The Speed of Custom Manufacturing is AI’s New Bottleneck

Read the article in SME / Advanced Manufacturing

Written by: Chris Cope, CADDi VP of Engineering

In April, the White House invoked Section 303 of the Defense Production Act and designated transformers, substations, high-voltage circuit breakers and electrical core steel as industrial resources essential to the national defense, on the finding that domestic capacity to produce them is "dangerously limited."

Industry tracker Wood Mackenzie put average U.S. lead times for large power transformers near 128 weeks, with the biggest units quoted at three to five years and medium-voltage switchgear sold out through much of 2028. Currence (formerly known as Sightline Climate) estimates that 30% to 50% of the U.S. data center capacity announced for this year will slip because operators cannot get the gear in time. An industry that spent three years worrying about chip allocation is now waiting on a coil-winding shop.

That inversion is worth taking seriously, because of what kind of manufacturing sits at the bottom of it. A large power transformer is configured to a utility's specification, built in small runs and qualified over long cycles. It is high-mix, low-volume work, which is the least automatable category we have. You cannot buy your way through a high-mix backlog the way a high-volume plant can, by purchasing faster machines and spreading tooling cost across a million units, because the dominant cost of the next job is the engineering that goes into it.

Throughput in that environment is measured in engineering decisions per week. Which means the expansion of artificial intelligence is currently rate-limited by how fast a few thousand engineers can make good calls about custom equipment.

The Constraint Is Not the Floor

Capital is not the shortage. The Federal Reserve Bank of New York's Manufacturing Survey, released in August, showed factory activity at its highest level since December 2021, with unfilled orders climbing and supply availability worsening.

Two days later, the Federal Reserve Bank of Philadelphia's Manufacturing Business Outlook Survey posted a five-year high alongside a future capital expenditures index at its highest reading since 1973. Orders are there and the money to expand is approved.

Deloitte and The Manufacturing Institute project a net need for 3.8 million manufacturing workers by 2033—roughly 2.8 million of that from retirements—with about half of those at risk of going unfilled. Engineers are the hardest part of that pipeline to rebuild because the education timeline cannot be compressed.

The realistic plan for the next three years is more output from roughly the engineering staff on hand today. That puts a lot of weight on where those hours currently go. Qualifying an alternate source for bushings because the incumbent is booked into 2029 means working out what the company built last time, what changed since and why, whether a clearance spec tightened in 2014 came from a customer requirement or a field failure, and what went wrong the last time they qualified a second source for the same part.

Those are engineering judgments, and the right people are making them. But most of the elapsed time is not judgment, it is retrieval. The engineer guesses part numbers, opens three or four systems that were never designed to talk to each other and goes looking for the one person who remembers the program. That person is retiring. The drawing he leaves behind records what was specified and says almost nothing about why, so the next engineer cannot tell which requirements are load-bearing and which are the residue of a problem solved 20 years ago.

In high-mix shops this compounds, because part counts are enormous, repeats are rare and the useful precedent is filed under a number nobody currently employed has a reason to know. This is why manufacturers in this position eventually point AI at the drawing archive—and it is the right instinct. The archive is the largest recoverable pool of engineering hours most of them have. It is also, for a growing number of them, about to stop being ordinary business data.

The Archive Is Changing Legal Status

The defense industrial base is widening deliberately. Roughly 10,000 new firms entered the defense market over the past two years and contract obligations to nontraditional companies more than doubled over the decade to $122.6 billion in fiscal 2025. The fiscal 2026 National Defense Authorization Act stripped out compliance requirements that used to keep commercial firms away and the Pentagon built a simplified entry platform and a reserve manufacturing network to recruit commercial capacity directly.

The April determination on grid equipment points the same way. The manufacturers being courted are exactly the high-mix, low-volume shops that the AI buildout is already waiting on. Most of them are evaluating that work as a customer question. The bigger change is what it does to their files. Under the International Traffic in Arms Regulations, technical data means the information required to design, produce, test, maintain or modify a defense article, and the regulation names the formats: blueprints, drawings, plans and documentation. The controlled thing is not a part in a crate. It is the drawing.

Releasing it to a external person is an export even when that person is an employee sitting in your own building in Ohio. Manufacturers who have been in defense for decades run a technology control plan for this but companies arriving through the new on-ramps generally do not, and they arrive carrying 30 years of drawings that were never segregated—because, until this year, there was no reason to segregate them.

Maintaining Data Boundaries

This is where AI evaluations go wrong, and it is a narrow technical point with a large consequence. Cloud storage of controlled drawings is only workable because of a specific carve-out in the export rules: Storing and transmitting technical data is not an export if it stays end-to-end encrypted and the vendor is never given the means to decrypt it.

That is the basis for most of the reassurance manufacturers get from software vendors and for storage it holds up. An AI system cannot read an encrypted drawing. To pull text off a title block, extract geometry or compare a part against prior designs, the tool has to hold that drawing in the clear, in its own environment, using keys it controls. The carve-out covers storage. It does not cover processing, which is the only thing you are actually buying.

So, a vendor who answers your compliance question with "encrypted at rest and in transit" has answered a question that was never the risk. The questions that matter are where the drawing is decrypted, who at the vendor can see it while it is readable and whether what the tool extracts from it—which is usually dimensions, materials and tolerances—is treated as controlled data, too. It generally should be, along with the index and the logs built from it.

None of that makes AI unusable on controlled drawings. It makes the data path something to settle during evaluation rather than after a pilot has been running on live files for six months. A vendor's export registration, mine included, is not the control that matters. Where the data is processed and who can touch it in the clear is.

Apply the same skepticism to the capability claims. Geometric similarity search on two-dimensional drawings is genuinely hard and has only recently become practical, and it changes the starting point by letting an engineer work from the drawing instead of a part number they would have to guess. What comes back is a search hit. It is not a finding that two parts are interchangeable and it is not a qualification. Any vendor blurring that line is describing something they have not built.

Where to Start

Find out what a decision costs you now. Pick one your team already finished, ideally where a requirement changed or a source moved, work backward from the final configuration, and write down what an engineer would need to explain why the part ended up that way. Count the systems, the searches and the people.

If the answer depends on an exact identifier or on someone within a few years of retiring, you have engineering capacity tied up in retrieval and you are about to need it back. Then find out where controlled data would actually travel in any tool you are considering. Not the storage diagram.

The processing path, the derived data, the access list. If you are heading into the defense industrial base, that boundary gets audited eventually, and drawing it early is far cheaper than reconstructing it after your archive is spread across an AI pipeline.

The chain here is worth saying plainly, because it is not the one the industry has been describing. The next gigawatt of AI capacity waits on a transformer. The transformer waits on engineering hours at a high-mix shop. A growing share of those hours are moving behind an export boundary that most of those shops have never had to manage. Export control is now on the critical path of the AI buildout, which is not where anyone expected to find it.

About CADDi

CADDi is a global manufacturing technology company driven by the vision to accelerate physical innovation tenfold. Headquartered in Chicago and Tokyo, the company was founded in 2017 by industry veterans Yushiro Kato and Aki Kobashi, formerly of McKinsey and Apple, respectively. CADDi currently serves manufacturing businesses in more than 20 countries, including many of the largest publicly traded companies in Japan and the United States. To learn more, visit caddi.com.