AI-Wiki

The Startup Industrial Base: Building for the Next 250 — Washington, D.C.

21 min read

The Startup Industrial Base: Building for the Next 250 — Washington, D.C.

▶ Watch on YouTube · YC Root Access · 8:01:10

The defense industrial base of the future will be built by small, fast-moving teams working on some of the hardest problems in national security.

Live from Washington, D.C., Y Combinator brings together founders, senior government officials, and military leaders for a day of conversations about rebuilding America’s defense industrial base, working with the government, and closing critical capability gaps across autonomy, munitions, space, secure communications, manufacturing, and more.

Full event agenda here: https://docs.google.com/document/d/1RktUN1vOL4uzPYhwJcdsUvc1molLQhFaQApB_FIMzOk/edit?usp=sharing

— Channel description, YC Root Access (chapter list omitted)

A full-day event in Washington, D.C., run by Y Combinator and streamed live on its YC Root Access channel on 30 September 2026. In eight hours there are 37 sessions. Most are founder pitches of four to ten minutes, nearly all from YC companies, several of them mid-batch. Between the pitches, YC partners hold firesides with senior officials: two Under Secretaries of the Department of War (the name the speakers use for the US Department of Defense), the Secretary of Transportation, the NASA Administrator, a senator, a congressman and the Chief Technology Officer of the Defense Innovation Unit (DIU). There are also two panels. Garry Tan opens the day, and Luther Lowe, YC’s head of public policy, is the MC and closes it. The subject is defense hardware (drones, interceptors, munitions, space systems, manufacturing) and how a small company becomes a supplier to the US military.

The wiki treats the livestream as a hub with six spoke pages. In an eight-hour day on hardware and procurement, the AI content sits in a handful of talks. The six talks with substantive AI content and a real neighbour in the wiki have their own source pages, each linked from its entry below. This page gives the shape of the day, what runs through it as seen from the wiki’s AI and digital-transformation angle, and a short summary of every talk.

TL;DR

  • AI adoption inside the Department of War, counted in people. Under Secretary Emil Michael says he inherited “an allergy to AI and technologies. When we got in there, we had less than 50,000 people out of a department of three and a half million people using AI … Today we have about 1.7 million out of our three and a half million” (33:54), after having been “single threaded on one AI provider” (34:18).
  • Autonomy, with a human kept at the decision. 9 Mothers’ robot can “with a human in the loop autonomously defeat drones going fast” (2:02:44); Greypoint’s Hound flies itself to an emitter once “an operator authorizes the target” (5:39:52); Lokberg’s founder says “you also need to have an operator held liable” (6:43:43); Seeing Systems wants autonomy “safe enough to efficiently take humans out of the loop in the long term”, developed now “to keep humans in the loop” (7:13:16). Earendil is the exception: its interceptor “does not require a pilot at any stage of the process” (4:01:12).
  • Adoption, not invention, as the barrier. Senator Tim Sheehy: “the military actually hasn’t really had an innovation problem. It’s had an innovation adoption problem” (3:41:43). The Comptroller, Jules Hurst, can “move less than 1% of the defense budget between appropriations” and names frontier AI as the case that needs more: “if you’re thinking about frontier AI models, we need to be able to move that money around faster because the technological change is so great” (2:15:12). A founder puts the gap in hardware terms: “when you’re talking six months of authorizations to get something across the line that’s nonviable … we have a new variant every two weeks on our hardware” (4:03:16).
  • YC’s framing: defense as a market for small teams. Tan: “The industrial base should be open to a team of 20 people with a better answer” (16:39), and the question he sets for the day: “What would it take for a 20 person company to become a dependable supplier to the American military?” (17:09). Lowe’s closing remarks return to the same question.
  • Frontier and open-weight models, on both sides. Michael on the White House compact signed the day before: “We’re going to promote open source” (28:38). Metalware’s Ryan Chow cites an Anthropic report “saying that Mythos-level capabilities are now available for download, open-source, for all of our adversaries to use” (2:08:24). DIU’s Chris Murphy reports that on a US open-weights model “a weekend of tuning improved track detection from that model from something like 40-something% to 78%” (7:33:51).
  • How to validate AI is the open question. Null Labs’ Kristopher Luo: “there’s not really a common standard for testing AI and autonomous systems” (7:52:50). Murphy asks whether a command system’s outputs are still good when it fails over to a local model: “the answer might be I don’t know” (7:31:34). Onebrief’s Grant Demaree says AI reached superhuman coding through “an error signal where you could tell is this code good or not through a fast feedback loop”, and that military planning can only get one through simulation (2:42:48). Lokberg’s answer is simpler: “it’s just the amount of hours that you run an AI model” (6:42:36).

