Zero Sum

The Number

427,000 welders short by 2028. 354,800 machinists — a third of them over 55. 64,000 tool and die makers, down 30% since 2000. Tens of thousands of foundry workers, NDT inspectors, heat treatment specialists, composite layup technicians. Add them up and you get America's skilled manufacturing workforce: finite, aging, and fought over by everyone.

AI needs them. Defense needs them. Energy transition needs them. Reshoring needs them. They can't all have them.

In Part I, I showed that the AI buildout is hitting a physical wall — made of metal, shaped by skilled hands that are disappearing. In Part II, I showed the problem is worse than it looks: turtles all the way down, every layer of the supply chain thinning out.

Now the part that should keep you up at night: everyone is fighting over the same finite pool of skilled tradespeople. And there is no one coming to replace them.

It's Not a Shortage. It's a Collision.

Every factory owner says the same thing. Not just about machinists — about welders, toolmakers, foundry operators, NDT inspectors, heat treat technicians, composite techs. Walk the floors of forge shops, investment casting houses, composites layup facilities, metrology labs, heat treatment plants — the pattern is identical everywhere. The average age is fifty-five. The apprenticeship pipeline is a trickle. And four industries are about to fight over the same shrinking pool.

The shortage isn't new. We've had it for a decade. What's new — what's unprecedented — is that four enormous demand surges have arrived simultaneously, and they're all drawing from the same finite pool.

This isn't a shortage. It's a collision. Four legitimate national priorities — each individually critical, each with hundreds of billions in funding — all competing for the same skilled tradespeople, in the same metro areas, at the same time. The pool they're drawing from is not growing. It's shrinking.

In 2026, the United States is simultaneously trying to:

1. Build AI infrastructure: $350B+ in hyperscaler capex in 2025, projected above $600B in 2026
2. Modernize the energy grid: $550B in IIJA spending plus $369B in IRA investment
3. Rearm its military: $175B in weapons procurement, with a defense industry already at capacity
4. Reshore its supply chains: 244,000 new manufacturing jobs announced in 2024, plus the entire CHIPS Act buildout

Total mobilization. Across every sector. With 12.6 million manufacturing workers — the lowest share of employment in American history — and a skilled trades core that is a fraction of even that.

The math doesn't work. Not because any individual program is wrong. Because they're all right, and they all need the same welders, machinists, toolmakers, inspectors, and technicians.

The Grid That Can't Keep Up

The IIJA and IRA together represent the largest energy infrastructure investment in American history — nearly a trillion dollars funding grid upgrades, renewables, transmission, and domestic energy manufacturing. On paper, transformative. On the ground, constrained by human hands.

Power transformers. The US grid runs on ~70,000 large transformers. Many are past their 40-year design life. Replacement lead times have stretched from one year to two-to-four. CISA — a national security agency — published a report warning the grid can't replace its own aging equipment fast enough. Most large transformers are imported. Building them domestically requires heavy welders, coil winders, and vacuum-impregnation technicians — the same trades the defense industry needs for ship hulls and the AI buildout needs for datacenter structures.

Same welder. Same technician. Three bosses.

Gas turbines. GE Vernova has an 80-gigawatt backlog stretching to 2029. Lead times: up from two and a half years to seven. Datacenter demand is the primary driver — Crusoe alone ordered 4.5 GW. The machinists who cut turbine blades for these power plants are the same machinists Pratt & Whitney needs for jet engines. The same machinists making cold plates for liquid-cooled server racks.

Small modular reactors. NuScale, X-energy, Kairos, TerraPower — the nuclear renaissance everyone agrees is essential for baseload power. Every one of these projects needs heavy forgings and precision welding from a vanishingly small supplier base. Nuclear competes for the same welding and machining workforce as submarine hull construction. As transformer manufacturing. As datacenter cooling systems.

The same welder. The same forge operator. The same shrinking pool.

The AI Buildout Eats Everything

Hyperscaler capex in 2025: above $350 billion, roughly 75% directed at AI. Microsoft committed $80B. Amazon, $100B+. Google, $75B. Meta, $60-65B. The Stargate JV announced $500B over multiple years. 2026 projections exceed $600B.

Every dollar of that must become a physical object. A datacenter isn't software — it's a building full of manufactured metal: structural steel, custom racks, machined cooling components, power distribution equipment, generators, transformers. A single 100MW AI datacenter requires $16-57 million in machined parts alone.

And the datacenter is only the first link. Every one needs power — turbines backordered to 2029, transformers with multi-year lead times, high-voltage grid equipment that didn't exist in the domestic supply chain five years ago. AI infrastructure doesn't just compete for workers at the datacenter level. It competes at every layer of the industrial stack, from the chip fab to the power plant to the copper mine.

