Every hot thesis in venture capital — AI infrastructure, defense, reshoring, energy — traces down to the same bottleneck. And almost nobody is funding it.
In 2026, the Big Five hyperscalers will spend [$602 billion](https://techblog.comsoc.org/2025/12/22/hyperscaler-capex-600-bn-in-2026/) on capital expenditure. Seventy-five percent of that — roughly $450 billion — goes directly to AI infrastructure. Hadrian just raised a [$260 million Series C](https://www.cnbc.com/2025/07/15/hadrian-raises-260-million-from-founders-fund-lux-capital.html) from Founders Fund and Lux Capital to build AI-powered machining factories for defense. EU defense spending hit [€381 billion in 2025](https://www.consilium.europa.eu/en/policies/defence-numbers/), rising for the tenth consecutive year. Poland alone is spending over 4% of GDP on defense. The Reshoring Initiative projects [nearly 240,000 manufacturing jobs](https://www.fcnews.net/2026/01/reshoring-update-projections-call-for-higher-u-s-jobs-onshored/) onshored to the US in 2025.
VCs see all of this. They're writing checks into AI labs, defense tech companies, climate startups, and industrial software. The macro thesis is correct: we are entering the largest physical buildout since the postwar era.
But here's the question nobody at the partner meeting is asking: who machines the parts?
Where does the cold plate come from? Who turns the 155mm shell casing? Which shop cuts the turbine blade? What lathe makes the valve body? Pick any thesis in venture capital right now — AI infrastructure, defense, reshoring, energy — and trace it down to its physical root. You hit the same bottleneck every single time.
Precision machining. CNC lathes and mills. The people who operate them.
This isn't a supply chain "challenge." It's a structural constraint on every growth thesis in the market.
Venture capital loves abstraction. Software eats the world. Platform plays. Network effects. But every one of the four dominant VC theses in 2026 terminates in a physical object that someone has to cut from metal.
Let's trace them.
The AI buildout is, at bottom, a thermal management problem. NVIDIA's GB200 NVL72 racks consume over 100kW per rack. At that power density, air cooling is physically impossible — you need direct liquid cooling, which means cold plates mounted directly onto every GPU.
A cold plate is a precision-machined micro-channel heat exchanger. Copper or aluminum. CNC-milled with internal channel structures that must maintain sub-millimeter tolerances to distribute coolant evenly across the chip surface. The GPU itself requires coolant temperatures [below 30°C](https://www.kenfatech.com/gb200-liquid-cooling-plate-design/) at the junction. Every degree matters. Every channel geometry matters. These are not stamped parts — they are machined.
The datacenter liquid cooling market is projected to grow from roughly [$2.8–4.8 billion in 2025 to $14.8–27.1 billion by 2030–2035](https://www.marketsandmarkets.com/Market-Reports/data-center-liquid-cooling-market-43706498.html), depending on which analyst you ask. The CAGR ranges from 18% to 33%. Nearly all of that market is dependent on precision-machined components: cold plates, manifolds, fittings, distribution blocks.
But cold plates are just the start. Datacenters also need machined power distribution hardware — busbars milled from copper, switchgear components, connector housings. Rack mounting systems and structural aluminum. And increasingly, they need dedicated power generation: natural gas turbines or nuclear SMRs, each of which brings its own machining demands (more on that below).
Rough estimate: even if machined components represent just 2–5% of total datacenter build cost, at $602 billion in hyperscaler capex, that's $12–30 billion in annual machining demand from datacenters alone. And it's growing at the rate of AI spending, which is growing at 36% year-over-year.
The defense case is starker.
Before Russia invaded Ukraine in early 2022, the US was producing [14,000 155mm artillery rounds per month](https://breakingdefense.com/2025/10/army-ammunition-production/). The Pentagon set a target of 100,000 rounds per month. By late 2024 — nearly three years later — actual production was [40,000 rounds per month](https://www.nationaldefensemagazine.org/articles/2024/155mm-shell-production). The 100,000 target has been pushed to mid-2026. We couldn't even triple production of a relatively simple munition in three years.
A 155mm shell casing is turned on a CNC lathe. The forging is straightforward. The constraint isn't the design — it's the capacity. Lathes. Operators. Shifts.
The F-35 program tells the same story from a different angle. As of early 2025, Lockheed Martin's assembly line had [over 4,000 parts shortages — double historic levels](https://www.gao.gov/products/gao-24-106909). Fifty-two aircraft were stalled in final assembly waiting for parts. The GAO flagged 238-day delivery delays. These aren't exotic components. Many are precision-machined fittings, brackets, housings, and structural elements sourced from a defense industrial base that has been [slowly consolidating and moving offshore for decades](https://www.heritage.org/press/new-heritage-report-warns-deteriorating-us-defense-industrial-base-amid-most-hostile-global-environment-since-world-war-ii).
