Excursion-control OS for GAA and hybrid-bonded chip lines that turn hidden-layer 3D scans into yield-saving process corrections.

1. Advanced-node structures now sit beyond the reach of conventional optical metrology, so fabs need software built around hidden-layer visibility rather than incremental dashboard upgrades.
2. Measurement precision has become a direct yield-economics issue because small errors can cascade across hundreds of manufacturing steps before a bad wafer is caught.
3. Real-time process feedback is becoming mandatory, which creates room for a startup that converts scan output into immediate hold and correction decisions instead of retrospective analysis.
4. The push to fund application centers and collaborative R&D suggests the bottleneck has moved from buying the tool to operationalizing it across real fab workflows.

Catalyst. Nearfield's rise shows fabs have crossed into architectures where seeing the defect is finally possible but acting on it fast enough is the new bottleneck.

Hidden Layer Yield OS connects new 3D metrology tools, existing optical inspection, SPC data, and process-step history for a single advanced module such as GAA or hybrid-bonded integration. It scores hidden-layer anomalies by likely downstream yield impact, then recommends whether to hold a wafer lot, run confirmatory sampling, or push a targeted process correction before more value is added on top of a bad wafer. The first product is not an autonomous fab brain; it is a human-in-the-loop excursion board purpose-built for the exact moment when expensive hidden-layer scans become available but action is still manual and slow. As engineers accept, reject, or modify recommendations, the system learns fab-specific defect signatures and becomes the operating memory for how one line protects yield at advanced nodes.

What's different. Incumbents largely own the measurement box or the generic process-control stack, but the new bottleneck is the action layer between them. This company is purpose-built for hidden-layer excursion decisions at advanced nodes: it links 3D scan signatures to step history, lot disposition, and recommended corrections for one module at a time. The defensible asset is a fab-specific corpus of accepted excursion decisions and defect-to-outcome mappings that gets more valuable as each node and module ramps.

Jobs to be done

flowchart LR
  Buyer[Process control leader] --> Pain[Hidden-layer defects escape until yield collapses]
  Pain --> Product[Hidden Layer Yield OS]
  Product --> Outcome[Faster holds and higher advanced-node yield]

Executive takeaways

• Why-now is real: GAA, CFET, HBM, and hybrid bonding have moved inspection into buried or stacked structures, while new 3D metrology finally makes those defects visible without yet making fab action fast or standardized [1][3][5][6][15][16][17][19].
• The highest-probability wedge is not broad 'AI for fabs' but a human-in-loop hidden-layer excursion board for one module, where hold/release/correct decisions are still made through war rooms, generic SPC, and vendor application support [6][7][10][11][12].
• Budget exists, but proof burden is severe: AI/HPC and HBM economics make escaped defects expensive, yet buyers will demand quantified impact on yield stability and disposition cycle time before expanding scope [13][27][31][32].
• Incumbents are strong but incomplete: KLA, Onto, PDF Solutions, yield platforms, and metrology OEMs each own part of the stack, but none of the fetched evidence shows a vendor-neutral product centered on hidden-layer cross-tool lot disposition as the default category winner [10][11][12][13][29][34].
• This is venture-viable only if the startup compounds from one module into process memory across modules and fabs; the adjacent market is multi-billion, but the initial decision-layer wedge is likely a few dozen high-value modules, not a mass-market software category [21][22][23][24][25].

Market definition

This category is best defined as hidden-layer process-control software for leading-edge logic and advanced-packaging modules: a decision layer that sits between 3D metrology/inspection outputs and MES/SPC/yield disposition so fabs can act on buried-feature defects before more value is added downstream [3][5][6][10][15][17].

Customer and buyer

Primary users are directors and senior managers in process control, yield engineering, and module integration at leading-edge foundries, IDMs, and advanced-packaging groups; economic buyers are VP-level process-control or fab-yield leaders because the pain spans metrology, lot disposition, throughput, and AI/HPC product economics [3][10][13][28][31].

Buying triggers

• A node or module ramp into GAA, CFET, 3D memory, or hybrid-bonded integration creates buried-feature measurement problems that legacy optical workflows cannot resolve cleanly. [1][5][19]
• Hybrid-bonding pitch moves below 10µm and wafer flatness, recess, contamination, and overlay tolerances tighten enough that small mistakes turn into bond failures or false holds. [6][7][8][15][17]
• A new 3D metrology tool install or applications-center rollout exposes a workflow gap between new measurement visibility and same-shift process decisions. [1][18][19]
• HBM and AI package economics worsen the cost of finding defects too late, making earlier disposition and sampling decisions materially more valuable. [27][31]

Willingness to pay

Willingness to pay should be high but proof-bound. The fetched market shows quote-led enterprise process-control software, very expensive downstream escapes, and public customer evidence that analytics platforms can improve yield and cut analysis time dramatically; however, procurement will be justified on avoided scrap, fewer false holds, and faster engineering decisions rather than generic 'AI' budgets [10][12][13][31][32]. [10][12][13][31][32]

