Happy Tuesday. Good things come to those who wait. Make it to the end of today’s issue and there’s a treat: a new section keeping watch over the cinematic universe of Tolkien-inspired hard tech startups. Consider it Per Aspera’s sequel to the software-focused HBO hit series of the 2010s, Silicon Valley.
IN THIS WEEK’S EDITION:
🏦 Inside the Pentagon’s bank
☢️ Around the world with big reactors
💥 Crusey off cuff: the OTA hubbub
🗞️ Indiana’s 2nd swing, 201 → the lab, and $ANT
🔮 Presenting: The Shire Watch (new!)
Forwarded this? Join Per Aspera today.
On Monday, the Office of Strategic Capital (OSC), the Pentagon’s in-house lender, signed a $150M conditional loan with our friends at Stoke Space (a PA Founding Sponsor 🇺🇸) to build a second launchpad at Cape Canaveral, sized for the bigger Nova Block 2. Nine days earlier, OSC signed an $80M loan with Antheia, which brews the starting ingredients for generic drugs from sugar and engineered yeast. Two loans in ten days, ranging from launch to pharma, from an office folks commonly associate with rare earths, piqued our interest, so we combed through every deal since OSC was stood up on Dec. 1, 2022.
There are 14 deals, every one a loan, and the 12 still standing add up to $6.6B. OSC itself owns shares in nobody, which is worth emphasizing, because Washington has been doing a lot of buying in lately. A loan gets paid back with interest and leaves the cap table alone, whereas equity makes the taxpayer your co-owner.
Here's what stands out:
OSC spent its first two years on strategy and paperwork. Congress gave it lending authority in the FY24 NDAA and $984M of lending room in March 2024. Its first call drew 200+ applications asking for $8.9B, about 9x what it could lend.
Then the OBBB showed up with $1.5B for the Pentagon’s credit account, enough to stand behind as much as $200B of loans (half for critical minerals). OSC’s first loan went out the door in August 2025: $150M to MP Materials for heavy rare earth separation at Mountain Pass. (This is also the only loan OSC has announced as executed; everything since is a commitment, and funds don’t move until the borrower clears its conditions.)
Eight of the 14 are minerals and magnets: MP, Vulcan Elements, ReElement, Korea Zinc’s Tennessee smelter, Phoenix Tailings, Energy Fuels, Niron Magnetics and Sunrise Energy Metals. Since July the menu has widened: drones (Performance Drone Works, $820M), batteries (Sila, $1.4B), chips (Wolfspeed, $1.5B), quantum (PsiQuantum, $150M), medicine (Antheia), and now launch (Stoke).
Congress hasn’t let OSC buy stock, so when it wants upside it takes warrants. Vulcan handed some over, ReElement did before it walked away from its $80M in July, and Wolfspeed’s cover up to 7.5% of the company as the loan funds. The 40% stake the USG took in the Korea Zinc JV sits next to OSC’s loans, but came through a different door.
OSC says it committed over $8.4B in FY26 through Aug. 7, more than the deals it has named add up to. So there are more borrowers we haven’t met yet.
And it has started lending to lenders. OSC's new National Security Fund Finance program will lend $500M-$1B apiece to private credit funds that then make secured loans to minerals companies.
"These commitments demonstrate that OSC's federal financing tools can successfully scale private capital investment in sectors vital to our economic and national security."
Our take
Factories, mines, reactors, launchpads, and the like rely on long, cheap, patient capital that doesn’t dilute the team (AKA, debt). Venture equity is a very expensive way to scale production and pour concrete.
The founders at the top of their game right now are capital-stack-maxxers, continuously optimizing their capitalization processes and calibrating the color of money to the stage of the company: equity for the risk nobody else will underwrite, debt once there’s real assets, long-term contracts, and proven cashflows in place, and government credit where there’s been a market failure and/or Washington has decided it needs the asset (as with MP Materials). Two other ready examples:
Hadrian’s Chris Power, who as we wrote in #010, “found a way to flip the script and skillfully tap the American capital stack with purpose: using debt to build factories, preserving equity for talent, and scaling with the velocity of a software business.”
CesiumAstro’s Shey Sabripour, whose Bee Cave, TX startup closed $185M of EXIM debt, a JPMorgan revolver, and a $270M Series C inside 18 days earlier this year (#039).
