SNI: WEEK 13

Welcome to all the AI news that matters. The wins, the fails and the somewhere in-betweens. Across tech, biopharma, medtech, advanced manufacturing and insurance this week.

tl;dr: Lots of tokens. Many large clients. But no more Sora

Jensen Huang popped up to emphasise that data centres should now be thought of as ‘token factories’. But most insiders were watching a different drama unfold. Entire theses could be written around the week OpenAI experienced. If you weren’t watching closely you could be forgiven for thinking it was the plot of a movie. With the protagonist's journey being a 'coming-to-terms-with-reality' story - the reality being that AI is only profitable when sold to large organisations.

Here’s the briefest ‘believe-it-or-not’:

Monday: ‘Leaks’ revealed that OpenAI is aggressively doubling its target headcount to 8,000 by the end of the year - to fend off Anthropic and Google. Later that day, Sam Altman took the stage at a BlackRock event to pitch billing users for token usage like water or electricity.

Tuesday: OpenAI announced it is shutting down the Sora social video app. Official messaging confirmed a wider pivot away from consumer markets to reallocate compute power toward enterprise solutions, coding and robotics.

Wednesday: The immediate fallout of the Sora shutdown hit, with the cancellation of the $1bn Disney deal.

Thursday: We learned that ChatGPT, Codex and Atlas will now live in a consolidated desktop superapp. Which won’t have an ‘adult mode’. OpenAI admitted it had indefinitely shelved its most controversial feature - after plenty of pushback from safety advisors, investors and staff.

Perhaps that decision was finally dragged over the line after a Los Angeles jury found Meta and YouTube liable on Wednesday for deliberately designing addictive algorithms that harmed a young user. Just a day after Meta was hit with a separate $375 million civil penalty in New Mexico over child safety and misleading consumers.

But Sam Altman likely consoled himself that OpenAI isn’t the only tech business in open retreat. After a year-long push to integrate AI into every corner of its operating system, Microsoft announced it’s actively cutting back on "unnecessary" Copilot AI features across Windows apps.

In other tech sector news: The EU backed delays to key AI Act deadlines, pushing back high-risk system compliance and watermarking rules

Kleiner Perkins raised $3.5bn across two funds, a 75% increase over its last raise, concentrating venture firepower on AI-native companies

OpenAI partnered with Broadcom on custom chip design, signalling a move to reduce dependence on Nvidia for inference workloads

Three individuals were charged with attempting to circumvent US AI chip sanctions, evidence that enforcement is tightening around export controls

GitHub reversed its position on AI training data, opting users in by default from 24 April – expanding Microsoft's training corpus across millions of repositories And now stay tuned for the sectors:

Biopharma:

Insilico Medicine published peer-reviewed phase 2a results for its AI-designed drug rentosertib, the first such milestone for a molecule generated entirely by AI Daiichi Sankyo and Tempus AI formed a strategic collaboration on ADC biomarker discovery, linking Tempus' oncology database directly to clinical differentiation

Earendil Labs secured $787m to expand its AI drug discovery pipeline of more than 40 candidates, one of the largest single financings in the space

Medtech:

Viz.ai and Alnylam partnered to build an AI-driven detection pathway for cardiac amyloidosis, integrating FDA-cleared echocardiography AI with EHR systems

Microsoft's Dragon Copilot AI scribe reached NHS trusts, with one specialist reporting two hours saved per week – now listed on the NHS Ambient Voice Technology Registry

Advanced Manufacturing:

Normal Computing raised $50m led by Samsung Catalyst to address AI hardware energy consumption through novel chip architectures

Panasonic disclosed that data centre batteries are now selling years in advance, and the company is shifting production capacity from automotive to compute

Sift, founded by two ex-SpaceX engineers, brought manufacturing data infrastructure to the factory floor, targeting the gap between AI planning tools and shopfloor execution

Insurance:

Chaucer and Ceto AI launched a Lloyd's marine MGA that integrates real-time vessel data directly into underwriting decisions

TPG and Allianz led a $350m investment in Cambridge Mobile Telematics to expand AI-driven road safety and real-time risk modelling

WTW's survey of 59 insurers found that those using advanced analytics achieved combined ratios six points lower and premium growth three points higher – yet only 16% currently augment underwriting with AI

But what set podcast tongues a-wagging?

Azeem Azhar revealed his personal AI usage went from 150,000 tokens per day in summer 2024 to 870 million in a single day this week - a 5,800x increase in under two years. His point? Look at how the inference economics compound: reasoning models have created 10,000x the compute demand per user compared to standard chat. Multiply that across enterprise deployments and you understand why Nvidia has $1 trillion in committed orders and why grid capacity - in Ireland, the Nordics, and everywhere else - is the binding constraint.

60% of companies are exaggerating AI's role in layoffs

A resume.org survey of 1,000 hiring managers, unpacked by Nathaniel Whittemore on the AI Daily Brief, found that 60% said they deliberately emphasised AI's role in redundancies because markets and stakeholders view it more favourably than admitting to overhiring or strategic misjudgement. Only 9% said AI had actually replaced any roles fully. Whittemore's distinction between 'efficiency AI' (cutting costs) and 'opportunity AI' (creating new capacity) means many companies framing cost cuts as AI transformation are borrowing credibility from a technology they haven't yet deployed at scale. The danger for policymakers and enterprise planners alike is that regulatory responses, workforce investment and strategic decisions are being calibrated to a narrative that the companies creating it know to be exaggerated.

Jensen's 'token factory' reframes what a data centre is - and does

Jensen Huang on Lex Fridman described four scaling laws - pre-training, post-training, test-time and agentic - which each multiply compute demand. His messaging echoed Azeem’s: data centres are no longer computers. They are factories. They produce tokens. And the architecture shifted from single GPUs to rack-scale computing because agents - which 'bang on tools' and spawn sub-agents - demand a fundamentally different infrastructure.

The submarine problem is a workforce problem

Chris Power - CEO, Hadrian - and Admiral Robert Goucher on the a16z podcast described American submarine manufacturing: decades of industrial decline, a skilled workforce that doesn't exist in sufficient numbers, and a technology intervention compressing the timeline. Power: 'Three years ago it was incredibly difficult and now... is pretty fast.' The mechanism isn't AI replacing welders. There aren't enough welders to replace. Software multiplies the output of the ones who exist. The defence context adds a structural accelerant - single-person authority replacing distributed multi-stakeholder accountability - that isn't available in commercial manufacturing. But the underlying logic transfers: when you cannot hire your way to the precision workforce your industry requires, verification and amplification of scarce human skill becomes the primary lever. Which may be as true in pharmaceutical manufacturing, semiconductor fabrication and specialist insurance underwriting as it is in submarine hulls.

Thank you for reading this week's report. Come back next week for all the AI news you need to know in your sector.