bool), limit. Sourced public estimates (analyst/press/earnings), human-reviewed; every record carries provenance.confidence_tier. wafer_price is intentionally omitted. Cite as "Silicon Analysts — Foundry Allocation".

Foundry IR ECONOMICS — per-foundry, per-process-node, per-fiscal-quarter wafer ASP (min/max/blended, USD per 300mm-equivalent wafer, $250-grained) and fab UTILIZATION (%), derived exclusively from PUBLIC IR materials (earnings releases/transcripts/decks, trade press) via a documented scaling calculation (rev-mix-v1): reported revenue × reported node revenue-shares × reported wafer shipments, allocated on pinned analyst prior ratios. Covers tsmc | umc | intel | samsung | smic | gf. Every row carries source_urls + release_dates + confidence (high/medium/low); utilization is 'stated' (company said it — UMC/SMIC style) or 'derived' (shipments vs capacity estimate, capped medium) and NEVER fabricated per node. include_facts=true returns the underlying evidence facts (verbatim quote + source per datum). Also returns node_margin_estimates for TSMC: per-node est. wafer price / est. wafer cost / est. GROSS MARGIN % with ranges (N3/N5/N7/N16/N28+/N2) — single-vintage Silicon Analysts ESTIMATES from public analysis, explicitly labelled (TSMC does not disclose per-node margin; company-level GM is quarterly IR). USE THIS for: "what does a TSMC 3nm wafer sell for and how has it moved by quarter?", "TSMC blended ASP trend", "UMC utilization last quarter", "N3 share of TSMC revenue over time", "estimated gross margin by node", node-economics history for models. DO NOT USE for: the current spot wafer price band only (use get_wafer_pricing — that is the live analyst-consensus band this dataset cross-validates against); allocation/lead-time/booking (use get_foundry_allocation); chip-level cost (use calculate_chip_cost / get_accelerator_costs). Filters: foundry, node (canonical token, e.g. n3 | 22-28nm | 18a), node_group (leading_3nm | class_5nm | ...), quarter (2026Q1 | 2025FY), from/to range, include_facts, limit. LATEST period per foundry is free; multi-period HISTORY (quarter/from/to) requires a Pro key — free callers are clamped to latest with an explanatory meta.note (never a

"What Changed" — recent MOVEMENTS in Silicon Analysts' public data over a 7d/30d window, derived from the daily snapshot ledger. Each moved metric returns direction (up/down), magnitude (pct_delta for value metrics, pp_delta for percentage metrics), old/new values, the two snapshot dates compared (as_of, prior_as_of), window_days_actual (the REAL lookback — the ledger is young, so a 30d window clamps to available history), and per-record provenance. Domains (the datasetId values): wafer_pricing, chip_cost, gpu_secondary, margin_benchmark, foundry_capacity, defect_density, nre_cost, packaging_benchmark, chip_archetype, electricity_price, cloud_pricing, llm_pricing, memory_spot, foundry_economics, market_prints, fab_capacity, hbm_market. USE THIS for: "what moved in semiconductor costs this week?", "did any wafer prices change recently?", "what changed since my last fetch on June 20?" (use since), building a market-change digest, monitoring deltas across the data layer over time. DO NOT USE for: current absolute values (use get_wafer_pricing / get_accelerator_costs / get_foundry_allocation); allocation lead-time trend specifically (use get_foundry_allocation with include_history). Filters: window (7d|30d), since (ISO date — compare the latest snapshot against the newest snapshot at/before it; overrides window for baseline selection), datasetId (one domain), minDelta (override the significance threshold), limit. N2/Apple omitted (conflict-safe). Returns an empty array when nothing moved past the significance gate — does not error. Cite as "Silicon Analysts — What Changed".