The challenge, of course, is the context-switching: staying “legible” to all three groups at once (with a cap table a VC can underwrite, collateral a bank can lend against, and a national-security case a program office can endorse). But a few of you reading this have managed just fine. And we’re optimistic that more of you are figuring it out, too!
THE WORLD LEADER. Fun fact: the most nuclear-powered country on Earth is about to be Slovakia. Five reactors already supply 60%+ of its electricity, and a sixth, Mochovce 4, reached the grid on Oct. 15. At full power it covers another ~13% of national consumption, which pushes Slovakia past France.
It is a slow, expensive way to take the crown: Mochovce 4 is half of a two-reactor completion job that was budgeted at €2.8B when work restarted in 2008 and finished at ~€6.7B, a 139% overrun, and that was to finish units whose foundations were poured in the 1980s. The unit took 18 years from that restart to first electricity, and nearly four decades from the day construction began, when the country putting in shovels was still Czechoslovakia.
And yet, by recent Western standards, Slovakia got a good deal.
Looking West, France’s Flamanville 3 took 17 years and €23.7B with financing (~€14.5K/kW). Vogtle 3/4, America’s only new build this century, took 11 years from first concrete and ~$35B (north of $15K/kW).
Looking East, South Korea’s first APR1400s came in around $2.4K/kW overnight, and China’s Sanmen AP1000s finished at ~$3.2K/kW all-in, by Beijing’s count.
The above figures mix currencies, price years, and financing, but this chart puts everything in one basis: overnight cost per watt in 2020 dollars. The direction of travel is clear. China’s line bends down from the 1990s, while America’s inflects up. Source: Roosevelt Institute / Nature.
The cleanest comparison would be Sanmen, where China built the same Westinghouse AP1000 that Georgia built, starting four years earlier. All four Chinese units were running before Vogtle 3. China then localized the design as the CAP1000 and currently has a dozen under construction. The furthest along, Haiyang 3, should connect by year-end, four and a half years after first concrete.
ALL EYES ON THE AP1000
For all of the excitement about new nuclear form factors (SMRs!), Washington and its allies are very keen to build the proven design (AP1000) again, at fleet scale, and brute-force our way back down the learning curve. Over the last year the pieces have fallen into place:
The $$$: A year ago, the government agreed to arrange financing for at least $80B of Westinghouse reactors. Japan earmarked up to $100B more, tucked inside its $550B investment pledge. On Oct. 1, Korea committed up to $120B for six AP1000s and two APR1400s on federal sites, with a $10B advance due by year-end.
The parts: Vessels and steam generators take years to forge, and no press in North America is big enough for the largest pieces. In June, DOE put up $17.5B in loans to order long-lead parts for up to 10 units.
Doosan Enerbility’s 4.3 km² works in Changwon: scrap steel goes in one gate, reactor pressure vessels and steam generators leave by the dock. Built different!
The builders: Korean industry makes those parts and builds these plants. Doosan Enerbility forged Vogtle’s steam generators and Sanmen’s reactor vessels at its works in Changwon (see above), and Korea, which has put eight APR1400s into service since 2016, will build up to two of its own on a federal site, with ~$2B a unit flowing back to Westinghouse.
The paperwork: On Oct. 14, the NRC renewed the AP1000 certification to 2066 with 180+ amendments folded in.
THE GREAT HOPE is that Vogtle was a one-off penalty: the price of dusting off the boots after 30 years, relearning the tribal knowledge we’d lost, and stop-starting a few times along the way. The government has called for 10 large reactors under construction by 2030, so the clock is ticking.
A CODA IN CARRS BOTTOM, KY: Last Wednesday, The Nuclear Company proposed Prosperity One, a site with up to six AP1000s on the Ohio River in Lewis County. At 6.6 GW, it would be the largest nuclear plant in America. Phase One calls for two units, with construction targeted for 2028. It has no cost estimate, no offtaker, no NRC application, and no PJM interconnection filing yet. What it does have: an Eclipse-led $51M Series A, a JV with Brookfield (which is also running the V.C. Summer 2 and 3 restart), and a chief nuclear officer, Joe Klecha, who spent the Vogtle construction years on site. And, hey, we like big swings!
