The AI revolution is real. It is arguably the most important technological shift since electrification. The trillion-dollar capital deployments, the workforce disruptions, the geopolitical scramble for compute — all of it is happening, and none of it is hype.
But the AI revolution has a physical dependency that almost nobody is discussing.
Every datacenter needs cooling manifolds machined to tight tolerances. Every F-35 has thousands of precision-milled components. Every semiconductor fab requires custom fixturing. Every satellite, every surgical robot, every autonomous vehicle — at some point in their production, metal must be cut to shape by computer-controlled machines. This process is called CNC machining, and the United States has a structural, worsening shortage of the people who know how to do it.
There are 354,800 machinists in America right now. The largest age cohort is 55 to 59. The Bureau of Labor Statistics projects the occupation will decline by 2% over the next decade. Every one of the 34,200 annual job openings is a replacement — someone retiring or leaving. Not a single one represents growth.
The occupation that physically produces nearly every precision component in the advanced economy is shrinking, at the exact moment demand is surging from every direction.
The Industrial Base Cannot Surge
There is a concept in defense planning called "surge capacity" — the ability to rapidly scale production when a crisis demands it. It is the foundation on which deterrence rests. If your adversaries believe you can outproduce them when it matters, they think twice.
America's surge capacity is broken.
The proof is straightforward. When Russia invaded Ukraine in early 2022, the United States was producing approximately 14,000 155mm artillery shells per month. These are not exotic munitions — steel tubes filled with explosives, a design fundamentally unchanged for decades. The Pentagon set a target of 100,000 rounds per month. Three years later, production sits at 40,000. The Army doesn't expect to hit 100,000 until mid-2026, and that timeline has already slipped multiple times.
The Center for Strategic and International Studies noted the Army's struggles are "linked to the state of the industrial base before the Russian invasion of Ukraine." This was not a surge problem caused by Ukraine. It exposed a capacity problem that already existed.
The story is worse for complex systems. The F-35 Lightning II had over 4,000 parts shortages on Lockheed Martin's assembly line as of early 2025 — double historic levels. Fifty-two aircraft sat stalled in final assembly, waiting for components. Average delivery delays reached 238 days.
The Heritage Foundation's assessment uses language you don't normally hear from a think tank known for measured analysis: the domestic defense industrial base has "slowly declined as domestic defense production has consolidated and American manufacturing has moved overseas." Decades of poor policymaking have led to "underinvestment, regulatory drag, labor misalignment and market distortions" — all in what the Heritage Foundation calls "the most hostile global environment since World War II."
The National Defense University Press was more blunt: "A growing chorus of U.S. defense analysts, lawmakers, and military officials has emphasized that the United States lacks the munitions production capacity to meet the demands of the contemporary strategic environment."
This is the backdrop against which the reshoring consensus is supposed to deliver results. Bipartisan agreement. Hundreds of billions in subsidies — CHIPS Act, IRA, defense appropriations. Political consensus, capital, demand.
But no one to do the work.
The Workforce That Isn't Coming
The aggregate statistics understate how dire this is.
The BLS counts 354,800 machinists and tool-and-die makers in the United States as of 2024. The median wage is $56,150 — decent, but not the kind of money that draws ambitious young people away from tech or finance. The three largest age cohorts are 55–59 (47,710 workers), 60–64 (43,797), and 50–54 (43,138), according to American Community Survey data. That means roughly 134,645 machinists — 39% of the entire workforce — are over 50.
The replacement pipeline is, to use a technical term, nonexistent. The BLS projects a net decline of 5,900 machinists over the next decade. There is no authoritative national count of how many CNC machinists graduate from training programs each year, which tells you something about how seriously we've been tracking this.
Zoom out to all of manufacturing and the picture gets worse. Deloitte and the Manufacturing Institute projected in 2021 that 2.1 million manufacturing jobs could go unfilled by 2030, at a cost to the economy of up to $1 trillion. That projection was calculated before the reshoring push, before the defense buildup, before $602 billion in datacenter capex. The actual gap is almost certainly larger.
There's a structural reason wages aren't solving this. Machinist wages have been relatively flat in real terms. In a functioning labor market, acute shortages drive wages up, drawing new entrants. That isn't happening because most shops can't afford to pay more — their margins are already thin, their customers won't absorb higher prices, and they're competing against overseas production where labor costs a fraction of the domestic rate. It's a death spiral: low wages → fewer entrants → deeper shortage → still can't raise wages → more people leave for easier careers.
