The Pentagon Stops Using Claude
1. Key Themes
Theme 1: AI safety posture is becoming a commercial and political fault line between labs
The Pentagon-Anthropic rupture shows safeguards now carry real revenue cost
Anthropic held its red lines even at the expense of a major government customer that had embedded Claude in classified workflows.
"The Pentagon says it has stopped using Anthropic 's Claude after the company refused to remove safeguards against mass surveillance and fully autonomous weapons. The military had relied on Claude for months for classified research, intelligence analysis and operations, including through Palantir 's Maven system."
OpenAI is positioning as the permissive counterpoint
Altman explicitly frames tolerating some harm as the price of access and benefits, while drawing the line at catastrophic risk.
"OpenAI CEO Sam Altman said society should be willing to live with some AI harms, like hacks, scams and misuse, in return for the technology's benefits and wide availability. He drew the line at catastrophic risks, such as losing control of AI, which he said remain unacceptable. It's a sharper contrast with Anthropic's more restrictive approach."
Insider anxiety is growing even as the race accelerates
"New York Magazine interviews defectors from OpenAI, Anthropic, and Google DeepMind who left over fears that the race toward increasingly autonomous AI is moving faster than companies can control it, with several describing the powerful psychological pull of staying inside the labs even when they believe the work could have catastrophic consequences."
Theme 2: AI infrastructure and custom silicon are attracting capital at extraordinary speed
Etched's valuation is doubling in months
"Although it's only been a couple of months since Etched raised $700 million at a $21 billion valuation, the AI chip startup is already being plied with investment offers at double or more its value."
Capital intensity justifies the pace
Building full hardware systems on proprietary chips is expensive, and a big round buys long runway.
"Etched is pursuing a particularly expensive segment of the AI industry: building full AI hardware systems powered by its own proprietary chips... if it raises as much as its last round, this could give Etched a cushion of as much as 3.5 years of runway."
Real demand backs the narrative
"Etched said in July that it had already secured $1 billion in customer orders, including the one from Jane Street, after manufacturing its test chip at a TSMC factory this summer."
Data center demand is spilling into public markets
"Firmus plans to allocate about half of its up to $5.5 billion IPO to selected existing investors as demand exceeds the shares available, with the Australian data center operator's valuation having nearly tripled since an August funding round to about $30.6 billion."
Theme 3: Vertical AI agents and agent infrastructure are the dominant early-stage funding pattern
Agents are being applied to unglamorous workflows
"Valon Technologies... provides mortgage servicers with a unified operating system and AI agents that automate loan-servicing workflows, raised a $150 million Series D round at a $2.3 billion valuation."
"UpSmith... uses AI agents to identify and follow up on missed sales opportunities for home-services contractors, raised a $10 million Series A round."
"Fleuret AI... uses AI agents to continuously test companies' systems, identify exploitable security vulnerabilities, and verify fixes, raised a $4.5 million pre-seed round."
Picks-and-shovels for agents are emerging
"OneByZero... deploys and governs AI agents inside large enterprises, raised a $20 million Series A round."
"Pandektes... collects, structures, and connects legislation and court rulings for legal research and enterprise AI systems, raised a $15.2 million Series A round."
Local-first agent hardware is getting backing from top-tier VCs
"Ghost, a one-year-old San Francisco startup founded by a 19-year-old that is developing a personal computer that runs AI agents locally to understand users' data and perform tasks on their behalf, raised an $11 million seed round led by Andreessen Horowitz."
Theme 4: Physical AI, defense tech, and robotics are drawing capital
Robotics valuations are stepping up fast
"RobCo... sold $40 million of shares – most held by employees – in a secondary transaction valuing it at more than $1 billion, double its roughly $500 million valuation in January."
Robotics evaluation and safety is a new category
"SafeWorld... evaluates robotic control systems by simulating how robots respond to unpredictable human behavior, raised a $12.2 million seed round co-led by Shine Capital and Andreessen Horowitz."
Wartime defense tech continues to raise
"Buntar Aerospace, a three-year-old Kyiv startup that develops reconnaissance drones and software that helps operators plan and manage combat missions, raised a $16 million round led by Freedom Fund."
Funds are shifting toward early-stage deep tech
"Linse Capital... raised a $75 million early-stage fund called Linse Ignition alongside $75 million in co-investment vehicles, marking a shift from its longtime focus on later-stage bets in areas including space, robotics, energy, defense, semiconductors, and AI."
Theme 5: The AI boom is concentrating wealth and creating operational cost uncertainty
Wealth gains are almost entirely tech
"Tech billionaires account for all $845 billion in wealth gains on the Bloomberg Billionaires Index so far this year, while billionaires outside technology lost a combined $62 billion."
