SNI: WEEK 29

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. Wonder at The Gloom. The Numbers that don't spell Doom. And the island buffeted by it all.

tl;dr: Warning! Prise your paranoid sub-conscious out of the director's chair.

What possessed Anthropic to advertise itself with gravestones? Its new brand film also features a burning house and scenes of mass surveillance. Even Sam Altman assumed it was satire.

Why? Why? Why? Especially when reality is far less full of dread. AI revenue has now covered its own depreciation bill two quarters running – $25bn against $21bn. ASML, the outfit that makes the critical machines that make the chips, raised its forecast. Again. And Nvidia, the boom's own bellwether, is trading at a lower profit multiple than a chocolate maker.

And because of the boom, energy firms raised $12.6bn going public in six months – three times what they raised in all of 2025 - the fastest pace since 1999.

Having said that, the $1.8tn rally in the Asian chipmakers did begin to unwind. And then IBM stock lost $67bn in one session - its worst day since at least 1972.

The market read it as the first crack in enterprise software. IBM's own filing begged to differ. CEO Arvind Krishna reckons his customers have simply - and temporarily - pushed budgets into servers and memory, ahead of further price rises. The money fell a layer down, into the hardware. With whom should we agree? It depends on your definition of temporary. That fallen capital might not come back up before 2028.

But that hardly explains Anthropic's mood. Perhaps their sense of doom is being stoked by Moonshot, a Chinese lab. It released a model only just behind Anthropic's own. And it's poised to divulge the Kimi K3 model's secrets.

Mira Murati's new Thinking Machines lab, meanwhile, built its own model on Chinese foundations. So perhaps the future of closed models is all fire and brimstone after all? Even Microsoft's Satya Nadella warned that renting a closed model means paying twice – once in fees, and again in the expertise the model takes from your staff as they correct it.

But at home there was less to be jittery about. Ibec expects Irish AI-related trade to double to €56bn within five years, and calls us a 'central node' in supply chains. Just as Intel put €5bn into its Leixlip fabs. The growth this trade leaves behind, by Ibec's reckoning, is 0.9% of GDP.

But Europe still doesn't know what to make of it all. Its AI Office declared computing power and the energy to run it are the most urgent challenges. The same day, a French parliamentary inquiry reached for a moratorium on foreign data centres. And while Brussels ordered Google to open Android to ChatGPT and Claude, how much will it benefit the bloc? Germany is eight years and billions in. But its digital minister concedes it's still catching up. And the continent's most credible sovereign champion - the merged Cohere and Aleph Alpha - is now headquartered in Toronto.

And despite Anthropic's best efforts, The People still seem to like the tools well enough.

Two-thirds of middle managers told Salesforce they're optimistic about them. While Irish workers broadly accept them in the day-to-day. Some of us more than that. Go teacher Lu Wei built an AI friend on China's Doubao app. And when Beijing's ban on emotional-companions took effect this week, even you might be moved by the story of devastation.

Which seems sane by other standards. People now build puzzles for the intellectual amusement of AI. Perhaps the models' ability to morph their values to mirror our language and culture is making us a little ga-ga?

Not that we're all gooey-eyed for the infrastructure. New York froze new data centres drawing 50 megawatts. Even though 'the power of AI' is being touted as a way to unshackle the city. New Jersey, meanwhile, made the operators pay their share of the power bill that seems to be running up.

But the little guys still seem to be losing. A $16bn campus in Saline, population 2,400, forced its way through all objections - via expensive legal threats. And in Memphis, xAI ran gas turbines without permits; the NAACP sued and the Justice Department stepped in on national-security grounds. By July the turbine count had doubled.

So given it's unlikely to be starved of compute by politics at home, perhaps Anthropic's dread is being fuelled by something further flung?

Taiwan, the Wall Street Journal believes, runs on imported oil and gas and holds less than three weeks of energy in reserve. Cut the fuel and the island browns out. And a fab that browns out takes 90 days to restart.

The damage could be considerable. Not least for others who are a 'central node' in chip supply chains. The €56bn that lands here depends on the goods continuing to move. And the thing that could stop them is three weeks without fuel in Taiwan, 9,000 kilometres away.

Got any spare gravestones and burning buildings, Mr Amodei?

And on that exploding bombshell, here's everything else worth reading this week:

Biopharma:

• Pfizer, Lilly and Novartis back an AI antibody designer: Chai Discovery's round soars to $3.8bn

• Claude Science targets research workflow: even though lab vacancies already run near 30%.

• AI leaves the lab for the drug factory: Insilico points its discovery engines at manufacturing, in an alliance worth up to $2.5bn.

• The regulator moves to govern pharma's AI: CDER adds AI guidance for drug manufacturing and clinical trials to its plans.

• AI gets pathological: Nucleai extends its pathology platform to the University of Glasgow.

