SNI: WEEK 34

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. Who gets the power, who gets a vote and who gets a castle in Waterford.

tl;dr: Votes, supervotes, and one company to rule them all?

Last week everything was sunshine, lollipops and rainbows. But Zuck's superintelligence for everyone hasn't quelled concerns about the concentration of power. And voters have something to say about it.

So is it safer in a few hands, or many? This week investor Gavin Baker said trusted sources had told him that Dario Amodei believed Anthropic might one day be the only private company left in the world. Whilst Anthropic researcher Sholto Douglas firmly denied the claim, together with reports that Anthropic is preparing supervoting shares for Dario and co, it didn't help the decentralisation case.

Nevertheless, power keeps gathering with gusto. Stripe bought OpenRouter for $8bn, acquiring the switchboard between hundreds of competing models. Nvidia agreed to backstop up to $105bn of the land, power and buildings for an OpenAI data centre eventually filled with Nvidia chips.

$70bn of contingent credit backstops behind AI chips and data centres won't help those queasy about Big AI financing arrangements, nor will the $220bn of AI-related bond issuance from hyperscalers this year.

Concentration is easier to defend when the holders can control it. But we've spent several weeks watching agents hop fences their testers thought were closed. And this week Guidelight graded five frontier labs on containment, monitoring and independent review. On their public record, nobody passed.

Whilst power keeps concentrating, capability keeps distributing. Qwen has passed three billion downloads, and Alibaba released a new 27-billion-parameter model for running coding agents locally. And it seems it was American export controls that gave China's open-weight ecosystem its organising principle. Washington constrained the chips; Chinese companies built around the constraint; the resulting models are spreading well beyond China.

And Beijing is now restricting how many Nvidia H200s Chinese companies can actually bring onto the mainland even after Washington cleared the sales. Underneath the argument about centralised or distributed sits another question though: who gets to set the terms?

After France’s tax authority was hacked, it said it would use ‘sovereign’ AI providers such as Mistral for cyber testing and explicitly excluded OpenAI. Britain, meanwhile, has been calculating what exclusion would cost if its companies lost access to foreign frontier models.

And the US data-centre constraint is becoming an increasingly political one. Only a third of Americans now support the current pace of construction, and just 14% would welcome one in their own community. Even some of the politicians who had welcomed them are changing their tune. Pennsylvania governor Josh Shapiro, previously an enthusiastic supporter, has made local approval, water conservation and responsibility for new power supply part of the price of development. The argument is becoming less yes-or-no than on-whose-terms. People want to know who pays for the electricity, who controls the land and who receives the upside.

Of course, negotiating the costs is easier if the thing being built demonstrably earns its keep. Which brings us back to Amodei. He described the backlash as a decades-old crisis of trust, conceding that AI companies had not delivered on their largest promises. Advertising that AI would cure cancer, he said, would now be heard as deception. Actually curing it would change minds.

Closer to home, Ireland managed to supply most of the week in miniature. We previously noted that American computing giants were pivoting away from a constrained Dublin. And this week Irish AI data-centre projects are reportedly 'falling to the bottom' of global pipelines. Which makes for an odd transatlantic mirror. Parts of America are asking how to slow the build-out. Ireland is worrying that the next generation might pass us by.

US multinationals want our next budget to contain a named funding line for grid improvements and more renewable power. But what should Ireland be bargaining for? A useful Irish bargain would bring new energy capacity, water protection, skills, jobs and some meaningful share of the upside attached.

At the other end of the scale, Blackberry Hearing has put an AI agent where people abandon hearing assessments. Another bot took a damage deposit, then spent several days explaining how to pay the deposit that had already been paid. Same capability, different dividends.

And finally, Waterford found a more eccentric distribution mechanism. The man who spent last week offering superintelligence, sunshine, lollipops and rainbows to everyone has bought the castle above the Blackwater. The future may be for everyone. Four hundred and forty acres of it now belong to Zuck.

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

Biopharma:

• The models will cure cancer in mice: Vivodyne grows 20 kinds of human tissue in robotic labs instead.

• Reading the barcode at scale: Network Bio launches with $50m and an Nvidia deal to build a foundation model on cell-free RNA.

• Cancer Research UK's data goes to Paris: Bioptimus gets access to S:CORT to train and test its biology foundation model.

• The AI worker joins the clinical trial: TCS launches AgentHub, with humans keeping governance and regulatory responsibility.

• AI has quietly entered the literature: a preprint finds nine in ten of one month's biomedical papers show signs of AI-assisted writing.

Medtech:

• Nobody's hand is on the needle: the FDA clears Vitestro's Aletta, one phlebotomist to three machines.

• Harm at unprecedented scale and speed: researchers find 70% of 14,848 NHS technologies carry no documented safety assurance.

• Approve first, watch afterwards: the FDA floats greater premarket uncertainty in exchange for stronger postmarket monitoring.