Every factory owner I've spoken to — and I've spent a decade in this industry talking to hundreds of them — gives the same answer when asked about their biggest constraint. Not materials. Not machine availability. Not demand. People who can program CNC machines.
The Bottleneck Nobody Talks About
CNC machining has a specific bottleneck that most people outside the industry don't know exists: CAM programming.
CAM stands for Computer-Aided Manufacturing. It's the process of taking a CAD model — a 3D design of a component — and turning it into the precise sequence of instructions that tells a CNC machine how to cut metal. What speed, which tools, in what order, how to hold the part, how to avoid collisions, how to stay within tolerances. The combinatorial space is genuinely astronomical — there are more possible ways to machine a typical component than there are atoms in the observable universe. The combinatorial explosion of tool selection, cutting strategies, sequencing, fixturing, speeds, feeds, and collision avoidance is one of the hardest optimization problems in industrial engineering.
It takes years of training and experience to get good at it. A skilled CAM programmer doesn't just know the software; they know how metal behaves under cutting forces, how tools deflect, how heat builds, when a process is drifting toward failure. It's deep, embodied expertise built over thousands of hours at the machine.
And the people who have it are retiring.
A senior Lockheed Martin Ventures executive put it to me this way: "They could not hire all the machinists in the world and be able to make what they're being asked to make by the Pentagon." This from someone inside the largest defense contractor on Earth — $67 billion in annual revenue, access to every lever of American industrial power — saying that hiring alone cannot close the gap. It's one data point, but it's consistent with what the BLS numbers, the Heritage Foundation, and every shop owner I've talked to are saying: the capacity problem is structural, not cyclical.
The bottleneck is not machines. America has machines. The bottleneck is not money — the defense budget is at historic highs. The bottleneck is the human capacity to program those machines.
The Drudgery Problem
What makes the workforce crisis worse is that the existing workforce is dramatically underutilized — not because they're lazy, but because their tools are terrible.
Most people, if they think about CNC machining at all, picture a skilled craftsperson carefully shaping metal into precision components. The reality is that most of a CNC programmer's day is spent fighting software. CAM software packages are legendarily bad. The interfaces are arcane, the workflows convoluted, and straightforward tasks become exercises in digital bureaucracy.
It can take five to fifteen minutes just to orient a part correctly in the virtual vise inside a CAM package. Not because the programmer doesn't know what they're doing, but because the software requires clicking through nested menus and configuring parameters that should be defaults. Multiply that by every operation on every feature of every part, and a programmer who could be thinking about process optimization is instead filling out digital forms.
The industry average for spindle uptime — the percentage of time a CNC machine is actually cutting metal, which is the only time it generates revenue — is less than 50% during operating hours. Not 50% of a 24-hour day. Fifty percent of the shift when someone is supposedly running it. The rest is setup, programming, waiting for programs, troubleshooting. Most shops can't run additional shifts because they don't have enough programmers to keep the machines fed.
This represents a staggering amount of latent capacity sitting idle inside existing machine shops. The constraint isn't that America needs more factories. It's that America needs to unlock the factories it already has.
Lead times are typically six to twelve weeks. Customers want to order more but can't because shops don't have the programming capacity. It's not a demand problem. The demand is there. The constraint is human attention.
The Reshoring Math
Everyone agrees America should make more things domestically — supply chain resilience, national security, job creation, reduced dependence on geopolitical adversaries. The problem is that reshoring, as currently conceived, doesn't work.
The reason manufacturing moved overseas wasn't that Chinese or Vietnamese workers are better. They're cheaper. Dramatically cheaper. The labor cost per part in Shenzhen versus Indianapolis is so large that companies happily pay to ship across an ocean, clear customs, and accept longer lead times. The shipping costs, logistics headaches, and inventory buffers are still cheaper than American labor.
Subsidies can close part of that gap but they're temporary and politically fragile. Tariffs can close part of it but raise costs downstream and invite retaliation. Neither addresses the fundamental issue: even if you wanted to make things in America, you don't have enough people who know how.
But there's a variable in this equation that most reshoring advocates haven't thought about carefully enough: the cost of the person in the part.
