Daily AI Brief

Meta will migrate major ranking and recommendation workloads for its social apps to Arm-based CPUs, aiming for better energy efficiency and performance per watt at scale. The shift highlights a broader diversification of AI compute beyond GPUs for specific tasks, even as Meta continues to invest heavily in GPU capacity. The announcement landed alongside broader infrastructure moves.

Meta will invest US$1.5bn in a new Texas facility designed for AI workloads, targeting a 1GW campus with renewable energy matching and liquid cooling; operations are slated to begin in 2028. The spend adds to the hyperscalers' ballooning AI infrastructure budgets, which remain a key market narrative into year-end.

UK-based Nscale said it will deliver around 200,000 Nvidia GPUs to Microsoft via data centres in Texas and Portugal, with additional capacity from a JV campus in Narvik, Norway. The Financial Times has reported the contract could be worth up to US$14bn based on comparable deals, and Dell is a partner on execution. It's another sign that compute constraints—not demand—remain the gating factor for AI adoption.

Anthropic overhauled its smallest model, Haiku 4.5, claiming performance on many tasks comparable to its mid-tier Sonnet at roughly one-third the cost—and 1/15th the price of its flagship Opus. The company says enterprise usage now represents ~80% of revenue and that smaller, cheaper models are unlocking broader deployment across large organisations.

The IMF told the FT that the surge in AI-related capital spending has helped shield the US from a sharper slowdown this year. The view will intensify focus on whether expected 2026–27 returns on AI capex materialise—and how power constraints and supply chains could be bottlenecks.

A 25-country Pew survey finds a median 53% trust the EU to regulate AI effectively, versus 37% for the US and 27% for China. Younger adults are relatively more trusting of China (and, to a lesser extent, the US) than older groups.

A new certification mark aims to identify books produced without generative AI—an early sign of industry-level labelling efforts to reassure authors and readers amid rising synthetic content.

The Bank of England updated its public materials outlining how it engages with AI-related innovation and risks across the financial system—useful context as UK regulators weigh competition, conduct and operational resilience questions around AI.

A collaboration between Google Research and partners introduces a Gemma-based model suite for single-cell analysis, including an efficient encoder and improved clustering/performance on benchmark datasets—another sign of foundation models diffusing into wet-lab workflows.

Google unveiled "Coral NPU," a hardware–software stack for on-device AI, targeting low-latency, secure inference at the edge with an updated toolchain and deployment story for developers.

Liberate secures funding from Sequoia and a16z for AI agents focused on sales and operations workflows.

Positioning against 'agentic ERP' incumbents, Campfire raises significant funding for AI-powered back-office automation.

Woz secures funding to help operations teams automate data wrangling and analysis tasks.

Index Ventures backs Planyear to bring AI-powered retirement and benefits planning to SMEs.

Aragorn AI raises seed funding to bring "desktop-grade" agentic automation to mobile devices.

Interactive Brokers launches a natural-language portfolio analysis tool inside Client Portal/Advisor Portal/Desktop. Early scope includes allocations, performance and activity queries, with more data domains promised.

Google enables new creative capabilities in the Gemini API—including Veo 3.1 for higher-fidelity video generation and improved controls for developers.

A sober look at where the AI trade could wobble (returns on capex, power constraints, circular deal risks) as Q3 earnings hit. Useful context for CIOs and FP&A leads budgeting 2026.

"Small models really help because they can be a more economical way of deploying into that."

— Mike Krieger, Chief Product Officer, Anthropic

New model iteration aimed at higher-quality, controllable video generation inside the Gemini API, with developer-facing updates for assets, prompts and safety controls. If you're piloting synthetic video for ads, explainers or product demos, this is one to test.