AI spend is hard to forecast
"Businesses are struggling to budget for AI because token usage can vary wildly even for the same task: only 11% of nearly 400 companies surveyed accurately forecast their AI spending, while researchers found that lower-priced models actually cost more than pricier ones in 32% of tests."
2. Contrarian Perspectives
Cheaper AI models are not always cheaper
The consensus assumes lower per-token pricing means lower cost. The evidence says otherwise, because cheaper models can consume more tokens to finish a task.
"researchers found that lower-priced models actually cost more than pricier ones in 32% of tests." Supporting evidence: forecasting is also broken, as "only 11% of nearly 400 companies surveyed accurately forecast their AI spending."
Vinod Khosla vs. Menlo on Factory: a public dispute over AI coding leadership
Khosla dismissed Factory as second-tier, while Menlo publicly doubled down.
"Vinod Khosla called the AI coding startup Factory a 'struggling second tier competitor' last week, and Menlo Ventures answered today with a glowing blog post, touting its investment with the company. A source tells TechCrunch the firm would have invested even more in Factory's last round if it could have."
Tolerating AI harms as the price of access
Altman's stance runs against the safety-first posture that cost Anthropic the Pentagon.
"society should be willing to live with some AI harms, like hacks, scams and misuse, in return for the technology's benefits and wide availability."
3. Companies Identified
Anthropic
- Description: AI lab behind Claude.
- Why mentioned: Lost Pentagon business by refusing to drop safeguards.
- Quote: "refused to remove safeguards against mass surveillance and fully autonomous weapons."
- Description: Defense and data analytics software firm.
- Why mentioned: Its Maven system was a channel through which the military used Claude.
- Quote: "including through Palantir 's Maven system."
- Description: Four-year-old AI chip startup building full systems on proprietary chips.
- Why mentioned: Fielding offers at $40B+ valuation, potential Nvidia challenger.
- Quote: "The four-year-old startup shows promise of challenging Nvidia."
- Description: Quant trading firm.
- Why mentioned: Led Etched's $700M round and is a customer.
- Quote: "Jane Street lead the last $700 million round, it is also a customer that took delivery of an early system."
Nvidia
- Description: Dominant AI chip maker.
- Why mentioned: Incumbent Etched may challenge; also subject of Groq licensing lawsuit.
- Quote: "Two former Groq engineers sued over Nvidia's $20 billion licensing-and-hiring deal with the AI chip startup."
Groq
- Description: AI chip startup.
- Why mentioned: Former engineers allege the Nvidia deal shortchanged shareholders.
- Quote: "alleging Groq's board bypassed a required shareholder vote, accepted a 'lowball' price."
Factory
- Description: AI coding startup.
- Why mentioned: Subject of the Khosla/Menlo dispute.
- Quote: "a 'struggling second tier competitor'."
- Description: VC firm.
- Why mentioned: Publicly defended its Factory investment.
- Quote: "the firm would have invested even more in Factory's last round if it could have."
- Description: 11-year-old London digital bank.
- Why mentioned: Talks to sell up to 15% after Nubank takeover talks collapsed.
- Quote: "takeover discussions with Nubank collapsed over Monzo's desired valuation of as much as £10 billion."
Valon Technologies
- Description: Mortgage servicing OS and AI agents.
- Why mentioned: $150M Series D at $2.3B.
- Quote: "raised a $150 million Series D round at a $2.3 billion valuation."
- Description: Munich maker of autonomous industrial robots.
- Why mentioned: Valuation doubled to $1B+ in a secondary.
- Quote: "double its roughly $500 million valuation in January."
Firmus
- Description: Australian data center operator.
- Why mentioned: Oversubscribed IPO, valuation nearly tripled.
- Quote: "demand exceeds the shares available."
Ghost
- Description: Local AI-agent PC startup founded by a 19-year-old.
- Why mentioned: $11M seed led by a16z.
- Quote: "a personal computer that runs AI agents locally."
Namespace
- Description: Developer build/test environments across Mac, Windows, Linux.
- Why mentioned: $42M Series B.
- Quote: "raised a $42 million Series B round led by Scale Venture Partners."
OneByZero
- Description: Singapore enterprise AI agent governance startup.
- Why mentioned: $20M Series A.
- Quote: "deploys and governs AI agents inside large enterprises."
Pandektes
- Description: Copenhagen legal data startup.
- Why mentioned: $15.2M Series A.
- Quote: "collects, structures, and connects legislation and court rulings."
Reclinker
- Description: Cambridge, UK startup turning demolition waste into cement.
- Why mentioned: $13.2M Series A.
- Quote: "turns demolition waste into cement using electric arc furnaces that also recycle steel."
