Hadrian was founded in 2022. It has since opened six factories, the newest 2.2 million square feet, and has about 725 employees, with plans to hire roughly 600 more in four months. Revenue is in the hundreds of millions and heading toward billions. Most of its customers are in defense and space.
Asked what it actually takes to scale manufacturing in the United States, Power's core answer is standardization: build factories out of modular, interchangeable stations that can be reconfigured as demand changes. Standard stations give you something consistent to automate against, and they let one factory serve many customers and products at once. That idea comes back in almost every part of the conversation, from factory layout to how Hadrian hedges its capital bets.
The rest of this piece works through what that means: the market Hadrian sells into, how it designs factories, how it thinks about capital and risk, how it scales people, and how Power operates as CEO.
Part 1The market: what you're really selling is trust
There's no capacity to rent
Imagine a startup designing a new low-cost interceptor missile. It would make no sense for them to build their own factory at worse unit economics than a specialist manufacturer could offer. But in the US there is almost no contract manufacturing capacity to hire.
Compare Shenzhen. A new humanoid robot company there never builds its own circuit board factory, because there's abundant capacity at scale that it can trust. In the US that capacity doesn't exist, so it's a chicken-and-egg problem: customers won't commit to a manufacturer that hasn't built capacity, and nobody builds capacity without customers.
Power's analogy is AWS. Amazon built huge cloud capacity for its own retail and engineering needs first, and only then sold it to others. Stripe could trust AWS because it was already running at scale. Renting GPUs from a data center that's half built is far easier to believe in than renting from one on a slide.
His rule is that you have to build about 1.2x your future sales capacity before you have the contracts. Otherwise you fall behind on optics and, more importantly, on execution. What a manufacturer really sells, he says, is trust that it will deliver. The problem isn't that people are exaggerating. The capacity simply doesn't exist, so someone has to build it ahead of demand, which takes a lot of nerve and capital.
What Power got wrong: customers don't pay for speed
Some quick background. American machining roughly splits into "print" and "no-print" work. The print is the engineering drawing. In print work, the customer sends a detailed drawing with its own tolerances and specs, and you have to hit them and prove it, which is typical for aerospace and defense. In no-print work, like the online sheet-metal service SendCutSend, the customer uploads a file, picks options from a menu, and gets the shop's standard tolerances. Hadrian has always been in the harder print business.
Power's original bet was that customers in this business would pay a premium for speed. Hadrian's first customer was a large rocket company, chosen on the logic that if they could satisfy the toughest customer, they could sell to the defense primes, and on the assumption that many other growth-stage companies would also pay for fast turnaround on complex parts.
Two things turned out wrong. At the time, there were no other startup customers of any scale like that rocket company. And almost nobody else valued speed. In Power's words, that rocket company is the only space or defense prime that can actually convert speed into value.
You can't be slow. But delivering in 8 weeks, and cheaper, when everyone else takes 12 is manageable. Delivering in 2 weeks costs orders of magnitude more in engineering and complexity. What most of the market pays for is flexibility, scale, and total program cost. Hadrian changed course accordingly.
Power also argues that large amounts of cheap capital follow execution, not fundraising skill. Better storytellers exist, but the companies that end up with the most capital have better businesses. And there's no such thing as "asymmetric manufacturing." Products can be asymmetric, like a cheap drone that destroys an expensive target, but no manufacturing method makes a Tomahawk casting for a dollar. What manufacturing offers is scale, flexibility, and agility.
Part 2The factory: build for flexibility and for failure
The factory as a GPU cluster
Power describes most production lines as working like a CPU, running things in sequence, and Hadrian's as working like a GPU. The clearer version of his analogy is a GPU cluster with a job scheduler: a pool of identical, interchangeable machines onto which any job can be dispatched. The goal is to run a high-mix, low-volume factory as efficiently as a low-mix, high-volume one, and to manage peak capacity the way you'd manage compute.
Design for failure, not the success case
Most factories are designed around things going right: 90% yield, smooth flow. Hadrian builds in slack capacity specifically for failures.
Power's image is a five-lane freeway full of Waymos. When one blows a tire, you want an off-ramp or a slip lane where it can wait, so it doesn't back up all five lanes at rush hour. Most factories, in their equipment, people, and processes, have no off-ramp. Building one requires stations standardized enough to swap like GPUs, and efficient enough that the slack doesn't eat your margins.