The National Defense University Press put it plainly in December 2025: ["The United States lacks the munitions production capacity to meet the demands of the contemporary strategic environment."](https://ndupress.ndu.edu/Media/News/News-Article-View/Article/4366531/)
And this is before Europe's rearmament wave really hits production. EU defense spending is approaching €400 billion annually. Poland, the Baltics, and the Nordics are spending at historically unprecedented rates. That money eventually converts into orders for munitions, vehicles, aircraft, and missiles — all of which require machined components, many of which draw from the same limited supply base.
The reshoring narrative is real. The CHIPS Act, IRA, and bipartisan industrial policy have triggered a genuine wave of manufacturing investment. The Reshoring Initiative tracked 287,000 announced manufacturing jobs in 2023 and projects 240,000 in 2025. Semiconductor fabs, battery plants, and EV factories are being built across the Sun Belt and Midwest.
But a factory is not a factory without machine tools — and the people who program and operate them. Every new fab needs support machining for fixtures, tooling, maintenance parts, and custom components. Every battery plant needs machined electrode dies, cooling systems, and structural elements. Every EV plant needs machined drivetrain and suspension components.
Kearney's 2025 Reshoring Index found that [most latent US manufacturing capacity has been absorbed](https://www.advancedmanufacturing.org/leadership-innovation/looking-ahead-manufacturing-2026/article_8ead3c3f-3020-4567-83dc-63838ddde59e.html) and is struggling to keep pace with domestic demand growth. The capacity isn't there. US metalworking machinery orders surged [40% year-over-year in October 2025](https://www.qualitymag.com/articles/99273-us-metalworking-machinery-orders-surge-40-year-over-year-in-october) — shops are trying to add capacity. But a new CNC machine without someone to program and operate it is a very expensive paperweight.
The AI buildout is also an energy buildout. Hyperscalers are planning gigawatts of new power generation — natural gas turbines now, small modular nuclear reactors in a few years. Gas turbine orders have exploded, and the three major manufacturers — GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries — are [overwhelmed, with delivery times pushed out to 2029 and beyond](https://www.ctvc.co/gas-turbine-gridlock-236/).
According to EPRI, the primary bottlenecks are [rotor forgings and hot-section blades](https://www.powermag.com/gas-turbine-supply-chain-bottlenecks-could-reshape-the-generation-mix-in-2030-and-beyond/) — constrained by limited suppliers and highly technical manufacturing processes. A gas turbine blade is one of the most demanding machining jobs in existence: single-crystal nickel superalloy, five-axis simultaneous milling, internal cooling channels, coatings. There are a handful of factories on Earth that can do it.
Nuclear SMR components are even more demanding — tighter tolerances, exotic materials, nuclear-grade quality requirements. Every megawatt of new capacity requires machined valve bodies, pump housings, flanges, pressure vessels, and pipe fittings.
The energy thesis traces to atoms the same way everything else does.
Here's what makes this different from a normal supply chain problem: all four theses hit the same bottleneck simultaneously.
It's not that AI needs machining. Or that defense needs machining. Or that reshoring needs machining. Or that energy needs machining.
It's that they all need machining, at the same time, from the same finite pool of machines and machinists. And that pool is shrinking.
When a hyperscaler needs cold plates for a new datacenter, they're competing for CNC time with a defense prime that needs shell casings, a turbine manufacturer that needs blade machining, and a reshored semiconductor fab that needs custom fixturing. They're all drawing from the same capacity.
This is the thing that doesn't show up in any individual sector analysis. It's only visible when you overlay the demand curves from all four sectors onto the same supply base. And the picture is ugly.
The numbers are brutal and they are not getting better.
The US employs [354,800 machinists](https://www.bls.gov/ooh/production/machinists-and-tool-and-die-makers.htm) as of 2024. The three largest age cohorts are 55–59 (47,710 people), 60–64 (43,797), and 50–54 (43,138). [Roughly 39% of the workforce is over 50](https://datausa.io/profile/soc/machinists). The Bureau of Labor Statistics projects a -2% decline in machinist employment over the next decade — not because demand is falling, but because the BLS is projecting that we simply won't replace the people who leave. All 34,200 annual openings are replacement demand. Zero growth.
Zoom out to manufacturing broadly: Deloitte's January 2026 analysis found that by 2033, the industry [may need 3.8 million new workers, with nearly 1.9 million of those roles at risk of going unfilled](https://www.deloitte.com/us/en/insights/topics/economy/spotlight/us-manufacturing-labor-impact.html) if current workforce challenges persist. In August 2025, there were approximately 409,000 unfilled manufacturing positions. The earlier Deloitte/Manufacturing Institute study projected [2.1 million manufacturing jobs unfilled by 2030, at a cost of up to $1 trillion](https://nam.org/2-1-million-manufacturing-jobs-could-go-unfilled-by-2030-13743/) to the US economy.