When a CNC programmer spends eight hours writing a program for a component, that labor cost gets embedded in every part the program produces. If the programmer produces one or two programs per day — typical for complex parts — then human programming time is a significant fraction of the per-unit cost. And it's the fraction where the US-to-overseas delta is widest.
Now consider what happens when AI handles the repetitive, time-intensive portions of CAM programming — roughing strategies, standard feature recognition, tool selection for common operations — while the human focuses on process planning, tolerance management, the judgment calls that require experience.
If that same programmer produces ten programs per day instead of one, the labor cost per part drops by an order of magnitude. The delta between domestic and overseas production shrinks dramatically. Suddenly the cost of putting things on boats — shipping, customs, inventory carrying costs, lead time risk — becomes the dominant factor. The economics of proximity start to win.
This isn't hypothetical. AI-powered CAM tools exist today and are deployed in hundreds of machine shops, producing documented programming time reductions of 60–95% for standard operations. The machinists using them aren't doing less-skilled work. They're doing more of the highest-skilled work — the judgment calls, the process optimization — while AI handles the drudgery that currently consumes most of their day.
One experienced machinist, augmented by good AI tools, can cover the programming output of a small team. That changes the reshoring math from fantasy to viable.
AI That Understands Metal
There's a widespread assumption that AI is primarily a knowledge-work technology — that it automates text, code, and images, not physical processes. This is wrong in an important way. AI doesn't need to physically operate a machine to transform manufacturing. It needs to automate the cognitive bottleneck: the programming step between a design and a finished part.
But you can't just point a large language model at a CAD file and ask it to generate a toolpath. An LLM has no understanding of cutting forces, tool deflection, chip evacuation, workholding constraints, or the thousand physical realities that determine whether a program produces a good part or destroys a $50,000 spindle. The primitives don't exist in its training data. Machining is not text.
What's required is purpose-built AI infrastructure: machining strategy engines, tool libraries, collision detection systems, optimization solvers that can navigate the combinatorial explosion of possible manufacturing approaches. Systems that understand the physics and practice of cutting metal, built by people who have actually done it.
This is a decade-scale engineering problem. The companies that have been working on it longest have the deepest moats — not because manufacturing AI is trendy now, but because the problem genuinely requires years of accumulated domain data, engineering iteration, and hard-won understanding of how metal actually behaves under cutting forces. There are no shortcuts.
The Money Is Moving
In the last twelve months, three manufacturing AI companies have raised a combined $510 million in venture capital.
Hadrian raised a $260 million Series C in July 2025, led by Founders Fund and Lux Capital, with participation from a16z and Morgan Stanley. Their approach: AI-powered factories built from scratch for precision machining, focused on defense and space — capex-heavy, hardware-plus-software.
Bright Machines raised $126 million in May 2025, led by BlackRock, with NVIDIA and Microsoft participating — software-defined manufacturing focused on robotic assembly.
Machina Labs raised $124 million in March 2026 for AI-driven metal forming, backed by the US Air Force.
The investors behind these rounds — Founders Fund, a16z, Lux Capital, BlackRock, NVIDIA — include some of the most disciplined capital allocators in technology. They see the same data the BLS publishes: machining capacity is structurally declining at the exact moment demand is structurally increasing. They see that the gap cannot be closed by training alone, because the pipeline is broken and the demographic math doesn't work. And they see that AI applied to the right bottleneck is the only lever that scales fast enough.
There are broadly two approaches. One is building entirely new automated factories — Hadrian's model. This requires enormous capital, years of construction, and novel hardware. It's a valid strategy, but inherently slow.
The other is making the existing installed base of CNC machines — hundreds of thousands of them, sitting in shops across America — dramatically more productive by augmenting the people who operate them. Not replacing machinists. Giving them tools that handle the repetitive work so they can do ten times as much of the high-value work only they can do.
The first approach builds new capacity. The second unlocks capacity that already exists. Both are necessary. But only the second can deliver results at the speed the current crisis demands, because the machines, the shops, and the machinists are already there — for now.
The $602 Billion Tailwind
If you think the machining capacity crisis is limited to defense, consider what the technology industry is building.
In 2026, the five largest hyperscale cloud companies — Amazon, Google, Microsoft, Meta, and Oracle — are projected to spend a combined $602 billion in capital expenditure. Roughly 75% — approximately $450 billion — is going directly to AI infrastructure. This represents a 36% increase over 2025 and has pushed capital intensity to 45–57% of revenue, historically unthinkable levels for companies this size.