Curated market-data TIME SERIES with per-point sourcing — the datasets behind siliconanalysts.com/market-data. Includes: hbm-pricing (HBM contract + blended $/GB by generation, HBM2→HBM4, anchors 2017→2026 — series_keys like 'hbm3e-contract'; NOTE: no public HBM spot market exists — HBM sells via LTAs, and the dataset says so rather than fabricating a spot curve), component-lead-times (CoWoS-S/CoWoS-L/HBM3E/TSMC-N3 lead times in weeks back to 2022), wafer-price-tsmc (wafer price by node back to 65nm), semiconductor capex, DRAM/NAND pricing, and more. Every point carries value_low/mid/high, confidence, data_type (Confirmed|Estimate|Projection), source_name, source_date, source_note — estimates are typed as estimates, never dressed as observations. USE THIS for: HBM contract price history by generation and basis ("what did HBM3E contract $/GB do through the 2023 shortage?" — note the revenue-implied vs per-stack bases are ~1.7x apart and must not be compared across series), lead-time trend series, wafer price history by node, memory price cycles — any question needing the dated SERIES rather than the current snapshot. DO NOT USE for: current HBM market snapshot (use get_hbm_market_data); current wafer price bands (use get_wafer_pricing); IR-derived per-node ASP/utilization (use get_foundry_economics); allocation status (use get_foundry_allocation). Params: dataset (id; pass 'list' to enumerate the catalog), series_key (optional filter, e.g. 'hbm3e-contract'). Tiering: anonymous → recent points; free key → recent + newest-3-per-series history preview; Pro → complete series. Each dataset returns `methodUrl` — a published page describing HOW the series was built (typing rules, derivations, deliberate gaps) — or null; read it before reasoning about modelled points. Pro callers can pull the whole series as flat CSV/JSONL in one call: GET /api/v1/export?dataset=<id>&format=csv. Cite as "Silicon Analysts — Market Data".

Market Intelligence — the freshest SOURCED semiconductor market briefs, generated daily from a Tavily + Claude scan of primary press, earnings, and trade outlets. Each brief returns title, severity (Critical/High/Medium/Low), confidence_score (0-100), quantitative_impact (e.g. "HBM4 12-Hi 36GB contract price: $600 (2026) → $1,300 (2027) per stack"), an executive summary, a short analysis, a category (Logic/Memory/Packaging/Connectivity/Power/Geopolitics), and a curated sources[] list — plus per-record provenance. UNIQUELY: each brief also carries `entities` (the chips/nodes/packaging/HBM-gen/companies it concerns), `impact` (when it's a component cost move, the per-chip BOM dollar deltas computed from Silicon Analysts' cost models — e.g. "HBM3E +20% for 2027 → +$650 on B200" — with each chip's line_usd → new_line_usd, the year and basis its HBM line is priced at (priced_at), and a pre-filled calculator URL; an HBM move is added only to lines priced BEFORE its year, so a rise our models already carry is never counted twice), `impact_assessment` (why impact is or is not modeled: status modeled | not_added | not_a_component_price | mix_driven | no_price_stated | implausible | year_unknown, the parsed move — kind, pct, year, HBM generation, absolute level + basis — and not_modeled[] chips with their reason: legacy | already_priced | different_basis | implausible), `related` (cross-links to the live datapoints + tools), and `novelty` (when the HEADLINE fact first became public as the scanner could VERIFY it — verdict fresh/dated/stale/unknown, the verified first_public_date + source, and dataset_match when the figure was already in Silicon Analysts' data, i.e. the brief is a recap; a Critical is only ever stored when the fact is verifiably ≤7 days old). No pure-news source does this. The machine feed returns ALL severities; `published` flags the Critical/High briefs that also have a public page. Public sources only; no insider data. USE THIS for: "what's the latest in H