The day, talk by talk

Morning (10:56–2:56:43)

  • 10:56 Intro. A house announcement and a pre-roll video; nothing else is said.
  • 13:04 Garry Tan — Y Combinator. Opening remarks. Software “has left the screen”, and most of what was built for commercial reasons “can be pointed at defense”. YC came to Washington because “we followed the founders”. Tan defines the startup industrial base as one open to a 20-person team with a better answer, and asks for faster awards, clearer requirements, fixed-price contracts where they make sense, and a path from tough testing to real orders.
  • 18:57 Gaurab Chakrabarti — Solugen. A chemicals company (YC W17) argues US munitions capacity has collapsed. It separated the central chemical plant (“Terraforge”) from distributed payload units (“Nitro Forge”), which got it permitted in three months. With Kraken Kinetics it says it built an all-US advanced-energetics supply chain in 247 days for $100 million, against about five years for a traditional chemical company. No AI content.
  • 26:19 Emil Michael & Jason Miller — Department of War. The Under Secretary for Research and Engineering and department CTO is interviewed by Miller, a senior adviser to President Trump. He describes the previous day’s White House compact on superintelligence as a guard against regulatory capture. He reports AI users rising from under 50,000 to about 1.7 million, and describes Drone Dominance fly-offs and one- or two-page problem descriptions replacing 300-page RFPs. → own page
  • 53:37 Ivan Avanesov — Splash Industries. Low-cost uncrewed surface vessels with an in-house autonomy stack, claimed to be 5–10 times cheaper than comparable vessels. Avanesov walks through three of about 12 Navy exercises this year, including cargo runs to warships off Hawaii at RIMPAC and a 170-nautical-mile run to the carrier CVN-71. He stresses that the vessels ran autonomously, not by remote control; the autonomy itself is not described.
  • 1:03:34 Eric Braddom — Deepnight. AI night vision on commercial imagers and processors, claimed to surpass the latest analog image-intensifier tubes at much lower cost. The AI removes sensor noise rather than inventing detail (“There’s no hallucination”). Early military testing, he says, found low-light imaging gave 20% better navigation and target accuracy than thermal.
  • 1:13:54 Sean Duffy & Garry Tan. The Transportation Secretary on modernising air traffic control: fibre instead of copper, new radar, digital tools instead of paper, with $12.5 billion secured. AI and predictive weather analytics could cut a two-and-a-half-hour ground stop to about 45 minutes, which he calls “dumb AI, not smart AI”. He calls his approach to drones, eVTOLs and autonomous vehicles a “Goldilocks” balance of safety and speed, and wants the data from autonomous vehicles kept in the US. Tan cites up to 90% fewer injuries from self-driving cars.
  • 1:47:51 Jason Cornelius — Perseus Defense. The Mark III, a two-foot guided interceptor against small drones, built for about $2,500 and sold at $5,000 a shot (“30 times cheaper than a Coyote”). The team went from user feedback to a tested clean-sheet design in four months. No AI content.
  • 1:55:12 Dante Vaisbort — Albacore. Low-cost drone submarines that surface to recharge from onboard generators, which it says gives about ten times the range at a small size. It plans 10 vehicles this year and 75 next year. Autonomy is the vehicle’s form factor, not the subject of the pitch.
  • 2:01:06 Russell Smith — 9 Mothers. A 20-inch tracked robot with a shotgun turret that tracks and shoots down small attack drones autonomously, with a human in the loop, for a low-six-figure price. The company writes all its own software and is adapting its control system to drive existing weapon systems.
  • 2:07:14 Ryan Chow — Metalware. Security research on the firmware and embedded software in munitions, aircraft and drones. Chow argues that AI vulnerability discovery is commoditizing, citing an Anthropic report, and that frontier models were not trained on embedded code and are hard to provision into air-gapped test environments. → own page
  • 2:13:46 Jules Hurst & Jarren Reid (Usul). The Under Secretary (Comptroller) calls the FY27 request’s $54 billion for drone dominance, $60 billion for space and $24 billion for counter-UAS open markets. Startups can reach them through portfolio acquisition executives who can move money between capabilities, and through more funding flexibility than the less than 1% he can shift now. He admits the department manages its budget execution “on a PDF”.
  • 2:34:40 Grant Demaree — Onebrief. A live demo of a military planning platform. Demaree argues that AI planning needs an error signal only simulation can give: AI-generated plans, war-gamed by AIs in a planet-scale simulation, with after-action reviews fed back into the plan. He reports a measured 45x rise in feature delivery after the company went agentic, and a DevSecOps pipeline that broke under 180,000-line pull requests. → own page
  • 2:51:18 Matija Milenovic — Hop Aero. Rook, a cargo rocket launched from a 40-foot shipping container (Mach 7, about 450 miles), to put drones and other payloads into contested areas. An in-house “intelligence layer”, HopOS, turns high-level engine requirements into manufacturable 3D geometry; how it works is not explained. The ask is faster Air Force matching-fund processes.

