One platform. Seven suites. 37 products.

One knowledge-graph spine for fab design and construction — every tool linked to its POCs, utilities, standards, documents and decisions, with AI agents on every gate, plus release↔PO↔need-date build alignment. Read-first over your BIM/CDE estate; write-back only via named-engineer approval.

Design ops runs at maturity ~2/5 even here: verification eats 60–70% of effort, EDA license spend is a black box, tapeout slips surface late. A telemetry-and-agents layer over your own exhaust — regression logs, coverage, license servers, Jira.

Standards and process windows treated as executable, not archival: recipes, SPC limits and operating states checked continuously against the current design and running fab. Simulation becomes governed evidence — every run pinned to its revision, stale results flagged.

Thousands of PCN/EOL notices a year, each traced by hand while 90–180-day last-time-buy windows close. Every notice is parsed and resolved against your BOM and qualification graph — blast radius across products, quals, WIP and customers, tuned for recall.

Greenfield plants hire thousands of freshers while the answer to “bonder error E-412 on QFN-48” lives in a veteran’s head. SOPs, travelers, MES/FDC history, e-logbooks and per-serial tool docs become grounded answers — every one cited, or it abstains.

A fab operations command center over your MES and FDC estate: live WIP with predicted cycle times, OEE accounting, an excursion feed ranked by value-at-risk, and a scenario twin for the what-ifs. Read-first — every write passes a named-human gate.

A yield excursion is a seven-figure event — one 2019 excursion scrapped $550M of wafers in a quarter. Wafer-map perception feeds a self-building genealogy graph; agentic RCA runs cross-stage over MES, FDC and STDF — ranked hypotheses, every claim cited.

OSAT margins run ~15–25% against foundry ~50%, yet thousands of package × test combinations are scheduled on Excel. A CP-SAT solver keeps a live horizon schedule; a sub-second dispatcher recommends the next lot. The planner approves; MES stays the record.

Calendar-based PM guards tools worth $5M–$380M each. FDC and sensor streams feed anomaly detection and remaining-useful-life estimates per failure mode — every alert an explicit economic decision, with the work order gated into SAP-PM.

Services run 22–35% of revenue at OEMs on tribal knowledge. Every ticket is triaged for remote resolution before a truck rolls; engineers get page-cited troubleshooting — plus an offline Engineer / 2 a.m. mode for legacy tools the OEM abandoned; every event is audited against entitlements.

Remote health, predictive maintenance and a per-serial digital twin for the installed base. Five connectivity tiers feed drift baselines and RUL models; anomalies arrive as one approvable package. Raw telemetry never leaves the fab’s data boundary.

First-pass yield on complex tool assemblies runs 85–90% — rework on machines priced $5M–$380M, while new-site ramps repeat the veterans’ apprenticeship. The traveler for a serial + configuration becomes interactive, cited, vision-verified steps at the clean bench.

Semi³ schedules the fab — Tool Build AI schedules the people who build the tools. CP-SAT finite-capacity scheduling of engineer-to-order builds across clean bays, test stands and skill-certified crews, headless on your SAP: the solver proposes, the master scheduler publishes.

Every configured-tool quote is an engineering exercise run from experts’ heads — and in a supercycle, quote latency is lost share. A CP-SAT rulebase decides validity, a deterministic engine prices the quote-to-BOM; the LLM drafts the narrative, never the numbers.

A fab installs ~1,200 tools from 100+ vendors; hook-up alone runs $200–500k per tool. Every order becomes one live thread — build → FAT → ship → hook-up → SAT → SL1–SL3 — with slip propagation and self-assembling dossiers.

Service-parts demand is intermittent and install-base-driven — the canonical hard forecasting problem, run today on ERP min/max. Forecasts come from fleet age, utilization and PM waves; the whole echelon is optimized with SLA penalties priced in. Planners approve every change.

Agentic sales for used, refurbished and legacy tools, where every serial is the SKU: the per-serial asset graph — configuration, attested condition grade, provenance, jurisdiction — drives export-screened quotes, refurb-delta CPQ and flash-matched trades. Rides the shipped SalesI substrate.

A $5M–380M tool changes hands and the buyer asks a simple question: what is actually inside this one? Not the model — this serial. Which parts were replaced, from which lots; which tests and calibrations it passed; who signed for the condition. That record usually exists as tribal memory across three systems. Here it is a single frozen, signed document per serial, and the export refuses to produce it when the evidence is not there.

