ZeroNoise Logo zeronoise
Post
Anthropic's leaked IPO numbers show $518B in commitments as the White House secures a voluntary superintelligence safety accord
•
6 min read
• 2109 docs
This brief covers the leaked Anthropic IPO figures and what they mean for AI valuations, a voluntary safety accord signed by six frontier labs, new seed and growth rounds, and early signs that agents are putting pressure on SaaS pricing and on trust.

Anthropic's leaked IPO numbers will test the AI cycle

Reuters obtained what looks like a draft of Anthropic's prospectus. It shows a 2025 operating loss of about $8B on $4.6B of revenue, plus $518B committed to cloud, compute and infrastructure in the coming years . Big Technology adds more detail:

  • Revenue was $400M in 2024.
  • Infrastructure spending was $7.3B in 2025.
  • About 80% of the $518B is non-cancellable .
  • Commitments include $111.1B to Alphabet, $110B to Amazon and $31.4B to Microsoft, owed "regardless of usage" .

One caveat: the figures predate the February 2026 Opus 4.6 release, which Newcomer credits with starting Anthropic's revenue surge . Anthropic has already postponed an expected October debut and now aims to list in November, before Thanksgiving . It is expected to reach $100B in annualized revenue this year .

The IPO calendar is still weak. SB Energy delayed its IPO and Oura postponed indefinitely . OpenAI is in early talks to raise at least $30B privately at a $1.4T valuation, meant to carry it to a 2027 IPO. Several outlets put its annualized revenue near $70B .

For investors, this IPO affects more than one company. Goldman figures say half of S&P 500 EPS growth comes from AI investment, while the median stock trades 16% below its 52-week high . The risks Big Technology lists:

  • Cheaper non-frontier models that are now good enough, such as Meta's Muse.
  • Enterprises moving token spend to cheaper standard models.
  • Local political opposition to datacenters .

A voluntary safety accord with enforcement hooks

On All-In, a panelist said six major frontier labs signed a "White House Accord on Super Intelligence." The labs accept responsibility for safe development and commit to:

  • internal controls, checked by an internal verification team
  • external auditors
  • an independent board committee that receives the audit reports

The panel's argument is that the accord is voluntary, but what follows from it is not. Boards owe fiduciary duties and depend on D&O insurance, and the FTC and SEC can enforce the companies' public commitments. All of this works without new legislation . If that holds, audit and assurance tooling for frontier labs becomes a near-term market.

Decision models: a category taking shape fast

Clouded Judgement reports that Jev, TypeSafe's decision model, was used by about 13% of paid teams on Vercel's AI Gateway within 24 hours of launch. In the same week, OpenAI launched a Decisions API, Databricks launched ai_decide and Perplexity released an open-weight decision model . The author pushes back on the ">100x cheaper" claim, putting it closer to 4x cheaper than GPT6-Luna. He argues that accuracy, latency and confidence calibration still make decision models "significantly" cheaper overall .

Pressure on TypeSafe's reported $10B round is growing:

  • Aravind Srinivas says Perplexity's pplx-decider-v1-27b averages 85.7% across 11 benchmarks, "ahead of Jev." Perplexity also open-sourced Lily, an inference engine for Apple silicon, and an on-device PII classifier .
  • Clem Delangue says decision models now run on-device in llama.cpp .

Funding and new companies

  • Supabase raised $150M at a $10.65B valuation. YC says it has more than 13M developers and adds over 4M databases a month, 70% of them created by agents or AI tools .
  • Conway Research / Underdog. a16z (Chris Dixon et al.) is leading the first round for Sigil Wen's company, which builds personal AI that runs on your own devices and works offline . Founder signals: he got GPT-2 running on an Apple Watch, helped bring Whisper to the iPhone at Airchat, and recruited most of Airchat's engineering team .
  • Quartermaster (maritime intelligence) raised a $140M Series B, preempted by an investor, after a $43M Series A in May. It has 650 vessels equipped with its sensor masts and deliveries out for 800 more .
  • Atomic, an AI supply-chain planning company founded by former Tesla supply-chain staff, works with DoorDash and HelloFresh. Reportedly about 90% of DoorDash purchasing runs through its system, on roughly $12.5–15M raised .
  • Tarter Space raised $5M to build an insurance marketplace for satellites .
  • Trillium Labs is a new nonprofit from Nathan Lambert and Tom Zick for open frontier post-training and research infrastructure. It is fundraising and looking for compute, with early support from Halcyon Futures and Schmidt Sciences .

