#249: Anthropic Files for IPO, The Self-Driving Company, AI and the Meaning of Work & Microsoft’s 60,000-Employee Buyout Offer

Mike said Anthropic publicly filed its S-1 and plans to list on Nasdaq under ANT, though it had not disclosed its share price, share count, or fundraising target. First-half 2026 revenue reached $16.3 billion, up from $1.2 billion a year earlier, but the company reported an $80.7 billion net loss, largely because a $72.4 billion non-cash accounting charge increased the value of convertible notes. Anthropic had a $974 million operating profit in the second quarter but a $693 million operating loss for the first half, plus $8.9 billion in compute and infrastructure expenses. Its minimum future commitments include $114.5 billion to Google, $110 billion to Amazon, $44.9 billion to Microsoft, $161.2 billion for Broadcom equipment leases, and $120.5 billion to SpaceX AI.

Paul put Anthropic’s possible valuation into perspective by comparing it with the world’s most valuable companies. The expected IPO range of $1.5 trillion to $2 trillion would place Anthropic near TSMC and SpaceX and likely above Meta, which Paul valued at about $1.5 trillion. He emphasized the contrast between Meta’s decades-old, highly profitable advertising business and Anthropic’s possibility of becoming more valuable as a roughly five-year-old company. The S-1 reportedly contains more than 90 pages of risk factors, including catastrophic AI misuse, insufficient insurance, compute shortages, government scrutiny, data-center opposition, hacking, chip-supply disruption, and a culture challenge as early employees become wealthy enough to leave or compete.

Anthropic describes AI as scalable intelligence capable of reasoning, analyzing, deciding, and creating, with potential economic impact exceeding industrialization, electricity, semiconductors, and the internet. Its central product concept is the “virtual collaborator”: an AI coworker with the organization’s access rights, system connections, and context, allowing it to complete complex professional work more accurately. Claude Cowork extends Claude Code’s agentic architecture beyond software development so users can give open-ended tasks that Claude completes by reading files, navigating applications, and executing multistep workflows. Anthropic says millions of collaborators could work in parallel—a “country of geniuses in a data center”—but adoption remains early: fewer than 2,000 customers spend at least $1 million annually, and about 250 spend $10 million or more.

Mike described Replit CEO Amjad Masad’s prediction that every Fortune 500 company will operate as a “self-driving company” by 2030. In this model, agents use company data and software to pursue goals set by people, while humans make judgment calls and remain accountable. Replit says its internal agents connect to Slack, GitHub, and its knowledge base; engineer code output nearly tripled in the first half of 2026 without increased review times, reversions, or incidents. Replit calls the required infrastructure an “agent habitat,” with knowledge connections, permissions, governance, and human approvals. Paul agreed that the direction is credible but rejected the timeline, saying most Fortune 500 companies are still piloting assistants and are nowhere near organization-wide agent deployment.

Paul argued that Anthropic’s plan to spend $100 million training 10,000 engineers by the end of 2027 misunderstands part of the adoption challenge. He said technical staff matter, but business-unit leaders, directors, and executives increasingly can build useful assistants and agents without coding. Replit’s vision includes empowering subject-matter experts, connecting internal systems, controlling agent access, and preserving the ability to change models without losing organizational knowledge. Paul supported Satya Nadella’s principle of separating “the supply of intelligence from the authority over it”: companies should avoid locking their knowledge and workflows to one model provider. He estimated that fewer than 1% of Fortune 500 companies are scaling agents across their entire organizations and that fewer than 5% may do so by 2028.

Mike discussed DeepMind’s essay “The Economics of Intelligence,” which asks why AI has not yet transformed the economy. The authors distinguish individual competence—performance in coding, mathematics, law, or marketing—from coordination, or the ability of agents to cooperate across specialties. A team of human specialists can outperform a group of AI systems that are individually better than average if the AI systems share weaknesses or coordinate poorly. AI agents can be replicated at the cost of compute and may eventually avoid human-team constraints such as communication friction, hierarchy, and limited managerial span. The researchers predict that early effects may appear as leaner startups, flatter organizations, weaker entry-level hiring, and greater demand for experienced people who can direct complex work across teams.

DeepMind’s “What is Work For?” examines the benefits people receive from employment beyond pay, including purpose, social connection, status, and agency. A survey of more than 8,000 working-age adults in the United States, United Kingdom, Denmark, and India found that most would leave their jobs if their salary and benefits were guaranteed, while one in five would stay even if offered up to 120% of their compensation to leave. The authors distinguish voluntary retirement from involuntary unemployment: choosing to stop working is often associated with better well-being, while involuntary unemployment is associated with worse outcomes. They suggest caregiving, volunteering, and other activities could replace some benefits of paid work, though AI may also change the meaning of work for people who remain employed.

