Why Anduril is qualifying submarine welds two years before its shipyard opens

At 160,000 square feet in Orange County, California, a production line is running a test the Navy has never run before. Robots carrying ultrasonic sensors are working the same welds as the welding equipment, feeding readings to software that decides whether the metal underneath is sound. This is not a laboratory bench. The parts moving through that line are Virginia-class submarine components, torpedo tubes among them.

The line belongs to Anduril. The robots belong to Gecko Robotics, a Pittsburgh company founded in 2013 that has spent four years selling inspection to the Navy's maintenance fleet. On October 8 the two companies announced that Gecko's platforms and its Cantilever data platform would move into manufacturing itself, embedded in Anduril's own factory software, inside the Navy's SUBSAFE certification regime for watertight-critical components.

That announcement landed two days after Anduril unveiled a 2 million square foot shipyard in Baltimore County called Arsenal-2, backed by $3.7 billion in private capital and a Navy contract worth up to $2.9 billion. The two announcements together represent $6.6 billion. The second one carries the actual story.

The rate that has not moved since 2022

1.1 to 1.3. That is roughly how many Virginia-class boats General Dynamics Electric Boat and Huntington Ingalls Industries have delivered each year since 2022. The Navy needs two a year, rising to 2.33 once AUKUS transfers to Australia are counted. The shortfall has usually been described in terms of suppliers, steel, and yard footprint.

A senior administration official pointed somewhere less comfortable on October 6, calling torpedo tubes "a gating component" for boats already under construction and estimating that Arsenal-2 would add roughly 9 million labour hours a year to the submarine industrial base, about a 15 percent increase in Virginia-class capacity. Floor space is the part of this problem that capital can buy.

The gate nobody names

The Navy's contract with Anduril is structured unlike most defence procurement. Payments track demonstrated production outcomes rather than passing milestones, and the company absorbs most of the execution risk. Military Times described the arrangement as part of a single $6.6 billion package in which the private capital and the government award move together.

That structure prices delay differently. A component sitting in an inspection queue, waiting on a qualified technician, is revenue the company has already spent capital to earn. Queue time stops being a scheduling inconvenience and becomes the line item most directly wired to cash.

Which is what makes the Gecko partnership more than a vendor selection. Non-destructive testing — NDT, the ultrasonic examination of welds and forgings for internal flaws, porosity and wall loss — has long rested on certified technicians who take years to train. Wall loss, porosity, and internal flaws in welds and forgings have to be located without cutting the part open, and the record of what was found has to stay intact for the life of the boat.

Under a fixed milestone contract, an inspection bottleneck shows up as a schedule problem somebody escalates. Under this one, it shows up as money that was supposed to arrive this quarter and did not.

It allows us to go in and assess and create that digital baseline of the actual, as-built condition of this critical component, and so it takes things like our robotic capability and actually gathers and inspects what is going on underneath the weld.— Troy Demmer, co-founder and president, Gecko Robotics

Anduril's vice president and general manager for shipbuilding, Bill Noderer, said the NDT technology is already embedded in live production workflows at the Orange County protofactory and integrated with ArsenalOS, the industrial software that will run Arsenal-2. His stated aim is to cut inspection wait times and establish quality processes the company can carry to Maryland from the yard's first day.

Two years of head start

2030 is when Arsenal-2 is scheduled to open. The 160,000 square foot California facility is not a satellite office; it is the rehearsal. Anduril's own release describes it as a two-year operational head start to develop and validate manufacturing methods, production workflows, and quality systems before Baltimore comes online.

The site will support prototyping and production of Virginia-class test assets and critical maritime components. It will recruit and train the early workforce on the systems and standards that eventually scale. In practical terms the sequence runs: qualify the line in California, prove the torpedo tubes, transfer the paperwork and the process knowledge to Maryland, then expand from basic modules into complex assemblies and eventually the bow.

That is the correct order of operations, and it is also the admission. Building a shipyard is a known problem with known contractors. Standing up a SUBSAFE production qualification regime at a site that has never made a submarine part before is not.