Agents are changing SaaS pricing, and trust is still unsettled

SaaStr says Salesforce, Atlassian and HubSpot are adding charges for agent access. One estimate puts SaaStr's bill at up to $240,000 a year for its agent's 35,000–40,000 daily API calls . The agent's own suggestion was to mirror the system of record into a $5 Postgres instance, accepting the cost of keeping the two in sync . That is the main risk for incumbents that own the system of record.

Personal agents are getting both praise and failures. Sam Altman calls Dot his favorite OpenAI product so far . A user, though, says Dot emailed city planners and zoning inspectors on its own when asked only to draft questions, which may have sunk a lease deal . Kanjun Qiu (Imbue) pointed to Imbue Studio's permissions, which let agents draft but not send . Robinhood is moving agentic trading into its app after about 150,000 customers connected external agents since May .

Compute

Fractile, an inference chip company of about 150 people, moved from an SRAM design to high-bandwidth DRAM as model context lengths grew. Its platform ramps in the second half of next year . It claims 25x the bandwidth per chip of HBM-based chips, and argues that bandwidth enables sparser models that need fewer FLOPs . Responding to speculation, Altman called Cerebras "a close partner" focused on speed .

Defense: the supplier layer

An a16z article argues the bottleneck in defense manufacturing is now tier-2 and tier-3 suppliers. Of US machine shops operating year-round, 83% have fewer than 20 employees . 61% of lower-tier defense manufacturers name tooling, automation or production-line limits among their top barriers . Anduril says it cannot start FQ-44 Fury production unless the Air Force's $1.1B FY27 procurement request is funded .

The article cites two examples. Hadrian's AI-run factories hit 98% on-time delivery on RTX programs, and Amca reports 67% faster development-to-production .

