1. Google out-raises the IPO field

Alphabet priced an $84.75B equity raise on June 2 (upsized from $80B), the largest single equity raise in US corporate history, to fund AI compute against demand it says is outrunning supply. The anchor was a $10B private placement from Berkshire Hathaway, Greg Abel's biggest deployment since succeeding Warren Buffett, with the most conservative name in investing validating the AI capex supercycle. It landed the same week the listing race turned concrete: Anthropic filed its confidential S-1 on June 1, and SpaceX (which owns xAI) is targeting a June 12 debut near $1.75T.

  • Alphabet raised more than SpaceX's entire IPO aims to (about $75B), and guided 2026 capex to $180-190B. The incumbents that can self-fund the buildout are separating from those that must tap public markets to pay for it.
  • Goldman, JPMorgan and Morgan Stanley ran the deal. The banks underwriting the IPOs are also financing the incumbent.

2. The AI PC era arrives

Two keynotes in one week put AI onto the device. At Computex, Nvidia said its next-generation Vera Rubin platform is in full production and unveiled Vera, a CPU built specifically for AI agents, with Jensen Huang arguing that "AI agents will be the largest users of computing." At Microsoft Build, the pitch was Windows as an operating system for agents: on-device AI APIs spanning CPU and GPU, a new Agent 365 toolkit, an always-on work agent called Scout, and the Surface RTX Spark Dev Box. Nvidia's RTX Spark machines, compact 128GB boxes that run sizable models locally, ship this fall from Asus, Dell, HP, Lenovo, Microsoft and MSI.

  • The shared theme is decentralization at the edge: inference moving off the cloud and onto the desk, which makes on-premise and data-sovereign AI practical rather than theoretical.
  • Read against story 1, the picture splits in two: capital and capability concentrating at the top, while the hardware to run AI spreads to every machine.

3. Anthropic asks for a brake

Anthropic published "When AI builds itself" on June 4, calling for a coordinated, verifiable mechanism to slow or temporarily pause frontier development if models approach "recursive self-improvement," the point at which an AI can design and train its own successor. The company disclosed that more than 80% of the code merged into its own codebase in May was written by Claude, and that its engineers now ship roughly eight times more code per quarter than two years ago. Co-author Jack Clark likened today's trajectory to a car with only a gas pedal. The plea is deliberately not unilateral: Anthropic says it would slow only if rival labs, in both the US and China, agreed under conditions everyone could verify.

  • The irony sets the frame: the company that just filed for a roughly $1T IPO is the one asking the industry to agree on a brake.
  • A voluntary pause has no committed takers, and competitive and geopolitical pressure runs the other way. The more realistic constraint arrives later and from outside the labs, through courts, regulators and public reaction after a high-profile failure, not through a voluntary accord.

4. Adoption is outrunning governance

Two flagship surveys landed the same week with the same conclusion: AI use has gone mainstream while the systems to manage it have not. BCG's fourth annual AI at Work survey of 11,749 employees found 74% of frontline staff now use AI regularly, up 23 points in a year, and 42% of regular users save a full workday a week, yet two-thirds get no guidance on what to do with the time and most do not redirect it to higher-value work. HBR's third annual usage study, drawn from 12,637 real use cases, found therapy and companionship is again the single most common use and warned of "thinkslop," people outsourcing judgment to models tuned to flatter them. BCG's sharpest line: strategic clarity lifts measurable impact far more than better tools, and nearly half of employees now spend more time directing AI than doing the work themselves.

  • Half of organizations say they lack clear governance for mixed human-and-agent teams, even as 61% of employees expect agents to do half their job within three years. Adoption is the easy part; the operating model is the bottleneck.
  • The two reports disagree on almost nothing. When a strategy house surveying enterprises and a usage study tracking individuals both land on "use is racing ahead of judgment," the gap is real, not a framing choice.

5. Meta ships agents, delays its model

Meta launched its Business Agent globally on June 3 across WhatsApp, Messenger and Instagram, pitched by Mark Zuckerberg as a tool that will "eventually help you run your whole business," and aimed squarely at the more than a billion business-to-consumer conversations that flow through its apps daily. In the same week, the Wall Street Journal reported that Meta's flagship frontier model has slipped again, from an April target to June and now later, on performance and infrastructure concerns. Notably, Meta has not said which model powers the new agent.

  • The strategy is distribution-first: Meta is betting that reach into billions of existing conversations beats raw frontier capability, the opposite of the benchmark race everyone else is running.
  • The open question is whether a business agent built on an undisclosed, possibly lagging model can hold enterprise trust once buyers ask what is under the hood.

More AI stories of the week

  • Sanders proposes a 50% public stake in AI firms. A Senate bill, the American AI Sovereign Wealth Fund Act, would impose a one-time 50% stock tax on the largest AI companies and route the equity, plus board seats, into a federally managed sovereign wealth fund.
  • Altman floated government equity stakes to the Trump administration. NOTUS reported early talks about AI firms voluntarily ceding shares to Washington, an idea Altman first pitched to Trump in 2025, putting the administration on oddly parallel ground with Sanders. Anthropic was not part of the conversations.
  • Florida becomes the first state to sue OpenAI. The state named Sam Altman personally in an 83-page child-safety complaint, the sharpest escalation yet in the legal backlash against frontier AI.
  • Canada publishes a national AI strategy. Ottawa laid out its plan for sovereignty, adoption and safeguards, joining the wave of governments formalizing an AI policy posture.
  • AI rivals unite on bioweapons screening. The CEOs of OpenAI, Anthropic, Google DeepMind and Microsoft AI, plus Meta's AI chief and gene-synthesis firms, jointly urged Congress to make screening of synthetic DNA and RNA orders mandatory, tied to the bipartisan Biosecurity Modernization and Innovation Act.
  • McKinsey: from protecting data to protecting judgment. A McKinsey/NACD panel of sitting directors and CISOs argues AI oversight has outgrown quarterly risk reviews, now demanding real-time monitoring and genuine AI fluency rather than slogans, with the standout line that after 25 years of protecting data, the job is now protecting judgment.
  • Anthropic expands Project Glasswing. The company widened access to its security-gated Mythos-class models, continuing to treat frontier capability as something released through vetting rather than a product page.
  • ChatGPT's memory gets a major upgrade. OpenAI's new "Dreaming V3" system synthesizes context across years of past chats in the background rather than from a saved list, deepening personalization and lock-in while raising fresh questions about what these assistants quietly retain about people and how that is governed.