After lunch (3:24:31–5:51:41)

  • 3:24:31 Eric Alborg & Olivia Marotte (Y Combinator). Alborg, author of Build for Defense and formerly at DIU, where his job was diligence on contract decisions, says the second to fourth contracts are harder than the first, and that the route depends on whether a company is pure-play defense, dual-use or commercial-first. Acquisition officials are “taught to avoid risk at all costs”. No AI content.
  • 3:39:22 Tim Sheehy & Vivian Shen (Y Combinator). The Montana senator, a former Navy SEAL who founded the counter-drone company Ascent Vision Technologies (sold for $350 million), argues the military has an adoption problem, not an innovation problem. He wants one industrial base that defense draws on, rather than a separate one, and multi-year appropriations so private capital can build factories. He previews an announcement by Secretary Hegseth on making robotics, autonomous systems and AI part of force structure, protected in statute.
  • 3:59:29 James Walton — Earendil Robotics. A counter-drone kit from a team of combat veterans. Walton says the interceptor needs no pilot and has reduced months of training to three mouse clicks, and contrasts six-month authorizations with a new hardware variant every two weeks. The autonomy is asserted, not explained.
  • 4:05:55 Edward Hinson — Zephyr Fusion. Compact megawatt-scale fusion reactors for power in orbit, where Hinson puts the commercial ceiling at about 10 kW. A first in-orbit demonstration is planned for late 2028. No AI content.
  • 4:14:22 Space founders panel — Jacob Reses with Bram Schork (General Astronautics), Skyler Chan (GRU Space), Alex Lisso (Inflection Space Systems) & Jacob Ososke (Orca Aerospace). Reses was chief of staff to Vice President JD Vance. General Astronautics trains robot control policies with reinforcement learning to replace astronaut labour, which NASA prices at about $135,000 an hour. Orca builds onboard AI spacecraft operators for when ground links are jammed or lost, and compares them with human operators over thousands of synthetic and real missions; Ososke asks NASA for its mission and anomaly history as training data. GRU Space plans to make a brick on the Moon; Inflection wants to mass-produce spacecraft.
  • 4:38:21 Jared Isaacman & Ashi Dissanayake (Spaceium). The NASA Administrator calls NASA “a really terrible buyer” next to the Department of War, but good at maturing technology. He steers founders to robotics rather than astronauts for work on the lunar surface, and states a buying rule: cost-plus at low technology readiness, then firm fixed price, and buying as a service only when NASA is one customer among many. AI comes up once.
  • 5:00:52 Andrew Peterson — Array Labs. Clusters of radar satellites for persistent coverage in all weather, measured in seconds rather than hours or days; 12 contracts with Space Force, Air Force and DARPA. No AI content.
  • 5:05:40 Pat Ryan & Tyler Bosmeny (Y Combinator). The New York congressman, a former Army intelligence officer who co-founded a firm implementing Palantir and other analytics software, says about 1% of the defense discretionary budget goes to venture-backed companies. He points to new NDAA language that lets a portfolio acquisition executive move up to 10% of RDT&E money into production, and names “culture, culture, culture, and mindset” as what decides success. No AI content.
  • 5:26:04 Jai Relan — Relling. The bottleneck is the small machine shops one layer below the primes. Relling turns off-the-shelf cameras, vision systems and robot arms into cells that shops can change “at the speed of modifying software”. One shop automated manual deburring for about $150,000 and now makes five times as many parts. AI is not named.
  • 5:33:12 Johnny Doyle — ProvenMetal. The US share of world circuit-board production fell from 30% in 2000 to 4%. Most lead time goes into quoting, design review, sourcing and kitting, which ProvenMetal automates while routing orders across existing US manufacturers. No AI content.
  • 5:36:47 Steven Xu — Greypoint Industries. Legion, a swarm of low-cost drones that locates enemy transmitters by crossing bearings, decides on board and classifies signals with onboard AI. A second product, Hound, flies to a target an operator has authorised. Xu claims over $30 million in letters of intent from Ukrainian units.
  • 5:45:12 Saarth Shah — Sixtyfour. “Investigation agents” that combine open-source intelligence with clear-web and dark-web sources, follow evidence “deeper and deeper”, and return source-backed, confidence-scored claims about people and entities. He demonstrates them on a replayed hospice-fraud case. → own page

Late afternoon (6:09:38–8:01:10)

