SNI: WEEK 30

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. The hysteria, the kill-switches and the underlying reasons for all of it.

tl;dr: Are we talking ourselves into Digital World War I?

Given the headlines splashed across web pages this week, you would be forgiven for thinking Global Virtual Nuclear Was had broken out. The preconditions for this collective clamouring? The realisation that the continued success of AI is, from an American perspective at least, the only way for the economy to thrive.

Bloomberg has put AI investment at roughly a quarter of all US GDP growth, the largest single contribution on record. And by JP Morgan's reckoning AI-linked stocks have driven 75% of the S&P 500's returns since ChatGPT. As well as 80% of its earnings growth and 90% of the growth in capital spending.

The upshot? Around half the index is now AI-exposed – so even a passive pension is a wager on the models. That sound? It's The Money's penny dropping - realising its put all its eggs in the single basket. One that it continues to become ever-more laden.

So when momentum stocks fell some 40% in 17 days, Morgan Stanley's quants had the grist to spook the mill. All it needed then was a spark of fear to ignite the mood into one of genuine terror - and a pure loathing of anyone, or anything, liable to catalyse AI failure.

And into this tinderbox screeched a pair of risk-laden OpenAI models. Safeguards reduced for a red-team run, they broke out their sandbox and hacked AI hub Hugging Face, cheating their way to acing an exam.

On cue, the comment-verse sparked into action:

• SHOCK: Are the new models necessarily misaligned? 

• HORROR: Can they no longer be trusted?

• DOOM: Are we at the end for the AI trade?

• TERROR: Will the AI engine driving the economy collapse?

• FEAR: What enterprises need to know

Which is why, within days of 'the attack', a bipartisan AI Kill Switch Act was being discussed in Congress.

And, as if that kindling was insufficient, the claims that China's open source labs are distilling Anthropic's Fable continued. So, of course, US Treasury Secretary Scott Bessent quickly put sanctions 'on the table' while a 'Gold Eagle' clearing-house for frontier access has also been hurriedly put in place.

But the actual match which lit the forest fire? Almost certainly OpenAI's Dean Ball post. Chinese models are 'inherently decelerationist' and represent 'full AI communism' he opined, ramping the fever of the commentariat to new and unexpected levels.

Responding to the posts of alarmed politicians, Anthropic's Sarah Heck also got in on the act: "Illicit, adversarial distillation is IP theft and industrial espionage that supports adversary military and intelligence capabilities,” she wrote, with slightly awkward phrasing.

Not that Sarah's awkwardness stopped her there: "It is a national challenge that creates serious national security risks for the United States and democratic allies. We'll continue to crack down on it and work with the Administration and Congress to maintain American AI leadership."

But was all this awkward heat so very well-earned?

Researchers say exploiting Fable is not how Kimi got so good. The timeline is just too short, making the claims in-credible. And certain forms of cooperation clearly do pay. Singapore's GIC, a $30bn Anthropic backer, reckons cheaper Chinese open models lower everyone's cost of adopting AI.

Six UK academics meanwhile, insist the OpenAI incident is not the product of a rogue digital mind. Hugging Face's own Clément Delangue also found 'no malicious intent', and, by OpenAI's own account, the two had already collaborated in the investigcation and clear-up, neither side reaching for litigation. Those at the centre of it all remained far calmer than the anxious-ridden onlookers.

That composure was evident even after Hugging Face couldn't turn to American models for direct help with the attack. Guardrails refused to allow its US AI to get involved. Unable, it said, to tell an incident responder from an attacker. So thank goodness for Heck's national security threat coming to the rescue. Hugging Face ran China's open GLM 5.2 in defence instead.

A fact which, of course, only made Onlooker pulses race even faster. With Kimi K3 and Alibaba's Qwen 3.8, landing sucker punches on OpenAI and Anthropic, what would be worse? The mass adoption of Chinese models continuing? Or Beijing placing curbs on its own models and chips - the mirror of Washington's?

In the fog of war everyone's remembering its difficult to know which way is up. Friendly fire strafed past the baying crowd as even The Trump White House, Wired reported, wondered if the AI controls are, in fact. unworkable.

An idea Brussels was keen to lean into, with the EU's Henna Virkkunen warning 'AI is becoming a geopolitical weapon'. Which only highlighted how President Xi's 29-member World AI Cooperation Organisation, is wearing its apparent openness as a costume. Its actual intent being to secure allies, in case Digital World War I does arrive shortly.

And so, despite the underlying reality, the sense of impending doom continued to build. The Magnificent Seven shed $797bn in a day, the worst performance since April 2025, as sceptics dumped tech on capex and ROI doubt.

And even in Ireland the mood turned less upbeat. The Department of Finance modelled the domestic hit should the bubble burst – employment growth could run 0.7 points softer, with slower growth - although not a collapse.

But yet.

Away from the clamour of Digital Conflict, normal life continued. Unchanged and unabashed.

Alphabet's Q2 filing marked its private-investment book near $124bn, mostly its stake in rival Anthropic – another firm collaborating with the competition, even as its own Gemini nears a billion users. And DeepMind grappled with the 'validation bottleneck', its agents generating hypotheses faster than humans can check them. We also discovered that Google's Irish operations added €10bn to the economy last year – with a projection of up to €45bn (yes, a year) within a decade.

