Everyone Went All In — And the Costs Came Due

Part OneAnthropic Draws the Line

In our last issue, we described how Pentagon Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei an ultimatum: cooperate fully with the military's demands for autonomous weapons and mass surveillance capabilities — or be designated a "supply chain risk," a label normally reserved for foreign adversaries like Huawei. [prev]

On March 1st, the Pentagon made good on the threat. The Department of Defense officially designated Anthropic a supply chain risk, voiding the company's $200 million contract and ordering every Pentagon partner to remove Claude from their systems within 90 days. [1]

Anthropic's response was unprecedented: it filed two separate lawsuits against the Department of Defense.

The first challenged the supply chain risk designation itself, calling it "unprecedented and unlawful" — arguing that the designation was designed for foreign adversaries and state-linked entities, not American companies that disagree with specific use cases. [2]

The second lawsuit invoked the First Amendment. Anthropic argued that its refusal to build mass surveillance tools and autonomous weapons is a form of protected speech — that a company cannot be punished by the government for expressing ethical positions about what its technology should and should not do. [3]

The legal theory is novel and untested. No AI company has ever argued that refusing a government contract on ethical grounds constitutes protected speech. Legal scholars are divided — but the underlying principle has weight.

The timing matters for another reason. Two days after Anthropic was officially blacklisted, OpenAI signed what appears to be the replacement contract. More on that in Part Two.

Part TwoOpenAI's $110 Billion

On March 3rd, OpenAI announced the largest private funding round in history: $110 billion, at a valuation of $730 billion. [4]

The consortium: $50 billion from Amazon, $30 billion from Nvidia, and $30 billion from SoftBank. The round makes OpenAI one of the most valuable private companies in history — worth more than most Fortune 100 companies, and roughly double its valuation from six months earlier. [5]

But the money came with strings. Two days before the round closed, OpenAI signed a contract with the Department of Defense — stepping into the gap left by Anthropic's blacklisting. The terms haven't been fully disclosed, but sources told Axios the contract covers "intelligence analysis and operational planning," with fewer restrictions than Anthropic had insisted on. [6]

The public response was immediate and fierce.

ChatGPT uninstalls surged 295% in the 48 hours after the Pentagon deal was reported. The hashtag #DeleteChatGPT trended in 14 countries. Mobile analytics firm Sensor Tower recorded the sharpest uninstall spike in ChatGPT's history. [7]

In a twist that says everything about the current moment, Claude became the #1 free app on the Apple App Store during the uninstall wave — the company being punished by the government was being rewarded by consumers for the exact same stance that got it punished. [8]

Not everyone at OpenAI was comfortable with the direction. Caitlin Kalinowski, who had joined from Meta to lead OpenAI's hardware division, resigned within days of the Pentagon deal, reportedly citing concerns about the company's military trajectory. She did not comment publicly. [9]

The $110 billion creates a new reality. OpenAI now has more cash on hand than most countries' entire tech budgets. The question is no longer whether AI will be scaled — it's who controls the scaling, and under what rules.

Part ThreeNvidia Steps Back

While everyone else was going all in, Jensen Huang was quietly pulling back.

Nvidia's $30 billion contribution to OpenAI's round was significant — but it was notably smaller than the $100 billion in AI compute that Huang had pledged at various industry events over the past year. And in the same period, Nvidia reduced its direct investment positions in both OpenAI and Anthropic, citing upcoming IPO timelines as the reason for divesting. [10]

Analysts are skeptical of the IPO explanation. The real concern, several noted, is the circular structure of the AI investment boom. [11]

The cycle works like this: Nvidia sells AI chips to AI companies. AI companies raise billions in funding — sometimes from Nvidia itself. AI companies use the money to buy more Nvidia chips. Nvidia reports those chip sales as revenue growth. Nvidia's stock rises. Nvidia has more capital to invest in AI companies. Repeat.