Assume every vendor will fail you
Part of Hadrian's software, called Opus, takes over the control software of its machines. Early on, the team asked a German machine vendor for API documentation, and the vendor assumed they were hackers, since nobody had asked before. On the floor, two brands of machine reported inconsistent data through their APIs. On one, about half the API calls came back "not implemented yet," in German. Integration took two weeks for one brand and twelve for the other.
The lesson: you're dealing with unknown unknowns. Plan as though everyone will let you down, build engineering systems around that, never trust a controller or robot until you've actually integrated it, and pad every timeline.
Give new parts their own fast lane
Before a new part goes into production, the shop has to prove it can make it to spec. This is called new product introduction (NPI): cut a first part, measure it on a CMM (a precision measuring machine), find what's off, adjust, and cut again. Something is almost always off, like a tool that bends slightly or a dimension that's 20 microns out, so it often takes 3 to 10 loops.
Each loop is maybe two hours of real work: an hour of machining, half an hour of measuring, half an hour deciding on a fix. But in a normal shop a loop takes two to five days, because the part spends most of that time waiting. It waits for the CMM, which is busy measuring production parts. It waits for the engineer to see the results. Then it waits for the machine, which was given a production job the moment it finished and now has to be set up again. Power's point is that NPI runs late because of this back-and-forth, not because machining is slow. The goal is to shrink the gap between measuring a part and fixing the process, ideally to one loop per hour.
Why software doesn't fix it. The obvious idea is to mark NPI parts as top priority so they skip the queue. That doesn't work, because a priority flag can't free up a busy machine. The CMM still finishes the production batch it's in the middle of. And fast iteration needs the cutting machine to sit idle, still set up, while its part is being measured, which is exactly what every scheduler, human or software, is designed to prevent.
Hadrian's fix is physical. A few machines and a CMM are set aside for new parts only, even if the area is just roped off. The part comes off the machine, gets measured a few meters away with the engineer standing there, and goes straight back onto a machine that's still set up. Loops drop from days to about an hour. It's the same reason engineers keep a dedicated machine for debugging instead of submitting every test run to a busy shared cluster.
Power draws a broader lesson: software tends to mirror the physical layout. A fast lane that physically exists is simple to model and easy to see on the floor. Priority rules layered on a shared pool get complicated, and people stop following them. So lay out the factory to match the workflow you want, and keep the software simple.
Overbuild what's expensive to change later
Two things in a factory are especially hard to change once they're built. Foundations are the reinforced concrete slabs the machines sit on; heavy precision machines need thick, vibration-isolated foundations or they can't hold tight tolerances. Electrical drops are the points where power is brought down from the ceiling or up through the floor so a machine can be connected. Together they decide where equipment can physically go.
Factory layout determines construction cost (foundations, electrical, clean rooms) and, less obviously, operating efficiency, including travel time and even where people end up talking to each other. Foundations and electrical drops are among the longest lead items in bringing up a factory. So when moving fast, Hadrian:
- Overbuilds foundations, spending something like $25M instead of $15M on vibration-isolated concrete everywhere. That buys the layout team four more months of simulation, and equipment can then go anywhere.
- Puts electrical drops everywhere, knowing about 30% will never be used. A contractor can install them all in two weeks, and a manufacturing engineer can connect any machine to any drop later.
The principle: pay extra on the expensive things that lock you in, so they don't have to wait on decisions that take a long time.
Part 3The money: every bet can kill you
Capex is binary, and not betting is worse
Suppose you need $1B of equipment to win a large contract, say circuit boards for a drone maker. Traditionally you'd want the contract before spending the billion. But no one awards a production contract to a company that hasn't started executing.
- Don't spend: 0% chance of the revenue. You die, or end up a roughly $100M acquisition.
- Spend: if you don't close the contract, you die. But your chance of winning the revenue is now above 50%.
Until a company reaches roughly $50B in scale, every major capex bet can kill it. Not making the bet guarantees it.
How to make those bets: the two variables are burn rate and execution risk. Execution risk is the easier one, because experience teaches you that everything slips. Buy a 3D printer, a CNC machine, or a robot, or build out a factory, and so much can go wrong that you plan for it to arrive late. The harder and more valuable skill is making capex flexible.
Turning one bet into three
If you're deciding between building a drone factory and a missile factory, it looks like two separate all-or-nothing bets, each needing its own equipment and people. With a flexible workforce and software-controlled equipment, Power says about 80% of capex can move between programs, the way a data center doesn't need to know whether consumer chatbots or enterprise AI will use more compute, because both run on the same GPUs.
