SNI: WEEK 23

Welcome to all the AI news that matters this week – across tech, biopharma, medtech, advanced manufacturing and insurance. Where the work started, where its moving to and who wants the gains it’s yet to create.

tl;dr: Compute? It's coming home.

Ker-laaaaang! You could almost hear the penny drop this week. Right as finance teams took cover under a blizzard of invoices.

The question du jour: If using cloud AI is becoming expensive and unsustainable, (which it is) why don't we move it closer to home? On machines that also keep our data locked down?

Which is why Nvidia spread its bets by taking its fight to the laptop. Not to be outdone, Microsoft's Surface Ultra now runs a big model on a desk, with no data centre in the loop; Because (whisper it discreetly ahead of OpenAI and Anthropic IPOs) there's a growing share of tasks for which you don't need the frontier at all; smaller models are often sufficient. Which is why Microsoft is building its own again – and why the whole shebang fits on a laptop in the first place.

Not one to leave a bandwagon un-jumped upon, Perplexity released a controller that routes each AI job to wherever it's cheapest to think. Either on your own kit – or beyond. So you can see what each watt of energy delivers.

Which is a genuine seam of concern. Uber capped its engineers at $1,800 of tokens a month. Because, while no one is disputing the engineers are quicker, that speed is often falling short before reaching the bottom line. Which is why a chorus of commentators spent the week shouting warnings about the AI bubble again.

But are they thinking in exponentials? Because that's what others are experiencing. Analysts had Micron earning $18 a share as late as December. Which has now risen to $58. Where will it be by this December? In Exponential Land, a half year makes quite the difference.

And this is why, of course, the big money still believes in bigness. SoftBank pledged €75 billion for Europe's largest data centre in northern France. Alphabet, which hadn't issued stock in more than two decades, sold $80 billion of it – to keep pace with $190 billion of spending this year.

And with so much demand Nvidia says it's still 'supply constrained'. But this may also be further fuel for the local computing trend to develop. And next week, there's an event in Cupertino some believe may move it along, at pace.

Which may come as a relief to many in Ireland. Not least because the UN branded our experience – we now use a fifth of electricity for server halls – as 'a cautionary tale'.

With so much at stake, it is no wonder governments seemed to wake up, en masse, and try to grab the controls. Except each reached out for a different lever to pull.

• Canada is buying shares in its own AI companies, so the profits come back to its citizens and not only the founders – a state stake, a C$25 billion fund behind it.

• The United States, can't decide how to react. Bernie Sanders wants a one-off 35% tax on the biggest firms, paid back to the public as a dividend. While Donald Trump did the hokey-cokey with his own oversight order – pulled, rewritten, signed.

• France is pouring tens of billions into data centres at home – €110 billion of pledges in a single week – and backing a frontier lab of its own while it's at it, on the wager that it's wiser to own the machines and the models than to rent them from California.

• Argentina flung the door open, its president making the pitch in the FT itself: a law to let an AI own and run a company, no person liable, next to no tax. Incorporate in Buenos Aires, and whatever the machine earns, you keep.

Is it just us or does it feel like the rules of the next economy are being drafted, in a dozen capitals at once, by people who have yet to read each other's efforts? The next few months could be interesting. Will consensus build?

And what will that look like given the fear is real: an expected 'jobpocalypse', the hiring algorithms quietly screening people out and a generation telling the FT that AI has gone from helpful to harmful.

But is all that fear sustainable? Bridgewater found fewer than a fifth of firms using AI, and nine in ten of those seeing no change in headcount. While Gartner found the companies that cut staff to pay for AI got no RoI.

Without further evidence, the future is far from settled. But the answers are being forged at a faster pace than ever before. Which, perhaps, makes everything else worth reading this week very much worthy of your time:

Biopharma:

• The Transformer's co-inventor now designs RNA drugs: Alnylam's first AI discovery deal takes $30 million up front – and up to $2 billion if it pays off.

• One billion protein structures, free to anyone: the Chan Zuckerberg Biohub maps more than a billion structures and edges past AlphaFold3.

• OpenAI's life-sciences model gets cheaper to run: the GPT-Rosalind update cuts genomics tokens 31% and signs Novo Nordisk.

• Sanofi wants to halve the cycle: a Snowflake data deal aimed at development, where roughly nine in ten drugs still fail in Phase 3.

• Incyte's AI molecule deal grows past $1 billion: the Genesis collaboration widens to physics and the small-molecule bench.

