Arda says it cuts factory re-planning from a week to two days. The long-term prize might be a faster idea-to-production loop

Per Arda CEO Bob McGrew, re-planning a 2,000-part product at Anduril after a design change took two days instead of about a week. That's what Arda does today: it turns design changes into updated assembly orders and floor instructions. The long-term prize might be a faster loop from idea to production, with more iterations, quicker feedback, better products and variants built for single customers. That loop is our thesis, not a measured result.

The question

On October 20, Arda Industries came out of stealth with $232M raised: a $72M seed led by Accel and Founders Fund, and a $160M Series A at a $1.5B valuation co-led by Altimeter, Accel and Khosla (Business Wire). Its CEO, Bob McGrew, was OpenAI's chief research officer and, before that, Palantir's second engineer. Its first customer is Anduril.

What Arda does today is narrower than the funding suggests. When a part changes, it reads the product's CAD, documents and planning data, recomputes the assembly order and drafts updated instructions, and an engineer approves each step; per Arda (McGrew, company-reported), a 2,000-part Anduril job took two days instead of about a week. The long-term prize might be a faster loop: if changing the factory gets cheap, a company can change the product more often, learn from each version sooner, and eventually build variants for single customers. That loop is our thesis, not a measured result, and approvals, inspection, suppliers and government change control still set its pace, as we cover in the closing section.

The easy reading is "an AI that writes factory paperwork." That undersells it, and it also skips the part worth understanding. The real question is: when a design changes by one part, why does it take a factory days or weeks to catch up, and which part of that work can software actually remove?

To answer that, you have to look at something most people outside a factory never see: the assembly order.

Every product has two descriptions

A physical product is described twice.

• The design says what the product is: the CAD model, the drawings, and the engineering parts list.
• The process says how to build it: which parts go together in what order, at which station, with which tools and fixtures, with what checks along the way, and what the technician reads at each step.

The process is not a mechanical translation of the design. When manufacturing engineers convert a design parts list into a build parts list, Dassault Systèmes' survey of 173 manufacturers found they reorganize about 76% of the components and add about 195 items that never appear in the design, such as thermal pads, consumables and kits. That step alone takes 12–18% of new-product-introduction resources, and 26% in aerospace (DELMIA; vendor survey).

At the center of the process sits the assembly order: the sequence in which parts go together and where each step happens. Everything else in the plant hangs off it.

Why redoing the assembly order is hard

1. Geometry rules out most orders. A part can go in only if there's a clear path for it, and for the tool that fastens it. A cover blocks everything beneath it. A taller component can block a screwdriver from reaching a screw that used to be easy. A half-built unit has to stay stable or be held by a fixture. Bolt patterns have to be tightened in a set order, and adhesives need time to cure. Anything that needs testing or inspection has to be checked before something covers it up.

2. The number of possible orders is huge in theory, but that isn't where the time goes. With no constraints at all, 20 parts can be put together one at a time in about 2.4 quintillion (20!) orders. Carnegie Mellon researchers Homem de Mello and Sanderson showed in 1988 that a complete planner would have to check 1,048,555 one-part-at-a-time splits, or about 1.74 billion if subassemblies are allowed. But that's the worst case, for a 20-part assembly where every part touches every other, and the authors themselves call it "very unlikely" (CMU-RI). The same report gives formulas for a chain, where each part touches only the next. For a 20-part chain that's about 190 splits one part at a time, or 1,330 with subassemblies (our calculation from their formulas). Real products sit somewhere in between.

Engineers never come close to enumerating orders. Contacts between parts are sparse. Products come in modules (boards, harnesses, a chassis), each a small problem of its own. Engineers think in disassembly: what can come out now, given the geometry? They write down a few precedence rules ("X before Y"), and De Fazio and Whitney showed in 1987 that a short list of such answers defines every feasible order (IEEE J. Robotics & Automation). And for a one-part change they repair the last plan locally rather than starting over.