One-way doors and buying time
Jeff Bezos distinguishes two-way doors, decisions you can reverse, from one-way doors, which you can't. Isaiah Taylor of Valar Atomics says moving fast means sprinting through some one-way doors. Power goes further: in manufacturing, capex and long engineering projects are all one-way doors. He uses two modes for dealing with that.
Mode one: spend more to buy time. A customer needs a working automated welding station prototype in month 10. Equipment takes 9 months to arrive. Specifying the station with complete accuracy would take the best team three months, putting you at month 13, which fails. The station costs about $1M, of which roughly $800K is single-purpose equipment that can't be changed, so spend three of four weeks getting that exactly right. Rough-cut the rest, like the robot arms. Guess wrong and it costs about $50K, since the arm gets reused elsewhere. Miss the deadline and the customer cancels and the company dies.
So it's worth spending around 30% more to build in hedges. Being three weeks late because of a botched software integration is survivable. Having nine months of equipment arrive and discovering on day one that it's wrong puts you 18 months out. Software and sales mistakes can be seen and corrected within a quarter. Capex mistakes can't. When you aren't sure, hedge, and treat the hedge as a financial position spread across programs.
Mode two: sometimes you just have to be right. Power calls it rolling a hard six, or a natural 20 in Dungeons & Dragons. In Battlestar Galactica, raiding the heavily defended fuel depot is enormously risky, but the alternative is running out of fuel. With long correction times, the only option is often to be very good and execute. The skill that matters most is telling apart what you can fix later from what you can't.
Why the risk of death stays around 80%
Two and a half years ago, Power put Hadrian's survival odds at about 20%. Today, with far more revenue, he says the risk of death is still around 80%. It isn't because the company is careless. It's structural: risk only goes away once growth slows to about 20% a year.
Hadrian has to keep growing fast, both for venture returns and because the mission requires capacity built ahead of demand. That means growing its "risk balance sheet" (capex plus R&D) at about twice the sane rate. Power's example: $2B of capex now makes $5B of revenue possible in 2028. Missing $5B means no venture returns, and the capex bill means the company is finished. So it has to do both. Forecasting still matters enormously ("we're not YOLOing a billion dollars"), but the risk only really goes away once growth slows to around 20% and cash flows catch up.
A company-killing week
A few years ago, Hadrian was well funded for a Series A company but not flush. As usual, equipment was arriving about two months ahead of customer contracts. One important delivery was four months out, and a fundraise was coming. Missing a delivery quarter would have undercut the whole claim that they could execute.
Two senior engineers found rust on new machines from a reputable vendor and believed it was terminal. The vendor said it wasn't a problem. Within 48 hours the team traced it: the vendor had quietly switched a subcomponent supplier, like still buying Camrys but with a different tire brand. Serial numbers 600 to 800 across 40 machines pointed to a batch problem that was spreading. The vendor's warranty plan was to rebuild the machines one at a time and finish by Thanksgiving, meaning about eight weeks of downtime. That would have been like a SaaS company missing a sales quarter, with burn effectively quadrupling.
Power checked the diagnosis himself over 48 hours, then spent a week pressing the vendor's president, explaining that the company would die if it wasn't fixed in four weeks. The vendor flew in 30 engineers from abroad and finished in four weeks.
His summary: something that could kill the company happens about once a month. It's manufacturing; that's the game.
Financing a capital-heavy business
Power says only two kinds of capital-heavy business work. In the first, customers prepay for your buildout, as NASA did for early SpaceX; the Pentagon, retail energy customers, and data center tenants generally don't. Otherwise, you must design the business so that it eventually qualifies for long-term infrastructure credit, and predict that path accurately over about four years.
Engineering, in his framing, is a subsidy of gross margin: automation is slower and more expensive for a year or two before it pays off, and investors accept that for software teams. Capex duration and cost work the same way, and a good CEO shouldn't miss that curve by more than about six months. Hadrian worked out the metrics that would make its factories financeable: contract length, stability, equipment life, uptime. It progressed from venture debt to equipment leasing and beyond, and its financing costs were underwater until its late Series B. Investors accepted that because Hadrian could lay out the path to cheaper capital precisely enough that credit funds would confirm it would be financeable later.
Every round, Hadrian also has to teach investors that its cost accounting is right. It isn't inventing rules; it takes a first-principles view of where costs really sit in manufacturing, which looks unfamiliar to investors used to software and product companies. The bar is to be so precise that investors come to trust your thinking in that area, and so that a public-market analyst would agree your business-unit scorecards are the right way to look at the company.