Historical Benchmarks — the bitemporal benchmark-observations ledger behind the Chip Cost Calculator: wafer cost by node/foundry (deflationary curves), defect-density (D0) learning curves per node, advanced-packaging costs incl. the broken-out CoWoS interposer entity, test cost, backend yield, and HBM $/GB. Each observation carries as_of (the date the reading reflects — curated backfill from dated public archives extends history), detected_at (capture time), and full sourcing metadata (source_type taxonomy: foundry_ir | wfe_vendor_earnings | government_filing | press_release | analyst_report | company_announcement | trade_press | public_web; source_url; confidence high/medium/low). grain=month|quarter returns median/min/max rollups per period; grain=raw returns per-source observations. Access tiers: free key → preview, Pro/Enterprise → full ledger, anonymous → none. USE THIS for: "how has TSMC N5 wafer pricing moved over 24 months?", "is our internal D0 ramp tracking the market's learning curve?", "CoWoS interposer cost trend", benchmarking product-lifecycle cost projections. DO NOT USE for: current point values (use get_wafer_pricing / get_packaging_costs); the daily PIT ledger replay (use /api/v1/snapshot-series); margin history (use /api/v1/margin-trends). Filters: benchmark_type (required: wafer_cost|defect_density|packaging_cost|interposer_cost|test_cost|backend_yield|hbm_cost_per_gb), entity_id, foundry, from/to (as_of bounds), grain (raw|month|quarter), limit. Access: a free API key returns a short preview (latest few observations); Pro/Enterprise unlock the full ledger; anonymous callers get none (empty + a get-a-key note). Cite as "Silicon Analysts — Historical Benchmarks".

Fab Capacity — Silicon Analysts' per-fab, per-tech-node-class capacity ESTIMATES from the fabs + fab_capacity_snapshots time-series (100+ fab and advanced-packaging sites: TSMC, Samsung, Intel, SMIC, GlobalFoundries, SK hynix, Micron and more; Frontend in kwspm, Backend advanced-packaging in k units/month). Current readings are Silicon Analysts estimates (source_type analyst_report, source_name Silicon Analysts; each entry's capacity_basis = 'sa_estimate'), not company figures (TSMC, Samsung and Intel do not disclose per-fab capacity) — describe them as Silicon Analysts estimates when quoting them. The foundry_ir / wfe_vendor_earnings / government_filing source types are for future dated observations (capacity_basis 'cited_source'). Default mode returns the LATEST state per (fab, node class) at/before as_of, joined with fab metadata (name, foundry, country, status) and availability_status — OUR allocation tier for the fab (fully_booked = our tightest tier, down to available; availability_basis 'sa_estimate'), not a statement by the company. For evidence-graded supply questions (who said a component is allocated or tight, and when) use get_bottlenecks. series=true returns the full dated series — node conversions appear as capacity shifting between node-class rows across effective_dates (e.g. 28nm shrinking while 7nm grows). Every reading carries its source columns (source_type + citation + confidence, where confidence says whether the fab RECORD is company-confirmed, not whether its capacity is disclosed) and period_relation (past / current / future vs as_of; filter on this, not is_projection, which is the stored flag and is also true on every reading of a fab whose record is still an estimate). Latest state does not carry forward an annual series that ended before the dataset's latest year at/before as_of (e.g. a fab that was sold or closed); meta.corrections lists the published corrections to the fabs the query matched. Latest state: all tiers (incl. anonymous). se

Forecast Vintages — third-party forecasts (research firms like TrendForce/WSTS/SEMI, and company capex/bit-growth guidance) archived with their ORIGINAL publication date. Query the REVISION HISTORY, not just the latest number: "what did TrendForce say about 2026 HBM bit growth in January vs July?". Each row: originator, originator_type, metric, target_period (e.g. CY2026, 2027H1), value (num or low/high), unit, as_of (publication date), a source URL, and a verbatim quote. This is the vintage archive of OTHER organizations' forecasts — distinct from our own scenario models. USE THIS for: forecast revision tracking, "how has the 2026 capex outlook moved across TSMC's earnings calls?", comparing what different firms projected for the same target period, building a consensus-vs-time view. DO NOT USE for: current cost/pricing values (use get_wafer_pricing / get_accelerator_costs); Silicon Analysts' OWN frozen and graded projections (use get_track_record). Filters: originator, originator_type (research_firm|company_guidance|government|bank|industry_body|other), metric, target_period, entity_id, limit (max rows). group='series' additionally returns a "chains" array — the rows already collapsed by originator + metric + target period, oldest print first, with the change between prints — which is usually what you want instead of reassembling them yourself. Latest slice for all tiers; full history (from/to/since/all/group=series) needs a free API key — anonymous callers get the latest slice with a note, never an error. Cite as "Silicon Analysts — Forecast Vintages".

