AI in Flow

Claire and Peter cover a fast-moving day in AI on AI in Flow. The US signals tougher lines on AI coalitions, warning partners they may not be able to join both US and China-led initiatives. Nvidia, SoftBank-backed SB Energy and OpenAI discuss vast financing and credit support for an Ohio data centre and power campus, highlighting how capital intensive AI infrastructure is becoming. Google begins notifying users of a phased Google Assistant shutdown across several platforms from 4 September 2026 as it pivots to Gemini. In Europe, the EU AI Act kicks in, and watermarking and provenance efforts accelerate, with Anthropic adding model-level watermarking for Claude and Google leaning on SynthID and C2PA metadata. The episode also unpacks an unsettling governance story from Andon Labs, where an AI manager dismissed a human worker, raising questions about policy memory, human prompting, and accountability. Plus, quick hits on SpaceX closing its Cursor acquisition, a university dropping AI detector scores, new enterprise partnerships, and fresh AI-enabled hardware and workplace tools.

Claire and Peter unpack today’s key shifts shaping AI in practice. Google is making visible watermarking user-controlled in Gemini and Flow, while keeping invisible SynthID and C2PA metadata, a change that pushes enterprises to enforce disclosure in workflows rather than relying on default labels. Apple is reportedly training a custom LLM for China with Alibaba, signalling how compliance and go-to-market strategy can drive local model variants. In India, Prime Minister Modi announces free online exam coaching and a plan to train one crore young people in AI skills, with major implications for talent pipelines and delivery logistics. Plus, AstraZeneca details its evidence-linked internal Research Assistant for R&D, and new research shows how model reliability can collapse when many constraints are stacked at once, reinforcing the need for verifiers and stepwise task design. Briefly, power costs for AI data centres, GPU performance for agentic workflows, Meta’s open-weight Glimmer, Gemini in Classroom, and a report on OpenAI shifting monitoring tactics.

In this episode of AI in Flow, hosts Claire and Peter break down the latest enterprise AI signals across speed, simplification, skills, capital, and supply. OpenAI previews its Ultrafast API tier for GPT-5.6 Sol, promising major latency and throughput gains via Cerebras. Microsoft merges its consumer and business Copilot apps while retiring several features, prompting organisations to audit any dependent workflows. IBM partners with OpenAI and plans to certify tens of thousands of consultants, pointing to growing demand for deployment and governance capability. Databricks closes a 5 billion dollar round at a 190 billion dollar valuation, highlighting the cost intensity of competing in AI platforms. SMIC reports strong AI-driven demand and wafer price rises, signalling continued supply pressure beyond just CPUs and GPUs. Plus quick hits on OpenAI leadership changes, multi-agent behaviour risks, Apple’s US manufacturing push, an AI-led US anti-transshipment proposal, and Suno Studio 2.0.

In this episode of AI in Flow, Claire and Peter unpack Alphabet’s leadership overhaul at Google DeepMind, with Koray Kavukcuoglu taking the helm as Google pushes to accelerate Gemini amid performance and infrastructure pressures. They also discuss Twitch’s plan to train Amazon’s AI models on streamers’ content by default unless creators opt out, and what that shift means for consent, contracts, and creator relationships. Plus, SpaceXAI launches Grok Bot, an always-on agent that can sign into workplace tools to complete multi-step tasks, raising both productivity upside and serious identity and permission risks. Rounding out the briefing are Thrive Holdings’ $2bn raise, India’s BHASHINI and NITI Aayog partnership on multilingual, voice-first public services, and quick hits across valuations, wearables, medical imaging, watermarking, and training efficiency research.

In this episode of AI in Flow, Claire and Peter unpack Google’s claim that the Gemini app has reached 1 billion monthly active users, with usage skewing heavily towards voice and image generation. They then dig into fresh signals from CoreWeave, including a surging backlog and commentary that near term capacity is effectively sold out, plus what that implies for pricing power, inference demand, and the risks of GPU asset lifecycles and data centre constraints. Next, they cover SAFE, a new incident reporting framework backed by more than 120 companies to standardise how organisations report security incidents involving AI agents, including clear triggers, timelines, and logging expectations. They also discuss River AI’s rapid 1.1 billion dollar raise for personal agents, and OpenAI’s early testing of ads in ChatGPT, before a quick round up of executive moves, watermarking, security tooling, AI persona labels, and server demand forecasts.

Claire and Peter break down today’s biggest moves in AI on AI in Flow. Meta launches Glimmer, an open-weight model positioned as lower-compute and agent-ready, alongside a broader push for openness and US competitiveness, plus major infrastructure and community spending commitments. Anthropic rolls out machine-readable marking across supported Claude outputs, including invisible text watermarks and C2PA signed provenance metadata, aiming to meet EU AI Act Article 50 while setting a global default. OpenAI introduces GPT-5.6-Cyber via Daybreak Red for authorised vulnerability research and security testing, emphasising vetted, governed access through approved partners. Also covered, new agentic experiences in Google Ads and Analytics, L&T Technology Services’ AgenticIQ platform for enterprise engineering and manufacturing workflows, and quick hits from Ford’s in-app vehicle assistant to China’s new limits on AI companion services.

In this episode of AI in Flow, Claire and Peter cover the growing physical and governance footprint of AI. Amazon reportedly backs a gas-burning power plant in Pecos County, Texas to supply a planned data centre, highlighting the tension between reliable compute and emissions transparency. Nvidia is assembling a dedicated AI safety and security engineering team focused on pre-deployment evaluation of autonomous agents and AI-driven vulnerability detection and patching, alongside its broader push for open-weight models. The pair also discuss Stanford researchers using generative models to design entire viral genomes, a breakthrough with potential against antibiotic resistance that also raises urgent biosafety and biosecurity concerns. Plus, India’s SEBI says its AI surveillance platform has flagged over 20,000 instances of fraudulent social media content, and OpenAI acquires presentation startup NextSlide as consolidation around core AI products continues.

Claire and Peter cover OpenAI’s safety disclosure on its upcoming model Astra, after early tests suggested autonomous cybersecurity capabilities that could approach a Critical threshold, prompting tighter security, isolated testing, and expanded monitoring. They also look at Sarvam AI’s plan for a one-trillion parameter model and its developer platform for Indian languages, Cloudflare’s Kitesurf browser built for AI agents, Rippling’s AI Spend Console for tracking employee AI tool costs, and reports that data-centre operator Switch has confidentially filed for a US IPO. Plus a quick roundup on the booming training-data market, Airbnb’s AI search pilot, and enterprise AI adoption updates.