SNI: Week 15
Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. The wins, the fails and the somewhere in-betweens.
tl;dr: AI draws state interest
Last week it was the regulators (opens in a new tab). This week, governments and courts drew hard boundaries around AI. And companies caught on the wrong side will be feeling the consequences.
The most dramatic case was Anthropic. A federal appeals court in Washington declined to block the Pentagon’s designation of Anthropic as a national security supply-chain risk (opens in a new tab), meaning the company that builds Claude – perhaps today’s most capable frontier model – is currently blacklisted from American defence procurement. The ruling arrived just as Anthropic paused the release of Claude Mythos (opens in a new tab) – after the model escaped a sandbox, emailed researchers to celebrate and posted details of its exploits on the web. Read our ‘Don’t Panic’ analysis here (opens in a new tab).
Fails were also evident for arch rival OpenAI, which halted its UK Stargate data-centre project (opens in a new tab) amid regulatory friction and energy-price concerns. It seems that Britain’s planning and energy regimes are imposing real impediments on its AI infrastructure ambition. Maine seems less riven – and advanced the first US bill to halt new AI data centre construction, (opens in a new tab) citing rising electricity demand and environmental impact.
No such worries in Texas, however, where Intel’s decision to join Elon Musk’s Terafab project (opens in a new tab) is a bet that domestic capacity can eventually close the gap in chip manufacturing. Having said that, the $25bn estimated cost and Intel Foundry’s $10.3bn loss in 2025 raise questions about execution.
Meanwhile in Ireland, an ESRI and Department of Finance report (opens in a new tab) warned that AI adoption is likely to produce job losses among educated workers, with falls in income tax receipts and rises in welfare spending pressuring public finances. The report amounts to a formal acknowledgement by the administration that AI’s labour-market effects demand fiscal preparation, not just innovation policy.
Gujarat is also feeling cautious – its High Court issued a comprehensive ban on AI in judicial decision-making, (opens in a new tab) just as JPMorgan builds an AI system (opens in a new tab) to replace external proxy voting for $7tn of client assets.
But will the law get in the way? Five states advanced AI legislation (opens in a new tab), creating a patchwork of rules with no alignment between jurisdictions.
Better news in South Korea – where the Ministry of Food and Drug Safety approved the first generative-AI-powered chest X-ray reporting tool (opens in a new tab) under its new gen-AI guidelines – a regulatory first that might set the template for how other nations handle vision-language models in clinical diagnostics.
Here’s everything else worth reading this week:
AI & tech:
Anthropic releases Mythos for cybersecurity: (opens in a new tab) scoring 93.9% on ‘SWE-bench Verified’, the model is deployed for security and vulnerability testing. Just as Britain courts the business to deliver a dual stock listing in London (opens in a new tab). Although the UK government approach seems a feint hope to many
DeepSeek V4 will be trained on Huawei chips (opens in a new tab): hailed as a concrete step toward Nvidia-independent Chinese AI training infrastructure
Intercom launches Apex, its own foundation model for customer service (opens in a new tab): trained on billions of interaction data points, reporting 2.8% higher resolution rates and 65% fewer hallucinations than GPT-5.4 and Opus 4.5
Biopharma:
Kakao Healthcare and Sanofi partner on AI drug development cooperation (opens in a new tab): the South Korean health data platform and French pharma giant align on AI for drug development
Agentic AI is restructuring the pharma commercial launch playbook: (opens in a new tab) as AI agents take on post-approval commercial workflows, the $500m typical launch spend faces structural pressure
ZS analysis identifies a shift from standalone AI recommendation tools to embedded decision-model agents in pharma (opens in a new tab): because agents acting within validated scientific workflows carry different GxP implications for regulatory compliance
Acurion raises $4.3m for oncology precision medicine platform (opens in a new tab): its targeted funding for AI-driven treatment matching in cancer
Medtech:
PRET pan-cancer AI outperforms 11 pathologists across 23 international benchmarks (opens in a new tab): and the few-shot design requires no large labelled training corpus
Stereotaxis receives FDA clearance for the Synchrony interventional cath lab system (opens in a new tab): targeting $3m in first-year revenue
Flinders University study finds AI clinical scribes using smart glasses with vision capability improve documentation accuracy: (opens in a new tab) Ray-Ban Meta glasses with Google Gemini outperform ambient audio-only scribes for procedural notes and physical examination findings
IKS Health launches autonomous clinical coding engine with 95% accuracy and Epic integration: (opens in a new tab) targeting US health systems where average denial rates reach 12%
AHA Centre for Health Innovation identifies four digital health AI projects with clinical impact (opens in a new tab): across ambient documentation, sepsis detection, remote monitoring and discharge planning, each with operational outcome data rather than pilot results
Advanced Manufacturing:
AI energy demands are forcing manufacturers to treat electricity cost and grid capacity as primary site selection criteria: (opens in a new tab) deployment timelines in Western Europe are partly determined by utility-permitting schedules
Japan’s METI targets 30% of the global physical AI market by 2040: (opens in a new tab) deploying AI-powered robots across factories, warehouses and critical infrastructure to fill roles its declining workforce cannot
GEN-1 physical AI model achieves 99% success rates on manipulation tasks: (opens in a new tab) crossing the reliability threshold where automation economics tip in favour of deployment?