So finding an order isn't the bottleneck. Proving it works is: every tool can reach, the fixtures still hold, the stations still balance, inspections happen before parts get covered. So is carrying the change through every document downstream. The combinatorics explain something else: why automating this needed physics to rule out the orders that physically can't work. A computer can't rely on an engineer's intuition, so it needs motion planning. For each pull direction, the geometry says which parts are blocked, and the planner peels off whatever is free (Wilson & Latombe, 1994). Even that has hard cases: deciding in general whether some group of parts can be separated is NP-complete (Kavraki, Latombe & Wilson, 1993). That's the approach McGrew says Arda uses.

3. The order is tied to the line. A line is balanced so that every station's work fits inside the takt time, the time each unit can spend at a station. Add one part or one torque step and a station can go over. Move that step to the next station and you may break an ordering rule, which forces a new sequence, which moves kits, fixtures, test points and instructions. Splitting work across stations under these rules is itself NP-hard (SALBP). MIT's Daniel Whitney puts it simply: "Sometimes a different sequence will have better balance… A cycle lost on the bottleneck station is a cycle lost forever" (MIT 2.810).

4. Changing it triggers requalification. A re-sequenced build is a process change. In aerospace and defense, that can mean a new first-article inspection and, for major changes, government approval (see below).

5. Most of the rules aren't written in the CAD. CAD stores geometry. Whether a technician can see a connector, where a cable gets routed, and which fixture holds a board live in PDFs, notes and people's heads. Anduril describes parts catalogs "living in one engineer's head or one binder on one bench" (Anduril).

Where does the week go? No one has published a time study of engineers re-sequencing an assembly. The closest evidence is about engineering change orders in general. In Loch and Terwiesch's 1999 field study of change orders on an automotive climate-control system, a change order "takes on average almost one and a half weeks… although it has on average only 5 hours' work invested in it." The authors model the change process as a network of queues. They attribute the gap mainly to waiting and congestion: a change order "spent most of its lifetime 'sitting on someone's desk'," and lead time grew as busy groups such as simulation became bottlenecks (Loch & Terwiesch, J. Product Innovation Management, 1999; INSEAD preprint). That's an automotive study from 1999, not a measurement of assembly re-sequencing. Our inference is that a re-sequencing change rides the same queues. On top of that comes the engineering itself: proving tool reach, fixtures and station balance, and rewriting the documents. Arda's "a week to two days" comes with no published breakdown of either part.

What the assembly order touches

The assembly order isn't one document. It's the shared key that links the design to the plant's schedule, inventory, quality records, equipment, people and contract.

Who uses it	What they take from it	What they redo when it changes
Manufacturing engineers	The build parts list, stations and sequence	Re-derive the affected sequence and re-check access and tests
Technicians and supervisors	Step-by-step instructions at each station	Learn new steps; risk building to a stale revision
Industrial engineers	Task times and ordering rules, which set the line balance and station count	Rebalance, sometimes add a station
Supply chain and purchasing	When and where each part is used, and what goes in each kit	New part numbers, new kits, leftover stock
Quality	Inspection points, what to measure, the per-serial-number build record	Partial or full first-article inspection, new inspection plan
Tooling	Which fixtures, nests and torque tools each step needs	Modify or build fixtures
Automation and test engineers	Robot, torque-tool and test-rack programs	Reprogram; Arda estimates months per automated station
Safety and training	Hazardous steps and operator certifications	Re-review and re-certify
Cost accounting	Standard labor hours per step	Re-cost the product
Government customer	The configuration baseline and inspection evidence	Approve the change, sometimes by contract modification
Depots and field support	Which units were built to which revision	Retrofit kits and repair instructions

Why the same change gets "re-keyed" in several places

Re-keying means a person reads data out of one system and types or redraws it into another, because the two systems don't share data. Factories run on several systems, each owned by a different team and usually sold by a different vendor:

• PLM, the product-data system, holds the design and its parts list. Engineering owns it.
• ERP, the business system, handles purchasing, inventory, routings and cost. Operations and finance own it.
• MES, the manufacturing execution system, runs work orders, instructions and build records on the floor. Production owns it.
• Station controllers run the robots, torque tools and test racks.
• The quality system holds the inspection plans and records.