CRUSEY HERE, cofounder of Per Aspera and GP/cofounder of IceNine. Ryan asked me to weigh in on the OTA hubbub.
The defense tech world was put in a tizzy last week over OTAs (Other Transaction Authority), the 1958 contracting shortcut that lets the Pentagon buy outside the Federal Acquisition Regulation. The spark was this op-ed by a Stimson Center analyst in Breaking Defense. It’s worth reading and makes some interesting points: prototyping and production OTs 10x’d between 2016 and 2024, past $18B a year, with half flowing through consortia whose vendors GAO says the Pentagon can’t fully name. The author’s proposed fix is to confine OTAs to R&D and put young companies back under Cost Accounting Standards.
Let’s start with the strongest point in the piece: nobody knows if OTAs are faster. The Pentagon doesn’t measure how long an OTA takes to put a capability in the field vs. a FAR contract, so founders & VCs who say OTAs are faster are going off of anecdata & experience. In mine, Castelion got from prototype to production through OTAs in months. It’s one data point, but until we are tracking award-to-fielding time for both contract types, neither side gets to settle the speed argument. And underneath the speed question is a fight over who pays for development:
Primes tend to be built for cost-plus contracting, with audited overheads, Cost Accounting Standards, and a guaranteed margin on whatever that work costs.
New entrants tend to raise private capital, build on their own dime, and sell the finished product at a fixed price.
In the first model the taxpayer funds development. In the second, investors do. OTAs are the contract form that let the second model take hold inside the Pentagon, because a startup can carry a prototype through to a production award without rebuilding its accounting around government cost rules.
Now let’s follow the proposed fix through. If we were to limit OTAs to R&D and put young companies under cost accounting, a startup that can’t get a production deal after its prototype has no reason to spend its own money on the prototype, and no investor has reason to fund it. Development spending would move back onto the government’s books, which is what the piece says it wants to clean up. Meanwhile, the larger OTAs the author objects to, for Golden Dome interceptors and space data networks, went to Lockheed, Northrop, RTX, and SpaceX. In sum, narrowing the authority would squeeze the smallest users and barely touch the biggest.
My ten cents: measure delivery times, audit the consortia, and leave open the door that Castelion and many other great companies have gone through.
IceNine is an early-stage fund for space, defense, and critical technologies. We back founders building for the new industrial renaissance. Learn more here.
001 / SECOND SWING, HOOSIER STATE… Indiana has never run a commercial nuclear plant. Its one attempt, Marble Hill on the Ohio River, was abandoned half-built in 1984 after $2.5B had been spent, at the time the most expensive nuclear project ever walked away from. Last week, AEP and X-energy agreed to study a 960 MW plant made up of 12 of X-energy’s Xe-100 reactors on the site of the old Breed coal plant in Sullivan County. Amazon is paying for the study and would buy the power under a contract that covers the plant’s full cost (which keeps AEP’s ratepayers off the hook). The Xe-100 is a gas-cooled pebble-bed SMR, and X-energy’s first unit, at Dow's Seadrift site in Texas, is still waiting on its NRC construction permit. Also last week: the Navy picked Radiant’s 1 MW Kaleidos microreactor for Naval Weapons Station Crane, with power targeted for 2028. Forty-two years after Marble Hill, Indiana’s first working reactor will be a small one!
002 / 201 → THE LAB… In June 2025, Kevin Weil and Bob McGrew were sworn in as Army Reserve lieutenant colonels in Detachment 201 (#006). Last week, both launched companies to advance physical progress in the real world:
On Oct. 22, Weil, who ran OpenAI's science unit until April, and formerly helmed product at Planet Labs, Twitter, and Instagram, and sits on the board at Stoke Space, unveiled Warp Sciences with a $215M Series A led by Lightspeed at an $865M valuation. Weil & co are building a contract lab for molecule-making. (Synthesizing thousands of variations of a drug candidate is grunt work that pharma long ago farmed off to Chinese contractors). Warp will do the work in automated American labs, with chemists wearing cameras and pressure-sensing gloves to generate data for its models.
Two days earlier, McGrew's Arda came out of stealth with $232M to turn product designs into factory plans. (See more in Shire Watch below.)
