Frequently asked
The questions, including the awkward ones.
Grouped by who usually asks them: the platform, the trust model, process IP, commercials and the company itself. Where the honest answer is unflattering, it is the answer we have written.
The platform
What it is and what it does
-
Chipira is an autonomous operations layer for semiconductor advanced packaging. It perceives the package flow through X-ray, CT, SAM, AOI and tool telemetry, decides what to do using grounded AI agents, acts on bonders, molding tools, handlers and testers within bounds you define, and verifies the outcome — closing the loop across vendor-siloed equipment.
-
No. Dashboards report; Chipira acts. A yield management system tells you what happened last shift. Chipira changes the parameter that produced it, inside an envelope your process owners authored, and writes the verified outcome back.
-
An equipment vendor’s software can only see its own tool. The decisions that matter in advanced packaging depend on data from instruments the vendor did not build. Chipira is vendor-neutral by construction, sells no hardware, and owns the sequence across tools rather than the tool itself.
-
No. Chipira is additive. It reads through interfaces your tools already expose and writes through interfaces they already accept. A total Chipira outage degrades your line to exactly the process you run today.
Trust and autonomy
Why a fab would let it act
-
You measure it. Shadow mode runs the agents against your engineers on your packages with no write path at all, producing a measured accuracy figure before anything is granted. Autonomy gates open on that measurement and close automatically when it drifts.
-
Four things in series: the action envelope bounds every parameter and step; the twin pre-validates the move; the autonomy gate determines whether a human must approve; and envelope breach triggers fail-safe stop with automatic rollback. Every one of them is auditable.
-
Only your process owners, through an approval flow that is itself audited. Chipira support cannot widen an envelope for you, including during an incident.
-
Instantly, by your organisation, at any level. It is also withdrawn automatically when rolling accuracy falls below the gate’s floor — without waiting for a review meeting.
-
Generation is grounded in retrieval over your recipes, specs, design rules and standards, with enforced citations; output that cannot cite a source is rejected before it reaches a decision. Beyond that, no proposal reaches a tool without passing the envelope and, where required, the twin.
The mechanism
Four gates between a model and your line
Envelope, twin, autonomy gate, audit. Every proposal passes all four, every time, whichever agent produced it.
Process IP and security
Where your recipes go
-
In your tenant, at your fab edge. Inference runs on fab-edge hardware inside your building. The control plane governs models and fleets and does not need your process data to do it.
-
Not outside your tenant, and never without a separately signed agreement. Not for product improvement, not for benchmarks, not as a side effect of an upgrade.
-
No. Tenancy is enforced by construction, retrieval is permission-aware, and there is no shared index across customers. Cross-fab learning exists as an explicit, contractual, revocable option — never a default.
-
Yes. An air-gapped configuration runs the control plane on your premises with signed offline model delivery. You give up fleet automation and cross-site learning; some lines should.
Commercials
What it costs and how it is bought
-
Line is $18,000 per tool or bonder line per month. Fab is $120,000 per month for the whole line and the full agent loop. Enterprise is custom, typically landing between $800k and $9M ACV. Annual prepay carries a 15–20% discount band.
-
No, deliberately. A free pilot has no internal owner and no honest measurement. A paid pilot with a signed success metric produces a decision either way.
-
It is documented as a miss and you do not convert. We would rather that than a design partner drifting inconclusively for two quarters.
-
On the highest-value workflows, yes — a defined share of measured yield, scrap, known-good-die or ramp improvement against a jointly agreed baseline, capped and floored so neither side is exposed to a measurement dispute.
Manufacturing operations directorOSAT, advanced packagingI have read forty vendor FAQs this year. Not one of them contained a sentence that made the vendor look bad.
Composite drawn from design-partner and industry conversations. Illustrative, not a customer endorsement.
The company
Stage, risk and honesty
-
Not yet. Chipira is pre-launch and building a design-partner programme. Nothing on this site describes a shipping deployment at a named customer, and any figure that is a design target is marked as one.
-
It should not take an unconditional one — which is why the architecture removes the need to. Shadow mode, envelopes, twin validation, immutable audit, revocable autonomy and documented exit all exist so adoption does not require trusting our goodwill.
-
If fabs decide no external software may ever write to a bonder, the ceiling drops sharply. If package physics does not generalise across families, the data advantage is far weaker than we believe. Both are live risks we track explicitly.
-
Funding status, team composition and hiring plans are discussed directly under NDA rather than published as badges.
Go deeper
Where each answer is expanded
-
01
The platform
The closed loop, the autonomy model and how a proposal becomes an action.
-
02
Security and process IP
Tenancy, edge inference, audit, envelopes and deployment options.
-
03
Pricing
Line, Fab and Enterprise, metering, and how a quote is actually built.
-
04
Case studies
The four wedge workflows and the measurement protocol behind them.
Still unanswered
Questions we cannot answer yet
Because the honest answer is “we do not know”, and we would rather say so.
- How well does the model generalise across package families?
- We believe substantially, because the physics is shared. We will not know how substantially until design-partner data across several families exists.
- How fast can a fab realistically reach bounded autonomy?
- It depends far more on your change-control culture than on our technology, and we do not yet have enough deployments to give you a credible distribution.
- What is the true integration effort on a closed tool?
- Highly variable and occasionally impossible. We assess it per tool during scoping rather than quoting an average that would mislead you.
- Will incumbents open their tools to us?
- Some will, for distribution reasons. Some will not. Our architecture assumes the pessimistic case.
Start narrow, expand relentlessly
Land one workflow. Own the loop.
A Chipira engagement begins with a single wedge workflow, a shadow-mode baseline and one signed success metric. Everything after that is expansion.