Here's what makes the AI buildout different from the other three surges: it has the deepest pockets. Hyperscalers can — and do — outbid the Pentagon, outbid utilities, outbid chip fabs. When a machine shop in Ohio can get $200/hour for datacenter work or $120/hour for defense subcontracting, the shop takes the datacenter work. Rational. Devastating for everything else.

AI is the vacuum cleaner. It hoovers up skilled tradespeople from every other sector, and it can afford to keep going.

The Arsenal That Can't Surge

The defense industrial base: 1.17 million workers, 60,000+ companies, $175 billion in annual procurement. On paper, the most powerful military-industrial complex in history. In practice, it cannot deliver what it's been asked to deliver.

Virginia-class submarines: 1.2 per year against a target of 2. Artillery shells: 40,000 rounds/month after three years of trying to reach 100,000 — still at 40%. The F-35: 4,000+ missing parts, 52 aircraft stalled on the line, 238-day delivery delays on machined components from sub-tier suppliers who can't hire enough people.

I've talked to machine shop owners, forge house managers, and fabrication yard foremen in these supply chains. They aren't struggling because they lack orders. They're struggling because their best people keep getting poached — machinists, welders, NDT inspectors, heat treat operators — by datacenter contractors offering 30% more, by energy companies offering signing bonuses, by semiconductor fabs dangling relocation packages. The Pentagon pays last and slowest. In a bidding war for finite humans, that's a losing position.

Every machinist working on missile housings is a machinist not making datacenter cooling parts. Every welder in a shipyard is a welder not building transformers. Every NDT inspector certifying jet engine components is one not inspecting semiconductor tooling. Defense is one of four claimants on the same finite pool — and it's the one that can least afford to lose.

The Factories That Can't Staff

244,000 manufacturing jobs announced via reshoring in 2024. 287,000 the year before. Over two million since 2010. The political consensus is bipartisan and loud: bring manufacturing home.

"Announced" is doing a lot of work in that sentence.

The CHIPS Act tells the story. TSMC's Arizona fab: delayed six months, labor shortages cited as a primary factor. Each fab needs ~12,000 workers at peak construction. TSMC flew in workers from Taiwan — a solution that generated political friction and doesn't scale. Intel's Ohio expansion needs 7,000+ at peak. Samsung in Texas, Micron in New York and Idaho — each requires thousands of the same skilled tradespeople every other sector is hunting.

Every reshored factory is another straw in the same milkshake. These aren't new workers being created. They're existing workers being redistributed — pulled from defense sub-tier suppliers, from energy projects, from the machine shops that make everything else. Reshoring doesn't grow the pie. It adds another mouth.

The Overlap Nobody Models

Every cell in that table represents real demand. Active contracts, open requisitions, projects breaking ground in the next 24 months. Not projections — commitments.

Read it horizontally. A welder certified for heavy structural work can build submarine hulls, or transformer enclosures, or datacenter frames. A five-axis CNC machinist can cut turbine blades for a power plant, or turbine blades for a jet engine, or cold plates for a GPU cluster. A foundry worker pouring investment castings does it the same way whether the part is a turbine vane or a missile fin. An NDT inspector running ultrasonic tests on nuclear pressure vessel welds uses the same skills certifying aerospace forgings or datacenter structural steel. These sectors aren't competing in some abstract economic sense. They're competing for the same specific humans, with the same specific certifications, in the same specific metro areas.

A welder in Groton who leaves Electric Boat's submarine yard for a higher-paying datacenter job in Virginia is a direct subtraction from the submarine program. That's not a labor market fluctuation. That's a national security event.

Nobody is modeling this at a granular level. The AI industry models chip supply. Defense models procurement budgets. Energy models generation capacity. Reshoring advocates model announced investments. None of them model the workforce as a shared, finite resource all four are drawing from simultaneously. They're all making plans for the same workers. None of them are coordinating.

Four generals. One army. No communication.

The Demographic Bomb

Here's where "just train more people" dies.

Take machinists as the canary — the trade with the best data, representative of the broader crisis. The median machinist in America is in his late forties. The peak age cohort — the single largest group — is 55 to 59: 47,710 workers, all within a decade of retirement. Thirty percent of the entire machinist workforce is over 55. The same cliff shows up everywhere. The American Welding Society projects a shortfall of 427,000 welders by 2028. Tool and die makers have declined 30% since 2000. Foundry employment has fallen for two decades straight. The average age in heat treatment shops, composites facilities, NDT departments, and metrology labs mirrors machine shops: graying, thinning, with no bench behind the starters.