Category dynamics

Growth signal 6.9% CAGR

Validation signals

• Nearfield’s $380M round explicitly funds applications centers, production capacity, and deeper customer R&D, signaling that fabs are spending to operationalize new metrology rather than just buy tools.
• imec has already demonstrated 2µm D2W hybrid bonding with good electrical yield and 200nm W2W hybrid bonding with full-wafer overlay results, which makes the process-control layer more urgent, not less.
• Public yieldHUB testimonials describe 4% yield improvement, daily PE/QE usage, and order-of-magnitude productivity gains, showing that semiconductor buyers reward analytics when ROI is explicit.
• HBM is becoming taller and more expensive, with defects increasingly costly to find late in the flow, which strengthens the case for earlier disposition and sampling decisions.

Regulatory & technical constraints

• Advanced semiconductor manufacturing and support are exposed to export-control regimes that can restrict customer mix and cross-border service models.
• Hybrid bonding requires extremely tight surface prep and overlay control, including <2.5nm Cu recess in one imec D2W example and ~50nm overlay targets on the 200nm roadmap.
• Many critical measurements remain hybrid or complementary across AFM, optical, OCD, e-beam, and destructive reference methods, which raises integration complexity.
• Buried-feature GAA metrology still requires seeing through optically opaque stacks, so false certainty is dangerous and module-specific calibration is unavoidable.

Strategically, the field splits into (a) measurement and equipment incumbents such as KLA, ASML, Bruker, EVG, Camtek, and Nearfield, (b) analytics and yield-management platforms such as KLA software, Onto Yield Optimizer, PDF Exensio, and yieldHUB, and (c) manual substitutes such as yield war rooms, destructive sampling, and vendor application engineering. The proposed startup only wins if it becomes the neutral action layer that turns multi-tool hidden-layer signals into same-shift hold/release/correct recommendations instead of trying to outbuild the metrology boxes themselves [6][8][10][11][12][13][14][19][29][34].

Why incumbents do not win by default

• Metrology OEMs. They already own the measurement step and customer access, but most are optimized around generating or interpreting their own tool data, not orchestrating vendor-neutral lot disposition across mixed toolchains.
• Process-control suites. KLA-class platforms already market run-time process control and excursion analytics, so the startup cannot win on 'dashboarding'; it must win on hidden-layer specificity, cross-tool orchestration, and faster human-trustable decisions.
• Yield analytics platforms. PDF Solutions and yieldHUB prove demand for semiconductor analytics, but their public positioning is broader manufacturing/test correlation rather than a purpose-built pre-bond hidden-layer decision workflow.
• In-house yield engineering. Large fabs can always extend SPC scripts, engineer notebooks, and expert review boards, which means the startup must deliver institutional memory and cycle-time gains that are hard to replicate internally.

Hidden Layer Yield OS is a vendor-neutral decision layer for leading-edge fabs that now can see buried defects with new 3D metrology but still cannot act on those signals fast enough inside production workflows. The initial beachhead is one yield-critical GAA module in a top-tier logic foundry or IDM where process-control teams still use war rooms, SPC dashboards, and vendor application support to decide whether to hold, sample, or continue lots. The product should land in advisory mode first, because trust and data access are the gating constraints, not model novelty. The commercial logic is coherent: a node ramp or new metrology-tool install creates the pain, the VP of process control or fab yield owns the economic loss, and a paid module deployment can be justified on avoided downstream wafer loss and faster disposition. The research supports a narrow but valuable market, with an estimated $250.0M TAM, $85.0M beachhead SAM, and $8.0M year-3 SOM, but it does not identify named design partners or an acceptable false-hold threshold for first deployment. The company should therefore defer full-fab rollout, closed-loop automation, and packaging/HBM adjacency until it proves lower disposition time and fewer false holds on one module without increasing escaped defects. Defensibility comes from a fab-specific corpus of accepted recommendations, process context, and downstream yield outcomes that general analytics platforms and single-tool OEM software will struggle to replicate. The board-level risk is simple: if read-only pilots cannot earn trusted operator usage and measurable disposition improvement within 12 months, this is not yet a venture-scale software company.

Problem

• Leading-edge GAA and 3D integration flows now contain buried structures that legacy optical workflows cannot inspect cleanly, so critical defects are visible only after expensive downstream value has already been added.
• Process-control teams still rely on manual excursion review across SPC, tool output, and expert meetings, which slows lot disposition and increases both escaped-defect risk and false holds.

Solution

• Hidden Layer Yield OS ingests 3D metrology output, process-step history, and existing SPC context for one module and recommends hold, sample, or targeted process-correction actions in the same shift.
• The first product is a human-in-the-loop excursion board with recommendation evidence and feedback capture, not autonomous process control, so fabs can build trust before granting broader authority.

