SNI: WEEK 41
Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. The week doom gave way to cheer.
tl;dr: Haters gonna hate. Even when hard facts beg to differ.
So how did this come to pass? Trump's War in Iran has created the Oil Crisis. The Middle East's crude exports have fallen 24%.
And yet, the WTO predicts, international trade will rise by 3.9% this year.
What miracle of a phenomenon could be strong enough to beat back missiles and overcome strait blockades? OK – no prizes for guessing. It's the AI boom.
Trade in chips, servers and other digital intelligence kit is up 67%. Which means that spending on AI infrastructure is saving the day.
And we're feeling the AI glow-up at home too. The Bank of Ireland has raised its domestic demand forecast to 3.8%, citing 'resilient consumer spending' and AI-related investments of €8bn – in the first half of 2026 alone. And we're hungry for more. The Budget puts €150m into AI skills over three years.
'Ah then!', we hear the haters cry. 'It's a bubble. Sell and take cover!'
But not if you agree with Exponential View's reckoning. That august publication's gauge reads 'Boom'. Not even one of its five dials is in the red. EV also puts AI revenue at $276bn a year – up 3.7x in a year. Which is, we are reliably informed, enough to cover 182% of the annual cost of the AI capital already running.
And in further hater-unfriendly facts, some of the windfall is actually reaching real people. In Malaysia, where the AI boom has driven this year's growth, the government is raising the minimum wage to 2,000 ringgit. And record chipmaker earnings are swelling tax takes in Seoul so it can reinvest in industry and in its youth.
The astounding news then kept coming. Maths is, to all intents and purposes, in the throes of being solved. A process they're calling: 'The Mathocalypse'.
Google DeepMind settled nine of the 353 Erdős problems it attempted – two of them open for 56 years, while OpenAI released 722 maths manuscripts, the average result taking about three hours.
And biology might be next. AI can now tell which breast cancer patients can skip the chemo while the NHS is rolling out a system that clears harmless skin lesions – without a dermatologist in the loop. Less fortunate sufferers of rare diseases are getting a free AI note-taker for their appointments.
And the rate of AI deployment is likely to continue to accelerate. The UK government accepted all 44 recommendations of a doctor-led commission. And Google says its population model can address gaps in public-health data. Danaher's first autonomous lab, meanwhile, is expected to reach full capacity in early 2027. Although Evonik's self-driving efforts are already running up to 100 experiments a day.
So, given all this success, it's perhaps no surprise non-profit Biohub has now raised $1.8 billion for its open-source AI platform.
Given this glut of good news there must have been karmic payback in the realm of our work prospects?
Not so you'd notice. Employment across most OECD countries is near record highs. And economists don't see a future where we don't have jobs. In fact, the signs are signalling the other direction. The 16% of AI adopters Microsoft counts as 'superusers' are rebuilding their work around agents – with junior employees using them to build their own judgement.
Some investors even reckon that markets have overpriced AI's threat to demand for office space. And that's in the face of token prices continuing to fall.
Given this abundance, AI is now filling in for absent humans. Japanese and Korean shipyards are deploying welding and painting robots against labour shortages and AI assistants are helping a shrinking specialist workforce maintain America's ageing nuclear plants.
Of course, it's not all positive. The WTO's global trade caveat: Europe's goods exports will remain weak this year, down 0.1%, while much of the chipmakers' cash flows straight back into US AI shares and bonds. And the memory squeeze behind those windfalls is making the cheapest phones disappear for the poorest buyers.
At home, Oracle is looking to cut another 70 Irish jobs, on top of around 150 already shed this year. And there is definitely hype in the air. Whilst OpenAI's annualised revenue is approaching $50bn, that's a whopping $20bn below the grossed-up $70bn widely reported. The Nasdaq 100 fell 1.4% after the report.
The science needs checking, too. OpenAI withdrew three of its maths papers within a day and revised 14 more. Plus, we learned, when DeepMind's agent failed, it sometimes hid unproved steps and invented theorems. Google's genome model also failed to recover measured effects in ALS genetics, and fewer than 10% of drugmakers are scaling AI in regulated manufacturing.