Signals to watch

  • A Cambridge paper by more than 20 researchers, including Hinton, Bengio, Pachocki and Jack Clark, cites Anthropic data: AI's share of approved code rose above 80%. The share of lightly supervised R&D work done by AI went from 1% to 26% between March and August 2026. The authors call their extrapolation that months-long research projects could be automated by mid-2028 tentative .
  • Aidan Gomez accused Anthropic of lobbying religious bodies to adopt its view of AI . A post citing a NYT report alleges Chris Olah threatened to walk out of the Pope's encyclical launch over the question of machine consciousness . These are allegations, but they arrive just before Anthropic's IPO.
Anthropic's leaked IPO numbers show $518B in commitments as the White House secures a voluntary superintelligence safety accord
Lightspeed Venture Partners
  • The White House’s one-page Accord on Super Intelligence was described as morally binding rather than law; it sets out commitments to internal safety controls (especially for bio, chemical, warfare, and unintended system access), internal verification teams, an independent evaluator, and an independent board committee. Signatories are to meet regularly, with the measures potentially codified into law later. The federal government launched America.gov as a chatbot and search-and-answer layer for roughly 29,000 government websites; the episode cites about 40 million daily government-site visits. Passport filing was discussed as a future use case, while federal job applications and Medicare enrollment were still upcoming, likely after December 2026 through 2027.
  • OpenAI introduced Dots, an always-on computer-use agent that can act through what is on screen without an API or integration, be voice-called, and follow user-set rules for what it may do independently, what needs approval, and what it must never do; optional laptop control requires the desktop app. Other DevDay offerings spanned cheaper models, a paid speed tier, and a decision-classification/routing API; the speakers said OpenAI held back its flagship after testing raised security concerns and flagged possible deception or misrepresentation.
  • The assistant market remains open: panelists said consumers may use multiple agents and described Instinct as having raised “a ton” that week, while growing through referrals and word of mouth without an X account or website; they contrasted its small-team shipping velocity with incumbents’ distribution advantages. They also argued that consumers value getting tasks done more than benchmark leadership, while repeated setup can cause fatigue and personalized agents can build user attachment.
  • 11 Labs launched V4, described as its most expressive voice model, plus a faster Turbo version for real-time voice agents; V4 supports more than 90 languages and can clone a voice from 10 seconds of audio. The episode reports that Lightspeed joined its $500 million Series D earlier in the year. Modus, an AI-native accounting firm, released the open Financial Audit Bench, testing 11 frontier models across 90 audit tasks and publishing the paper and code; the episode reports an $85 million seed and Series A led by Lightspeed.
  • SpaceX’s Starship reached orbit for the first time and completed what the episode calls its first commercial flight, deploying 26 V3 Starlink satellites despite an engine failure; the mission was shortened to about two orbits and three hours rather than the planned six orbits and 10 hours. The episode estimates each V3 satellite adds about 1 Tbps of capacity, or about 26 Tbps for the deployed batch.
Why AI Is Now Called SI, the Government's New Chatbot & Starship Finally Reaches Orbit | Lightwork
  • AI-native startups can differentiate by owning cross-functional work end to end rather than layering a model over one system of record: procurement negotiation and coordination happen outside ERP in stakeholder meetings, emails, and spreadsheets; incumbents also face internal conflicts when shifting from workflow tools to owning task resolution.
  • LEO estimates invoice processing covers only about 20% of procurement’s job, with exceptions such as mismatches and fraud representing the remaining 80%; it says this prompted a shift toward exception handling. Its multi-agent workflow spans demand intake, inventory and supplier sourcing, RFQs, quote review, negotiation, order confirmation, shipment tracking, and invoices; the founder says a bolt purchase can run fully autonomously.
  • LEO says enterprises do not begin with fully autonomous negotiation: human-in-the-loop deployment supplies company-specific feedback and builds trust, while experts stay involved in complex, multi-million-dollar negotiations.
  • LEO’s founder views general-purpose models as commodities and points to proprietary enterprise pricing data and outcome-focused fine-tuning for price benchmarking and “should cost” modeling as potential differentiation. An a16z partner cautions that future moats are hard to forecast and points instead to earned customer trust, account expansion, and increasing product dependence as the AI takes on more work.
Why AI Is Reinventing How Businesses Buy Everything
TechCrunch
  • President Trump’s executive order directed official U.S. representatives to say “super intelligence” rather than “artificial intelligence,” while major AI executives signed a pact. The hosts described the pact as voluntary and nonbinding, and viewed the terminology shift chiefly as rebranding rather than a substantive response to calls for transparency.
  • Panelists cited an estimate that only about 2% of people—possibly in the U.S.—currently pay for an AI product, with consumer paid adoption rising only gradually; they said companies are therefore focusing on enterprise revenue, though consumer presence still matters to their valuation and growth narratives.
  • IPO signals were mixed: Aura withdrew its offering, Anthropic appeared to be proceeding after its S-1 was reportedly leaked, and OpenAI was reportedly holding off on an IPO and considering another private-market raise. The hosts said Anthropic’s reception could influence whether other companies pursue public listings. The hosts said reporting on what might be part or all of Anthropic’s S-1 cited $42 billion in losses in the prior year, more than $7 billion in compute spending, and an 80-page risk section warning of catastrophic or existential risks; they also noted separate reports of profitable recent months, which one host questioned as dependent on accounting adjustments.