Why we win

• The company targets the action layer between metrology OEMs and generic process-control suites, where the workflow pain is urgent but no vendor-neutral default appears established in the research.
• A fab-specific dataset linking hidden-layer signatures, accepted decisions, and downstream yield outcomes compounds with every reviewed excursion and becomes operational memory that is hard to reproduce internally or by a single tool vendor.
• The wedge is narrow enough to prove ROI on one module yet expandable into cross-module excursion management, recipe transfer, and broader fab process intelligence.

Milestones

flowchart LR
  Wedge[One GAA module wedge] --> MVP[Advisory mode excursion board]
  MVP --> Proof[Lower disposition time and fewer false holds]
  Proof --> Expansion[More modules, sites, and packaging workflows]

Founding team

Experiment roadmap

Risk assessment

What must be true

• One module workflow exists where scan-alert-to-disposition delay is materially painful and owned by a VP-level buyer.
• Read-only integration into metrology, SPC, and lot-history systems can be completed without a multi-year IT project.
• Historical replay can identify a recommendation policy that cuts false holds or disposition time without increasing escaped defects.
• Engineers will use a recommendation board every shift and feed back accepted or rejected actions often enough to build proprietary decision memory.
• OEMs and ecosystem partners tolerate or support a vendor-neutral action layer instead of blocking data access or bundling it away.

Open diligence questions

• Which exact GAA module workflow consumes the most senior engineering time today between scan alert and lot disposition?
• What false-hold and escaped-defect thresholds would a first customer accept for shadow-mode evaluation?
• How much data access can the startup get from mixed metrology environments without OEM resistance?
• Is the first faster path to revenue front-end GAA control or hybrid-bonding packaging control?
• Does the buyer require on-prem or customer-VPC deployment from day one?

Model sanity

• Revenue engine. Revenue is driven by sequential paid-pilot-to-$2M-subscription conversion at $500K/module/quarter; the base case needs four modules under contract by Q4Y3 to hit $6.5M annual revenue and a $8M ARR run-rate consistent with research.yaml SOM.
• Must go right. Module 1 pilot must convert to a production subscription before Q1Y2 month 15 — that single conversion swings Q1Y2 EBITDA from roughly -$500K to -$24K and keeps the company funded without a bridge raise.
• Model breaks if. Integration into metrology, SPC, and MES data stacks slips beyond 9 months per customer — the downside scenario shows this delays Module 4 to Y4, collapses Y3 revenue from $6.5M to $4.2M, and drives the cash trough to $2.5M, below 6 months of runway.
• Next-round proof. Three paying production module subscriptions with documented disposition-time improvement plus at least one second-fab pilot by Q3Y3 justify a Series A milestone and demonstrate repeatable enterprise motion beyond a single design-partner relationship.

Scenarios

Sensitivity

flowchart LR
  Leads[Target Fab Modules\n~34 SAM] --> Pilot[Paid Pilot\n$350K / 6 mo]
  Pilot -->|60pct conversion| Sub[Annual Subscription\n$2M per module]
  Sub --> Revenue[Quarterly Revenue\n$500K per module]
  Revenue --> COGS[COGS\n20pct subscription\n45pct services]
  Revenue --> GP[Gross Profit\n75-78pct]
  GP --> Opex[Opex\nHeadcount + Overhead]
  GP --> EBITDA[EBITDA]
  EBITDA --> Cash[Cash EOP\nrolls forward quarterly]

Flags: Four-customer revenue concentration in Y3: a single non-renewal in Q4Y3 removes $500K revenue and ~$390K EBITDA; model does not include a partial-churn hedge. · Module 1 pilot-to-production conversion is a single point of failure for Y2 cash flow; any slip past M15 turns Q1Y2 cash-flow negative by ~$450K and may require a bridge or accelerated seed extension. · Subscription COGS of 20% assumes shared cloud infrastructure; region-locked customer-VPC deployments (required by some fabs per BP operations) may push subscription COGS to 25–30%, reducing Y3 EBITDA by $130–200K. · CAC of $450K relies on founder-led direct sales efficiency throughout Y1–Y2; post-founder handoff to Sales/BD hire in Q1Y2 may increase marginal CAC toward $600K+ for Modules 3–4. · Y3 gross margin of 78% exceeds BP target of 70%; this is achievable only if subscription revenue exceeds 85% of Y3 mix — valid by Q4Y3 but not in Q1–Q2Y3 when two pilots are still in services mode.

• Integration friction. Semiconductor fabs have fragmented, high-security data environments, so connecting metrology, SPC, and process-history systems may slow pilots. Mitigation: Start with one module, one tool family, and read-only integrations that produce recommendations before attempting any deeper workflow embedding.
• Incumbent pull. Metrology or process-control incumbents could extend their software stack once the action layer proves valuable. Mitigation: Own the cross-tool decision workflow and fab-specific feedback corpus, which are harder for a single equipment vendor to replicate across mixed environments.
• Slow enterprise adoption. Yield leaders may hesitate to trust new recommendations on production lots without clear proof that the system reduces false holds and missed defects. Mitigation: Land with an advisory mode tied to one excursion class, publish recommendation evidence for every call, and prove cycle-time and disposition improvements before asking to expand scope.