The split is formal. ISA-95, the main standard for factory software, separates the business layer from the shop-floor layer, and the design system sits outside both (ISA). Each organizes parts differently: engineering groups parts by function, while the factory groups them by build order. Anduril's own description is designs "re-keyed, re-drawn, and re-interpreted across disconnected systems."

It also happens in more than one place. A second plant keeps its own version of the process. A contract manufacturer building the circuit boards re-enters the drawings into its own systems. A repair depot needs to know which fielded units were built to which revision. Because AS9102 also triggers re-inspection when the location of manufacture changes, each site carries its own quality paperwork too.

How a new build order reaches the floor

A new assembly order doesn't take effect when an engineer saves a file. It moves through a chain of systems. Each step updates data that someone first has to edit:

1. Engineering change order with effectivity. The change is released under a change number that says when it applies: from a date, or from a serial number. In SAP, "the effectivity parameters Time interval (DATE), Serial number interval (SERNR), and Material number (MATNR) are delivered as standard" (SAP Learning). Units already in process normally finish on the old revision. An SAP support note is titled "Change of the master data is not reflected on the existing released production order" (SAP KBA 3328953; title only, behind login). In Siemens Opcenter, a production supervisor "can decide whether to adopt the new revision, based on the actual status of production," including for in-progress work orders after approval (Opcenter Execution Discrete white paper, §4.3.4).
2. ERP updates the build parts list and the routing. Each part is assigned to an operation at a station.
3. MRP regenerates pick lists for new work orders. The warehouse learns where to bring parts from each station's supply area: "To stage products for a production order, a warehouse must know where it has to take the products… the PSA [production supply area] contains this information" (SAP Help). So the kit that used to be parts A, B and E to Station 4 becomes A, C and G to Station 9 (illustrative). The software generates that only after someone has assigned G to the Station 9 operation and given it a staging rule.
4. MES pushes revised step instructions to station screens. Opcenter can pre-assign parts to each operation, and tasks can require the operator's acknowledgement, data entry or an electronic signature (Siemens electronic work instructions; Opcenter white paper §4.3.8). Training sign-off goes with it.
5. Paper travelers in less digital shops. The same change becomes a revised traveler that moves with each unit. Collins Aerospace requires changes to "work instructions, travelers, routers" to be "recorded, dated, and traceable to a qualified person making the change" (Collins supplier quality requirements, §5.9.1).
6. Torque tools and test stations get reprogrammed. Torque controllers hold numbered programs, and the line computer selects one per step (Atlas Copco Open Protocol). A new torque value means a new or edited program (our inference from the protocol). Test racks get new procedures and limits.

ISA-95 splits these systems into a business layer (steps 2–3) and a shop-floor layer (steps 4–6), which is why each is a separate update.

What first-article inspection is

First-article inspection (FAI) checks the first unit built in a new configuration (new part, new process or new site) against every requirement on the drawing and specifications, and files the evidence. Lockheed Martin UK's supplier guide defines it as objective evidence that "the combination of material, tooling, processes, documentation and personnel is capable of producing compliant hardware" (LMUK SQAG 001). In other words, it tests the process, not just the part. The governing standard is SAE AS9102, which is paywalled, so we quote Lockheed's supplier guides, which follow it.

• Three forms. Form 1 lists the part numbers covered and records whether the FAI is full or partial. Form 2 covers materials, special processes and functional tests, down to the test procedure number and any software revision loaded during test. Form 3 records every design characteristic (each one numbered, or "ballooned," on the drawing) with its requirement and measured result.
• Partial (delta) FAI. This covers only what changed. Lockheed requires one when "a change in design potentially affects form, fit or function," or when "a change in manufacturing source, process, inspection method, location of manufacturer, tooling, or material potentially affects form, fit or function," and likewise for machine-program changes. A full FAI is needed for new parts. Lockheed's RMS business requires "a New Full FAI… for a 2-year lapse in production" (LM RMS FAI guidelines).
• Teeth. "Material received without an approved FAI is subject to immediate return" (LMUK).

