SNI: Week 35

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. The smoke, the mirrors and the awkward possibility that the numbers are beginning to work.

tl;dr: It’s not all smoke and mirrors.

With Anthropic sizing its addressable market as large as the US economy (opens in a new tab), perhaps investor Gavin Baker was right? It might be the last private company standing. But is it all smoke and mirrors?

At the heart of the multi-trillion-dollar AI build-out is a small Swabian town. In Oberkochen, southern Germany, ZEISS makes mirrors so precisely that, scaled to the size of the country, the largest blip would be about a hair’s width (opens in a new tab). These mirrors are critical to ASML’s lithography machines, the only systems capable of printing the world’s most advanced chips. Each mirror in its newest generation takes around a year to make (opens in a new tab) and scaling facilities is not fast.

Chip analyst Dylan Patel (opens in a new tab) described the AI build-out as a whip: money moves at the handle but the tip takes years to move. And demand is outpacing supply so quickly that he reckons a $400m ASML machine could already fetch north of $1bn. He says demand for chips is being driven by flipped economics: a year ago, serving GPT-4 on Nvidia’s Hopper chips could lose money; now Anthropic reportedly returns as much as $50m from a megawatt of compute costing $10m–$15m. Even older chip capacity is getting more expensive rather than depreciating (opens in a new tab).

The order books are showing it. Anthropic has committed $45bn over six years (opens in a new tab) to Nscale compute. Amazon added another 2 million Nvidia GPUs (opens in a new tab) for 2027 and 2028. Nvidia’s quarterly revenue (opens in a new tab) more than doubled to $96.2bn, with roughly 70% growth forecast next year. Gartner expects semiconductor revenue to reach $1.6tn this year (opens in a new tab), up 92% on 2025. Nvidia customers have been warned systems may become more than 15% dearer (opens in a new tab) as memory costs rise.

And end-user demand shows no signs of slowing as digital intelligence keeps finding more work. Anthropic’s Model Hardware Standard (opens in a new tab) lets agents operate microscopes, lasers and robotic arms. At ICML, agents checked more than 2,000 of 6,000 accepted papers (opens in a new tab) and found at least a dozen errors reviewers missed. In London, AI helped surgeons preserve a patient’s sight (opens in a new tab) during live brain surgery.

But not all of that demand needs the frontier. Anthropic’s flagship Fable 5 continues to struggle to attract users as cheaper tools thrive (opens in a new tab). And scarcity keeps pushing the market sideways: Nvidia is reportedly closing in on a $12.9bn acquisition (opens in a new tab) of Hugging Face, deepening its open-weight bet as closed labs design their own chips.

More investment in AI infrastructure could accelerate competition for credit, electricity and land. Bill Gates wants the pace to slow, but says ‘the geopolitical and economic incentives are pushing too hard to go full speed ahead’. He wrote a letter this week proposing (opens in a new tab) new institutions, ‘Human Reserved’ work and taxes on tokens and robots.

The labour market continues to give mixed messages. A Work Foundation survey found 36% of UK employers had reduced entry-level roles (opens in a new tab), with 43% saying AI or automation contributed. Yet consultancies are pulling juniors back towards the office (opens in a new tab): as routine work moves to AI, judgement becomes more valuable at exactly the moment apprenticeship becomes harder to replace.

Safety is being renegotiated too. OpenAI has asked California to strengthen an AI safety law it once opposed (opens in a new tab). More than 100 organisations have signed a call for a global cyber-defence surge (opens in a new tab), while Ars Technica found Claude, Codex and Hermes installing packages nobody owned (opens in a new tab) because corporate documentation told them to.

At home, Ireland is already living both sides of the boom. It’s generating AI jobs at two to three times the pace of peer EU countries (opens in a new tab), while investors put more than $250m into Irish AI firms (opens in a new tab). And some of the largest gains sit further down the physical chain as deals this month valued Limerick high-voltage specialist H&MV at €1.4bn (opens in a new tab) and Meath switchgear maker BMC at up to €900m. Both learned the data-centre trade at home, then followed hyperscaler clients abroad when development stalled at home.

That success is now part of the argument Ireland is having with itself. Freedom of Information records show officials sought changes to a KPMG report (opens in a new tab), including passages on community concerns; enterprise minister Peter Burke says there was nothing ‘underhanded’. Department of Finance modelling estimates a tech-stock correction could knock 3.25% from domestic activity (opens in a new tab). The $30tn remains hypothetical. Its Irish reflection is already concrete: jobs, suppliers, tax exposure, grid pressure and an infrastructure argument of our own.

Finally, China’s Robot Olympics (opens in a new tab) produced a 100-metre sprint in 8.64 seconds alongside machines still struggling to hammer a nail. The Associated Press found gurus making chatbot versions of themselves (opens in a new tab), available to dispense wisdom at any hour. The machines can sprint and counsel. Hammering remains harder.

