SNI: WEEK 10

Welcome to all the AI news that matters this week. The wins, the fails and the somewhere in-betweens. Across biopharma, medtech, complex manufacturing and insurance.

tl;dr: AI moves inside the firewall

Location is the thing this week. Specifically, the location where AI models run, whose infrastructure they sit on and who controls the data they touch.

From biotech licensing on-prem research platforms to insurers discovering that most deployments now involve AI, enterprises are pulling AI workloads behind their own firewalls.

A theme also neatly developed at our own event featuring Apple Silicon this week.

Why is it in vogue? Because companies that control inference, build proprietary data loops that compound in value over time. And we can also see growing tensions with the alternatives. A tension that is now starkly visible in Pentagon procurement, as well as pharma R&D and insurance underwriting.

AI & Tech

• OpenAI released GPT-5.4 in Standard, Thinking and Pro variants, intensifying model-layer competition weeks before anticipated IPOs

• Anthropic resumed Pentagon negotiations over its collapsed $200m defence contract, as CEO Dario Amodei framed the breakdown in political terms

• Nvidia CEO Jensen Huang said his company's investments in OpenAI and Anthropic are likely its last in both, citing anticipated IPOs

• Seven hyperscalers signed the White House Ratepayer Protection Pledge, committing to fund their own power generation for AI data centres

• Broadcom reported record Q1 revenue of $19.3bn, up 29% year on year, reinforcing that custom silicon demand from hyperscalers shows no sign of slowing

Biopharma

• KALA BIO licensed an on-premises AI research platform called Researgency, explicitly designed to keep biotech data behind the firewall

• Insilico Medicine and Liquid AI partnered on a 2.6B-parameter drug discovery model small enough to run on local infrastructure

• Tempus expanded its multi-year collaboration with Merck to accelerate precision-medicine biomarker discovery in oncology

Medtech

• Sectra acquired Oxipit, which holds the first CE Class IIB certification for autonomous AI in chest X-ray analysis

• RadNet agreed to acquire Gleamer for €230m, adding CE-marked and FDA-cleared radiology AI devices to its imaging network

• Perimeter Medical Imaging's 'Claire' became the first FDA-approved AI-enabled imaging device for intraoperative breast cancer surgery

Manufacturing

• Bob McGrew, OpenAI's former chief research officer, is reportedly raising $70m for a manufacturing-focused AI startup

• Cisco's 2026 industrial AI report found cybersecurity is the top barrier to expanding AI in manufacturing, with unreliable networks disrupting operations

  
  

Insurance

• Generative and agentic AI accounted for 68% of insurance AI deployments in Q4 2025, up sharply from prior quarters

• Bank of America estimated $15bn in insurance commissions are at risk from AI-driven disintermediation, targeting brokers and intermediaries

• Hong Kong's four financial regulators jointly launched a GenAI Sandbox++ to accelerate supervised AI adoption across banking, securities and insurance

And if you're still hungry for more, here's the detail on each:

AI & Tech

This week crystallised around two questions: who controls the infrastructure layer and who gets to sell to governments. Both reflect the broader theme of AI moving inside the firewall – whether that firewall belongs to a hyperscaler, a nation state or the Pentagon's procurement office.

Broadcom posts record $19.3bn quarter on custom AI chip surge: The chipmaker reported fiscal Q1 2026 revenue of $19.3bn, up 29% year on year, and approved a $10bn share buyback. The results were propelled by surging demand for custom AI accelerators – application-specific chips designed to the specifications of individual hyperscalers rather than sold as general-purpose GPUs. CEO Hock Tan told analysts that Broadcom will soon deploy 'multiple gigawatts' worth of custom accelerators at major cloud customers, and said AI companies 'can't make their own silicon any time soon' despite widespread industry discussion of in-house chip efforts.

OpenAI releases GPT-5.4 with Pro and Thinking variants: Describing it as 'our most capable and efficient frontier model for professional work'. The model ships in three configurationsand comes weeks before OpenAI's anticipated IPO and positions the company to demonstrate continued model-layer momentum to public-market investors. With Anthropic also expected to go public later this year, and Nvidia CEO Jensen Huang saying this week that his company's investments in both OpenAI and Anthropic are likely its last, the model-provider layer is entering a new phase of capitalisation. The shift from private to public capital will change the accountability structure - model providers will face quarterly earnings scrutiny for the first time.

Anthropic resumes Pentagon negotiations after $200m contract collapse: Its $200m contract with the Department of Defense collapsed after the two parties failed to agree on the degree to which the military could obtain unrestricted access to Anthropic's AI. The Pentagon subsequently signed with OpenAI instead. CEO Dario Amodei then resumed discussions with the Pentagon, reopening negotiations that multiple observers had assumed were dead.