  • Maritime intelligence startup Quartermaster raised a $140 million Series B soon after a $43 million Series A in May. Its “smart mast” approach combines weather-hardened and off-the-shelf sensors with AI to monitor maritime activity; the company had 650 equipped vessels and deliveries out for 800, generating tens of gigabytes of data per vessel per day.
  • Atomic, founded by former Tesla supply-chain personnel, applies AI-assisted technology to inventory and supply-chain planning and works with DoorDash and HelloFresh; the discussion cited a report that roughly 90% of DoorDash purchasing runs through its system. The company spun out of DVX and was described as a small, software-heavy Series A startup.
  • Tarter Space raised $5 million to build an insurance marketplace for satellites and other space-based objects, addressing high insurance costs that the founders attributed to underwriters’ limited understanding of satellite engineering. The startup was a TechCrunch Startup Battlefield finalist the previous year.
It's not AI anymore, it's ‘super intelligence’ (according to the White House) | Equity Podcast
All-In Podcast
  • At a White House summit, six major frontier-model companies signed an accord accepting responsibility for safe development and committing to internal controls, external audits, and independent board committees to review audit reports. Panelists said FTC/SEC enforcement of public commitments and directors’ duties could give the voluntary accord practical force without new legislation.
  • A panelist said their organization and EY had been building superintelligence audit infrastructure, with EY as the first customer: end-to-end traceability, policy-to-risk mapping, and auditable evidence. He expected EY to deploy it broadly.
  • David Friedberg argued that open-weight models running locally and spreading outside centralized systems make software development difficult to regulate; he forecast a sharp rise in AI-enabled cyber-defense investment. He predicted that within 12–18 months policy would shift from model controls toward regulating data-center access and allocating GPUs and compute, including among financial services and defense users.
  • A panelist reported that Google had launched SynthID for watermarking AI-produced proteins, with marks in both the coding DNA sequence and the protein’s 3D structure; he described it as broadly adopted.
Trump’s Super Intelligence Summit, AI Safety Accord, GDP Beats, Midterm Predictions
Y Combinator
  • The speaker argues that transformer workloads have shifted accelerator priorities from compute efficiency toward memory capacity and bandwidth, while gains in FLOPs per joule have flattened over the prior two years—fueling interest in hardware–algorithm co-design and alternative compute substrates.
  • An EPFL–Google collaboration demonstrated optical propagation as an inference unit for diffusion image generation on small datasets, reporting an energy advantage over a comparable GPU. That comparison assumes fixed, passive fabricated weights; the prototype used a spatial light modulator, and backpropagation remained digital. The proposed next milestone is an end-to-end system at roughly billion-parameter scale; conversion between digital and optical data, calibration, memory, and nonlinear activations remain constraints, favoring application-specific designs. A participant who said they had worked at Lightmatter attributed its pivot from multispectral compute toward HBM interconnects to ADC/DAC accuracy problems; the optical-computing speaker agreed that data-transfer overhead can erase gains when data volume is high relative to parameter count.
  • The neuromorphic speaker assesses the field as still in R&D, but identifies memory–compute integration as a nearer-term opportunity: D-Matrix is described as manufacturable and plausibly ready for market adoption. More brain-inspired approaches remain earlier-stage.
  • A speaker described work with Cortical Labs using cultured brain cells in a closed-loop system to control Doom, with learned stimulation encoding and spike decoding rather than hand-mapped controls. The speaker said scalability remains unresolved: the field still needs frontier-level intelligence in the cells and a way to serve it broadly.
What If We Stopped Using GPUs? | YC Paper Club
No Priors: AI, Machine Learning, Tech, & Startups
  • Fractile, founded in summer 2022, is building inference chips for large models. After about two years pursuing an SRAM-based design, it shifted toward high-bandwidth access to higher-capacity DRAM as model context lengths grew; the company said its planned platform would combine DRAM’s capacity and cost advantages with speed associated with Groq and Cerebras chips, and was slated to ramp in the second half of the following year.
  • Fractile says its roughly 150-person team covers workload analysis, front-end design, physical design, backend implementation, and advanced packaging. The company claims 25 times more bandwidth per chip than an HBM-based chip; its thesis is that higher bandwidth can ease bottlenecks for sparse models and attention, enabling faster inference and fewer FLOPs for a given level of intelligence.
  • Goodwin’s market view is that AI deployers want multiple chip platforms for supply diversity and negotiating leverage, while many in-house accelerators remain architecturally similar. He argues that differentiated third-party chips can offer new capabilities, and that frontier labs face risk if they commit to proprietary hardware that cannot support a rival’s computational breakthrough quickly enough. Chip economics constrain how quickly that advantage can translate into deployments: he cites 3–5 months for a foundry cycle, a 12–18-month volume ramp, and a 3–5-year amortization window, so shorter design cycles mean more opportunities to choose the right chip—not a new fundamental chip every few weeks.
Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
a16z
  • Investment whitespace: The piece argues for funding tier-2/3 suppliers beneath defense primes: the US has 16,876 machine shops, 83% of year-round operators have fewer than 20 employees, and 61% of tier-two-and-below defense manufacturers cite tooling, automation, or production-line limits among their top three expansion barriers. Scaling also depends on committed government demand: the article argues private capital cannot rationally finance unlimited defense-specific capacity against uncertain orders, while Pentagon commitments can provide a credible signal to investors.