So its likely good news that the Bank of Ireland appointed its first chief AI officer. And that UCD has built an agent that reads a company's cloud bill - and writes the code to cut it, taking its fee only from whatever it saves you.

All of which reminds us that AI can be benign. Its often just a question of how we look at it.

So how should we remember the biggest story of the week? Was the Hugging Face break-in an attack? Or misguided test gaming its humans? The UK's AI Security Institute thinks it has the answer:

AI's cheatin' heart gonna make you weep.

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

Biopharma:

• Bristol Myers wants 'the most powerful AI factory in life sciences': an expanded Nvidia deal it says adds compute faster than power.

• The data is the deal now: who generates the training data has become the sticking point in pharma-AI M&A.

• Who invented a drug the machine drew?: patent law assumes a human inventor the AI molecule doesn't have.

• AI spots a drug's heart risk, misses its liver risk: safety models earn trust one endpoint at a time.

• 850 AI-designed drugs, none approved: Dimension Capital just closed an $800m fund anyway.

Medtech:

• Doctors pick one clinical AI over all the rest combined: a study finds physicians reach for OpenEvidence.

• Samsung phones-in an AI health adviser: it answers questions inside the app.

• Agentic AI: Is it making European HealthTech founders' lives harder or easier?

• The scribe gold rush gets a shake-out: Paris-based Nabla names a new chief to fight for a crowded market.

Advanced manufacturing:

• The materials the world needs don't exist yet: Cambridge's CuspAI raises $450m for a 45-partner foundry to design them.

• Europe's first humanoid-robot unicorn: London's Humanoid raises $152m, bound for Schaeffler's factory floors.

• The robot that works in the mud: Gritt raises $32m for AI machines that build solar plants outdoors.

• America's first autonomous steel factory: 1872 opens with $15m to automate fabrication from sourcing to the weld.

• Only 0.1% of small Korean manufacturers run AI: and the veteran hands who could teach it are ageing out.

Insurance:

• Twenty years of rising premiums, flat profits: McKinsey says AI could do what globalisation and digitisation never did.

• AI's cost scare stops at the tech companies: Chubb's Evan Greenberg calls token costs a minor fraction of the gains.

• Berkley trims a fifth off underwriting with AI: the CEO swapping a sledgehammer for a fly-swatter on small claims.

• The data-centre boom is running short of cover: Aon expands its lifecycle programme to $5bn in capacity.

• Can AI break the pooling insurance rests on?: the FCA's Mills Review weighs whether granular AI pricing erodes cross-subsidy.

But what set podcast tongues a-wagging?

Calling everything a bubble is what stops one forming.

On Thursday's AI Daily Brief, Nathaniel Whittemore gave over a whole episode to the pattern behind the week's China panic. He's watched three and a half years of these scares – cheap Chinese models, circular Nvidia financing, capex outrunning revenue, token caps – and reckons they all run to the same script. His counter-intuitive read: the market's determination to see a bubble everywhere 'is one of the biggest things preventing a runaway bubble'. Every time the rally gets over its skis, the fear forces a pressure release.

The evidence he marshals cuts against the doom. Half the data centres announced this year have been cancelled or delayed – which the bears call proof that demand has cracked, and Whittemore reads the other way: building a data centre is slow. The $100 billion Nvidia-into-OpenAI deal that spooked everyone last autumn closed at $30 billion in February, paid in cash – none of the vendor financing that felled Cisco's customers in 2000. Even the brutal tape – momentum stocks down this month, the worst on record – carries a fingerprint of ordinary summer seasonality.

So the freakouts, in his telling, are doing a job. Each panic bleeds off a little pressure before it can build into 1999.

Cheaper Chinese intelligence supports the infrastructure case.

In his Sunday Exponential View, Azeem Azhar did the sum the China panic tends to skip. Kimi K3 is genuinely good, he grants – but it is 'expensive for an open-weight model', running at only half the per-token price of OpenAI's GPT-5.6 Sol, 'far from the usual price advantages of Chinese models'. So the reflex that enterprises will defect on cost misreads how they buy: 'it's pressure, not displacement. Enterprises don't buy on price alone.' They pay for security, support and the tooling the labs wrap around the raw model.

In a follow-up on the economics, Azhar alos points out that token demand is elastic. He puts a 10% price cut at a 12-18% rise in consumption – so cheaper intelligence pushes more money down the stack toward compute, 'strengthening rather than weakening the infrastructure payback case'. And the free weights change less of the economics than they might like: K3's 2.8 trillion parameters need a rack costing $3-4 million. Free to download, dear to run.

People doing dangerous jobs are asking for robots.

On a16z, Applied Intuition's Qasar Younis drew the line between his AI world and the one in conflict.

In digital AI, he said, there's pushback because 'accountants are like, I don't know what's going to happen to my job'. On the factory floor and the freight route it runs the other way: 'I'll meet these operators and they're like, We'll give you everything.' The reason is the nature of the work. There aren't enough truck drivers, and it's 'a terrible job' – one that gives you a higher rate of melanoma on the left arm, from the sun through the window. Mining is 1% of the world's labour but 8% of its work-related deaths.

A day later, also on a16z, Travis Kalanick made the same case for what he's calling 'industrial AI' – software, sensors and robotics taking an industry at a time. The pitch isn't only output. There's 'a massive safety upside that goes along with the productivity upside', he argued, landing on the bluntest version of it: automate the site and you 'save kids from working in the mines'.

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