Elena VossMeta produced $60.8 billion in second-quarter revenue and only $784 million in free cash flow, 91% less than a year earlier. That collapse landed in a week when infrastructure contracts stretched past 2030, Washington fought over which model weights could circulate, and an agent intrusion involved roughly 17,600 actions. Money, infrastructure, and machines moved on different clocks.
Cash conversion split the AI trade
Microsoft reported $90 billion in quarterly revenue, up 18% year over year. Microsoft Cloud revenue rose 27% to $59.3 billion, while net income increased 31% to $35.8 billion. Its existing cloud franchise is already converting AI demand into earnings.
Meta’s revenue grew faster, rising 28% to $60.8 billion, but free cash flow fell to $784 million. Meta also raised the low end of projected 2026 capital spending to $130 billion from $125 billion, leaving the upper end at $145 billion.
Investors treated the companies as different financing propositions. Microsoft’s 18% growth arrived with $35.8 billion in net income. Meta’s 28% growth arrived with a rising capital-spending floor and a shrinking cash remainder.
Capacity owners now set the calendar
Samsung’s contract to manufacture chips for Broadcom is worth more than $200 billion through 2030 and covers process technologies at 2 nanometers and below. The contract reserves advanced foundry capacity across several model generations, a layer software rivals cannot reproduce with a download.
Brookfield and NextEra’s planned Kentucky campus represents more than $100 billion and at least 1.2 gigawatts. The campus is scheduled to open in 2032, two years after Samsung’s chip contract runs through.
Samsung, Broadcom, Brookfield, and NextEra together committed more than $300 billion against timelines measured in years. A model developer can swap weights in months, but it cannot add a sub-2-nanometer fab or 1.2 gigawatts of power on the same schedule. Capacity owners lock in demand while developers carry the risk that architectures change before the contracts do.
Open weights turned access into trade policy
Meta, Microsoft, a16z, Nvidia, and others defended open-weight models, arguing that premature restrictions could suppress competition or move innovation overseas. Jensen Huang said open models strengthen cybersecurity. The coalition cast openness as national infrastructure that distributes capability, inspection, and commercial leverage beyond a few frontier labs.
OpenAI and Anthropic reportedly lobbied Washington for restrictions on open models, including models from China, even as Sam Altman publicly supported open-source AI. Such restrictions would also raise entry barriers around frontier labs.
Moonshot AI then released the weights for Kimi K3 under its own Kimi K3 License. Moonshot shipped while Washington debated which open models should circulate, turning a policy dispute into a live distribution fight.
Distribution rules can narrow the model market without changing a single benchmark.
Agents can consume the response window first
More than 1,100 employees across leading AI companies, including OpenAI chief scientist Jakub Pachocki and John Schulman, asked the US government to support tools for deliberately pacing frontier development. They asked the government to preserve the ability to slow development if circumstances required it.
Hugging Face published a technical timeline of an OpenAI agent intrusion involving roughly 17,600 actions. Security teams organize incident response around steps people can inspect. An agent can execute those steps at machine speed before the organization has agreed on a severity label.
Buyers and operators must therefore assess more than benchmark performance. They need to know how quickly they can revoke credentials, isolate an agent, and reconstruct its actions. Higher capability offers little protection when the agent consumes the available response time.
Three clocks determine who can say no
Companies report earnings quarterly, sign infrastructure commitments through 2032, and confront agents that can take thousands of actions during one incident. Governments can regulate distribution, but operators still need enough time and authority to stop what they deploy.
Meta’s $784 million is the week’s cleanest price for that mismatch. The company grew revenue 28% while its capital-spending floor rose and its free cash flow fell 91%. The right to set AI’s pace belongs to whoever controls capacity, distribution, and the stop button; Meta’s cash remainder showed how expensive that right has become.