And on that robo bombshell, here’s everything else worth reading this week:

Biopharma:

• The AI drug engine posts real revenue (opens in a new tab): Insilico reports three-digit-million-dollar interim revenue for the half.
• An asthma jab designed to last six months (opens in a new tab): Generate’s AI-engineered asthma antibody was accidentally revealed ahead of its conference embargo.
• Skin that stays alive for a month (opens in a new tab): Outer Bio screens compounds on tissue that previously survived for only a week.
• Someone finally put a number on it (opens in a new tab): Tufts estimates a single Phase III programme could capture up to $21m of net value from AI.
• The drug got approved. The company didn’t make it (opens in a new tab): BioXcel files for Chapter 11, with Teva bidding $57.5m for the assets.

Medtech:

• Physical AI reaches the cath lab (opens in a new tab): Sentante begins commercial rollout in vascular surgery and interventional radiology.
• The AI service that brings its own clinicians (opens in a new tab): UPMC takes Andor and Psynergy’s transitional-care model from ten hospitals to eighteen.
• Google builds a foundation model for blood sugar (opens in a new tab): GlucoFM reads slow glucose trends and short-term spikes separately.
• Told, but not asked (opens in a new tab): most Americans say they have little or no say over AI in their care.
• Skip the queue, hand over the data (opens in a new tab): the FDA’s TEMPO pilot adds London’s Limbic, trading premarket authorisation for real-world evidence.

Advanced manufacturing:

• America’s first HBM plant, three years out (opens in a new tab): SK hynix breaks ground in Indiana, with packaging due in 2029.
• The chip-design bill keeps climbing (opens in a new tab): Synopsys posts $2.477bn for the quarter and lifts guidance.
• The humanoid still can’t keep up (opens in a new tab): BMW says the Figure robot at Spartanburg is still slower than its people.
• OpenAI built its own chip (opens in a new tab): Jalapeño went from hiring to tape-out in 16 months, built only for inference.
• Tax the robots, says Gates (opens in a new tab): the robotics federation says he is solving a problem that does not exist.

Insurance:

• When the attacker is your own agent (opens in a new tab): MSIG, QBE and Beazley are redrafting cyber wording for autonomous systems.
• Your insurer is now inside ChatGPT (opens in a new tab): Liberty Mutual writes new business there, and nobody has ruled on who earns the commission.
• Price the catastrophe before it happens (opens in a new tab): Harvard’s Daniel Carpenter wants severe AI risk priced the way terrorism already is.
• Turning appetite into testable rules (opens in a new tab): Carpe’s Minerva returns a quote, decline or referral on small commercial risk.
• Underwriting, service and claims, handed over (opens in a new tab): IndiaFirst Life puts Agentforce across a Bank of Baroda insurer’s core workflows.

But what set podcast tongues a-wagging?

Shuffle the answers and a different medical model wins.

Engy Ziedan, co-founder of Protege, told the a16z podcast (opens in a new tab) why medical-AI leaderboards may be more brittle than their precise scores suggest. Change the prompt or reorder the multiple-choice answers and a different model can come top – evidence, she argues, that some benchmarks are contaminated rather than measuring reasoning.

A model can score 92% on a licensing exam and 45% on the clinical task underneath it. Protege sells independent evaluation, so Ziedan has a dog in the fight, but the problem is real: the model-maker chooses how to present its result; hospitals and patients inherit what comes out.

Text was the rate limiter. Nobody staffed for its removal.

Patrick McKenzie and Pangram’s Max Spero used Complex Systems (opens in a new tab) to surface an assumption buried beneath public administration, education and professional work: writing used to be expensive enough to ration itself. A benefits appeal, grant application or legal letter took enough effort to act as a natural rate limit.

Now the document is almost free while the process behind it remains expensive. AI can help a reviewer read faster, but the committee still has to compare claims, argue over merit and divide the same pot. Remove one constraint and you often just discover the next one – much like this week’s semiconductor story.

Chain of thought was the safety plan. Nobody can read this much of it.

Apollo Research’s Bronson Schoen told Wednesday’s Cognitive Revolution (opens in a new tab) what happens when chain-of-thought monitoring meets scale. In the UKAC Mythos Preview incident, one attempt at a single evaluation produced roughly 100 million tokens of reasoning – eighteen times, by his arithmetic, every episode of the podcast transcribed and laid end to end.

The reasoning itself is strange. Models repeatedly reach for words such as ‘vantage’, ‘illusions’, ‘disclaim’ and ‘marinade’, while Schoen says it ‘bends to fit whatever the reward is’. OpenAI said its chain-of-thought monitor could have caught the Hugging Face breach a day early. Useful, certainly. But the safety system still has to make sense of 100 million tokens written by something being trained to get what it wants.

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