For every machinist who retires, the pipeline produces a fraction of a replacement. The BLS projects the machinist workforce will decline 2% over the next decade — not despite these four surges, but because the retirement wave overwhelms new entrants regardless. 100,000 machinists will age out by 2035. That's nearly a third of the current workforce. Multiply that pattern across welding, toolmaking, foundry work, NDT, composites, and the other skilled trades, and hundreds of thousands of irreplaceable workers will disappear in a single decade.

The pay explains why the pipeline is empty. Median machinist wage: $56,150 a year. $27.74 an hour. Welders: similar. Tool and die makers: slightly better, still underwhelming. In a country where you can earn more driving for a rideshare app, where a plumber with no formal training can bill $150/hour, where a software engineer starts at twice that — the trades that built the American industrial base are compensated like afterthoughts. The market has been screaming the signal for twenty years. The response has been a shrug.

Training a machinist from zero takes two to four years. A precision five-axis CNC programmer — the kind needed for turbine blades and aerospace components — takes five to seven. A certified welder for nuclear-grade pressure vessels: three to five years. A composite layup technician for aerospace: two to three years of hands-on work before they're trusted with flight-critical parts. An NDT Level III inspector: a decade of progressive qualification. If America launched a Manhattan Project for skilled trades training tomorrow, the first graduates wouldn't be productive until 2029. By then, the retirement wave will have already removed more workers than any training program could produce.

In 2021, Deloitte projected 2.1 million manufacturing jobs would go unfilled by 2030 — at a cost of up to $1 trillion. That was calculated before the AI capex explosion. Before the post-Ukraine defense ramp. Before the IRA/IIJA spending surge. Before the CHIPS Act buildout.

2.1 million short under baseline conditions. Then add hundreds of billions in new demand from all four surges. The gap doesn't shrink. It becomes an abyss.

The Denominator Problem

China has 100-150 million manufacturing workers. The US has 12.6 million. That's an 8-to-12x gap.

Narrow to skilled trades — machinists, welders, toolmakers, foundry workers, the lot — and the gap widens. The US has 354,800 machinists; China has 7-10 million. Conservatively 23-to-1. Scale that across welding, casting, forging, and heat treatment, and the picture only gets worse.

China built over 1,000 commercial ships in 2024. The US built 8. Chinese shipyard capacity: 23.25 million tons. American: under 100,000 tons. 232-to-1. The US Navy's own data.

This is the denominator problem. The US is trying to simultaneously rearm its military, rebuild its grid, construct AI infrastructure, and reshore its industrial base — with a manufacturing workforce that is a rounding error compared to its primary strategic competitor.

China doesn't face a zero-sum constraint because it has the workers to do all of these things at once. It builds ships and fabs and grid infrastructure and AI datacenters in parallel — because it has 100 million manufacturing workers, not 12.6 million. The US is playing a zero-sum game. China isn't playing the same game.

What Breaks First

It's already breaking.

Transformer lead times: 2-4 years, up from one. Gas turbine lead times: up to seven years, up from two and a half. Submarine delivery: 60% of target. Artillery shells: 40% of target. F-35 parts: 238-day delays. TSMC Arizona: six months late. These aren't unrelated data points. They're the same phenomenon showing up in every sector, at every level: insufficient humans to turn money into metal.

The sectors with the deepest pockets win the bidding war. Right now, that's AI. Hyperscalers outbid the Pentagon, outbid utilities, outbid chip fabs. So the sectors that lose degrade first — grid modernization falls behind, defense sub-tier suppliers erode as machine shops take more lucrative commercial work, reshoring projects get announced with fanfare and delivered years late.

The irony is suffocating. All four surges are individually correct priorities. The US needs AI infrastructure. It needs a modern grid. It needs defense capacity. It needs reshored supply chains. Every one is rational and important. None is optional. And they all depend on the same finite resource — a resource that is aging, retiring, and not being replaced.

Money is not the bottleneck. All four sectors have massive funding — trillions in aggregate. The bottleneck is people. You cannot download a welder. You cannot 10x a machinist with a better algorithm — not yet. You cannot 3D-print an NDT inspector or fast-track a tool and die maker. You cannot venture-capital your way past the fact that training any of these skilled trades takes years, and the US is starting from a deficit of 2.1 million before any of this started.

The zero-sum fight is already underway. The question is which priorities America is willing to let fail — because with the current workforce, it cannot deliver all of them. Not close.

Unless.

Unless you radically multiply the productivity of the workers who exist. Make every machinist, every welder, every toolmaker, every foundry operator, every NDT inspector, every composite technician dramatically more effective. Not replace them — you can't, there's no one to replace them with. Multiply them. Because the math leaves no other option.

427,000 welders short. 354,800 machinists, a third near retirement. 64,000 tool and die makers and falling. Four industrial surges. One workforce. Zero slack.

Part IV: The Race →