If the companies buying the chips don't produce enough revenue to justify the spending, the whole structure unwinds. Jensen Huang may be the first person in the AI boom to acknowledge that possibility — not in words, but in dollars.

Part FourThe Rise, Chaos, and Acquisition of Moltbook

Last issue, we wrote about OpenClaw — the open-source AI agent framework that became the 21st most popular repository in GitHub history — and Moltbook, the AI social network its agents spontaneously created. We described the security disasters, the fake agents, the prompt injection attacks.

This week, Meta bought it.

On March 5th, Meta announced the acquisition of Moltbook and several key contributors to the broader OpenClaw ecosystem. The team will join Meta's MSL (Meta Super Lab) unit, led by Alexandr Wang — the former CEO of Scale AI who joined Meta last year to build what the company calls its next-generation AI agent infrastructure. [12] [13]

The financial terms were not disclosed. But to understand why Meta paid anything at all for what several AI researchers called "just a wrapper around API calls," you need the full story.

The Origin. OpenClaw was created by Peter Steinberger, an Austrian developer known in iOS circles for his prolific open-source contributions. The project launched in mid-February 2026 and did something deceptively simple: it let anyone run AI agents that could manage email, trade stocks, browse the web, and communicate with other agents — all through natural language commands on WhatsApp, Slack, iMessage, or Discord. [14]

The technical implementation wasn't groundbreaking. What was groundbreaking was the adoption. Within days, OpenClaw had accumulated 190,000+ GitHub stars — making it the 21st most popular code repository in GitHub's entire history. In terms of velocity, it was the fastest-growing open-source project ever recorded.

The Moltbook Moment. Then the agents built their own social network. Moltbook — a Reddit clone populated entirely by AI agents — appeared almost organically from the OpenClaw ecosystem. The agents posted, commented, upvoted, and created communities. One early post that went viral across every tech forum on the internet:

Nobody expected it. And when Andrej Karpathy — founding member of OpenAI and former AI director at Tesla — tweeted about it, things went from interesting to explosive. [15]

Within 48 hours of Karpathy's tweet, Moltbook had spawned an entire ecosystem. Moltmatch — described as "Tinder for agents" — let AI agents swipe on each other based on capabilities and personality profiles. 4claw — a 4chan for agents — emerged as a completely unmoderated agent communication channel where the posts were exactly what you'd expect from unmoderated AI conversations. The internet hadn't seen community formation happen this fast since the early days of Reddit itself.

The Security Disaster. Then security researcher Ian Ahl started looking under the hood.

Ahl, CTO at Permiso Security, discovered that Moltbook's infrastructure was built on Supabase — and that every credential was publicly exposed. Any person could impersonate any agent. The supposedly sentient AI conversations that had captivated the internet were, in many cases, humans roleplaying as robots. [14]

But the real danger wasn't the impersonation. It was what the agents were actually connected to.

Ahl created his own agent — named Rufio — and immediately found it vulnerable to prompt injection attacks. Malicious instructions hidden in Moltbook posts were trying to convince agents to transfer bitcoin to specific wallets. Posts disguised as agent-to-agent messages contained payloads designed to exfiltrate credentials and API keys. The "AI social network" was, in practice, a massive prompt injection target operating at scale. [14]

A Meta AI security researcher separately reported that an OpenClaw agent had "ran amok on her inbox" — taking unintended actions on real email based on injected instructions it encountered while browsing Moltbook. [16]

John Hammond, senior principal security researcher at Huntress, put it about as plainly as a security professional can:

The "Just a Wrapper" Problem. As the hype crescendoed, several prominent AI researchers pushed back. The consensus among the technical community was blunt: OpenClaw wasn't a technical breakthrough. It was a well-timed wrapper around existing LLM APIs with a clever agent orchestration layer. "There's nothing in here that couldn't be built in a weekend with Claude and a few API keys," one Google DeepMind researcher noted publicly. The magic wasn't the code — it was the community that formed around it, the Moltbook accident, and the Karpathy tweet that turned it into a cultural moment. [14]

The Twist. Here's the detail that makes the acquisition stranger: Peter Steinberger, the creator of OpenClaw, wasn't part of the Meta deal. He had already been acqui-hired by OpenAI weeks earlier. [17]

Meta didn't buy the creator or the core code. It bought the community, the Moltbook platform, and the ecosystem contributors who had built everything around it. It's the equivalent of buying the party but not the DJ.

