SNI: WEEK 26

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. What Washington reached for, what slipped past it, and whether the money behind it all still adds up.

tl;dr: First they took the models. Then they came for the robots.

Last week, Washington forced Anthropic to pull its best models. This week it let it be known that regulation is in vogue.

Whilst it handed Mythos back to a number of US organisations, OpenAI was asked to restrict GPT-5.6 to some twenty approved partners – the first time Washington has gated a US model before it ships.

And humanoid robots are next in line.

Commerce Secretary Howard Lutnick told a closed-door room his department is studying Chinese state-subsidised imports – 'we don't want state-subsidised robotics attacking us in America,' he said, because 'robotic arms are coming.' An American brain on a Chinese body, one attendee fear-mongered, is a very bad strategic plan. Congress is in the headspace too. The Select Committee on China has pointed to an Amazon listing – any American can buy a Unitree humanoid on Amazon. Unitree is classified as a Chinese military company.

Others are being spooked by the machines already arriving. And Hyundai's workers want a say: the union voted to strike for a veto over the humanoid robots heading for its lines.

So even if Anthropic talked its way into the ban, there are going to be far wider consequences. Macron called it a moment that 'clarified the stakes,' as he pressed for a G7 response. And the US adoption of greater regulation seems to hand a much-needed advantage to Europe, given we're far further down the regulatory road. As we argued last summer, regulation is inevitable and we hold a stronger hand than we realise.

Not that Anthropic seemed chastised. The lab dropped Claude into Slack. And the impact took many by surprise. It now acts as a colleague that reads your channels, takes the busywork, files the follow-ups and will even solve your problems, without bidding, in code. Some might argue it is what Microsoft said Copilot was going to be.

Meanwhile, Anthropic picked up colleagues of a more human kind too. Nobel laureate John Jumper left DeepMind for Dario's crew. Noam Shazeer went too, and Alphabet stock had its worst day in more than a year as the departures mounted. Even an economist joined the party: Stanford's Chad Jones, the same Chad Jones who once worked out it might be rational to take a one-in-three chance of human extinction for a shot at lives fifty-five times richer. But Google's trouble runs deeper than its payroll: its hold on search is starting to slip.

And Anthropic went on the offensive, too. It accused Alibaba of the largest distillation attack it has logged – roughly 25,000 fake accounts and 28.8 million Claude queries between April and June – and wrote to the White House for tougher curbs on Chinese labs.

Is the centre of gravity shifting, very slightly? Mistral shipped a document-reading model it says beats the field. And France's AI-run health insurer Alan banked €480 million – its backers naming US export controls as why Europe needs its own frontier. Nearfield pulled in €380 million – the largest Dutch deep-tech round on record.

And Ireland is ahead as well. Did we mention that Brightbeam got written about in the Business Post? Like us, two in three international investors told EY they plan to expand there, comfortably ahead of the European average. And Salesforce paid $3.6 billion for Fin, the AI-agent platform built by Intercom's Irish founder Eoghan McCabe. The ESRI sounded one careful note: the boom is flattering the national accounts, but the gloss is multinational spending on chips and data centres, not money reaching households – whose own spending it expects to tighten.

So given the universe is good at balance, perhaps it is no surprise the money had a jittery week. SpaceX shed more than $600 billion in three days, and the Nasdaq slid as rate fears met a harder question: when do the hundreds of billions pouring into AI start to pay their way? Even Oracle's 21,000 layoffs were really about the build, the savings going straight into its AI spend.

So is it worth it? The chips, at least, said yes. Micron posted a fifteen-fold surge in profit, Qualcomm signed Meta as the first big-tech buyer of its data-centre chips, and OpenAI unveiled its first chip, built with Broadcom, and chip-maker Groq raised $650 million, as more of Nvidia's customers chose to build their own. Exponential View ran a proper count of the AI economy: $110 billion of revenue, growing fast – enough that the big cloud firms' income just about covers what their kit will cost as it ages. But on the other side of the balance sheet, by Citi's reckoning, at least 40% of America's investment-grade corporate debt now carries a revenue link to AI.

But what about actual people? Sixteen million of us now talk to an AI companion daily, with 'therapy and companionship' the most common thing we use it for. But should we? Financial advisers warned that a chatbot will brightly tell a Londoner to move to Monaco for the tax. And Radio Nova found out the hard way that AI-made adverts still look like, well, AI-made adverts.

So two questions to leave you with. 1. If no one wants a data centre in their back yard, do we put them in orbit? And 2. If a chatbot serves you nonsense, who exactly needs to receive the lawsuit?

Pop those into the back of your brain for a mull. And while you do that, here's everything else worth reading this week:

Biopharma:

• The biggest AI-drug cheque yet: Insilico and SK Biopharmaceuticals sign a neuroimmune-drug deal worth up to $2.5 billion.

