The fights this week were about plumbing: who routes attention, which protocol connects agents to business software, and where the electricity comes from.
Search flipped its default
As of 10 July, Gemini 3.5 Flash is the default model behind Google Search globally, and the primary experience is an AI-written answer page with embedded source links rather than ranked blue links. Commentators are calling it the end of the 25-year content-ranking deal, and they're right: if the AI answer cites you, you exist; if it doesn't, your old rankings are irrelevant. This is the biggest change to how internet attention gets distributed since Google launched, and most site owners haven't processed it yet. Google
The candidate replacement for the visit-and-ad economy showed up four days later. Cloudflare opened the waitlist for its Monetization Gateway, built on HTTP 402 (the "Payment Required" status code that sat unused for three decades): an agent hits a site, reads a machine-readable price, pays programmatically, proceeds. Cloudflare fronts about a fifth of the web, so this is not a side experiment. Unrot
The standards war went public
Per The Information: Google, Microsoft, Salesforce, Snowflake and ServiceNow have agreed to back a shared rival standard for connecting AI agents to business software, a direct move against MCP, which has become the de facto default over 18 months. The twist is that all of them (plus OpenAI and Anthropic) simultaneously sit in a Linux Foundation group building open agent standards. Committee standards move slowly and MCP's install-base head start is real, but this much enterprise muscle on one side means anyone building agent integrations should hedge; watch who ships a spec versus who signed a press release. The Information
China's shutdown went live
The anthropomorphic-AI rules bit on 15 July. ByteDance shut Doubao's agent features for its 345M users (read-only until 15 October, then deletion); Alibaba killed Qwen's humanlike agents with no migration path, so user-built agents and their histories are permanently gone. Enterprise agents are untouched; Beijing is regulating parasocial AI, not work AI. For anyone building on someone else's agent layer, this is the live case study in platform risk. SCMP
The constraint moved to power
The compute crunch stopped being an insider anecdote this week: Google capped Meta's Gemini API access after Meta asked for more capacity than Google could supply, delaying Meta's internal projects. Compute is now the binding constraint even between trillion-dollar companies, and vertically integrated Google serves itself first. TSMC posted record Q2 revenue (about $39.6B, up 36% year on year) on AI demand. Blackstone, Apollo and KKR together put $5.34B into behind-the-meter data-centre power, a bet that sidestepping years-long grid queues is where the scarcity premium now lives. South Korea committed roughly $880B over a decade to chips, data centres and robotics. Power is the constraint now, and the biggest names in private capital just said so with money. Unrot · BuildFastWithAI
The labs got graded
The Future of Life Institute's 2026 AI Safety Index landed, and the best grade on the board is a C+ (Anthropic), with OpenAI and Google DeepMind at C, Meta at D+, and xAI, DeepSeek and Mistral effectively failing. The headline finding: several labs have quietly walked back earlier safety commitments. The same week, the EU announced pre-market model testing with ENISA for critical sectors. Independent scorecards and certification gates are converging on the same conclusion about self-regulation; governance is becoming a procurement column rather than a blog post. Future of Life Institute
Also shipped
Meta rolled out its Business Agent Platform globally (agents built on the billion-plus conversations flowing through WhatsApp, Messenger and Instagram) and started renting excess infrastructure as "Meta Compute", like a fourth hyperscaler. NVIDIA and ServiceNow launched Project Arc, a persistent desktop agent that learns a worker's routines over days and keeps data out of the frontier labs; if it stays reliable over weeks, the persistent agent becomes the default enterprise form factor. And Anthropic is reportedly preparing an S-1 for as early as October, on roughly $47B annualised revenue and reported profitability, currently the cleanest financial story in frontier AI. AI Business · Fortune
What to watch
Friday, 17 July: Gemini 3.5 Pro was expected (2M context, Deep Think, around $1.25 per million input, none of it confirmed by Google) on the same day Xi Jinping opened Shanghai's World AI Conference. A frontier launch and a superpower AI summit on one calendar square. Whether it landed, and at what price, is next week's story; if the leaked specs are real, enterprise model pricing resets for the second time in a month.