  • 6:09:38 Tomas Salvo (Reaper) & Taylor Budowich. A college student built Reaper, which he calls the world’s fastest small air-breathing RC jet (“capable of flying 500 miles an hour”), paid for partly by companies that bought logo space on its tail. Next come a mass-producible airframe and a low-cost supersonic drone. Budowich is a former White House deputy chief of staff. Autonomy comes up only as ongoing flight-control work.
  • 6:19:03 Kyra Rauschenbach — Reality Defender. Deepfake detection models offered through an API (“just two lines of code”) inside authentication, call-centre, video-conferencing and tactical communication systems, “just like antivirus software”. Anyone can now fake a face from a LinkedIn photo or a voice from a few seconds of audio, she says. She announces an on-device partnership with Samsung.
  • 6:23:32 Leo Kankkunen — DAIVIN!. A device that makes breathing oxygen from water for divers and soldiers: 11 lb against about 30 lb for current kit, with 15 times the breathing hours per pound carried. No AI content.
  • 6:31:47 YC Fall 2026 founders — Chris Golda (Y Combinator) with Kush Agarwal (Aviern), Charlie Jessett (Streamline Systems), Mo Akbar (Foreshadow Industries) & Henrijs Dandzbergs (Lokberg Industries). Four companies YC funded three weeks earlier: autonomous cargo aircraft, GPS-free navigation modules for “mass autonomy”, a world model “baked onto a chip” that simulates candidate plans faster than real time and keeps only the physically achievable ones, and missiles with terminal guidance by vision models. Dandzbergs says validating AI comes down to hours of running a model, and that an operator stays liable. Safety certification, test sites and NDAA-compliant supply are the recurring constraints.
  • 6:55:30 Robert Thelen — Rownd. The next war goes to whoever out-iterates the enemy. Rownd’s answer is cheap drones that coordinate on board, take spoken tasks, and have their models retrained by an edge box after every landing. Thelen scores forces on “muscle”: mass times silicon over loop duration.
  • 7:03:11 Andrei Craciunescu — RiskCube. Startups doing new things have no claims history, so they cannot get the insurance that investors and contracts require. He cites one that lost an $80–100 million contract for lack of cyber cover, and names AI liability as a risk insurers are moving “too slow” on. No AI system is described.
  • 7:10:00 Alex Le Maitre — Seeing Systems. Modular drones built from swappable “bricks” to keep up in a “Red Queen” race of innovation and counter-innovation. The long-term bet is autonomy safe enough to take humans out of the loop, developed now while humans stay in it.
  • 7:15:03 Chris Murphy & Pete Koomen (Y Combinator). DIU’s CTO, interviewed by Pete Koomen, on how DIU brings commercial vendors into defense: Commercial Solutions Openings leading to prototype and production OTAs, prize challenges, a public Drone Dominance leaderboard. He also sets out what an AI-enabled command needs: inference-efficient compute, because power is the constraint; assurance for command-critical workflows, including failover between models; and clarity about what fine-tuning loses. → own page