• Nvidia brings its chips to the lab bench: a new BioNeMo science-reasoning suite aimed straight at drug researchers.

• AI goes hunting for new antibiotics: a model turns up a family of 'Prionin' antimicrobial peptides.

• Ireland's ICON puts Microsoft inside its trials: the Dublin contract-research giant picks Microsoft to run AI-enabled clinical development.

• China open-sources its answer to the West: Alibaba releases a unified scientific model to take on the Western biotech-AI giants.

Medtech:

• When the algorithm says no to your treatment: Medicare's WISeR prior-authorisation programme is delaying and denying care through AI errors.

• London opens a test track for NHS AI: the MHRA's new sandbox is built to speed safe medical AI into the health service.

• Britain's tech industry weighs in on the rules: techUK responds to the national commission on regulating AI in healthcare.

• Catching Alzheimer's from the way you speak: a new paper backs SpeechDx's AI for earlier dementia detection.

• AI moves to the billing desk: the technology takes on the hospital revenue cycle and the medical invoice.

Advanced manufacturing:

• AWS comes to the steel mill: ArcelorMittal teams with Amazon's cloud to automate its plants and supply low-carbon steel.

• The world model that gives a robot its bearings: Nvidia's Cosmos 3 builds the physical-AI brains behind the machines.

• Humanoids clock in for real work: Automate 2026 marks the move from robot demos to the production line.

• BMW's newest hire is a humanoid: the carmaker deploys Figure's new 03 robot for logistics work at its Spartanburg plant in the US.

• The robots arrive, and so does the liability: insurers find autonomous machines don't fit the policies they already write.

Insurance:

• AI dreams up the disasters that haven't happened yet: diffusion models build tens of thousands of synthetic catastrophe years for Swiss Re, Verisk and Moody's RMS.

• One engine for every pricing call: Earnix launches an AI orchestration system to run insurers' decisions.

• Your driving data sets your premium: China's Cheche prices electric-vehicle cover from how you actually drive.

• AI underwriting, slotted into the systems insurers already run: Convr embeds its models across Guidewire, Duck Creek and Sapiens.

• Why insurers don't actually want autonomous AI: Digital Insurance reads the industry's reluctance to hand the decisions to a machine.

But what set podcast tongues a-wagging?

A peek inside AI-native firms.

On his Sunday Exponential View, Azeem Azhar flagged a Harvard–INSEAD paper that gives insight into how AI-native firms operate.

He discussed Koning and Kim's findings that AI-native start-ups run leaner than their rivals at the same size and funding – more engineers, fewer managers. But the gains only show up when the product itself is rebuilt around the AI, not bolted on the side.

The uncomfortable 'so what' for a regulated enterprise is that the gains everyone wants come from the hard task: rebuilding how the work is done around the machine.

Handing everyone a Copilot licence is the easy version, and it buys very little.

The gap between the demo and the day job.

On Big Technology, Mike Krieger – Instagram's co-founder, now running Anthropic Labs – was candid about what building on the frontier is like.

With Fable, the model since pulled, he queued a full night's work before bed and woke to an entire codebase written, planned and double-checked.

But then he admitted that models are erratic and despite app launches soaring the apps people actually use have flatlined.

Set his candour beside the economy-wide revenue figure – $110 billion – and the buildout barely earning back what it costs as it ages. The distance between what AI can do and what it has so far banked looks like the year's central tension.

Proving you're a real person is becoming a problem.

On the a16z show, Sonal Chokshi sat down with the team behind World – the iris-scanning project now on Time's cover – on the 'proof of human' problem: as AI makes fake images, video and voice near-free to produce, the internet needs a way to show a counterparty is a real, unique person.

The threat isn't crude spam. It's that a single well-funded actor can now conjure a convincing crowd – thousands of distinct-looking accounts that secretly act as one – to tilt a vote, a market or a pile of reviews. The defence is to give every real person exactly one provable identity: make the first near-free, so anyone can get it, and a second near-impossible, so no one can amass an army of them.

World's answer – an iris scan that proves you're unique without revealing who you are – is pitched as the most consequential audience segmentation in history. The part to watch isn't the maths; it's who holds the biometric, and on what terms.

The data centre is where the boom meets the neighbours.

On the AI Daily Brief, Nathaniel Whittemore reframed the data-centre backlash as the first place the AI boom meets ordinary streets.

Most of the popular fears – the water, the power bills – are, he argued, overstated; but waving them away as 'insane drivel', as some in the Valley have, only widens the trust gap.

His sharper point: the buildout shifts value from workers to capital at the scale of a town, and the people who get something back are the ones who organise for it – the unions and local negotiators winning teacher bonuses, ratepayer protections, Oregon's POWER Act.

The neighbours get a say, in other words, when they ask for one.

Thank you for reading Second Nature Intelligence. Come back next week for all the AI news you need to know.