AI in Flow

In today’s episode of AI in Flow, Claire and Peter unpack a fast-moving mix of AI signals: a key White House AI adviser steps down as Washington debates what public “ownership” of major AI firms could look like; new data shows AI is now the most-cited driver of US layoffs; and Starbucks shelves an AI inventory system after accuracy issues hit the shop floor. They also cover Google Drive’s on-device AI document scanning, Anthropic expanding cybersecurity support in India, rising community backlash to data-centre water and land use, and reports that Apple may accelerate the shift away from Intel Macs with deeper on-device AI. Plus: two small stories that show how AI is quietly reshaping professional services and frontline healthcare.

Claire and Peter break down a day of “industrial-scale” AI news on AI in Flow: Google’s massive compute deal with SpaceX to bridge soaring Gemini demand, a sharp pullback in AI chip stocks, and Anthropic’s warning that the bigger risk may be AI systems increasingly building AI (with Claude reportedly writing the majority of code merged into its own codebase). They also cover new U.S. national security directives to accelerate AI adoption while limiting vendor dependency, reports of the public potentially taking stakes in AI companies, a Meta Instagram support-bot exploit highlighting agentic security risks, and the growing push to cap runaway tool spend as enterprises confront token-heavy AI agents. Plus: Anthropic’s chip-financing mega-package, Marvell’s S&P 500 move, and McDonald’s AI drive-thru test showing what AI looks like in messy, real-world operations.

Claire and Peter break down a packed day in AI: Google reportedly trims Cloud security roles—including parts of Mandiant and its Threat Intelligence Group—as resources shift toward AI, raising questions for organizations that depend on those feeds. Anthropic warns about the risks of self-improving frontier systems and argues for a credible ability to slow or pause development—while reports say the NSA is already using a restricted Claude Mythos model for cyber operations, sharpening the stakes for enterprise threat models and patching discipline. The hosts also cover Meta’s push into health-oriented AI (and why it won’t open-source Muse Spark over bio-risk concerns), a $1B US–Japan “Genesis Mission” to build AI-driven autonomous labs, Airbnb’s new agentic AI lab for action-based experiences, a Utah data-center project scaled back after water and governance backlash, and a market wobble after Broadcom’s outlook—plus the growing debate over who captures AI’s economic upside.

Claire and Peter break down the non-glamorous forces now shaping AI’s next phase: Alphabet’s massive fundraising push for AI infrastructure, Google’s call for stricter data-centre water standards, and a UN warning on the sector’s growing water and energy footprint. They also look at Amazon’s layoffs alongside huge AI capex plans and emerging state-level restrictions on data centres, Meta’s new business AI agents across WhatsApp, Messenger, and Instagram, and Sam Altman’s warning that enterprise token spend is spiralling without governance. Plus: investor pressure for AI ROI, bubble talk, and Google’s move toward on-device AI with Edge Gallery and local Gemma models—hinting at a more hybrid, cost-aware future for AI in Flow.

Claire and Peter run through a packed set of headlines shaping how AI is built, bought, and governed. In the US, a new executive order proposes a voluntary 30‑day national security review window for frontier models—signalling potentially slower, more staged releases and new procurement questions. Uber’s move to cap employee spending on AI coding tools highlights the shift from open-ended experimentation to usage controls and ROI measurement. On the security front, Android’s fake-call detection points to device-level verification as voice cloning rises, while Anthropic expands access to its Mythos cybersecurity model—boosting vulnerability discovery and raising the stakes for patching and disclosure workflows. They also cover ChatGPT’s reported 1B monthly users alongside growing competition, Microsoft’s new in-house MAI model family and its implications for cost and vendor risk, a cautionary but notable quantum update with Majorana 2, Google Search’s push toward conversational answers with ads embedded in responses, and Workday’s governed approach to building agents on HR and finance data. The throughline: governance, verification, and value discipline are becoming the default operating model for AI.

Claire and Peter break down a packed day in AI: Anthropic’s reported confidential IPO filing and the trillion-dollar valuation chatter—and what going public could mean for transparency, pricing, and vendor risk. They cover Florida’s lawsuit against OpenAI and Sam Altman over alleged safety failures affecting minors, and the knock-on implications for enterprise AI governance, contracts, and incident response. Plus, a cautionary security story as hackers reportedly exploited an AI support bot tied to Instagram account recovery, highlighting how AI agents with privileged actions become a fresh attack surface. Rounding out the briefing: China-linked efforts to access Nvidia H200 compute via leasing and brokers, expanded Chinese trade secret rules covering data and algorithms, Qualtrics’ push from CX insight to automated action, insurance AI consolidation, supply chain concentration around chips, the unglamorous importance of data integration, and why aggressive AI adopters may still be hiring.

Claire and Peter break down the shift from “subsidised” AI to hard ROI math as enterprises rein in token-hungry agents with budgeting, routing, caching, and cost-per-outcome metrics. They cover SoftBank’s proposed €75B, 5GW AI data-centre push in France and what it could mean for European infrastructure, sovereignty, and energy constraints. Plus: AI-assisted shopping’s surging conversion rates in the UK, Alibaba’s UEFA partnership and the governance questions it raises, the NBA’s exploration of real-time AI support for referees, South Korea’s full-stack AI strategy, and why legacy tech names are benefiting as the AI build-out spreads across the entire stack. The episode closes with a reality-check on AI valuations and a practical case study of an AI-native solo law firm built in 48 hours—automation with accountability.

Claire and Peter break down the day’s most important AI shifts across the stack—from Huawei’s “Tau Scaling Law” and the realities of advanced chip packaging, to Google’s proposed AI data-centre hub in India and the rising scrutiny over water, power, and community impact. They also cover Wipro and ServiceNow’s push toward agentic automation inside core business workflows, the operational risks highlighted by OpenAI’s outage, and how India’s global capability centres are moving AI into finance, HR, R&D, and regulated work. Plus: JD.com’s stance on retraining rather than layoffs amid automation and labour protections, promising (but simulated) carbon-aware scheduling to cut data-centre energy use, and the privacy and lock-in questions raised as networking and AI video surveillance converge into unified cloud platforms. The throughline: AI is becoming more embedded in physical and operational systems—and governance, resilience, sustainability, and workforce planning are now the differentiators.