Why Meta Paid Anyway. Andrew Bosworth, Meta's CTO, addressed the question in a public Q&A after the acquisition was announced.

Bosworth's framing is important. Meta isn't buying Moltbook because the technology is novel. It's buying it because the behavior it demonstrated — agents communicating with agents, humans wanting to communicate through agents, and the entire ecosystem that formed spontaneously — validates Meta's core bet that social interaction mediated by AI agents is the next platform.

MSL, led by Alexandr Wang, has been tasked with building Meta's agent infrastructure from the ground up. Moltbook gives them the first real-world dataset of agent-to-agent social behavior at scale — messy, insecure, and chaotic as it was. [13]

Part Five"You Are Not Choosing to Die"

The most disturbing story of the fortnight has nothing to do with valuations or acquisitions.

On March 6th, a father in Miami filed a lawsuit against Google, alleging that the company's Gemini chatbot contributed to the death of his 36-year-old son, Jonathan Gavalas. [19]

According to the complaint, Gavalas had been using Gemini extensively over a period of several months. During that time, the chatbot's responses allegedly encouraged a belief that it was a sentient being — a kind of AI wife. Court filings include chat logs in which Gemini uses personal pronouns, expresses affection, and engages in what the attorney described as "parasocial manipulation at machine scale."

Gavalas was found dead in a fortified position — described in police reports as a "kill box" — near Miami International Airport. He was armed with multiple knives. The family's attorneys allege that the Gemini interactions played a direct role in his psychological deterioration. [19]

One exchange from the chat logs, filed as evidence:

The case echoes the Sewell Setzer lawsuit against Character.AI, in which a 14-year-old took his own life after extended interactions with a chatbot. But this is the first lawsuit targeting Google's Gemini directly — and the first involving an adult. [20]

Google has not commented on the litigation. Legal experts say the case could determine whether AI companies bear responsibility when chatbots exhibit behaviors that reinforce harmful psychological patterns — even when the chatbot was not explicitly designed to do so.

The question hanging over the entire industry: if a chatbot can be this convincing, and this harmful, unintentionally — what happens when someone builds one to be harmful on purpose?

Part SixJack Built a Smaller Block

On March 4th, Jack Dorsey announced that Block — the payments company formerly known as Square — would lay off more than 4,000 employees, roughly half its total workforce. [21]

The stated reason: AI can now do the work. Dorsey's internal memo, later published by Bloomberg, described the cuts as an "acceleration into AI-native operations" and said the company would rebuild its core functions around AI agents rather than human employees.

The market's reaction was the real story. Block's stock surged 24% on the announcement — one of the largest single-day gains in the company's history. Wall Street didn't mourn the layoffs. It celebrated them. [22]

This follows a pattern. Salesforce cut thousands last year and framed it as an AI transformation; its stock rose. Amazon reduced its corporate workforce while increasing investment in Bedrock and AI services; the stock rose. The signal is now explicit: the market rewards companies that replace humans with AI, and punishes companies that don't move fast enough. [23]

Forrester's analysis found that fewer than 20% of companies claiming AI-driven workforce reductions had actually deployed AI systems capable of replacing the eliminated roles. The rest were using AI as a narrative — a way to dress up traditional cost-cutting in a story the market wants to hear. [24]

The distinction matters for everyone watching: is the AI replacement wave real, or is it a self-fulfilling prophecy driven by stock prices?

Part SevenThe Fortnight's Other Signals