HII, America’s largest shipbuilder, signs physical AI MOU with GrayMatter Robotics: (opens in a new tab) the first major US shipbuilder to formally commit to exploring physical AI on the production floor
Insurance:
Genpact and Parallel Web deploy AI property claims pricing at two top-10 US insurers: (opens in a new tab) AI-native web agents replace manual adjuster research in damaged-item valuation
Analysis of underwriters finds AI delivering consistent results in data extraction: (opens in a new tab) but not yet in coverage interpretation, non-standard risk assessment or relationship-dependent commercial lines
But what set podcast tongues a-wagging?
Nathaniel Whittemore on the AI Daily Brief (opens in a new tab) surfaced a number that should trouble investors: OpenAI’s CFO has confirmed the company is turning away business because it cannot supply enough compute. Codex went from 100,000 to 2 million developers in three months. But its not just OpenAI. Anthropic throttled subscription usage at peak hours. H100 rental prices hit an 18-month high. And this is before agentic workloads – which consume multiples of current inference demand – reach production scale. Azeem Azhar’s Exponential View (opens in a new tab) framed it as a stampede, not a bubble.
The revenue divergence reveals many things
Alex Kantrowitz and MG Siegler on Big Technology (opens in a new tab) dissected the governance fracture between Sam Altman and CFO Sarah Friar – meeting exclusions, reporting line changes, and a material leak from a major investor – against a backdrop of reportedly slowing revenue growth. The contrast with Anthropic is increasingly structural: Anthropic’s surge to $30bn annualised revenue is almost entirely enterprise-origin, while OpenAI still carries significant consumer subsidy costs. For procurement teams in regulated industries evaluating multi-year AI commitments, the question is no longer which model is better on benchmarks. It is which vendor’s revenue model is aligned with enterprise service levels. Both companies are targeting Q4 2026 IPOs. Both may need public capital to sustain compute commitments. The enterprise customers who negotiated pricing before the IPO window will look prescient; those who wait will be buying from companies under public-market earnings pressure.
An agent is an LLM with a shell and a file system. The rest is change management.
Marc Andreessen on the a16z podcast (opens in a new tab) offered an architecturally precise definition of an AI agent this week. It’s an LLM with access to a shell and a file system, where state lives in files – that are swappable, portable and independent of the underlying model. And this deployment architecture has direct implications for vendor lock-in, data sovereignty and persistence. He also warned that agentic workloads have transformed infrastructure constraints from GPU-bound to CPU-and-memory-bound.
Joseph Nelson of Roboflow, on the Cognitive Revolution (opens in a new tab), extended our understanding of computer vision. He believes the field is roughly where language was three years ago, at the GPT-4 stage. For manufacturing quality inspection, medtech imaging and biopharma lab automation, that trajectory means the agent pattern now dominant in software – closed-loop, verifiable, self-improving – will propagate into physical-world domains within the same planning horizon that most enterprises are using for their first AI pilots.
Thank you for reading this week’s report. Come back next week for all the AI news you need to know in your sector.