An industry official told USNI News that Anduril has already begun building prototype torpedo tubes in California. Bryan Clark, a naval analyst at the Hudson Institute and a former submariner, told the same outlet that Anduril cannot substitute commercial parts into certified nuclear hardware, because "the supply chains are very tight."

Strip out the press release and the sequencing says one thing. Anduril is buying the option to qualify early, and it is paying for that option with two years of work on the most constrained process in the yard.

What the 90% figure does and does not cover

Gecko states that its systems have cut inspection timelines for forged components by up to 90 percent at defense industrial base customers, and that a worker with no prior NDT background can become productive in roughly an hour. Both numbers recur in coverage of the Anduril deal. Neither has been independently validated, and the announcement does not say which robotic platforms, sensor modalities, or inspector certification levels will apply at Orange County.

The 90 percent figure comes from prior work, principally trials with BPMI, a contractor supplying components for Navy nuclear-powered aircraft carriers and submarines. It is a real customer result on real forged parts. It is not a measurement of a Virginia-class production line with a first-time torpedo tube supplier on it.

The distinction matters because the two environments differ in exactly the way that tends to break throughput claims. A production line under a payment-for-output contract cannot afford a false negative, so acceptance thresholds stay conservative. Automation that accelerates inspection while generating a denser audit trail is useful. Automation that accelerates inspection and then fails acceptance is a new queue.

Three separate programmes are now automating work inside the same supply chain, with different scopes and different levels of public evidence:

Programme figures as announced by each party. Throughput claims are vendor-stated and not independently validated.

Two details from that table deserve more weight than the headline numbers. First, only one of the three programmes has publicly claimed a measured inspection-speed improvement, and it is the one being deployed on the newest hardware. Second, the HII agreements are explicitly milestone-gated, with work awarded only as suppliers hit technology-readiness, manufacturing-readiness, cost, schedule, and quality targets.

That is a different procurement philosophy from a payment-for-output contract. HII has built its demand signal so that unqualified machinery cannot bypass naval production controls. Anduril has accepted the risk in exchange for the revenue. The same automated inspection technology carries very different consequences depending on which side of that trade a company is standing on.

It is also worth noting what has not been disclosed at all. The financial terms and duration of the Gecko-Anduril agreement were not released, and Anduril's Maryland and Baltimore County incentives covering up to 13 percent of the planned investment still require approval from the state legislature and county council. Construction still depends on permitting approvals that have been announced but not granted.

What SUBSAFE will not bend for

SUBSAFE is the Navy's certification regime for the components that preserve a submarine's watertight integrity and its ability to recover from flooding. It does not scale its expectations for new entrants. A first-time torpedo tube supplier meets the same acceptance criteria as General Dynamics or Huntington Ingalls.

That constraint is what makes the Orange County sequencing unusual. Rather than building a large yard and then discovering whether the process works, Anduril is running the certification path on a smaller footprint while it still has the option to change the process. The California site is cheap enough relative to Baltimore to be a genuine experiment instead of a sunk monument.

Weld procedures, consumables, joint preparation, heat input, penetration, distortion, and inspection evidence all have to stay inside approved limits. Automated inspection can accelerate throughput while generating the traceability record that certifying authorities actually demand. Those are two separate products, and only one of them is a robot.

Gecko has been inside this ecosystem before. As we wrote in August 2026, its robots were scoped to fleet maintenance — hulls, ballast tanks, flight decks, ships already in service. Since November 2023 the company has performed robotic weld inspection on the $132 billion Columbia-class program through BlueForge Alliance, the nonprofit integrator supporting the submarine industrial base, alongside maintenance work on Virginia-class boats and aircraft carriers since 2024. In March the Navy awarded it a five-year indefinite-delivery, indefinite-quantity contract with a $71 million ceiling and up to $54 million of initial work across 18 Pacific Fleet vessels.

None of that work was production acceptance of new-build hardware. It was inspection of assets already in service, where a missed defect gets caught later. The Orange County deployment is the harder version, and it is the first time the company has said publicly that the systems are running in live production workflows.