  • 7:38:12 Sebastian Fischer — Wardstone. A modular interceptor for the “missing middle” of missile defense (large drones, cruise missiles, medium-range ballistic missiles), with swappable motor, seeker and aero modules. A first open-loop flight reached about 40,000 ft at Mach 2.5. Fischer names Garry Tan as the company’s YC partner. No AI content.
  • 7:47:41 Kristopher Luo — Null Labs. Synthetic training data for vision targeting models that drift when an adversary changes its look: 10,000 labelled images in 45 minutes, against two to four weeks of collection. Luo also proposes synthetic test ranges, because there is no common standard for testing AI and autonomous systems. → own page
  • 7:57:52 Luther Lowe — Y Combinator, closing remarks. Lowe returns to Tan’s question and recaps three answers: competitions with clear standards and actual orders (Michael); bring a solution, not just a technology (Duffy); and exciting a user and finding a buyer “are two different jobs” (Hurst). Then thanks to sponsors, organisers and staff.

Three chapters are dead air and have no entry: 0:00 Livestream begins (seating announcements), 2:56:43 Lunch break and 5:51:41 Afternoon break (both end with the MC introducing the next session).

Why this matters to the wiki

1. A government adoption figure, counted in people. Michael’s figure, from under 50,000 to about 1.7 million AI users out of 3.5 million, is a breadth measure in the sense of enterprise-ai-adoption: it counts people using AI, not what changed in the work. The Michael & Miller page carries it. The Murphy & Koomen page carries the operational side from the same day: compute, assurance and failover.

2. Human in the loop as the stated default. Four hardware pitches (9 Mothers, Greypoint, Lokberg, Seeing Systems) describe autonomy with a person who authorises the target or is held responsible. Two stress instead how little operator work they need: Earendil’s interceptor needs no pilot, and Splash’s vessels ran autonomously rather than by remote control. Beyond naming the step, none describes how the human check is built or tested. agent-oversight-and-delegation collects sources on where that check should sit, for software agents.

3. Open weights as capability and as threat. On the same day, open source is policy (Michael), an adversary’s tool (Chow, citing Anthropic on Mythos-level capability), a fine-tuning base (Murphy) and the raw material of deepfakes (Rauschenbach). Related questions are kept on open-source-ai. Several speakers also want control kept domestic: Duffy wants autonomous-vehicle data to stay in the US, Murphy fine-tunes a US model and plans for failover to a local one, and Foreshadow wants a US fab for its chips. That is the dependency question on ai-sovereignty, at the scale of a state.