Amodei framed the original breakdown in political terms, telling reporters the government relationship collapsed because 'we haven't donated to' or 'given dictator-style praise to Trump'. Five unresolved questions remain, according to CNBC's analysis: the precise guardrail terms Anthropic demanded, whether OpenAI's deal included the same restrictions, whether the Pentagon's position reflects technical requirements or political preferences, what role national security considerations played and whether Anthropic can secure any federal contract without a political relationship. The episode illustrates that for frontier AI companies, the firewall is now as much political as technical – and the cost of being on the wrong side is measured in hundreds of millions.

Hyperscalers sign White House pledge to fund their own power generation: Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed the White House Ratepayer Protection Pledge on 5 March, committing to pay for the new generation capacity required to power their AI data centres rather than passing costs to residential electricity customers. The pledge follows growing public and political concern that data centre energy demand – now forecast to exceed tens of gigawatts within the decade – could raise household electricity bills.

The pledge is voluntary and contains no enforcement mechanism, but it carries political weight. President Trump had told AI companies earlier in the week that infrastructure expansion had 'angered Americans' and that they needed to address public concern.

Biopharma

The week's most consequential biopharma developments share a common architectural choice: moving AI inference and research data on-premises, behind organisational firewalls, rather than routing them through third-party cloud infrastructure. The pattern signals that pharma and biotech companies view data sovereignty as a prerequisite for competitive AI deployment.

KALA BIO licenses on-premises AI platform for biotech R&D: KALA has entered into a platform development and exclusive licence agreement for Researgency, a proprietary AI research platform designed to run entirely on-prem. The platform is purpose-built for biotech companies that need AI capabilities and will not send proprietary molecular, clinical or preclinical data to external cloud providers.

The strategic logic is explicit. Biotech firms handle some of the most commercially sensitive data in any industry – compound structures, assay results, clinical trial data – and the risk of inadvertent exposure through cloud APIs has become a material concern. KALA BIO's bet is that the market for behind-the-firewall AI infrastructure is large enough to justify a dedicated platform business.

Insilico Medicine and Liquid AI partner on lightweight drug discovery models: Insilico Medicine and Liquid AI hope to deliver what they describe as lightweight scientific foundation models for drug discovery. The key specification: a single 2.6B-parameter AI model that the companies say achieves performance comparable to far larger models across drug discovery tasks. At 2.6B parameters, the model is small enough to run on local hardware rather than requiring hyperscaler-scale compute – a deliberate design choice that aligns with the broader movement toward on-premises inference.

The partnership pairs Insilico's drug discovery pipeline with Liquid AI's model architecture expertise. For pharma companies evaluating AI adoption, the availability of a compact, high-performing model that runs locally may lower the barrier to deployment while maintaining data control.

Tempus expands multi-year Merck collaboration on AI biomarkers: Aimed at accelerating the discovery and development of precision-medicine biomarkers, the deal has an initial focus on oncology. Tempus brings one of the largest clinico-genomic datasets in the industry – structured clinical records linked to genomic sequencing data – which Merck will use to identify and validate biomarkers that predict treatment response.

MedTech and digital health

This week, three acquisitions and a first-of-its-kind FDA approval reshaped the competitive structure of AI in diagnostic imaging. The common thread: established medtech platforms are buying their way into AI capabilities and regulatory clearances rather than building them from scratch – pulling AI inside their existing commercial firewalls.

Sectra acquires Oxipit, holder of first CE Class IIB autonomous radiology AI: The Swedish medical imaging IT company is buying Oxipit, a Lithuanian company that holds the first CE Class IIB certification for autonomous AI in chest X-ray analysis. The certification is significant: Class IIB is the highest risk class for diagnostic software under the EU Medical Device Regulation, and autonomous operation means the AI can read and report on chest X-rays without a radiologist reviewing each case.

For Sectra, which operates one of Europe's largest picture archiving and communication platforms, the acquisition adds an AI layer that can autonomously process normal chest X-rays – freeing radiologists to concentrate on complex cases. The regulatory clearance is the moat. Obtaining CE Class IIB certification requires extensive clinical validation data and a quality management system that meets the EU MDR's highest standards.

RadNet acquires Gleamer for €230m: The largest US outpatient radiology provider, is buying French radiology AI company Gleamer for €230m. Gleamer's products carry both CE marking and FDA clearance across multiple radiology applications. The deal gives RadNet – which already owns the DeepHealth AI subsidiary and recently secured a CE mark for its TechLive remote scanning platform – a second AI asset with independent regulatory clearances and a European commercial footprint.

RadNet's strategy is now visible in outline: vertically integrate AI across its 400+ imaging centres, use proprietary clinical data to refine models, and sell the resulting AI capabilities to third-party radiology providers. Each acquisition adds cleared devices, training data and commercial reach. The €230m price tag reflects the premium attached to regulatory-cleared radiology AI with demonstrated clinical performance.

Perimeter Medical Imaging's 'Claire' wins first FDA approval for intraoperative breast cancer AI: Claire became the first FDA-approved AI-enabled imaging device for use during breast cancer surgery. The system has demonstrated a statistically significant reduction in patients with residual cancer post-surgery – addressing the re-excision problem that affects roughly one in four breast-conserving surgery patients.

