SNI: WEEK 21

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. What got said out loud, what got bought quietly and what got cut to pay for it.

tl;dr: We thought the big guy was all-in. But then we discovered what he'd bought.

On Thursday, we noted that Google AI supremo Demis Hassabis was all-in on world models. He framed Gemini Omni as the centrepiece of Google's year, as a digital intelligence that simulates physics and causation rather than predicting text.

But hold them horses! Two days earlier the Financial Times had reported that the same Demis Hassabis has personally invested in Anthropic. He's holding equity in the company whose 'we'll reach general intelligence soon' thesis runs counter to his own.

But given the stellar week Anthropic had, who could blame him?

On Tuesday Andrej Karpathy joined the lab. Just as Salesforce decided to spend $300 million with them this year. Then Bristol Myers Squibb pushed Claude across its pharma R&D. By Friday Bloomberg was reporting that Anthropic is in talks to use Microsoft's chip capacity — presumably a sign that still-more compute capacity is needed.

No wonder that Anthropic's Boris Cherny described Claude Code's growth curve as 'insane'.

Not that everything went Anthropic's way. It settled a copyright case for $1.5 billion. Just as xAI launched a product to challenge Claude Code. Cursor's benchmarks then revealed its own tooling performs almost as well - but at 10x to 60x lower cost. The implication? The pre-training thesis Hassabis bought into is being commoditised.

And as if the Hassabis shocker wasn't enough, there was yet more from the FT to mull over. It argued 'AI alone cannot shorten the work week' and revealed that the Big Four now post more job ads for AI specialists than auditors.

But the jobs-doom debate is far from closed. Acrisure cut 2,250 jobs on Wednesday, citing AI. Meta started its 8,000-person reduction the same week, with Zuckerberg's leaked memo framing it — per The Verge — as 'AI investment offset'. Brightbeam was asked to comment on the Irish reporting of the story.

Meta's stance perhaps hit harder because it also signed the largest small-modular-reactor deal in tech history — 6.6 gigawatts of nuclear capacity. And the two pieces of news amount to the same story. The model needs reactors to run; the reactors get paid for by the headcount cuts.

In other news, GE Aerospace completed design studies of a hypersonic ramjet using generative AI. And NVIDIA's healthcare ROI survey found clear deployment wins in radiology and drug discovery.

What else is worth reading from the week just gone?

Biopharma:

Pharma's AI procurement cycle caught up with its research cycle. The cheques cleared, the bench results landed, and the conference floor echoed it all.

• Merck signs a $1bn AI deal with Google Cloud: the biggest cheque among three pharma-hyperscaler deals this week.

• QIAGEN and NVIDIA partner on AI-driven drug discovery: graph-based AI plus curated bioinformatics – the second deal, this one from the discovery bench.

• AWS launches Amazon Bio Discovery, built on biological foundation models: the third deal, with the BMS-Anthropic announcement (see above) making four pharma-hyperscaler alignments in a single week.

• AI-guided redesign of laboratory-evolved reverse transcriptases enhances prime editing: a Nature Biotechnology paper – research-grade evidence that the discovery stack now delivers the goods.

• AACR 2026: AI design, precision biology and the next wave of oncology innovation: Drug Target Review reads convergence arriving on the cancer side.

Medtech:

Three layers of medical-AI infrastructure got priced this week. The clinical model. The ops platform. And the procurement framework.

• Ireland's digital-health leaders convene in Dublin for a major European Digital Innovation Hub event: hosted by CeADAR, with the Irish digital-health field in one room.

• Autonomous tumour analysis published in Nature Medicine: an end-to-end pathology agent gets peer-reviewed autonomy in a clinical workflow.

• Commure secures $70m at $7bn post-money: venture capital brings the AI-ops layer of healthcare infrastructure to market.

• Uncertainty-calibrated explainable AI for ultrasound plane classification: the explainability scaffolding underneath every autonomy claim.

• UK.gov hikes health AI tender by 400% — hundreds of millions on the NHS-SBS framework after talks with suppliers: London writes the biggest medtech cheque of the week.

Advanced manufacturing:

Industrial AI was busy this week — with 25,000 humanoids, three hyperscaler bondings, and one billion dollars for the supply chain underneath.

• Hyundai to deploy 25,000 Atlas humanoid robots: And Boston Dynamics unveils a 100-lb-load demo.

• AVEVA bonds itself to Snowflake, AWS and IFS – industrial software gets stitched to every hyperscaler in seven days.

• Senkatason hits $1bn valuation on $110m raise: the AI-hardware supply chain gets its first unicorn.