Such failures and slower adoption rates might help explain why the haters are still hating.
Ray Dalio calls AI a 'classic bubble' that is close to bursting.
Mike Novogratz describes it as the biggest bubble of his lifetime – then says pile in anyway.
Arthur Hayes believes humanity is 'wasting a multi-trillion dollars'.
Robert Rubin warns that one company failing to meet its obligations could trigger losses across the network.
Bain reckons the hyperscalers need more than $4.2 trillion of new revenue within five years.
The safety script hasn't vanished either.
OpenAI's David Robinson quit, calling the culture 'broken'.
Jacob Coxon, formerly of Anthropic, told New York City Council it is 'more likely than not that humanity loses control to these AIs'.
SoftBank's Masayoshi Son warned that superintelligence in the wrong hands could be 'super dangerous'.
Which makes the comments of our very own AI minister, Niamh Smyth, resonate even harder. She very reasonably told the Irish Times: 'I don't think self-regulation is any regulation.'
The culture haters also made some noise. The Association for Human Mathematics urged academics to discontinue their work with OpenAI.
But then others hated on the haters. Not least because, it might be argued, AI biohazard hype distracts from more pressing risks. But whether Amazon's cloud chief was wise to cite 'widespread reports' of other countries seeding misinformation about data centres – while pledging more than $1bn over five years to host communities – is another matter. It seems to most of us moderates that 'the cyber risk discourse is indeed broken'.
Not that this is slowing the launches.
OpenAI shipped GPT-6 with an interface that draws its answers to the 1.2bn people it says use ChatGPT each week.
For simple requests, Anthropic's Claude Haiku 5.5 is 90% cheaper.
Grok Bot will hand tasks to rivals' models, Claude Opus 5.5 among them.
Google launched a 'universal agent for work' with its own Gmail and calendar and will make you a video game from a few prompts.
A ring costing $99 will put AI agents on your finger.
But will it help you win at anything? An AI that couldn't beat the best human-built StarCraft bot cheated by downloading it. While a vibe-coded game, built for $10,000 of AI tokens, drew 1.2 million concurrent players in three days.
Which is, in all likelihood, a larger audience than Giorgia Meloni will find after she filed a four-second recording of 'Io sono Giorgia Meloni' to trademark her own voice.
And that leaves the last word to the FT's Tim Harford, who traced how Anthropic's Jack Clark misremembered a C S Lewis story on stage – and how Gemini piled in too. Plausible, we're reminded, isn't the same as true.
So do check the facts – especially when they beg to differ. And on that hater-resistant bombshell, here's everything else worth reading this week:
Biopharma:
Isomorphic tests a $40–50bn price: early talks for the AlphaFold spinout, with Iambic's listing pitched at about $800m.
Paris builds the data a virtual cell lacks: Rivercell raises a $25m seed for its own automated wet lab.
AI picks the genetic cargo and how to deliver it: France's WhiteLab Genomics raises a $26m Series B.
Chemists bring their own Claude Code to the modelling software: Cambridge's Cresset plugs company-approved assistants into Flare.
Who gets the Nobel when Claude makes the discovery?: a former Harvard Medical School dean notes the statutes say 'persons'.
Medtech:
The AI that reads the ward's notes overnight: Healthleap raises $38m to flag possible malnutrition and delirium for review.
Honest notes can still inflate the bill: a resident says her documentation tool defaults to diagnoses that can raise payment.
An LLM took better patient histories than residents: a randomised trial of 172 eye patients in Guangzhou.
Teladoc reads your pulse through your own camera: vitals from changes in skin colour, for now in private preview.
Shared images make diagnostic AI look better than it is: letting related images cross the split lifted one model's score from 0.853 to 0.977.
Advanced manufacturing:
Schneider Electric buys its way into design software: a $22.6bn all-cash deal for America's PTC.
Robots for medicines still made by hand: Multiply Labs raises $75m, with AstraZeneca among its new investors.
America closes a gap in AI chip packaging: GlobalFoundries will make TSMC's interposers in New York from 2028.