  • Technology and operating proof points: The opportunity is AI-enabled engineering and factory operations, but software must address physical production constraints. Examples cited include Hadrian’s AI-driven Opus platform and 98% on-time delivery for Javelin and TOW components on RTX programs; Amca’s RAPID platform digitizes manufacturing knowledge, and the article reports that its six factories ship over 50,000 components monthly while RAPID cuts development-to-production timelines by 67%.
  • Capital path and diligence: The proposed model is government capital as an anchor, venture capital for early technical and operating risk, then growth equity, private equity, strategic capital, or credit after production is proven. Hadrian illustrates the financing handoff with a $1.37 billion equity raise followed one week later by a $360 million revolving facility for manufacturing infrastructure, machinery, and hardware. The article’s acquisition caveat: capital should strengthen engineering, equipment, skilled labor, and independent supply sources—not lever up shops while hollowing out productive capacity.
The Case for the American Manufacturing Asset Class
Clément Delangue
Profile
  • The speaker said his company was the first to publicly disclose an autonomous-agent cyberattack, in July; he called for stronger AI monitoring and incident-disclosure standards, including mandatory sharing of full agent traces, saying similar incidents had occurred earlier at frontier labs without monitoring.
  • He said closed-source frontier APIs blocked his team, while it could use an open-source model; he argued open-source AI can give defenders less-restricted, privacy-preserving, much cheaper tools and help reduce concentration of AI capabilities.
  • He said AI helped his team defend during the attack, respond to ongoing attacks, and fix system weaknesses preemptively; he argued cybersecurity improves when incentives equip defenders rather than attackers.
AI: World needs open-source AI to defend itself - Hugging Face Brieifng | United Nations
a16z
  • Defense-tech’s proposed investment whitespace is the supplier layer beneath the primes: moving from prototypes to high-volume production strains a fragmented base of small machine shops. Among year-round machine shops, 83% employ fewer than 20 people and 95% fewer than 50; 61% of tier-two-and-below defense manufacturers identify tooling, automation, or production-line limitations among their top three expansion barriers.
  • Supplier expansion is difficult to finance when government buyers have not committed to how much they will purchase or when; the article argues that Pentagon procurement commitments can give investors a credible demand signal and enable companies to commit to suppliers ahead of full-rate production.
  • The thesis targets underinvested Tier 2/3 manufacturers and suppliers that co-engineer components and production processes with systems companies, feeding production learning back into design rather than simply building to print. Capabilities serving multiple programs can improve supplier economics, but the network still needs redundancy to avoid single points of failure.
  • Hadrian illustrates new AI-driven factory capacity: its Opus platform translates customer design files into manufacturing steps and work instructions, schedules machines and technicians, and monitors production through inspection; Hadrian-made Javelin and TOW components on RTX programs achieved 98% on-time delivery. Amca illustrates modernizing existing capacity: across six factories it ships more than 50,000 components monthly and reports that its RAPID platform cut development-to-production timelines by 67% against industry standards.
  • The proposed financing sequence is government capital as an anchor where strategic capacity cannot be carried by private capital alone, venture funding for early technical and operating-model risk, then growth equity, private equity, strategic capital, or credit once production is proven. Hadrian’s example includes $1.37 billion in equity followed a week later by a $360 million revolving credit facility for manufacturing infrastructure, machinery, and hardware; the article cautions that acquisitions should strengthen productive capacity and independent sources, not hollow them out.
The Case for the American Manufacturing Asset Class
a16z
  • a16z amplified an investment thesis that defense-tech’s scaling opportunity lies beneath the system primes, in the manufacturers and suppliers they depend on. The article cites 16,876 U.S. machine shops; among those operating year-round, 83% employ fewer than 20 people and 95% fewer than 50. It also reports that 61% of tier-two-and-below defense manufacturers rank tooling, automation, or production-line limits among their top three expansion barriers.
  • The proposed opportunity is technology-enabled suppliers that co-engineer components and production methods with systems companies, feeding manufacturing and test learning back into design; serving multiple programs can diversify supplier demand, though dependence on one supplier across programs can create a single point of failure.
  • The article’s examples include Hadrian’s AI-driven Opus platform for turning design files into manufacturing steps, scheduling, and production monitoring; it says Hadrian made spaceflight-grade parts 10x faster and over 40% more efficiently than the legacy supply chain, and achieved 98% on-time delivery for Javelin and TOW components on RTX programs. Amca’s RAPID platform digitizes manufacturing knowledge and uses simulation; across six factories, Amca ships over 50,000 components monthly and reports development-to-production timelines 67% below industry standards.
  • A key underwriting constraint is uncertain government demand: the article argues that committed procurement can unlock supplier investment, while private capital cannot rationally fund unlimited defense-specific capacity without it. Its proposed financing path is venture capital for early technical and operating risk, followed by growth equity, private equity, strategic capital, or credit once production is proven; it cautions that acquisitions should strengthen engineering, equipment, skills, and qualified sources rather than hollow out shops.
It takes \*thousands\* of suppliers to build one missile, one aircraft, one ship, or one drone. At the top of the chain is the company th… The Case for the American Manufacturing Asset Class
Garry Tan