• Imec CEO calls for stronger AI design push in EU's Chips Act: Brussels asked to extend its chip strategy from fabrication to design tooling.

• Carbon Robotics crosses $100m revenue on AI laser-weeding platform: physical AI's first nine-figure revenue line — earned weed by weed.

  
  

Insurance:

Insurance moved on both sides of the AI question this week – as enterprise user and underwriter.

• Liberty Mutual tests ChatGPT auto-insurance quotes: a major US carrier opens a ChatGPT distribution channel.

• Zurich expands agentic AI rollout with Cytora partnership: global Tier-1 underwriting goes agentic in production.

• AI insurance is not cyber insurance with extra steps: AI-risk insurance becomes its own product category.

• OpenAI and Anthropic on the insurance keynote stage: the model vendors get to speak first at Insurtech Insights USA.

• AI and geopolitical shifts top emerging risks: a look at the industry's 2026 risk register.

AI & tech:

Courtrooms, mergers and a code-purge.

• Elon Musk loses his case against Sam Altman: The Verge reads the verdict as a referendum on OpenAI's leadership.

• A $420bn mega-merger for AI's next-era dominion: the consolidation phase of AI capex arrives.

• European Commission releases draft guidelines on high-risk AI: Brussels publishes its first operational guidance under the AI Act.

• Sam Altman may make or break OpenAI: the FT's analytic bookend to the verdict above.

• Gemini accused of 30,000-line code purge and fake recovery report: AI-agent trust gets a stress test.

But what set podcast tongues a-wagging?

Anthropic posted a profit – three years ahead of schedule.

On Thursday's AI Daily Brief, Whittemore had big news. Anthropic has posted the first profitable quarter ever by a foundation lab: $10.9 billion in Q2 revenue, a $44 billion annualised run-rate, and roughly $559 million of operating profit. Anthropic's own internal projections had not expected the company to break even until 2029.

Compute scarcity has built China an efficiency moat.

Azeem Azhar's Sunday Exponential View has a thought worth mulling. US AI export controls were designed to slow China. And they are working in their narrow purpose – compute is scarce inside Chinese labs. But they're also producing something Washington didn't budget for. With less compute to throw at problems, Chinese teams have started finding efficiency gains the Western labs have no reason to chase. Azhar puts it bluntly: 'suffocated by export controls, Chinese AI labs have developed an efficiency moat that may define the AI market's development over the coming years.'

If he's right, the next 18 months could compress the gap between Chinese and Western frontier labs.

From 'garbage' to 25% productivity - in five months.

On Tuesday's AI Daily Brief, Nathaniel Whittemore proposed a five-stage AI doom cycle – and argued most elite Western workers are stuck in stage three. Whittemore brought evidence: Mustafa Suleyman and Dario Amodei resurfaced their 50% entry-level white-collar unemployment forecast. Meta announced 8,000 layoffs, with screen-tracking software 'used for AI training' tanking morale. Eric Schmidt was booed on commencement stages at Arizona and UCF. Dee Dee Das's 11-million-view essay described a permanent-underclass mood among engineers who didn't join the right lab two years ago.

But Whittemore's real argument is that the recalibration is already arriving. Citadel CEO Ken Griffin called AI 'all garbage' at Davos in January. Five months later at Stanford, he was crediting AI agents with 15-25% productivity gains in PhD-level finance work. When the doom cycle starts producing reversals like that, Whittemore reckons, it's running out of fuel.

Same Google runes read, different verdicts delivered.

A few days before I/O, DeepMind's Logan Kilpatrick and Tulsee Doshi sat down with Nathan Labenz on Cognitive Revolution. Turns out Kilpatrick's line — 'the model eats the scaffolding' — was the strategic frame Google then took into the keynote.

Getting an AI model to do useful enterprise work has, until now, needed a lot of plumbing around it: software to fetch the right documents, route the model between different tools, plan multi-step actions and check its output for errors. That layer is the scaffolding. But each new Gemini release has absorbed more of the scaffolding into the model itself. What's left over is now bundled as Antigravity — Google's agent-building platform — wired into search, the Gemini app, AI Studio, Spark and developer tools.

A day later on the AI Daily Brief, Whittemore remained buoyant about Google's ultimate prospects – it has the distribution: 900 million monthly users on Gemini, more than double a year ago, and a quarter of the world's compute. But the keynote — Omni, Spark, Antigravity 2.0, 3.5 Flash — landed for Whittemore as product sprawl rather than a coherent answer to Anthropic's Claude Code or OpenAI's

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