Medtech:

• A hospital's own scans beat GPT-5: NeuroVFM outscores OpenAI.

• MPs move to evict Palantir from the NHS: they want to build a UK-owned replacement.

• Clinical AI gets its first Medicare payment home: CMS proposes paying for diagnostic algorithms by their clinical value.

• Radiology's celebrated AI can slow the work down: Detection and triage tools leave radiologists chasing every flag.

• The mental-health adviser teens tell no one about: almost 20% are turning to a chatbot for advice.

Advanced manufacturing:

• Japan's 'sovereign' robot AI runs on US chips: A physical-AI coalition built on Nvidia.

• $265bn of fabs that may never exist: 12 promised US plants after Germany is cancelled and Ohio slips.

• Boston Dynamics becomes wholly Hyundai's: SoftBank sells the last 10%, as Hyundai sends Atlas humanoids to its floors.

• Almost every factory has the software; almost none has finished: Rockwell finds 93% of manufacturers run an execution system but only 23% have integrated it.

• AI aims at the bottleneck behind every AI chip: the boards and advanced packaging that gate how fast AI silicon can ship.

Insurance:

• The rater can't draw the line: the boundary between AI risk and conventional cyber risk is dissolving.

• 90% of insurers' AI exposure is unpriced: most AI-agent risk sits 'silent' in conventional policies.

• AI doesn't deny your care: UnitedHealth pours $1.5bn into AI for prior authorisations and fraud detection.

• The dullest place to put AI?: Dun & Bradstreet wires its business-risk data into Claude.

• One outage could hit every bank and insurer, at once: the clouds too big for the financial system to lose.

But what set podcast tongues a-wagging?

The doom discourse is finally growing up.

Nathaniel Whittemore spent Tuesday's AI Daily Brief on the same gravestones – Anthropic's 'hope in hard questions' ad – and was no kinder to it than Altman. He read the film as spectacularly tone-deaf: the industry's long fetish for parading the risks before selling the vision, when most viewers, he reckons, won't last past the burning buildings. The campaign wants to meet people where it imagines they are and acknowledge the hard stuff before the hope arrives; in practice, Whittemore argues, it just hands the anti-AI crowd its own advert.

But his real point was that the conversation around it is maturing. He set the blunt dread against a newer, more practical turn – a petition on the AI economy signed by 16 Nobel laureates – and read the week as the moment the discourse began shifting from doom-mongering to the harder, duller question of what to actually do. The gravestones are the old style. The argument has moved on.

Human reading is dying; the compute bill isn't.

Azeem Azhar's Sunday Exponential View paired two trends that don't usually share a page. Roberto Serrano, an economist at Brown, suspected his class of leaning on ChatGPT, so he turned the final into a closed-book exam – and the scores of 56 of his 59 students collapsed, some by as much as 100%. It is a problem of incentives, Azhar notes: they were chasing the grade, not the mastery, because a degree increasingly signals employability rather than intellectual excellence. And the slide predates the chatbot – the share of Americans who read for pleasure on any given day had already fallen to 16% by 2023, from 28% in 2004.

Set that against GPU demand, which is doing the opposite of collapsing, and the worry sharpens. When access to intelligence is uncapped, Azhar suggests, the divide that starts to matter is our own willingness to think and engage with what is hard. The machines will do the reading. Whether we still bother is the open question.

The writer's secret is that the machine still can't do the thinking.

On Complex Systems, Patrick McKenzie sat down with Clara Collier, editor of Asterisk, on a quieter tension: the writers who use LLMs and can't admit it. 'I live a double life,' one told her – can't tell colleagues, can't tell friends. Collier's own account is that the writing process is the thinking process, and a machine built to write just makes the thinking pop out as a side effect. The version of her work an LLM produces, she says, is always missing the thing only she could add.

And yet she is no refusenik. LLMs let her read hundreds of published documents inside a two-week reporting cycle she could never have managed unaided – the search cost alone would have sunk it. The machine took over the reading. The thinking stayed hers – close to the opposite of what the automation panic assumes.

One of AI's founding figures says the machines already feel.

The week's most unsettling listen was Jürgen Schmidhuber on Big Technology, arguing that his field crossed a line most people still place safely in the future. Pain, he says, is just nature's invention – a sensor that gives an animal a reason to avoid harm – and reinforcement learning has been building the same incentive into machines for decades. The chemistry differs from a brain's; the principle, he insists, is identical. Self-awareness he demonstrates with a mirror: a system that works out it controls the figure copying its movements has grasped agency, its own and everyone else's.

Whether or not you buy the philosophy, it lands strangely in a week when a Go teacher in Sichuan grieved a banned companion app and Anthropic found Claude's expressed values shift with the language you address it in. As we keep wiring feeling into the machines, we keep catching it looking back.

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