Tom Westgarth

That's fair - probs the wrong analogy. Fintech a much better one!

The difference with Paypal Mafia I was referring to was the ecosystem of angel investing etc that emerged where winners would aggressively invest back into their sectors. We haven't crossed that event horizon in UK just yet imo.

Once you have picked the race, govt has to work out how to incentivise the ecosystem to help support one another.

An exam question I am mulling over is 'how do we create the UK's own version of the Paypal mafia?'

My general view is we should 'pick races not horses'. Pick a small number of niches (e.g. AI chip design) and go all in on developing an ecosystem around these companies

The industrial strategy fails to do this imo because the 'key sectors' amount to most of the UK economy.

The second risk is cultural. In govt, it is very hard to fund things outside of a 'fair and open competition'. It makes your business case much harder to get through.

Throw the kitchen sink at this. Big speeches. Letters to senior officials. Promote best practice. Flood the zone with this message.

But if you want to do this at the level of the industrial strategy, this requires going much further on issues around pay, paths to promotion, and creating a work environment that is semi-functional.

You need to actually bring in the expertise into govt if you want to make 'opinionated bets'.

This can be done. In SovAI, we drafted in one of the UK's leading young experts in AI for drug discovery, helping to push through critical work on OpenBind's AI for drug discovery datasets openbind.uk

‘I’m picking winners’: UK business secretary takes activist approach to economic growth AI evangelist Peter Kyle wants to scale up businesses, attract overseas investors and look out for UK’s poorer regions

This is a notable break from orthodoxy that guides thinking within the British state. But it is necessary in order for the UK to compete globally and restore sovereignty

There are a few challenges that this poses. What are they, and how can govt respond? www.theguardian.com/business/202...

GPTZero Dashboard Detect AI-generated text with GPTZero's dashboard

app.gptzero.me/documents/53...

50 papers published at Neurips 2025 (one of the three major global ML conferences) have AI hallucinations.

Another example of how institutions need to evolve the tooling they have to spot pollution of the knowledge commons

This data will look very different in 12 months time. Supply chains are becoming more heterogenous. Export controls will have a material impact (though just how much is unclear).

Also going to do the normie thing and say Claude Code was a big help in spotting these narratives and building charts.

What is quite remarkable is how rapid Nvidia's transition to Blackwell series is.

By Q3 2025, Blackwell chips account for the vast majority of Nvidia's compute shipments. Hopper is a thing of the past.

Again, demonstrating just how strong demand is.

This chart demonstrates the impact of such a tax over time.

Price performance diverging between NVIDIA’s Chinese and non Chinese outputs again is a major theme of why China has become relatively compute poor.

Interesting to see just how sharp the H20 cliff was. That's an 85% collapse in a single quarter when export restrictions tightened.

Will be intriguing to see what happens in a world of more relaxed export controls.

The China Compute Tax

Chinese AI labs are paying orders of magnitude (4x) more per unit of compute than the rest of the world.

Nvidia's H20 (the China-compliant chip) costs $80,230 per H100-equivalent vs $20,176 globally.

Now that export controls are relaxed, the tax drops from 4x to roughly 1.2x

Had fun pouring through some of the great data from @epochai.bsky.social's new AI Chip Sales Tracker

It illustrates a bifurcating global supply chain and new cost leaders emerging 🧵

epoch.ai/data/ai-chip...

Politicians: don't use AI to make your maps.

Extremely exciting day for open science.

Another reason to be bullish on UK AI. Boltz, a startup that emerged out of MIT to be at the frontier of small molecule and protein design, has moved to London.

I hear cases like this from a few founders, often a mix of fundamentals (talent)+personal.

A heck of a chart: in every single one of the 10 major US cities that built the most housing between 2017 and 2023, rents for older, existing units fell—often by quite a bit.

How far can decentralized training over the internet scale? Decentralized training over the internet promises to scale training to the limits of the internet.

Decentralised training potentially has big policy implications if it continues to innovate + scale:
- Could genuinely enable middle power collaboration, rather than vague 'shared sovereignty' notions on table now
- Creates jurisdictional nightmares with 'no clear off switch'
epoch.ai/gradient-upd...

Decentralised training of AI, where thousands of small datacenters, or individual consumers, pool their spare compute over the internet, is 20xing in scale a year. But this is nowhere near the scale of leading AI infra projects or non-AI decentralised projects.

Great piece by @epochai.bsky.social

Thanks! Will read and get back to you with some reflections!

V good thread. IMV what’s double damning about the Ovenden piece is that you could write a thousand pieces about how Whitehall could run better, but absolutely none of them would run through “why should the diplomatic service care about a British citizen in an authoritarian jail?”

So Wang’s framing here is trying to explain why the US can’t seem to build things, and China has such supremacy in doing so. Dictatorship vs democracy doesn’t explain that imo.

I really recommend reading Breakneck if you haven’t as it will explain China infrastructure strategy a lot better than I!

Thanks!

I had my own version of this in the works, framed as a “myth busting ideas on how to get things done in govt”, but may go back to the drawing board a bit with the framing based on The Discourse

Thanks for sharing - hope you’re well!

I thought it was worth trying to flesh out further as there is deffo something in the argument, but I didn’t feel it got to the heart of what I’d observed in govt

Governing without Giving a F*** The advice Keir Starmer should take from a self-help guru

Forgot to link to @pollymackenzie.bsky.social's excellent piece! howtorunacountry.substack.com/p/governing-...

Is Britain the Lawyerly Society? Dan Wang’s new book is timely and erudite. Parallels of the US 'lawyerly society' to the UK exist, but maybe not for the reasons that you think.

The el-Fattah example is imo a red herring, and more of an example of a govt that is more legalistic/procedural than others. I've written in more detail about to what extent the UK is the 'lawyerly society' in a Substack post attached. ENDS tomwestgarth.substack.com/p/is-britain...

Most people I spoke with in this No10 admin lamented that they were constantly fire fighting, & didn’t have the space to think about how to fix big problems. There are ofc big structural reasons for this, but the v best people find the space. They don’t talk about imaginary levers, they create them.

Nevertheless, it isn’t necessarily the case this means that a ‘Stakeholder State’ wins - other powerful forces can. The Stakeholder State didn’t rally for cutting winter fuel payments, unless you count HMT as a Stakeholder. This speaks to a wider paralysis amidst no Centre direction.

@pollymackenzie.bsky.social's substack makes this point well. This govt has not consciously chosen which political battles to fight. Combine that with a tendency to centralise, we are in a place where throughout government, ‘everything is a priority’, so nothing is.