Part 4The organization: scaling people and capabilities
Capability bricks
Hadrian thinks of its factories as built from a library of reusable capabilities, which Power calls Lego bricks.
When demand outruns the forecast
Hadrian's forecast went up every month, which is why it raised a Series C led by Founders Fund, a follow-on round, and a Series D in fairly quick succession. Its pipeline grows by about $2B a quarter with a federal sales team of only eight people. (Keller Rinaudo Cliffton of Zipline, whom Power rates among the best deep tech CEOs, had a similar experience when Uber pulled his timeline in by two and a half years.) This creates two problems.
Talent. Nobody knows whether a particular contract will close and require 50 more robotics engineers. But across 10 contracts, there's about a 95% chance Hadrian will need at least 100. So it aggregates forecasts and bets on the trend. In a fast-growing company you can always pause hiring for a quarter, but you can't restart it quickly. Recruiting has its own lag: a great recruiter takes about 90 days to hire and then makes about four great hires. By the time you realize recruiting needs to double, you have about six months of slack.
Capex. With multiple programs, equipment becomes shared bricks bought ahead of need. Hadrian standardizes on Fanuc robot arms in a few sizes and buys them in the hundreds, and program teams pull from inventory. That's inefficient for any single program and very efficient for the company, cutting both risk and timelines. About 20% of equipment, the program-specific pieces, can't be pooled.
This is part of why Power says risk decisions get easier as you get bigger, even though running the company gets harder. Materials get cheaper, debt gets easier, and you get to the front of vendor queues because you're everyone's biggest customer. The market adjusts to whoever is growing fastest. Things that used to be enormously painful, like bringing up a new line of machines, now follow a playbook.
Designing an org that can nearly double in four months
- Small modular units. Capabilities and programs run as teams of at most 30 to 40, each with strong engineering leadership. The weld engineering team and its software team grow from 30 to 70 under two leaders, the chief weld engineer and the chief software engineer. There are about 80 such units. That avoids the problem of 100 backend engineers suddenly needing eight managers.
- Clean interfaces between teams. Boundaries are defined well enough that the weld team doesn't need to know who runs scheduling.
- High bars from the first hires. The first five engineers on a team set the standard and are told not to hire anyone below it.
- Centralized hiring tests. New managers don't invent their own interview process. The company invests heavily in, say, the weld engineering test, so anyone who passes is good unless they cheated.
- Very few sacred values. Most companies start with about 10 cultural values. Power says cut to two, clear enough to say at onboarding that people will be fired for breaking them. Keep pace, methodology, and core product principles consistent, and let everything else vary. Many values are "luxury beliefs," luxuries of being small. Deciding which to cut is judgment, such as noticing that 20 new hires struggled with a rule that wasn't written clearly.
Choosing what to pursue
Hadrian turns down a lot of commercial opportunity. Its filters:
- Mission first: re-industrialization and defense.
- Doors that won't open again. Some programs with the Department of War or the primes are the first time in 50 years a customer has let a contractor into a division, with around $20B behind them. Those come first.
- What can wait. Low-cost interceptor manufacturing for many different designers would be efficient for everyone, but there will be plenty of those programs later, so they're below the line for now.
- Brick coverage. 19 of 20 existing is easy to say yes to; 1 of 20 is an existential decision.
What limits how fast to grow, when demand is effectively unlimited, is a careful, constantly updated sense of how overextended the whole company is.
Part 5The operator: judgment under pressure
Your priors don't scale
You know your own abilities, and those of your ten best people, very well. The problem is that the team running the tenth program is people you've never met. Power's analogy: if you can personally make seven good podcasts a week, the eighth gets shaky, and the interviewer for the tenth isn't you. The quality drop won't show up in the numbers for six months. So put your own priors aside, let the organization show you where reality is, and judge by the worst case.
He's been badly overextended twice: the first 18 months, when Hadrian raised $1.3M and set out to build what he describes as $200M worth of software, and the last nine months. He says he's better now at catching himself, noticing when he's overloaded a team that was doing well.
Ulysses contracts
Judgment degrades when you're tired, and even more when you're on a high. It's like drinking: by the sixth beer you can't judge whether to leave your keys at home, so you decide before the third that they stay home. Power writes down rules while clear-headed that he won't break when sleep-deprived, manic, winning, or losing.