A re-sequenced assembly is a process change, so quality has to judge whether it could affect form, fit or function. If it could, the next unit gets a delta FAI before it ships. In defense, a major ("Class I") change must also be approved by the government's configuration board and authorized with a contract modification (MIL-HDBK-61B). Software can draft the forms. Someone still has to build and measure the unit.

Before and after: one part change, traced

Here's one supplier-driven change on a 2,000-part missile avionics box. The box and its timeline come from McGrew. The specific details below are an illustrative example we made up to show the ripple, not a description of Anduril's product.

The supplier of a power module can't scale. Engineering picks a second-source module that does the same job but is slightly taller, has a different connector pin-out, and calls for a different screw torque.

Line A, without Arda.

1. Engineering releases the new CAD revision and parts list.
2. Configuration management decides whether the government has to approve it.
3. A manufacturing engineer re-keys the change into the build parts list: new module, longer screws, and a thermal pad that never appears in the design.
4. They redo the assembly order by hand. The taller module now blocks the driver from reaching two existing board screws, so those screws move earlier, and the shield installation moves after a new continuity test.
5. Station 3 now runs about 90 seconds over takt, so one step moves to Station 4, which changes that station's kit.
6. Purchasing re-enters the part, supplier, routing and kit changes into ERP.
7. Instructions for Stations 3 and 4 get rewritten. Anduril's baseline was about 10 hours for basic instructions and up to 30 for complex workflows.
8. The torque tool and the test rack get new programs.
9. The board fixture needs a relief cut for the taller module.
10. Quality updates the inspection plan and runs a partial first-article inspection.
11. The floor switches serial numbers to the new revision and deals with half-built units.
12. Operators at two stations are re-certified, and the cost is recalculated.
13. The board supplier repeats most of this in its own systems, and the depot gets retrofit instructions.

That's one module, but about 14 documents, more than 10 roles and at least three sites. Per McGrew, this kind of whole-product job took about a week.

Line B, with Arda. Steps 1 and 2 are the same. Arda reads the CAD, PDFs and planning data, works through the affected parts, stations and instructions, recomputes the assembly order with motion planning, and drafts the new instructions: about 15 minutes for basic ones and about 2.5 hours for complex workflows. An engineer reviews and approves every step. Per McGrew, the same 2,000-part job took two days. Anduril's ArsenalOS software then pushes the change through purchasing, quality and build records inside Anduril.

Steps 2, 8 (if a robot is involved), 9, the physical inspection in step 10, and the training in step 12 are still done by people or physical work in Line B. That's the honest limit.

Every Arda and Anduril figure here is company-reported. The instruction speedups are framed as "Anduril found that in many cases," with no published method.

The missing middle, and why it persists

Every design carries a fixed cost that has to be paid back over the units built before the next change. That cost includes tooling (dies, molds, fixtures), engineering (the process work above), automation programming, and requalification. Unit cost is roughly that fixed cost divided by the number of units, plus the cost of building each one.

• In mass production (millions of units), the fixed cost disappears into volume.
• In prototypes, the customer pays it up front as engineering charges.
• In between (hundreds to tens of thousands of units per design, many designs, frequent revisions), it dominates. That's the missing middle, in the volume sense Machina Labs CEO Edward Mehr uses: "making hundreds to thousands of complex parts fast, affordably" (LinkedIn). The phrase has other meanings in economics and policy. This is the one we mean.

Why it doesn't fix itself in the US comes down to evidence MIT's Suzanne Berger and Ben Armstrong gathered visiting American suppliers (MIT):

• The suppliers "are high-mix, low-volume manufacturers," and robots "are still very difficult to reprogram," so "it's not economically rational to buy a robot that can do only one operation."
• Only 9.5% of US manufacturing plants have even one robot.
• Orders "vary unpredictably, so heavy investment in equipment for making any particular product is unwise."

Those causes feed each other. Without steady orders, suppliers don't automate. Without automation, productivity and wages stay low and workers leave. Without a dense supplier base, manufacturers source abroad, which removes the orders. Lowering the fixed cost per design is the one lever that doesn't need anyone to bet on volume first. Two companies are pulling it from different ends, and one of them is in LA.