The median machinist wage is [$56,150 per year](https://www.bls.gov/ooh/production/machinists-and-tool-and-die-makers.htm). Tool and die makers earn $63,180. These are not poverty wages, but they're not competing with software engineering or even HVAC in many markets. The economics create a death spiral: margins in machining are thin → shops can't aggressively bid up wages → fewer young people enter → the shortage deepens → but customers still won't pay more → repeat.
Meanwhile, the people who actually know how to do this work are retiring. Each one takes decades of institutional knowledge with them — not just how to program a machine, but how to read a drawing, understand design intent, select the right tooling, build a robust process, hold tolerances. You cannot YouTube your way to that knowledge. It compounds over years of making real parts on real machines. When a 60-year-old machinist retires, you don't lose one headcount. You lose twenty years of compounded expertise.
The apprenticeship pipeline is broken in ways that compound the problem. Traditional programs still teach manual G-code with pencils, spend years on manual machines before touching CNC, and fill time with skills that graduates will never use again. Four-year programs produce people who are barely ready for production work. We're training machinists for 1985 in 2026.
This is the bottleneck behind the bottleneck. Capital is available. Demand is exploding. The machines can be bought. But the hands are not there, and they take years to develop.
If you're a VC and you've read this far, you're probably wondering: okay, so what do I fund?
Four categories. All of them unsexy. All of them critical.
The most leveraged intervention in the entire manufacturing stack is making existing machinists dramatically more productive. There aren't going to be enough machinists — that's a demographic fact. So the question becomes: what can each machinist accomplish?
Average spindle utilization in US machine shops is below 50% during operating hours. Most shops can't run unattended shifts because they don't have enough people to program the work. Lead times are 6–12 weeks. Customers want to order more and can't. There is a goldmine of latent capacity sitting in machines that aren't running because there's nobody to program them.
CAM software — the programs that generate toolpaths telling CNC machines how to cut — is the critical chokepoint. Programming a part can take hours to days. It requires deep expertise. And the tools are, to put it charitably, not great. AI-assisted CAM is starting to change this. Companies like [CloudNC](https://cloudnc.com) are using AI to accelerate CNC programming from hours to minutes, allowing each machinist to handle dramatically more work. When you take a machinist from 1–2 CAM programs per day to 10–20, the math changes for the entire industry.
This is the most capital-efficient play in manufacturing. You're not building factories ($260M+). You're not buying machines ($500K–$2M each). You're multiplying the output of the most constrained resource — skilled humans — with software.
The current apprenticeship model needs to be rebuilt from the ground up. New machinists should start on CNC machines and CAM software from day one, using simulators, AI tools, and modern workflows. The goal isn't to preserve tradition — it's to produce people who can build production processes on automated equipment as fast as possible.
There's a venture-scale opportunity in modern machining education: platforms that compress the 4-year apprenticeship into 12–18 months of intensive, production-oriented training using simulation, AI tutoring, and real machine time. The economics work because the demand side is insatiable — every factory owner in America will tell you their #1 problem is hiring.
The US is no longer a leading machine tool manufacturer. The top producers are Japan (Mazak, Okuma, DMG Mori), Germany (DMG Mori, Trumpf), and increasingly China. American manufacturing depends on machine tools built elsewhere.
There's an opportunity in next-generation machine tools designed for the workforce that actually exists: machines with better interfaces, more automation, less setup time, and deeper integration with AI-assisted programming. The Datron Neo won awards because it had a camera on the probe, a joystick interface, and a UI designed for humans rather than for 1990s electrical engineers. That should be the baseline, not the exception.
Every machine shop in America runs on a combination of spreadsheets, whiteboards, phone calls, and ERP systems designed in the Clinton administration. The scheduling, quoting, and supply chain management layer for precision manufacturing is a wasteland of outdated software and manual processes.
This is the classic "boring SaaS" opportunity, but with unusually strong tailwinds. Every factory is capacity-constrained. Every factory is desperate for more throughput. Software that reduces scheduling friction, automates quoting, or optimizes machine utilization doesn't need to sell a vision — it just needs to show shops how to ship more parts with the people they already have.
There's a meme in VC about "picks and shovels" investing — don't mine for gold, sell the tools to the miners. NVIDIA is the canonical example: they sell compute to AI companies.
But NVIDIA's GPUs need to be cooled. The cold plates are machined. The machine tools are built in factories. The factories are run by machinists. The machinists are trained in programs. The programs use software.
We are several layers deeper than the picks and shovels. We are at the picks and shovels of the picks and shovels — the atomic layer of physical infrastructure that everything else depends on.