Forecast Track Record — Silicon Analysts' OWN projections, frozen monthly into write-once vintages and graded against outcomes. Every row: model, scenario (bear/base/bull), series, target period, the predicted mid + low–high band frozen at vintage time, and once the period matures, the realized value with a correct/partial/incorrect resolution and error %. Vintages cannot be backfilled or edited — a changed projection that was never frozen is gone, which is what makes this a track record. USE THIS for: checking how Silicon Analysts' HBM/DDR4/CoWoS projections have scored, citing our prediction accuracy, comparing what we projected for a period across successive vintage months, auditing the frozen assumption set behind a projection (include_assumptions=true). DO NOT USE for: third-party forecasts (TrendForce/WSTS/company guidance — use get_forecasts; ours are graded, theirs are archived); current market values (use get_wafer_pricing / get_hbm_market_data). Filters: model (hbm-pricing-model|dram-ddr4-model|cowos-capacity-model), series (e.g. hbm3e), include_assumptions (bool). Fully public at full fidelity for every tier including anonymous — the scorecard is deliberately ungated. Cite as "Silicon Analysts — Forecast Track Record".

AI Hardware Bottleneck Board — eight points where AI accelerator supply can stall (CoWoS-L packaging, HBM, TSMC N3, the HBM base die on foundry logic, large ABF substrates, T-glass, test equipment, InP lasers), each with a judged state: loose / balanced / tight / allocated / sold_out, or watch. Every node returns the evidence behind its state (with the company's own words where they exist), five standard numbers — lead time, price change, share under contract, customer deposits, capacity and timing — each with an evidence grade and source links, what would change the state, and the next dated checkpoint. Also returns the customer-deposit table (TSMC, Samsung, Micron, SK hynix balance-sheet lines) and claims that failed against primary sources. Grades: A = a filing or the company's own words; B = two or more independent outlets; C = one outlet; D = echo. A state moves only on A or B evidence; judgments are weekly, dated, and never overwritten. USE THIS for: "is CoWoS / HBM / ABF substrate supply constrained right now, and on what evidence?", "which AI hardware inputs are sold out or allocated?", "who is prepaying for capacity?", checking whether a widely repeated shortage claim (e.g. "CoWoS sold out through 2026") is in the company's own words, finding the next event that could change a bottleneck. DO NOT USE for: per-node lead-time or customer-share time series (use get_foundry_allocation); fab-level capacity (use get_fab_capacity); HBM prices and vendor share (use get_hbm_market_data). Filters: node (id or name fragment: cowos-l, hbm, n3, hbm-base-die, abf-substrate, t-glass, ate, inp-laser), state, include_history (bool — every earlier judgment with its date, rules version and reason). Missing numbers are null, never estimated. Fully public at full fidelity for every tier including anonymous. Cite as "Silicon Analysts — AI Hardware Bottleneck Board".

WFE Order-Book — wafer-fab-equipment maker disclosures (ASML, Applied Materials, Lam Research, KLA, Tokyo Electron) from quarterly IR: bookings, backlog, segment revenue, and guidance, in the STATED currency (EUR for ASML, JPY for TEL, USD for the rest — never converted). The 12–24-month leading indicator for fab capacity. Each row: company, fiscal_period, metric, value(s), unit, as_of (release date), source URL, verbatim quote. USE THIS for: "is the equipment order-book turning up or down?", ASML bookings trend, AMAT segment revenue by quarter, reading WFE demand ahead of fab-capacity changes. DO NOT USE for: fab capacity itself (use get_fab_capacity); foundry wafer ASP (use get_foundry_economics). Filters: company (asml|amat|lam|kla|tel), metric (bookings|backlog|deferred_revenue|segment_revenue|guidance_revenue|lead_time_weeks), segment, fiscal_period, limit (max rows). Latest slice for all tiers; full history Pro (never a 403). Cite as "Silicon Analysts — WFE Equipment Order-Book".