Two ways to cut the fixed cost per design

Machina Labs: replace the die with robot hours. A stamped part costs a share of the die plus a few dollars of press time. Machina's robots form sheet metal with no part-specific die, so each part pays for its own time on the robot cell instead. That's what "turning the die into a variable cost" means: the die doesn't get cheaper, it gets replaced by machine time you pay for on every part.

What drives that per-part cost is mostly forming time. The tool traces the part one contour at a time, stepping down a little each pass, so time grows with the part's formed area and depth and shrinks with a larger step and a faster feed. The tradeoff is finish: halving the step doubles the time and cuts the ridge height between passes to a quarter (derived from the published process literature). Then you add scanning, trimming, heat treatment, finishing and inspection.

Mehr says a part takes "somewhere from half an hour to a few hours" (Modern Machine Shop), and a Machina patent gives one example of 4 to 8 hours for a 1 mm stainless part, depending on forming force (US 11,865,716). Machina hasn't published what an hour of cell time costs. Under our own labeled assumptions (a $1M die and $150–$400 per cell-hour), dieless forming wins below somewhere between about 300 and 13,000 parts, depending mostly on how many robot hours each part takes. That range is the honest answer. The single number that would narrow it is Machina's real cost per cell-hour.

Arda: replace engineering weeks with compute plus review. Arda goes after the engineering term. Its value isn't out-searching engineers for a build order. It reads the constraints scattered across CAD, PDFs and notes, checks that a sequence is physically feasible, and carries the change through the build parts list, the routing and the instructions. Arda also says reprogramming conventional automation for one station "can take three to six months and cost close to $1 million," while a new phone stays at full production for only about four months. That's Arda's description of the status quo, with no source, but the logic holds: if reprogramming outlasts the production run, the automation never pays off.

Why now, and why not Anduril or the big vendors

The planning math is old. What changed is that it can read real inputs. McGrew says Arda builds instructions with "robotic motion planning to figure out how all the parts are supposed to come together" (CNBC/Altimeter interview, 11:28). The 1980s planners above stayed in the lab because the rules they needed were scattered across documents and people. Current AI models can read those documents. So the new part isn't the search. It's turning messy real inputs into the constraints a planner can check, then writing out everything downstream.

Anduril did half the work and bought the other half. Anduril spent years connecting its systems into ArsenalOS, which it describes as combining its own Forge platform "with commercial third-party systems, so every stage... runs on the best available tool." It calls building that layer "hard even for software-native companies." It worked with Arda's founders before the company was incorporated and runs Arda inside its own controlled environment. An agent needs a connected system to plug into, and Anduril had built one.

The big vendors are building toward it inside their own systems. Siemens released a planning assistant for "early adopter evaluation" in January 2026 (Siemens). Dassault calls a fully generated process plan its "destination." PTC generates manufacturing structures from existing parts lists and templates. All three work on data already in their own systems. Arda reads CAD, PDFs and data from any of them, and starts from geometry.

The novel angle: the instruction book is the robot program

What Arda sells today isn't a better search. It's the constraint-reading and feasibility-checking an engineer does by eye, plus the propagation of the change into every document in the chain above. The queues remain: approvals, delta FAI and physical checks still happen at human speed, so calendar time shrinks only as far as those steps keep up.

Today's Arda results are on human-staffed lines. But the motion plan that tells a technician how parts come together is also the starting point for telling a robot. So every instruction set Arda writes for people builds up the plans its next product, agents that program robots station by station, will need. Arda's goal for that product is automated lines "updated in days rather than months." It's still a prototype.

If it works, the most expensive per-design item, reprogramming automation, starts to look like the engineering hours do now. Fixtures, requalification and supplier changes stay physical.

That matters for defense right now. The Army says it wants at least a million drones over two to three years (Reuters), with designs that change as threats do. Deloitte and The Manufacturing Institute estimate 3.8 million manufacturing jobs to fill by 2033, with 1.9 million possibly unfilled (Manufacturing Institute). Anduril's COO Matt Grimm: "We would be behind on our production goals without it."

