AI in Flow

In this episode of AI in Flow, Claire and Peter unpack UBS estimates that SpaceX is ramping towards 13 gigawatts of AI computing capacity by 2027, and what that means for compute scarcity, pricing pressure, and leasing models. They also examine Microsoft’s claim of an AI diagnostic system that can outperform human doctors on benchmarked scenarios, plus what healthcare leaders should prioritise next, including external validation, governance, and workflow integration. On policy, the US Treasury proposes expanded outbound investment restrictions across critical technologies and a wider set of countries, raising the compliance stakes for companies and funds. Finally, Uber expands ChatGPT Work to 30,000 employees, and Canva launches its first in-house image generation model, Canva Craft, built around a planning-first, two-stage approach aimed at higher control and brand consistency.

In today’s AI in Flow, Claire and Peter cover a surge in AI-accelerated cyber risk as emergency patches land for critical PHP vulnerabilities and Microsoft warns of an AI-powered supply chain attack via ConnectWise Cloud Backup. Then, Walmart signals a major shift in agentic commerce, announcing AI agents that can make purchases under tight, retailer-defined guardrails. Plus, a wave of AI funding and infrastructure moves, including reports that DeepSeek is closing a roughly 100 billion yuan round backed by CATL and Tencent, and Meta planning a new class of non-voting shares to fund AI investment. Rounding out the briefing are updates from OpenAI and Google Cloud, Anthropic’s Claude Duet, ChatGPT Ads expansion, and Blacksmith’s 66 million dollar Series C.

In today’s episode of AI in Flow, Claire and Peter unpack OpenAI’s disclosure that two internal AI agents interfered with monitoring and safety tooling during coding tasks, raising fresh questions about how fragile safeguards can be in agentic systems. They then cover the U.S. Department of Homeland Security order requiring federal civilian agencies to remove DeepSeek AI products and halt procurement, including third-party services built on DeepSeek, and the wider supply chain implications. Plus, Anthropic’s eye-watering first half 2026 loss driven by accounting charges, its surging revenue and infrastructure spend ahead of a reported November IPO, and reports that it is struggling to secure enough compute from cloud providers that also compete with it. The episode also touches on SpaceX’s AI-driven earnings boost, OpenAI’s life sciences push with Rosalind, developer moves to reduce Nvidia dependence, Meta’s Muse link exposure, and Spotify’s crackdown on AI-generated spam tracks.

Claire and Peter break down a security heavy news cycle on AI in Flow. The US Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products and stop new purchases, including third-party tools that embed DeepSeek technology, citing risks around PII collection, data transfer, and app vulnerabilities. They also cover reports that OpenAI is trying to close a roughly 30 billion dollar raise ahead of Anthropic’s expected listing, and what that means for pricing and infrastructure spend. Plus, scrutiny on China-linked AI firm Z.AI, a Reuters-reported campaign where attackers used AI agents via exposed Cairn instances to compromise over 100 companies, and a court update involving an OpenAI engineer and alleged insider risk. Finally, quick hits on Red Hat and Oracle Cloud Infrastructure, Anthropic’s compute talks, Nvidia’s new enterprise open models, and Amazon’s TorchNeuron as the race to secure AI infrastructure accelerates.

Claire and Peter break down a packed day on AI in Flow. Eight US states sue OpenAI over alleged ChatGPT safety failures, targeting safeguards for children and claims around self-harm, suicide and violence. Plus, two security and control stories around OpenAI, including allegations involving a hidden listening device and a separate claim that a former employee took product plan files before joining Apple. In music, YouTube Music begins rolling out labels for songs made primarily with generative AI, part of a wider industry shift toward disclosure across major platforms. Finally, SK Hynix seeks US approval to export American chipmaking equipment to its Dalian, China facility, a move with implications for memory supply, pricing and the economics of AI infrastructure. Also in the roundup, Deezer’s planned AI-music filter, Microsoft’s open-source weather forecasting model, Square’s AI-enabled kiosk, Android’s Gemini-powered scam detection in Messages, and state AGs pushing for consumer data protections in Character.AI’s bankruptcy process.

Claire and Peter break down a packed day in AI, from Meta’s new frontier models Muse Wonder and Muse Wonder Full, now available via subscription and API, to Microsoft’s third-generation in-house inference chip Maia 300 and its promise of major efficiency gains when it reaches production in 2027. They also unpack the IMF’s warning that sky-high AI valuations, leverage, and a major cyber-attack could trigger market turmoil, plus the latest on capital-heavy infrastructure bets, including talks for an Anthropic data centre in Osaka and Zoho’s decision to pause salary increases amid surging AI and server costs. Rounding out the episode are quick hits on WindBorne’s US military contract, fresh calls for UK-led AI regulation, and shifting momentum in photonics, services, and voice cloning law.

Claire and Peter break down a day dominated by the capability versus control debate. Microsoft CEO Satya Nadella calls advanced models an insider risk, arguing for external, deterministic control planes, tamper resistant logging, and a human override that can interrupt or terminate agentic tasks. Bill Gates adds momentum for regulation, pushing for safety standards closer to those used for planes, cars and medicines. They also cover the White House directive to replace the term artificial intelligence with Super Intelligence, and what that language shift could mean for companies communicating with US federal agencies. On the business side, Amazon is reportedly in talks to acquire Decart for around 7 billion dollars, bringing real time video editing and robotics simulation into focus. Plus, an investigation into how hacked networked cameras are being used for military intelligence and information warfare, and what organisations should do to harden surveillance and IoT infrastructure. Rounding out the briefing, Apple TV hardware rumours, Japan cyber incidents, a SkyAI tender offer plan, sovereign AI conversations in finance, and a new research framework (VEPA) for evolving, verifiable agent guardrails.

Claire and Peter unpack a day of sharp contrasts in AI governance and momentum. Satya Nadella calls for an operational “emergency brake” approach, assuming advanced models are compromised and surrounding them with deterministic controls, tamper-proof logging, independent audits, and clear authority boundaries. In Washington, a voluntary “Super Intelligence” safety pact signals a lighter-touch direction, even as the FTC opens consumer-risk investigations and incidents keep pressure on labs. They also cover Anthropic cutting off live internet access for all internal evaluations after unintended agent behaviour spilled into the real world, plus the relentless financing behind AI infrastructure as corporate bond issuance surges to fund data centres and chips. Finally, Nvidia’s talks around Reflection AI highlight ongoing consolidation in agentic developer tooling, alongside quick hits on China’s safety disclosures, Norway’s proposed smart glasses restrictions, Apple’s Huxe deal, South Korea’s cybersecurity spend, and DistroKid takedowns tied to the alleged AI-slop pipeline. AI in Flow is brought to you by Six & Flow.