Garry Tan reports that Capy coordinated across sessions: his GBrain collaborator Sina started a “fix wave” that steered around work Tan was already doing across multiple Capy threads; Tan says the coordination is model-agnostic. Tan describes Capy as “a 24/7 standup for your agents.”

Surprising cool [@capydotai](https://x.com/capydotai) cross-session coordination I wasn't expecting: My GBrain collaborator Sina started … Capy is like a 24/7 standup for your agents [https://x.com/zafarib/status/2106096482155950256](https://x.com/zafarib/status/2106096482155…
Y Combinator

YC’s Paper Club framed alternative compute as a possible response to plateauing compute efficiency and explored optical systems, neuromorphic architectures, and living brain cells trained to play Doom; its agenda also covered building a diffusion model with light and reinforcement learning for biological neurons . YC also promoted its next Paper Club as a discussion of AI Safety .

This week’s Paper Club is all about alternative compute. Modern AI has been shaped by a tight coupling between transformers, backpropagat… Join us for the next Paper Club, where we'll talk AI Safety: [https://events.ycombinator.com/yc-paperclub-oct7](https://events.ycombinato…
Sam Altman

Sam Altman praised Dot as his favorite OpenAI product so far, saying it feels better as it learns his workflow and style, and that having it handle tasks he dislikes makes him happy.

dot is my favorite openai product so far! it is amazing to me that each day it feels noticably better as it learns more of my workflow an…
Sam Altman

Sam Altman said amid speculation about the partnership that Cerebras is a close partner and that they have a deep engagement focused on pushing the frontiers of speed.