A hedge fund investor gave him a phrase for why this matters: real estate cycles keep happening because each new wave of 25-year-old investors has never felt the gut punch of losing everything. A risk model saying you're overleveraged doesn't help if you've never felt the loss. Nurses, he notes, are good at following rules like this.
Learn by doing, like lifting weights
Power had never recruited before starting Hadrian. In the first 12 weeks he read a few blog posts and then did about 20 screening calls a day, plus outbound outreach and interviews. By week eight he understood it. Product development, engineering, manufacturing, sales, recruiting, and finance are all learnable. The trap is tying self-worth to not being great at something after two weeks. If you've never benched 200 pounds, start with 20 and add weight.
Be the second-best person at every function
Your CFO should be the best at finance and your head of talent the best at recruiting. But the best people want to work for the best CEOs, so you need to understand each function well enough to partner with its leader and make real trade-offs on the spot, for example with a staff engineer, instead of asking for a spreadsheet. Power isn't the best engineer at Hadrian, but he considers himself one of the best engineering CEOs people have worked for. You should be able to write some code so you can tell good from bad, the way you can tell a world-class jazz musician from a poor imitation without playing yourself.
Earn trust instead of performing it
Power reverses the usual idea of building relationships in order to win contracts. First create value, and be trusted enough that customers bring you their hardest problems. If you then design a solution that can credibly work, the contract follows.
In practice, Hadrian spent about two and a half years, during the Biden administration, telling the government it wasn't ready for these problems yet and offering advice on manufacturing policy instead, to the point of frustration. They probably could have won contracts sooner. Because it was honest rather than a sales tactic, people believed them when they finally said they were ready. Hadrian also declines contracts it doesn't think will create value for the customer.
A slow way to build a billion-dollar company, and a very fast way to build a hundred-billion-dollar one.
He describes "false doors." At a party, some people maneuver their way onto the balcony with the host and are sure they're in. But governments, large customers, great hires, and long-term investors all develop defenses against people who are angling for access, and the people angling usually don't realize they've taken the wrong door. If you act as an honest expert, and bid only where you truly believe you're the best, the market comes around to your view. It compounds: every Hadrian round has gotten easier, not from hype, but because investors who passed came back after watching 80% of what sounded implausible come true.
Maximize velocity instead of forecasting
Power calls his forecasting "all vibes." With effectively unlimited demand, the job is to find where the organization isn't at full speed and fix it, then let the market and the company show you how fast you can grow. You can always slow down, find more customers, or redirect engineers. You can't quickly speed up an organization. He points to Sam Altman raising money ahead of what the organization needed, and to Ramp: rather than committing to arbitrary release dates, keep product and engineering velocity as high as possible, have good taste in what to build next, and announce things when they're ready.
What deserves your attention shifts constantly. Onboarding used to be a "fix it eventually" item. With thousands of people about to join, an extra hour spent on onboarding now shapes the productivity and focus of every one of them. Sometimes the highest-leverage thing is a company-wide performance-management document, and sometimes it's managing one engineer four levels down. The question is what can safely run imperfectly for 8 to 12 weeks, and what has to be fixed tomorrow because the damage compounds.
Manage energy, not hours
Some people can run on a treadmill for an hour and some prefer deadlifts. Power can do complex product management every day, but has only about three days of heavy context-switching in him before he's worn out. He arranges his work around which tasks give him energy and which drain it, and thinks about his people the same way.
Do the hard things immediately
From Sam Altman he took this: the instinct with a hard conversation, like firing someone, is to schedule it for Thursday and lose the whole week dreading it. Instead, send a short note right away ("we need to talk tomorrow about X"), and copy someone to put it on the calendar so you can't back out. Power often works until midnight drafting every hard email, so that in the morning all he has to do is hit send.
His short version: pull the hard decisions forward, do them kindly, but do them.
The common thread
Almost everything Power says follows from one property of manufacturing: correction times are long. A software mistake can be seen and fixed in a quarter. A wrong equipment order shows up two years later. That's why capacity has to be built before contracts, why so many decisions are one-way doors worth hedging, why flexibility (standard cells, shared capex, reusable capability bricks) is worth paying extra for, and why the risk of death stays high as long as the company grows fast.
The other constant is trust. Customers can't easily check a manufacturer in advance, so they buy a track record: compliance earned over years, deliveries made on time, and honesty about what you can and can't do yet. Power's bet is that honesty and execution compound faster than anyone expects.