Medtech:

• AI that grades the trainee: an RCSI graduate's OnWard gives medical students instant feedback on clinical placements.

• The surgical robot wants more operations: Medtronic files to expand its Hugo system into general and gynaecologic surgery.

• The hospital that wants to own its own: Mayo Clinic and Microsoft are building a frontier model for healthcare.

• The first AI tool you meet is silent: GE HealthCare wins FDA clearance for AI radiotherapy contouring – another clearance in the class.

• Pathology's data, pooled and given away: Tempus opens a digital-pathology consortium with Yale New Haven Health and Memorial Sloan Kettering.

Advanced manufacturing:

Computex turned the demo layer up to full volume – physical-AI models, humanoid platforms and whole-factory blueprints.

• Amazon's robot fleet passes a million: a generative-AI model called DeepFleet now coordinates the lot across 300-plus sites.

• A 5-week chip-design loop now runs in a day: Cadence's ChipStack agent runs design and verification at what it calls Level-5 autonomy.

• Siemens picks one of its own factories to go fully adaptive: an industrial AI operating system, blueprinted first at the Erlangen electronics plant in Germany.

• A VW plant in Tianjin reports 16% more output: 105 AI scenarios across stamping, welding, paint and assembly, with quality issues down a fifth.

• TSMC puts AI inside the fab: NVIDIA computing across lithography, process control and visual inspection, with an Omniverse 'FabTwin' to simulate the line before it's built.

Insurance:

• The algorithm that quietly cancelled your cover: aerial-imagery models are driving home-insurance nonrenewals with no human inspection.

• Agentic underwriting graduates to production: CFC weighs widening its Lane Assist – it scrapes the submission, fills the model and proposes a quote, with the underwriter signing off.

• Lloyd's writes cover for the AI build-out: Navium and Overhaul's Helix consortium offers up to $75m in transit cover for the chips and servers crossing the world to fill the data centres.

• The compliance apparatus becomes the product: Sapiens wraps agentic claims and underwriting in a single audited ontology, with Abu Dhabi's ADIA backing the move.

• AI moves onto the live risk register: WTW counts AI-related incidents up roughly 50% a year, with the exposure spreading across multiple lines.

But what set podcast tongues a-wagging?

The scarcer and dearer the cloud gets, the more sense your own hardware makes.

On the All-In podcast, OpenAI's CFO Sarah Friar was unusually candid about the squeeze. Compute is scarce and stays scarce through 2026 and into 2027 – if you want to buy more, she said, good luck. If the cloud is that dear and that contested, the question writes itself: why send every job so far to be thought about?

Daniel Miessler has an answer: he runs on a Mac Mini under his desk. On Cognitive Revolution, the security researcher walked Nathan Labenz through his Personal AI Infrastructure: a one-gigabyte local database holding five years of his life. With 'employee' agents working it on his own hardware, nothing leaves the building.

It's the 'own-your-data future' – and Miessler is honest about the debt that comes with it, since prompt injection is nowhere near solved and the only real defence is to concentrate trust in a few hardened machines.

Which is roughly the case Satya Nadella took to Microsoft Build. On No Priors, he framed the whole conference around a single line: The model tuned on a firm's own data and judgement is a new defensible asset that creates competitive advantage. This type of agent, he argued, belongs on the balance sheet.

He also said, out loud, that the world is very sceptical of tech companies – so the energy, the jobs and the tax had better show up. Indeed.

Which also begs the harder question of who keeps the gain once the digital intelligence has delivered its work?

The starting point of Dwarkesh's two hours with economists Alex Imas and Phil Trammell was a number that has barely moved in two centuries: labour's share of national income, steady at around 60% for two hundred years, which they think automation could push toward zero.

But given the value has to land somewhere, who will capture it? Is AI more like electricity, where the gains spread out to everyone who plugs in? Or more like social media, where the rents pool with whoever owns the platform?

The two walk the redistribution menu – negative income tax, universal basic income, universal basic capital – without pretending any of it is settled. It's a calm conversation about a violent possibility. It asks the same questions all politicians are now scrambling over.

Which is exactly where Nathaniel Whittemore picked it up. On the AI Daily Brief he noticed the AI discussion has shifted from who gets to use it to who gets to own it – and bank what it earns.

Our conclusion? The gains are coming and they can be directed. Either to a tiny group of people. Or humanity in general. The choice is ours, as societies, to make. But our first answers may persist for many decades to come.

Thanks for reading. Join us again next week for all the AI news you need to know.