Fab Milestones — the dated construction & ramp event log: announcement → groundbreaking → equipment move-in → risk production → HVM, plus expansions and SCHEDULE SLIPS recorded as their own events (a delay never overwrites the original plan). Back to 2020. Each row: foundry, fab_name, event_type, event_date, announced_date, a summary, source URL, verbatim quote, and is_projection for forward-dated milestones. USE THIS for: "which fabs hit a milestone recently?", tracking TSMC Arizona / Samsung Taylor / Intel Ohio / Micron / SK hynix timelines, "which projects have slipped?", validating fab-capacity projections against construction reality. DO NOT USE for: current capacity numbers (use get_fab_capacity); allocation/lead-time (use get_foundry_allocation). Filters: foundry, fab_id, event_type (announced|groundbreaking|topping_out|equipment_move_in|risk_production|hvm_start|expansion|delay|cancellation|conversion), country, limit (max rows). Latest slice for all tiers; full history Pro (never a 403). Cite as "Silicon Analysts — Fab Construction Milestones".

Trade-Policy Timeline — the dated log of semiconductor trade-policy actions (export controls, entity listings, license policies, tariffs, subsidies, retaliation) anchored to GOVERNMENT PRIMARY documents (Federal Register / BIS, USTR, MOFCOM, METI, EU, Netherlands…), back to the Oct 2022 BIS advanced-computing rule. Each row: jurisdiction, agency, event_type, title, published/effective dates, affected_entities, node_threshold, a document reference (e.g. Federal Register cite), source URL, verbatim quote. USE THIS for: "what export-control rule changed in December 2024?", building a policy timeline, "which actions named SMIC?", grounding a geopolitics/supply analysis in the actual published action. DO NOT USE for: analysis/commentary on policy (use get_market_intelligence); rumored or anticipated actions (only PUBLISHED actions are recorded). Filters: jurisdiction (us|china|japan|netherlands|korea|taiwan|eu|uk|other), agency (BIS|USTR|MOFCOM|METI|EU-COM…), event_type, limit (max rows). Latest slice for all tiers; full history Pro (never a 403). Cite as "Silicon Analysts — Trade-Policy Timeline".

Search Silicon Analysts — returns up to 10 matching records as {results: [{id, title, url}]}: curated market datasets and their individual series (HBM contract $/GB by generation, DRAM/NAND pricing, TSMC wafer price history, CoWoS capacity and lead times, capex), the reference pages and tool domains behind them (wafer pricing by node, AI accelerator costs, foundry allocation, HBM market and qualification, fab capacity, packaging costs, forecast track record, what changed), and published analysis articles. Every url is a citable https://siliconanalysts.com page. USE THIS for: connector hosts that only speak search + fetch — ChatGPT deep research and ChatGPT company knowledge — and any agent that needs to discover which record answers a free-text question before reading it with fetch. DO NOT USE for: structured or filtered queries when you can call tools directly — the specialised tools (get_market_dataset, get_wafer_pricing, get_foundry_allocation, get_hbm_market_data, ...) take filters and return typed fields; search returns ids, titles and urls only, never values. Deterministic keyword match over published catalogs (no live data read), identical for every tier. An empty or unmatched query returns {results: []}, not an error. Pass a result's id to fetch.

Fetch one Silicon Analysts record by the id search returned — returns {id, title, text, url, metadata}. text is a compact, citable plain-text rendering of the record's CURRENT values (a dataset's points with value, range, data_type, confidence, source and date per point; a tool domain's current rows; an article's summary, takeaways, body and sources). metadata carries as_of, freshness, basis, data_types/confidence, source_count, cite_as (the house citation string) and license_url. USE THIS for: ChatGPT deep research / company knowledge and other search + fetch connector hosts reading a record found with search; getting a citable text summary of one dataset, series, tool domain or article. DO NOT USE for: filtered or structured pulls when you can call tools directly — metadata.structured_tool names the specialised tool (and arguments) behind the record, which returns typed fields and accepts filters. Id formats: dataset:<dataset_id> | dataset:<dataset_id>#<series_key> | article:<slug> | tool:<tool_name>. Unknown ids return INVALID_PARAMS. Tiering is exactly the underlying tool's: anonymous callers get current values only (no history), a free key adds a history preview, Pro the complete series — the metadata history_note says when a record was clamped; as on the website, a source note that restates a value your tier cannot see is withheld. Cite as metadata.cite_as.