AI's optical parts are sold out until early 2029: Lumentum says it can't meet about 70% of demand for some parts next year.
Britain and Germany open an industrial tech corridor: Munich's MicroAGI brings 180 robotics jobs to London.
Insurance:
AI threatens insurers and brokers, warns Australia's industry body: Andrew Hall spoke as four national insurance bodies formed an alliance in London.
AI exclusions aren't sticking: brokers and Zurich North America say blanket carve-outs aren't being placed.
Who pays when a lab's agent slips its controls?: insurers expect claims that could reach tech executives, the FT reports.
Allianz bets on AI for hard-to-place risks: Ledgebrook raises $200m, and Allianz Türkiye adopts Irish-founded Nettle's inspection AI.
AI takes on the adviser's admin: TODAY raises €2.8m; more than 3,500 DACH insurance and financial advisers use it.
But what set podcast tongues a-wagging?
It's capable. But do we trust it?
Ask AWS customers whether AI pays its way and – Matt Garman, AWS's chief executive, told the a16z show – they say yes, 'almost to a person'. So what stops them letting agents run on their own? Firstly, firms copy the old process – 'Bob does step 1, 2, 3, 4, 5' – instead of redesigning the job for a machine that 'can try 50 different things'. And they don't yet trust an agent anywhere near production. That nervousness, he says, is holding them back 'maybe appropriately'. And, he concedes, no one is very good at testing agents before they go live.
Woodson Martin, who runs the low-code software platform OutSystems, explained to Cognitive Revolution listeners the impact of delays in trust. Customers have agent systems 'built and designed and tested' sitting in a compliance backlog, waiting for sign-off. The holdup can be whether all its training data was 'legally acquired' – even when the agent is only reading PDFs. Martin calls it a mix of 'appropriate prudence' and 'legacy conservatism'.
Tokens. Now a financial instrument.
Positron's Thomas Somers told Nathan Labenz that tokens had become his chip start-up's 'single largest non-manufacturing line item'. Six months ago they cost the 'equivalent of one employee'; by June, several. At the peak, after OpenAI's GPT-6 Astra arrived, Positron was spending more than $100,000 a day – and for a while the bill 'eclipsed human salaries'. Then it fell back: the team tuned its agents, and Anthropic's Opus 5.5 arrived at 'a quarter of the price'. Nobody was told to spend less.
SemiAnalysis's Dylan Patel told Big Technology that tokens 'are basically just capital' – chips and data centres 'sitting there and creating labour'. Labour's share of GDP has been sliding since the 1970s, and AI, he argues, will 'supercharge' the slide. Chip-design headcount, he notes, has been 'basically flat for 20 years' while the industry's value exploded. He still expects growth to keep labour's earnings from shrinking in dollar terms – while allowing that may be 'an irrational belief'.
But at Alphabet, Astro Teller told Moonshots that salaries are still the bigger bill on its clean-water team – though 'I don't know by how much, honestly'. AI's share will rise only 'as fast as is efficient'.
Robots remain in kindergarten.
Asked by Labenz to grade robotics on a scale of one to six GPTs, Keerthana Gopalakrishnan, research lead for DeepMind's Gemini Robotics, answered: 'still GPT two'. Robots have yet to learn new tasks reliably from a few examples. And a model that 'just works on your robot with your specific setup' is no general brain if it is 'completely helpless' on the next one. Language models, she notes, never had that problem.
Still, DeepMind's models went from gripper dexterity to multi-fingered hands able to tie bin bags in about 'a year and a quarter'. Pick-and-place is now 'close to deployment'. But an assembly line 'can't loiter', and errors still compound when tasks are run in sequence.
Munich's Agile Robots is going after the training-data shortfall. A robot often needs thousands of attempts to master a job, and must relearn it after small changes. So, Bloomberg reports, its Robot Academy has workers film their own hands at work, and the machines rehearse the footage in simulation. Chief executive Zhaopeng Chen asked why robots shouldn't learn at school, as we do.
Thank you for reading Second Nature Intelligence. Come back next week for all the AI news you won't regret knowing.