Every service organisation wants to know whether it is actually good, and none of them will hand over raw numbers to find out. The usual answer is an industry survey a year out of date. This is the other answer: consent-gated cohorts on a small set of ratios, where membership is opt-in per cohort and a figure is published only once the cohort is large enough that no member can be reverse-engineered out of it. The guard is code, not policy — the system refuses to return a cohort minimum or maximum, because those are attributable.

Your ERP records the buy; nothing tells you the fair price. Semiconductor-native sourcing intelligence for the $0.5–3B tool program and nine hyper-concentrated commodity categories: should-cost on every quote, PO slips predicted early — cited, gated, read-first on SAP.

Qualifying a second source takes 6–18 months — in categories like ABF film with one effective source. Discovery, JEDEC/AEC-Q qualification projects and risk monitoring run as durable agents, so the second source exists before the crunch.

Where should the buffer live — wafer bank, die bank or finished goods? Solved stochastically across echelons and forms, MSL and shelf-life clocks automated per lot. When supply tightens: bank die now, finish to order — gated.

An entire fab arrives through customs — one missing annexure parks an etch cluster in bond. Import files assemble ICEGATE-ready, SEZ/bonded/AEO scheme logic applied, delay risk tracked per tool. The agent prepares; your licensed broker files.

Data incumbents describe parts; this acts on your BOM: price and lead-time foresight with regime detection, a live risk register per line, form-fit-function alternates always flagged “requires qualification”, counterfeit screening before any gated shortage buy.

Every other product on this page faces a fab, an OSAT or a tool maker. This one faces the distributor — the tier that moves parts between them and thousands of smaller buyers, and that runs today on spreadsheets and a shared inbox. Moisture-sensitive parts carry a floor-life clock that most systems lose at the first hop; certificates of analysis arrive as PDFs in a folder. Here the lot is the unit of record: MSL level and floor-life hours travel with it, its CoA stays in custody receipt-to-shipment, and the quote desk cannot price what it cannot certify.

A customer escape costs days of your scarcest engineers — under sub-PPM automotive expectations. 8D, CAPA and PPAP packs are drafted straight from the root-cause evidence trail; containment-to-closure runs on state machines. Cite-or-abstain, named approvals.

Takes a returned unit from RMA intake back through genealogy to its wafer, lot, tool and recipe; drives the FA-lab queue; classifies the mechanism; decides the highest-net-recovery disposition — and catches the epidemic cluster before it becomes a recall.

Aggregates field-failure populations — FA results, burn-in escapes, customer DPPM trends — into calibrated FIT/Weibull models per device, package and lot-window; catches drift against the qualified AEC-Q/JEDEC baseline; routes each learning to its lever: design, process, screen or derating.

The most trade-regulated goods on earth — entity lists +42 then +23 within a year. Classification is retrieval-first and cited; the order book re-screens the day a rule changes; declaration chases run as durable campaigns. Prepare, never auto-file.

Rolls every lot’s actual cost up from the floor — material, consumables, tool-time — yield-adjusted to cost-per-good-die, NRE amortized, take-or-pay loaded. Variance decomposes as a waterfall that sums exactly; margin by product and customer, live. SAP CO stays the record.

Validates every invoice against its cited contract terms and runs the semiconductor claims stack — ship-and-debit, price protection, rebates, POS, consignment, OTIF/LD penalties. The Promise Ledger turns a broken promise into an adjudicated penalty, fab evidence chain attached.

Nobody decides what to book — planning takes allocation as a given. Booked wafer, packaging and memory capacity becomes a portfolio of contracts and options: take-or-pay exposure projected into P&L, draw reconciled, ranked book/hold/release calls. You book; it never writes.

The referee, not a player: every recommendation tracked through identified → accepted → implemented → realized — flipping only on a matched ERP signal, a PO actually paid, credited against the would-have-happened counterfactual. Modelled and realized never blur.

Not another chart wall — a ranked feed of decisions worth money. The cross-suite spine becomes a CEO cockpit: recommendations ranked by value-at-risk × evidence, KPI tiles carrying evidence badges, a narrated “what changed this week” drilling to source records.

Planning in 2026 means co-booking wafer starts, CoWoS and HBM as one constraint — CoWoS booked out 52–78 weeks, HBM sold out. Demand and supply modeled against the real constraint set; shocks get costed scenarios in hours, not war rooms.

Order management runs over EDI and RosettaNet with SAP SD as the record — nothing reasons about it. Exceptions triaged across thousands of lines, promise dates served by Delivery AI’s ATP engine, fair-share administered with audited rationale. Every write passes a named gate.

A fab is a promise machine — it commits good die it hasn’t yielded yet. Four checks per promise — Available, Capable, Yielded, Profitable — logged to a Promise Ledger. It never bluffs: low confidence routes to a human.