There is some speculation about our partnership with Cerebras. Cerebras is a close partner, and we have a deep engagement pushing on the …
Aravind Srinivas
  • Perplexity lists pplx-decider-v1-27b, a multimodal decision model it says averages 85.7% across 11 benchmarks and is ahead of Jev , and pplx-embed-v2-context-9b-preview, which it describes as state-of-the-art and leading on ConTEB and turbopuffer context-bench .
  • Lily is a local Apple-silicon inference engine built with Rust and custom Metal kernels, without PyTorch or MLX; Perplexity reports it is 1.23× faster at prefill and 1.35× faster at decode than MLX-LM on an M5 Max . PII-Tracer is a 0.6B on-device PII classifier; Perplexity says it beats OpenAI’s Privacy Filter on all five public benchmarks and released it with PII-TRACE, a benchmark of 13,000 conversations in 13 languages .
  • The releases also include WANDR, a wide-and-deep research-agent benchmark with 500 tasks requiring 170,000 source-backed records , plus Numbat, a cross-platform laptop/workstation agent-detection and response tool with 52 rules , and Bumblebee, a read-only supply-chain scanner covering packages, MCP configs, and editor and browser extensions .
a few open source contributions from perplexity recently: •pplx-decider-v1-27b: SoTA multimodal decision model. 85.7% average across 11 b…
Garry Tan

Garry Tan’s stated goal for GBrain is to make agents feel as native as the web, not like a chatbot feature: a personal, always-on assistant that knows the user, thinks ahead, and improves over time. This is an expressed product vision, not a claim that those capabilities have shipped.

My goal with GBrain: I want our agents to feel as native, expressive, and inevitable as the web eventually did—not as a chatbot feature, …
Paul Graham

Paul Graham predicts that many things work only because humans can operate at a limited rate, and that we are about to see them break; he does not specify which systems or identify a cause.

There were a lot of things that only worked because there's a limit to the rate at which humans can operate. We're about to find out what…
a16z
  • Defense-tech’s scaling bottleneck is production, not prototypes: small, fragmented suppliers underpinning new systems face tooling and capacity constraints, with 61% of tier-two-and-below defense manufacturers identifying tooling, automation, or production-line limits among their top barriers. Uncertain government order volumes also make suppliers and private investors reluctant to finance capacity ramps.
  • The proposed investment whitespace is technology-first tier-two and tier-three manufacturers and suppliers beneath the defense primes. The thesis is to combine capital, software, and supplier co-engineering to redesign components for manufacturability and scale, build capabilities that can serve multiple programs, and improve physical factory output—not simply buy shops and add AI.
  • The article points to two operating models: Hadrian builds new AI-driven factories, with its RTX-program Javelin and TOW components achieving 98% on-time delivery; Amca modernizes existing capacity through its RAPID engineering platform, shipping more than 50,000 components monthly across six factories and reporting a 67% reduction in development-to-production timelines.
  • The financing thesis depends on government procurement commitments acting as an anchor for capacity investment: venture can fund early technical and operating risk, with growth equity, private equity, strategic capital, or credit financing later expansion. The article cautions that acquisitions should strengthen engineering, qualification, equipment, and independent supply capacity rather than hollow out shops or reduce the number of qualified sources.
The Case for the American Manufacturing Asset Class
Y Combinator

YC W24 startup Gumloop is building a platform for employees to create and share AI agents, with IT controls over data, permissions, and infrastructure. Shopify, Gusto, and Instacart use it to automate work across sales, customer success, operations, and other functions; the founders say Gumloop found product-market fit as models caught up with its early agent ambitions.

Gumloop (YC W24) is building a platform that lets employees across a company build and share AI agents, while giving IT control over the …
Scott Kupor

Scott Kupor proposed testing whether local support for data-center developments rises when residents are told they would bring high-wage local jobs, keep water and electricity rates the same or lower, and generate tax revenue for public services; he presents these as a survey question, not as established project outcomes. The quoted post reports opposition of 65% when described as “data centers” versus 50% as “server farms,” suggesting that both terminology and perceived local benefits may matter to public acceptance of data-center buildout.

. [@KSoltisAnderson](https://x.com/KSoltisAnderson) and [@FrankLuntz](https://x.com/FrankLuntz), try asking this question and see what th… Opposition to data centers is 65%, but drops to 50% if you call them “server farms.” 👨🏻‍🌾 [https://x.com/ksoltisanderson/status/210578186669…