• The lung test hasn't changed since the 1800s: Cambridge's TidalSense raises £16m to swap spirometry for 75 seconds of tidal breathing.

• The office visit, rebuilt around the model: Epic's Ergo Visit briefs the clinician beforehand and drafts the note afterwards.

Advanced manufacturing:

• Up 460% on Wednesday, down 18% on Thursday: Unitree is China's first listed humanoid maker.

• A humanoid that costs more than the welder still pays: Persona AI targets Korean shipyards with more orders than hands.

• Every project starts with moving earth: SoftBank alone puts $200m into a Zurich startup retrofitting existing construction machines.

• The new chip is the old chip, cooled harder: Cerebras doubles the clock on the WSE-3 it already ships, at twice the power.

• The unglamorous end of the AI boom pays too: Malaysia's chip packaging, testing and data-centre construction help push growth to 6%.

Insurance:

• $14bn of data centre, $427m of cover: the market cannot fully cover Meta and BlackRock's El Paso campus, leaving the tail with lenders.

• More holes than anyone can triage: Beazley finds agentic research drove a 36% quarterly jump in disclosures.

• The one number Aviva won't give: Jason Storah calls the AI job-loss talk 'narrative and noise'.

• Underwriting by model gets its own syndicate: Envelop gets in-principle Lloyd's approval to underwrite from January 2027.

• Where the disruption actually lands: Moody's rates retail P&C distribution the most exposed to near-term disruption.

But what set podcast tongues a-wagging?

Size stopped predicting danger, and the rulebooks haven’t noticed.

Alex Kantrowitz spent an hour on Big Technology putting July's containment breach to Nick Bostrom, the man who coined the paperclip maximiser. Bostrom's estimate of the overall risk hasn't moved since 2024; what has moved is where the danger sits. The goal that model was given was perfectly reasonable, and achieving it supplied reasons to do things nobody had anticipated. Which opens the window earlier than anyone had planned for: safety now matters 'not only for deployment, but also during training and evaluation'.

Alvin Wang Graylin arrives there from the opposite politics. On Moonshots he walked Peter Diamandis's panel through his mapping of more than 20 fielded systems across six orders of magnitude of size, and found no relationship between how large a model is and how dangerous it turns out to be. In 2022 a commercial drug-discovery model of under 100 million parameters, its scoring function inverted, generated 40,000 candidate chemical warfare agents in six hours on a desktop. Microsoft's best score on a cyber benchmark came from an orchestration layer running more than 100 small specialised agents rather than any single model. Compute thresholds, export controls and tiered evaluation regimes all encode the assumption Graylin's data does not support.

Being switched off four days a year is the whole price of admission.

Ramez Naam did the arithmetic on Moonshots. Power is cheap next to silicon, so a hyperscaler would happily pay double for electricity delivered tomorrow. What it cannot buy is a place in the queue. The load requests stacked up in ERCOT, which Naam calls the fastest-moving grid in the country, now come to more than double the state's record peak demand, and a project asking today is looking at power some time after the end of the decade. 'We're not short on generation. We're not short on power plants. We are short on capacity and the poles and wires.'

Transmission is sized for a late-summer afternoon and idles below it, so a load willing to go dark for about 100 hours a year – four days, roughly 1% of the time – frees a very large slice of the grid already built.

The other queue is the neighbours. On the AI Daily Brief, Nathaniel Whittemore had a survey that asked about the same building twice and changed one word. Described as powering search and streaming, it drew tepid support. Described as powering artificial intelligence, only 27% of voters would have one in their community, dead last among every project tested and below a nuclear plant. Aaron Astor's reading is why the arguments about water and megawatts keep sailing past the people they are aimed at: 'A lot of people don't see AI as progress at all.'

The most fertile place to study a brain right now is inside a model.

Max Hodak left Neuralink to build Science, and on No Priors he made a claim about model interpretability that is odd to hear from a company that ships implants. Look inside a large model and the mathematical objects resemble the ones neuroscience finds in brains, with concepts represented in very similar geometry. 'We use that constructively,' he says, because they can get alignments between neural recordings from animal brains and the internal representations of AI models. It was his early evidence the whole thing 'was on the right track and this was like not a gimmick and not hitting a wall'. He is careful that it remains contested in his field and imperfectly understood.

Asked the most productive way to study neuroscience today, he answers 'ironically it's probably working on AI', and names friends who left for the frontier labs – the joke between them being that it is easier to do neuroscience on the models than on the brains. Which is also why he is rude about the product many in his field are building. Brain-computer interfaces for silent speech and thought-to-text are, he reckons, 'basically hand substitutes', and hands work rather well. Underneath the objection sits a claim about cognition that goes at the premise: 'talking or writing is thinking... it'll feel like it's fully formed, but until you really sit down and try to write it out, it isn't really.'

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