003 / WHAT IS THIS, AN IPO FOR ANTS? Anthropic filed its S-1 on Monday, applying to list on the Nasdaq as “ANT,” with Goldman Sachs lead-left and Morgan Stanley beside it. The number that everyone can’t shut up about is the $80.7B net loss for H1 2026… but $72.4B of that is a non-cash charge from marking up convertible notes as the valuation climbed. The operating loss was $693M, on $16.3B of revenue (up from $1.2B a year earlier…!). Naturally, we went looking for the risk factors and hardware outlays.
Risk factors: ANT lists its Pentagon fight in there, as well as June’s episode where Commerce slapped worldwide export restrictions on the Fable 5 and Mythos 5 models. And then, very much on brand, Anthropic warns that advanced AI “could… pose catastrophic or existential risk to humanity,” which we suppose might weigh on earnings.
Hardware: This year, Anthropic has signed ~$417B in leases and purchase commitments for datacenters, compute equipment, and cloud capacity, with another ~$125-150B of TPU capacity on order for 2028-2031. ANT has a $110B take-or-pay deal with Amazon through 2036 and $114.5B with Google through 2033 (both are investors in Anthropic), plus up to $120.5B with SpaceX for Nvidia GPUs into 2029. Claude runs on four chip families (Trainium, TPUs, Nvidia, and AMD GPUs), and the company says it has "consistently operated in a compute-constrained environment."
Gather ‘round the hearth, friends, for there is news from the far West, where an unlikely fellowship of companies has taken its names from the old tales, and lately has been doing deeds worthy of the songs.
In the old tales, Arda, the world itself, was sung before it was built. Bob McGrew, OpenAI’s former chief research officer, keeps the same order of operations. Arda is conjuring an AI agent that reads a product’s CAD scrolls and lays out the whole forge: assembly order, stations, robots, work instructions. Its first customer is Anduril, where the smiths’ basic work instructions that took 10 hours now take 15 minutes. “Imagine a fully automated version of Shenzhen in West Texas,” quoth McGrew.
Nearby in The Shire, the Flame of the West is being reforged. Anduril is raising its own Grey Havens: Arsenal-2, a $6.6B submarine yard at Sparrows Point (#066). Back in Ohio, its first forge, Arsenal-1, has a new commission: the circuit boards for BlueChip, a battle-hardened laptop unveiled today with Qualcomm and built on Snapdragon chips wrought in TSMC Arizona’s forge. About 90% of the parts come from The Fellowship Of The Free Peoples (the U.S. and allies). “These computers are the anti-surveillance computer,” quoth Palmer Luckey. Sauron would hate this.
The Shire has always kept watchdogs, if only to guard Farmer Maggot’s mushrooms from thieving young hobbits. But we have bigger fields to guard: Warhound, a venture Luckey dreamed up with Ben Lamm of Colossal (of dire-wolf fame), launched Oct. 22 with $200M in seed gold to breed and gene-edit hounds of war, clad in Anduril armor and fitted with its Eagle Eye. More than 80% of the Pentagon’s hounds come from kennels across The Sea in Europe, and so few finish training that it must raise about 200 to field 100.
Under the Lonely Mountain, Erebor, the Luckey- and Lonsdale-backed bank founded in 2025, now guards a hoard of more than $7B in deposits, up from $1.1B at the end of March, and a mere eight months after opening its gates in February. Almost none of that gold has left the mountain: at the last count (June 30), loans came to ~2% of deposits, and ~90% of the hoard sat in cash. Smaug would approve.
Some gold did move: Erebor led a $200M credit line for fellow Shire member Valar, the Powers of the world, which just lured Andy Lambert, the smith who ran Falcon 9 production, away from SpaceX. Now Valar would raise a city of reactors with Project Beehive, a hive of 456 SMRs on 9,000+ acres of federal land in Utah (about 9.6 GW all in), with the first due to be lit in 2028.
And Varda, named for the Queen of the Valar who kindled the stars, has three capsules among those stars at once: two bearing hypersonic trials for commerce and the armies of the West, one carrying a test elixir against cancer. All three are due to fall back to Middle-earth in the Mordor-ish wastes of the Australian outback around Christmas, there and back again, one more leg of Varda’s quest for potions brewed in the heavens.
One does not simply walk into hard tech.
Not all who wander through those last 6 paragraphs are lost. Some of us are, though.
And with that, we’ll see y’all back here next week. 🛡️⚔️
|