Three manufacturing AI companies raised over $510 million combined in the last twelve months: Hadrian ($260M), Bright Machines ($126M from BlackRock, NVIDIA, and Microsoft), and Machina Labs ($124M). The investors include Founders Fund, a16z, Lux Capital, BlackRock, NVIDIA, and Microsoft. Smart money is starting to see this.
But $510 million across three companies is a rounding error compared to the scale of the problem. AI startups raise that in a week. The manufacturing AI category needs 10x more capital, 10x more companies, and 10x more attention from the people who allocate it.
Every VC portfolio has exposure to atoms whether the partners know it or not.
Your AI infrastructure company needs cold plates. Your defense tech company needs shell casings. Your climate tech company needs turbine blades. Your reshoring play needs machine tools and the people to run them. Somewhere in every supply chain, there's a CNC machine, and someone has to program it.
The VCs who understand this will make investments that look boring and turn out to be foundational. They'll fund manufacturing software, workforce development, and machine tool innovation — the unsexy infrastructure that every sexy thesis depends on.
The VCs who don't understand this will watch their portfolio companies miss delivery timelines, blow through budgets on physical prototyping, and discover that you can't ship hardware at software speed when you're competing for CNC time with the Department of Defense and every hyperscaler on Earth.
The macro is right. AI is real. Defense is real. Reshoring is real. Energy transition is real. The demand side of every major thesis is firing on all cylinders.
But demand without supply is just prices going up.
And the supply side terminates in 354,800 machinists, median age approaching 50, declining at 2% per decade, with 1.9 million manufacturing positions at risk of going permanently unfilled.
That's the trade. That's where the alpha is. That's what the fuck VCs should do now.
[Theo Saville](https://countingatoms.com) is the co-founder and CEO of CloudNC, where he's spent 10 years building AI for CNC manufacturing. He is not a neutral observer. He is an extremely interested party who happens to have data.
- [CreditSights/IEEE ComSoc — Hyperscaler capex $602B projection (Dec 2025)](https://techblog.comsoc.org/2025/12/22/hyperscaler-capex-600-bn-in-2026/)
- [Bureau of Labor Statistics — Machinists and Tool and Die Makers (2024)](https://www.bls.gov/ooh/production/machinists-and-tool-and-die-makers.htm)
- [Data USA / ACS — Machinist age demographics (2023)](https://datausa.io/profile/soc/machinists)
- [Deloitte — Manufacturing workforce outlook (Jan 2026)](https://www.deloitte.com/us/en/insights/topics/economy/spotlight/us-manufacturing-labor-impact.html)
- [Deloitte & Manufacturing Institute — 2.1M skills gap projection (2021)](https://nam.org/2-1-million-manufacturing-jobs-could-go-unfilled-by-2030-13743/)
- [GAO Report GAO-24-106909 — F-35 parts shortages](https://www.gao.gov/products/gao-24-106909)
- [Breaking Defense / National Defense Magazine — 155mm production](https://breakingdefense.com/2025/10/army-ammunition-production/)
- [NDU Press — US munitions production capacity (Dec 2025)](https://ndupress.ndu.edu/Media/News/News-Article-View/Article/4366531/)
- [Heritage Foundation — Defense Industrial Base assessment](https://www.heritage.org/press/new-heritage-report-warns-deteriorating-us-defense-industrial-base-amid-most-hostile-global-environment-since-world-war-ii)
- [EU Council — EU defence expenditure 2024-2025](https://www.consilium.europa.eu/en/policies/defence-numbers/)
- [Reshoring Initiative — 240,000 jobs projected 2025](https://www.fcnews.net/2026/01/reshoring-update-projections-call-for-higher-u-s-jobs-onshored/)
- [Kearney Reshoring Index 2025 — Latent capacity absorbed](https://www.advancedmanufacturing.org/leadership-innovation/looking-ahead-manufacturing-2026/article_8ead3c3f-3020-4567-83dc-63838ddde59e.html)
- [Quality Magazine / AMT — Machine tool orders surge 40% YoY (Oct 2025)](https://www.qualitymag.com/articles/99273-us-metalworking-machinery-orders-surge-40-year-over-year-in-october)
- [POWER Magazine / EPRI — Gas turbine supply chain bottlenecks (Jan 2026)](https://www.powermag.com/gas-turbine-supply-chain-bottlenecks-could-reshape-the-generation-mix-in-2030-and-beyond/)
- [CTVC — Gas turbine delivery times pushed to 2029+](https://www.ctvc.co/gas-turbine-gridlock-236/)
- [KenFa Tech — NVIDIA GB200 liquid cooling plate design](https://www.kenfatech.com/gb200-liquid-cooling-plate-design/)
- [CNBC — Hadrian $260M Series C (Jul 2025)](https://www.cnbc.com/2025/07/15/hadrian-raises-260-million-from-founders-fund-lux-capital.html)