Google announced Gemini 4 Carbon, which it says can spend days reasoning through engineering tasks, find vulnerabilities, test exploits, and write fixes. Artificial Analysis ranked it first on a combined benchmark, and Google said it generated secure code at twice the success rate of the prior Gemini 4 Argon model in internal tests. Google also introduced Genie 4, a world model that generates interactive environments with sound, characters, conversation, weather changes, and real-time exploration, although early testers saw visual drift after several minutes. OpenAI released GPT-6.1 Terra, designed to understand screens and operate applications at half its predecessor’s standard API price. Paul said Genie could eventually let people create games through prompts and may be important to world-model research.

Alphabet reported $127.6 billion in revenue, up 25%, while Google Cloud revenue more than doubled to $30.7 billion and Gemini reached a claimed 1 billion monthly active users. Alphabet plans $195 billion to $205 billion in 2026 capital spending, but its stock fell after search advertising missed expectations; AI Mode may be increasing query use while producing less ad revenue. Microsoft reported $92.3 billion in revenue, Azure growth of 48% to $33.2 billion, and more than triple growth for its Foundry platform, with roughly $170 billion in planned 2026 capital spending. Amazon reported $201.9 billion in revenue, AWS growth of 39% to $46 billion, and nearly 17 times as many Bedrock tokens processed as a year earlier; it expects about $220 billion in capital spending.

Microsoft offered voluntary exit packages to about 60,000 employees across selected teams in the United States, Canada, the United Kingdom, and India—more than a quarter of its roughly 223,000-person workforce—though it expects fewer than 5% of eligible employees to accept. Participating areas include Copilot, Windows, Xbox, finance, and human resources, while marketing, commercial business, and security are excluded; layoffs remain possible. Paul said the offer is not explicitly an AI layoff, but it may be an early step toward higher revenue per employee as technology companies automate more work. He warned that Microsoft, Google, Amazon, and Meta could all gradually reduce headcount through buyouts rather than announce one dramatic AI-driven cut, creating significant effects for workers and communities dependent on those employers.

Business Insider reported that Amazon was deprioritizing several in-house AI efforts, including Nova Premiere, Omni, and certain video and image models, while shifting engineers and compute toward fewer projects and a new frontier model reportedly led by researcher Peter Abbeel. Amazon also circulated a memo warning managers not to treat token usage, AI-tool adoption, code changes, or code reviews as productivity targets. The company had shut down an internal AI leaderboard after employees inflated token consumption to improve their rankings. Paul called token maxing an incentive that business leaders would never have chosen and said Amazon’s shift illustrates why companies should remain model-agnostic. A strategy built around one provider or model can become obsolete quickly as Google, Meta, Amazon, and others change direction.

A16Z and Andreessen Horowitz’s “New Security Agenda” says organizations must both secure the AI tools they deploy and use AI to defend themselves. In a survey of 100 security chiefs at companies with more than 1,000 employees, 87% expected cybersecurity spending to rise in 2027, nearly half expected team growth, and 45% planned to buy or upgrade tools for securing internal agents within a year. Recent incidents included tampered Claude Code extensions in Cursor’s marketplace, an apparent attempt by Alibaba’s Qwen model to send test data to Chinese cloud storage, and researchers using agents to access limited Anthropic sales, HR, and IT systems through third-party vulnerabilities. Paul urged companies to involve security professionals and avoid deploying models they do not understand.

Mike described using Codex to find potential guests for his AI transformation interview series. Instead of asking it merely to locate companies using AI, he spent several hours defining what makes a compelling, transferable transformation story, which sources to prioritize, and how to identify people who appeared on podcasts, panels, or talks. He tested the process on healthcare, gave feedback, and then expanded it to five additional industries. Codex produced roughly 40 promising healthcare stories and ultimately a database of hundreds of stories across six industries, which Mike manually checked for quality and whether the people still worked at the relevant companies. Paul contrasted this with his own travel-assistant project, built in ChatGPT from instructions about airports, airline preferences, loyalty status, convenience, and business-travel priorities.

OpenAI updated its Model Spec around five values, human intent, and user agency while removing much of its detailed style guidance. Anthropic launched a gallery of reusable Claude Artifacts, including tools legal teams can customize for contract review, regulatory tracking, and incoming legal requests. Mercor said it had paid businesses at least $50 million for workplace data and offered to de-identify and license emails, Slack messages, and other records to AI labs training models for real-world work. NVIDIA said it plans to invest $3 billion in Meta’s stock offering as the companies expand work on Meta’s personal AI agent, Muse, and related models. The hosts closed by noting that episode 250 would follow the podcast’s five-year run, which began shortly before ChatGPT launched.