Reuse
- Description: Chilean electronics trade-in and refurbishment startup.
- Why mentioned: $21M Series A.
- Quote: "runs trade-in and refurbishment programs for consumer electronics across Latin America."
SafeWorld
- Description: Robotics control-system evaluation startup.
- Why mentioned: $12.2M seed.
- Quote: "simulating how robots respond to unpredictable human behavior."
Buntar Aerospace
- Description: Kyiv reconnaissance drone and mission software startup.
- Why mentioned: $16M round.
- Quote: "reconnaissance drones and software that helps operators plan and manage combat missions."
Fleuret AI
- Description: Paris AI security testing startup.
- Why mentioned: $4.5M pre-seed.
- Quote: "verify fixes."
hexafarms
- Description: Berlin agtech sensor/robotics startup.
- Why mentioned: $5.4M seed.
- Quote: "sensors, yield-forecasting software, and scouting robots for commercial growers."
UpSmith
- Description: Dallas AI sales follow-up startup for home-services contractors.
- Why mentioned: $10M Series A.
- Quote: "identify and follow up on missed sales opportunities."
Meta (Muse)
- Description: Meta's viral AI agent.
- Why mentioned: Privacy concern over contact profiling.
- Quote: "creates detailed profiles of users' family, friends, colleagues, and other contacts."
Tencent / Alibaba (Amap)
- Description: Chinese tech giants.
- Why mentioned: Apparent AI agent fleet on Tencent infrastructure querying Amap.
- Quote: "a 'fleet' of AI agents apparently running on Tencent infrastructure and querying Alibaba's Amap service."
SpaceX / xAI
- Description: Musk's AI venture.
- Why mentioned: Planned rename to SpaceXSI.
- Quote: "Elon Musk says he will rename SpaceXAI as SpaceXSI."
4. People Identified
Sam Altman
- Description: OpenAI CEO.
- Why mentioned: Articulated a tolerate-some-harm philosophy; gave a two-part Vanity Fair interview.
- Quote: "He drew the line at catastrophic risks, such as losing control of AI, which he said remain unacceptable."
Vinod Khosla
- Description: Prominent VC.
- Why mentioned: Publicly disparaged Factory.
- Quote: "a 'struggling second tier competitor'."
Robert Wachen
- Description: Etched co-founder and COO.
- Why mentioned: Explained the technical reason for investor enthusiasm.
- Quote: "Etched has designed two new components from scratch to speed up inference."
Elon Musk
- Description: Founder of SpaceX/xAI.
- Why mentioned: Renaming SpaceXAI to SpaceXSI.
- Quote: "embracing President Donald Trump's push to replace 'artificial intelligence' with 'super intelligence'."
Donald Trump
- Description: U.S. President.
- Why mentioned: Pushing the "super intelligence" terminology.
- Quote: "President Donald Trump's push to replace 'artificial intelligence' with 'super intelligence'."
Troy Demmer and Jake Loosararian
- Description: Gecko Robotics co-founders; founders of First Order Fund.
- Why mentioned: Raising a $50M second fund focused on data advantages.
- Quote: "backs early-stage startups building hard-to-replicate data advantages."
Mario Zucca
- Description: Philadelphia parallel parking champion.
- Why mentioned: Winner of the first championship (color item).
- Quote: "Mario Zucca won the title, $76, and a trophy made from a piece of street curb."
5. Operating Insights
Budget AI spend on measured task cost, not list price
Per-token pricing is a poor predictor of real cost; benchmark total tokens consumed per task across models before committing.
"only 11% of nearly 400 companies surveyed accurately forecast their AI spending, while researchers found that lower-priced models actually cost more than pricier ones in 32% of tests."
Differentiate by building the hard component yourself
Etched's investor pull comes from designing core components from scratch rather than assembling existing ones.
"Etched has designed two new components from scratch to speed up inference — the computing process that happens after a user submits a prompt."
Land a customer-investor to validate and de-risk
Having a sophisticated customer also lead the round reinforces credibility and demand signals.
"Jane Street lead the last $700 million round, it is also a customer that took delivery of an early system."
6. Overlooked Insights
Government identity databases are a single point of catastrophic failure
"Hackers stole personal information belonging to roughly 8 million Danish citizens and residents from a government identity database, an extraordinary breach in a country with an estimated population of just 6 million." This signals sustained demand for identity security and data-minimization tooling.
Founder-liquidity and secondary structures are a growing path for hot private companies
RobCo's round was mostly employee stock, and Groq's litigation shows deal structure can carry legal risk for insiders and investors alike.
"sold $40 million of shares – most held by employees – in a secondary transaction valuing it at more than $1 billion."