The approval creates a first-mover regulatory advantage. Competitors must now demonstrate at least comparable clinical outcomes to achieve their own clearance, using Perimeter's published data as the benchmark. For hospitals evaluating intraoperative imaging, Claire is the only FDA-approved option – a position that carries both commercial and reputational weight during the product's period of exclusivity.

Advanced manufacturing

Manufacturing AI this week produced two stories with a shared implication: the talent and capital that built frontier AI models are now turning toward the factory floor, and the primary obstacle is not the models themselves but the infrastructure required to connect them to physical operations.

Ex-OpenAI research chief reportedly raising $70m for manufacturing AI: Bob McGrew, OpenAI's former chief research officer, is reportedly raising $70m for a startup focused on bringing AI to manufacturing. Details of the company's specific approach, product and target manufacturing segments are not yet known.

The signal is in the person and the capital. McGrew has led research at one of the two leading frontier AI labs; his decision to focus on manufacturing rather than another software or consumer category reflects a conviction that the sector's data density, physical complexity and relative AI immaturity represent a large and defensible opportunity. A $70m raise at this stage – before a product is publicly described – suggests investor confidence is tied to the founder's track record and network rather than demonstrated product-market fit. The move may also accelerate talent migration from frontier model labs into industrial AI, a trend visible since RLWRLD's $26m raise last week.

Cisco finds cybersecurity is the top barrier to manufacturing AI adoption: Cisco's 2026 industrial AI report, published this week, found that while AI has already improved productivity, quality and resilience in manufacturing, unreliable networks and cybersecurity concerns are the primary barriers to further expansion. The finding reframes the manufacturing AI adoption challenge: the constraint is not model capability but the underlying network and security infrastructure that connects AI to production systems.

For manufacturers evaluating AI deployment, the implication is that investment in network reliability and cybersecurity must precede or accompany AI rollout – not follow it. The report positions Cisco's own networking and security products as the prerequisite infrastructure layer, but the finding is consistent with broader industry evidence. Factories that move AI workloads on-prem – behind their own firewalls – face a compound challenge: they must secure both the AI models and the operational technology networks those models interact with.

Insurance

Insurance produced the week's clearest quantitative evidence that AI adoption is accelerating – and that the resulting competitive reordering is now being priced by equity analysts, not just discussed in conference presentations.

Generative and agentic AI account for 68% of insurance deployments: According to research firm Evident, generative and agentic AI together accounted for 68% of all AI deployments in the insurance industry during Q4 2025. The figure represents a sharp increase from prior quarters and signals that insurers have moved past pilot-stage experimentation into production deployment at scale.

The composition of the 68% matters. Agentic AI – systems that take actions autonomously rather than merely generating text – is now a meaningful share of deployments. Carriers using agentic AI for underwriting triage, claims routing and policy servicing are building operational data that feeds back into model improvement. Those still evaluating pilot programmes face a compounding disadvantage: every quarter of delayed deployment is a quarter of operational data their competitors have and they do not.

Bank of America estimates $15bn in insurance commissions at risk from AI: Bank of America Global Research published a report estimating that more than $15bn in insurance industry commissions are at risk from AI-driven disintermediation. The analysis targets brokers and intermediaries – the distribution layer that sits between carriers and policyholders – arguing that AI agents capable of quoting, comparing and binding policies reduce the economic rationale for human intermediation in personal lines and small commercial.

The report landed weeks after broker stocks suffered their worst single-day decline since 2008 following Insurify's ChatGPT-powered insurance shopping tool. The $15bn figure gives institutional investors a specific number to model against broker revenue – and names the major listed brokers - AJG, AON, Marsh McLennan, WTW - as companies whose commission streams face structural pressure. The counter-argument, articulated last week by Clear Group CEO Mike Edgeley, is that complex commercial lines require relationship-driven advisory that AI cannot replicate. The question is where the boundary between automatable and non-automatable intermediation actually falls – and BofA's estimate suggests it falls further into broker territory than the industry has assumed.

Hong Kong regulators launch GenAI Sandbox++ across financial services: The island's four principal financial regulators launched a supervised environment for financial institutions, including insurers, to test generative AI applications under regulatory observation before full production deployment.

The joint launch across all four is unusual. Most regulatory sandboxes operate within a single agency's remit. The cross-regulator design reflects an understanding that generative AI applications in financial services – particularly those handling underwriting, claims and customer interaction – span regulatory boundaries. For international insurers operating in Hong Kong, the sandbox provides a structured path to deploy AI workloads locally, behind a regulatory firewall, with supervisory clarity that reduces compliance risk. The initiative positions Hong Kong as a testing ground for AI governance frameworks that other Asian financial centres may follow.

Thank you for reading this week's report. Come back next week for all the AI news you need to know in your sector.