That money isn't just buying GPUs. It's building physical infrastructure — and critically, cooling systems. As GPU power density increases (NVIDIA's GB200 draws over 1,000 watts per chip), air cooling becomes physically inadequate. The industry is in the middle of a forced transition to direct liquid cooling, and every liquid-cooled GPU rack requires precision-machined cold plates, manifolds, and fittings.
The datacenter liquid cooling market is projected to grow from roughly $3–5 billion in 2025 to $15–27 billion by the early 2030s. That market is almost entirely dependent on CNC-machined components — copper and aluminum cold plates with micro-channel structures machined to tight tolerances.
Beyond cooling: machined rack hardware, power distribution busbars, cable management systems, structural components. Many hyperscalers are building dedicated power generation — gas turbines now, small modular nuclear reactors eventually — all requiring precision machined parts: turbine components, valve bodies, pump housings, flanges.
Conservative estimates suggest $30–100 million in machined components per $1 billion of datacenter construction. If even 2–5% of the $602 billion in 2026 hyperscaler capex flows to machining, that's $12–30 billion in annual demand from datacenters alone — on top of defense, aerospace, automotive, medical, and industrial demand.
The AI revolution, in other words, is generating enormous demand for the very manufacturing capacity that is structurally declining. The revolution has a dependency, and the dependency is in trouble.
The Knowledge Window
There is a dimension to this crisis that gets less attention than the workforce numbers but may matter more.
The 47,710 machinists in the 55–59 age bracket are five to ten years from retirement. The 43,797 in the 60–64 bracket are leaving now. Every one of them carries decades of accumulated knowledge — the tribal knowledge of how to hold a difficult part, the intuition for when a tool is about to fail, the sixth sense for a process drifting out of tolerance. This knowledge is nowhere in a manual or a corporate database. It exists only in the heads and hands of people heading for the exits.
AI-powered CAM systems that learn from thousands of successful machining operations can encode some of these patterns. But this only works if the technology is deployed while the experienced machinists are still there to teach it. Once they're gone, the knowledge goes with them. This isn't just a labor shortage. It's an information loss event.
The Deloitte projection of 2.1 million unfilled manufacturing jobs by 2030 was published in 2021. It's 2026. More than halfway to the deadline, and the gap has gotten worse. The BLS outlook is negative. The training pipeline is inadequate. The wage dynamics are stuck in a structural trap.
The reshoring consensus — the one area of genuine bipartisan agreement in an era of political paralysis — depends on something that physically cannot happen without a force multiplier. You cannot reshore manufacturing to a country that is losing machinists. You cannot rebuild a defense industrial base that can't surge because it can't staff. You cannot spend $602 billion on AI infrastructure without dramatically increasing demand for the precision machining that infrastructure requires.
AI-powered manufacturing is not a nice-to-have. It is not a "digital transformation initiative" for some future planning cycle. It is the load-bearing wall of every major economic and security initiative America is currently pursuing.
What Comes Next
There are 354,800 people standing between the United States and industrial irrelevance. They are overwhelmingly male, disproportionately over 50, skilled beyond what most Americans imagine, and they are not being replaced.
The conventional responses — more training programs, higher wages, better marketing of manufacturing careers — are correct but insufficient. We should pay machinists more. We should fix apprenticeship programs that teach G-code with pencils when they should be teaching AI-assisted CAM from day one. We should tell young people that CNC machining is intellectually demanding, increasingly enhanced by AI, and desperately needed.
But the demographic curve doesn't negotiate. The 2.1 million job gap by 2030 won't be closed by training programs that take years to produce graduates for an occupation that takes years more to master.
The viable path is making each machinist dramatically more productive. Not by working them harder — they already work hard. Not by cutting corners on quality — the tolerances exist for a reason. By giving them AI tools that handle the repetitive, time-consuming portions of their work so they can apply their irreplaceable expertise to more parts, more processes, more challenges.
This is the last American advantage that matters. Not that the US has the most machinists — it doesn't. Not the most machines — China leads there. Not the cheapest labor — it never will. The advantage is the convergence of deep manufacturing expertise, leading AI capability, and the institutional urgency to deploy both — while the generation that carries the knowledge is still here to encode it.
"While they're still here" has an expiration date. The largest cohort of American machinists turns 60 this decade. The knowledge is perishable. The window is real.