The workforce argument

The labor dimension is doing more work in these announcements than the robotics. Certified NDT technicians take years to train, and the naval industrial base does not have them in surplus. The Navy's long-range shipbuilding plan, released in May 2026, names supply chain stabilisation and skilled workforce growth as near-term priorities on the way to what the service calls the 1+2 rate, one Columbia-class and two Virginia-class boats per year.

Chris Brose, Anduril's president and chief strategy officer, described the operating philosophy as taking very complex tasks performed by very skilled workers and breaking them into a larger number of simpler tasks that a larger workforce can contribute to, according to DefenseScoop. An industry executive told USNI News that hiring at Arsenal-2 would draw on commercial aerospace, the car industry, vocational schools, and high schools.

Gecko's one-hour training figure is the most aggressive expression of that idea, and it depends on a specific technical premise: that the inspection expertise lives in the robot and the software rather than in the operator's head. Cantilever converts raw scans into thickness maps and flags suspect areas. A person with an hour of training reads the output and adjudicates it, rather than knowing how to acquire and interpret an ultrasonic signal cold.

If that premise holds, it changes what a shipyard can hire. If it does not, the constraint simply moves: more operators qualify, then more of them fail acceptance review, and the queue gets a second layer instead of losing one.

Hadrian, under a separate $900 million Navy contract announced in March, has set a public bar that is easier to read. Its goal is to automate 80 percent of the work and keep processes simple enough to train workers in 30 to 40 days while humans stay in the loop. That is a deliberately modest claim about how much of a certified trade can be compressed, published by a company that wants the next award.

Gecko is claiming a harder version of the same idea from a smaller base. One hour sits well outside the range anyone else in this supply chain has put in public.

Everyone else is queuing for the same technicians

The submarine production problem is not a shortage of ambition. A Congressional Research Service estimate put Navy investment in the submarine industrial base at roughly $9.8 billion since 2018, and the most recent tranche runs through several programmes at once. HII plans to outsource more than 2.5 million hours of shipbuilding work in 2026, a 30 percent increase over 2025, while expanding its network of assembly partners completing structures outside its principal yards before final integration.

Each of those moves pulls on the same finite pool of people who can weld a nuclear-pressure boundary to standard and inspect it afterwards. A robotic welding cell at a Path Robotics pilot and a robotic inspection cell at an Anduril protofactory compete for identical hires.

There is also a network effect inside the venture-backed cohort that is rarely mentioned in the announcements. Anduril co-founder Trae Stephens joined Gecko's board through Founders Fund in December 2023. The two companies have worked together for more than a year, and Demmer said Gecko has worked with Anduril on at least three other programmes at different facilities, including airworthiness work.

Gecko is not an arm's-length supplier selected from a tender. It is a portfolio company inside the same capital stack. That explains the speed of the integration and the vagueness of the commercial terms. It also means the competitive pressure other suppliers face here is structural rather than commercial.

What would break the bet

Anduril has roughly two years before the answer becomes expensive. If Gecko's systems cannot produce an acceptance record that Navy certification authorities will sign, the inspection queue survives into Arsenal-2 with a robotic layer added and none of the capacity recovered. The facility was built to surface that failure early, which is the point of building it early.

The second failure mode is quieter. Payment tied to production means the pressure to accept faster rises as the schedule slips, and acceptance criteria that tighten under that pressure produce a queue that no amount of sensor automation shortens.

The third is supplier-side. Anduril's release notes it still has to build its supplier base for qualified materials and fittings, which means new orders for makers of components that go into torpedo tubes and later modules. A yard cannot inspect its way out of a missing qualified supplier.

The Navy projects two Virginia-class deliveries a year by 2031 or 2032. Arsenal-2's initial operations are scheduled for 2030. The Orange County line is the first test, and it runs long before anyone knows whether the answer holds.

Torpedo tubes were described as a gating component for boats already under construction. Automation is now aimed at the step after the tube exists, on the theory that the queue, not the factory, is what has been holding the line.

Floor space was never the scarce input. Time with a certified pair of hands was.