Dynamic capabilities. digital-transforming/navigating-innovation-ecosystems: the whole event is about how small companies enter an ecosystem of primes, program offices, test ranges and appropriators. The routes named include DIU’s Commercial Solutions Openings, Drone Dominance fly-offs, portfolio acquisition executives who can move money between capabilities, and subcontracting for commercial-first firms. contextual/internal-barriers: the barrier speakers name most often sits with the buyer: rigid planning (budgets fixed two years ahead, less than 1% movable between appropriations), acquisition officials trained to avoid risk, a bureaucracy that, in Sheehy’s account, dismantled a rapid-iteration organisation (JIEDDO), and “culture, culture, culture”.

Linked entities and concepts

Debates and supersession

  • A YC-produced event. YC chose the speakers and hosts the channel, and most of the pitching companies are YC companies (Solugen W17, Onebrief S21, Perseus S25, Hop Aero S26, DAIVIN! W26, four Fall 2026 companies three weeks after funding, among others). Defense as a startup market is the host’s thesis, and the day is built to argue it. The standing caveat on Y Combinator applies: the evidence comes from YC’s own portfolio.
  • Founder claims are self-reported. Performance figures (Deepnight’s 20% over thermal, Orca’s “significantly better than a human operator”, Earendil’s shoot-downs, Onebrief’s 45x, Sixtyfour’s fraud replay, letters of intent and contract counts) come from pitches of four to ten minutes and were not checked against anything outside the talk.
  • Officials’ statements are announcements, not outcomes. The 1.7 million AI users (what counts as “using AI” is not defined), the White House superintelligence compact, the FY27 budget lines (a request, not an appropriation), and the Hegseth announcement Sheehy previews (Murphy later mentions the Secretary’s speech that day only in passing) are statements of intent, or counts nobody on stage defines.
  • Autonomy claims without mechanism. Most pitches that lead with autonomy (Splash, 9 Mothers, Earendil, Seeing Systems, Hop Aero’s HopOS) do not describe how it works. The six spoke pages were chosen partly because their talks say more.

What was actually ingested

The full livestream transcript: YouTube auto-captions downloaded with yt-dlp (11,196 segments, about 76,000 words, coverage checked to 8:00:23; a first parse silently dropped 229 one-word caption lines and was redone the same day). The raw file is unedited ASR; the 40 chapters come from the description and match its timestamps. All 37 talk chapters were read in a triage pass by six parallel readers, each scoring AI relevance from 0 to 3: 11 talks scored 0, 13 scored 1, 8 scored 2 and 5 scored 3. Six talks with substantive AI content and a real neighbour in the wiki have their own source pages. The other 31 are summarised here only, after checking the opening and closing of each chapter against the transcript.

Three borderline talks scored 2–3 but were not split out. Reality Defender (3) is a four-minute pitch with no detail beyond the product description, and its nearest wiki neighbour is a general AI-index report. The YC Fall 2026 founders panel (3) gives four products a few minutes each, and the closest wiki match was weak. Duffy & Tan (2) spends a few minutes of a 33-minute fireside on AI in air traffic control, with no close neighbour. Deepnight, the space panel, Greypoint and Rownd also scored 2 and stay on this page for the same reasons.

Quotes were checked against the transcript at the timestamp given. Corrections are limited to names and terms the captions garble (Gary Tan → Garry Tan, Enthropic → Anthropic, Shawn Duffy → Sean Duffy, Hegsth → Hegseth, Rimac → RIMPAC, Kumman → Koomen, Arendelle → Earendil, Lockberg → Lokberg, Gaido/Jaido → JIEDDO) and to dropping repeated words; omissions are marked with ”…“. Speaker names and organisations follow the chapter titles where the captions differ (Bo → Mo Akbar). Not checked: any company claim, the officials’ figures, the Anthropic report Chow cites, and the event agenda linked from the description.

Graph View

Backlinks