The so-what for LA

• The robot half is being built here. Launchpad Build AI in El Segundo launched a model in April that builds automation from a photo, video or CAD file, aimed at high-mix, low-volume manufacturers (PR Newswire). Arda starts from human-staffed lines and moves toward robots; Launchpad starts from the robot cell.
• The tooling half is here too. Los Angeles-based Machina Labs attacks the die. Arda attacks the engineering hours. LA has companies on both terms of the fixed cost per design, plus the defense customers who need it lower.
• LA hardware startups redesign faster than their manufacturing teams can grow. If re-planning costs compute plus one approving engineer, a small team can keep a line current.

The long-term bet: a faster idea-to-production loop

Here is our thesis, and it is unmeasured. If changing the factory gets cheap, a company can change the product more often. Each version gets to users sooner, feedback comes back sooner, and the next version is better. And once a change costs compute plus a review, building a variant for a single customer starts to look practical.

Research on product development supports the direction of that loop, though none of it is about assembly lines:

• Cheaper iteration cut effort. In a chip-design study, the approach that used many prototype iterations beat the alternative by a factor of 2.2 in person-months, and "over 43 percent of that difference" came from experimentation strategy (Thomke, 1998). The outcome measured was engineering effort, not product quality.
• Faster cycles can mean better products. Faster, cheaper experiment cycles can lead to "improved learning and ultimately products of higher quality" (Thomke, von Hippel & Franke, 1998, p. 321). The same paper notes that cycles only go as fast as people can learn between rounds.
• Earlier feedback went with better projects. Across 29 internet-software projects, "earlier feedback on product performance from the market" was associated with better-performing projects (MacCormack, Verganti & Iansiti, 2001). That's a correlation, in one industry.
• The edge comes from change speed. "If product designs could be changed instantaneously at low cost, the competitive impact of a new product would be sharply reduced" (Clark, Chew & Fujimoto, 1987, p. 731). If everyone can change cheaply, the advantage moves from having a good design to changing faster than rivals.

The loop has gates that Arda doesn't touch:

• Approval queues. In Loch and Terwiesch's study, most of a change order's lifetime was waiting, not work.
• Re-inspection. A change that could affect form, fit or function needs a partial first-article inspection before parts ship (LMUK SQAG 001).
• The physical world. Fixtures, tooling and supplier lead times move at their own speed.
• Government change control. In defense, a Class I change needs government approval and a contract modification (MIL-HDBK-61B).

There are also limits on how much faster iteration helps:

• It depends on uncertainty. Eisenhardt and Tabrizi's study of 72 computer-industry projects reportedly found that iterating sped up uncertain projects, while planning and overlapping tasks suited mature, predictable ones (1995; we have this only from secondhand summaries).
• Most of the cost of variety sits elsewhere. At a GM plant, day-to-day swings in option content raised overhead hours, downtime, rework and inventory, but not direct assembly labor (Fisher & Ittner, 1999). That cost falls mostly outside the step Arda speeds up.
• Tracking what got built. Per-customer variants also mean tracking exactly which configuration each serial number received. We think that's the next thing to automate.

The loop runs only as fast as its slowest gate. Faster re-planning is necessary but not sufficient. No one has yet measured whether it leads to better products.

What we still don't know

• How Anduril measured the 10-hours-to-15-minutes and 30-to-2.5-hours numbers, and whether review time is included.
• Where the "three to six months, close to $1 million per station" figure comes from.
• How Arda's "a week to two days" splits between engineering work and waiting in approval and inspection queues. No one has published a time study of assembly re-sequencing.
• Whether faster re-planning actually shortens the idea-to-production loop, or whether approvals, inspection and suppliers absorb the gain. No study links assembly re-planning speed to product improvement.
• Machina's real cell-hour cost and cycle time on a named part.
• When Arda's robot-programming product ships, on what hardware, and with whom.
• Whether Siemens, Dassault or PTC will partner, compete or buy.
• No SEC Form D or federal contract for Arda Industries turned up as of October 23, so the funding figures rest on the company's release and investors' posts.

All Arda and Anduril performance figures are company-reported.