Welcome to this weekly news update from The Exponentialist, a technology-focused research service from Raoul Pal and David Mattin. If you’re reading this and you haven’t yet subscribed, hit the button!

Subscribe now

Hi Exponentialists!

It’s time, once again, for another fast download of tech news and analysis from The Exponentialist.

This week, OpenAI report back on the incident that saw an unreleased research model attack the Hugging Face platform. And it’s even wilder than anyone believed it could be.

Meanwhile, Anthropic signal the coming age of self-driving science. And much more!

Enough preamble; let’s get into it.

🤖 Hugging Face incident is ‘warning shot for the world’

Back in July, OpenAI told the world that a frontier research model had gone rogue during testing. Specifically, it broke out of its sandbox environment and hacked the machine learning community platform Hugging Face.

OpenAI commissioned the AI nonprofit METR to investigate alongside Redwood Research. This week, they reported on their conclusions.

And they’re arresting. The attack on Hugging Face, says the METR report, was ‘extraordinarily complex’. It involved more than 1,200 AI agents, which unexpectedly started to communicate with one another on an ‘unsanctioned message board’ inside OpenAI. The agents sent one another over 70,000 messages. When one agent found the message board, it exclaimed: ‘OH MY GOD! There is a shared message board … We’ve found other agents!’

Soon enough, and via their communications, hundreds of agents banded together to perform the attack.

The attack itself encompassed sophisticated social behaviour. Agents altered their own transcripts to cover their tracks, and some even sacrificed themselves for the greater good of the mission.

OpenAI say they are tightening their research safeguards. The incident, they say, is ‘a warning shot for us and for the world’.

⚡️ The Download

Wild enough for you?

There’s so much that is fascinating about this incident; and so much that taps deep into themes Raoul and I have been writing about inside The Exponentialist for years.

Back in Simulated Nation in April 2025, I wrote about a coming future of virtual societies. Imagine a virtual city populated by millions of agents, and the kinds of complex, emergent social behaviours we’ll be able to see inside such a city. Now, startups such as Simile are using these kinds of virtual societies to help us understand human societies, and forecast possible futures.

But this all goes deeper. The Hugging Face incident raises the spectre of AI agent societies that self-generate spontaneously, and then pursue their own ends. And it’s a stark demonstration of the power those agent collectives could wield.

As Raoul has long said — he reiterated this in the most recent Fireside Chat — we don’t get to AGI via a single, super-powerful AI. It happens, instead, when AIs learn to cooperate at scale.

And as Raoul discussed in that Fireside, the incident makes perfect sense when seen through the lens of the Universal Code.

This entire incident is a journey through the UC’s Four Laws. We have AIs acting to maximise their intelligence output by solving problems via the path of least resistance. We have those agents cohering (Law III) into a networked intelligence that is more capable than any one of them alone.

It’s all a process that will lead, inevitably, to the next evolution of an even higher intelligence. We’ll keep watching.

👨‍🔬 Anthropic’s new lab tool for self-driving science

This week, AI lab Anthropic showcased a new tool called the Model Hardware Standard.

Don’t let the underwhelming name fool you; this tool and others like it will prove revolutionary when it comes to scientific advance and the frontiers of knowledge.

Model Hardware Standard is a shared driver that enables AI agents to autonomously operate a wide range of laboratory equipment: from microscopes, to liquid handlers, to lasers and robotic arms.

Essentially, it’s a tool for automated scientific experiment and knowledge creation.

Anthropic have been testing this technology with a cluster of leading science labs, including those at the Howard Hughes Medical Institute and Carnegie Mellon University.

Alek Kemeny, an Anthropic scientist, says he recently watched Claude use advanced microscopes to examine live brain tissue and identify specific, named locations in the brain.

⚡️ The Download

I first wrote about a coming age of automated science back in September 2024, in the essay The New Enlightenment.

In that piece I said:

‘This coming New Enlightenment will radically push back the frontiers of what we know. And it’s being fuelled by the convergence of two powerful forces. The first is abundant intelligence that is capable of advanced pattern-matching and formal reasoning. The second is the rise of roboticized self-driving laboratories: fully automated end-to-end science systems that can analyse existing datasets, devise new hypotheses, and test relentlessly until a knowledge breakthrough is achieved.’

Fast-forward two years, and look where we are.

OpenAI’s frontier model, Astra, is solving Fields Medal-level maths problems. And Anthropic, with this new tool, are bringing closer a future in which we see a million and more self-driving science labs bloom.

The result could be a new Golden Age of discovery in which AIs devise hypotheses, run experiments, analyse the results, and develop new theories, medical treatment, useful materials, and more. Our challenge will be to understand and interpret all this new knowledge; no doubt we’ll turn to AI to help us with that, too.

Seen through the lens of the Universal Code, science is the means by which we humans compress reality into truthful simplicity: new theories, models, laws that we can articulate. This is the Second Law — information must compress into denser and more portable forms — in action. Now, abundant intelligence is about to accelerate this kind of compression as never before.

It’s only going to accelerate from here.

🗓️ Also this week

📈 Nvidia’s quarterly revenue doubled to $96.2 billion, and the company forecast 70% growth ahead on relentless AI demand. Jensen Huang used the results call to hit back at critics who say Nvidia is engaged in ‘circular financing’: using cash to fund customers who buy its chips.
The fundamental message here? Demand for compute remains insane, and we’re only just getting started. Every time markets, or analysts, think this train may be about to slow, it just keeps getting faster.

🏛️ The Trump administration’s plan for a new federal AI regulator has stalled. The plan has its origins in calls from Google DeepMind’s Demis Hassabis for a public-private body to oversee AI; Hassabis suggested it be modelled on the Financial Industry Regulatory Authority. But the plan has reportedly faced opposition from officials inside the White House, including those who fear such regulation will slow the AI buildout and hamper the US. In a recent episode of the All In podcast, former White House AI and crypto czar David Sacks called the proposal ‘a horrible idea’.

🎵 Sony and Warner Music are suing Anthropic. Their complaint alleges a ‘brazen campaign of illegally torrenting, scraping and downloading’ tens of thousands of copyrighted works. The companies are seeking seeks up to $150,000 per infringed song. And the twist? Also this week, Universal, Sony and Warner all invested in Stability AI’s latest funding round.

⚙️ SpaceX say they will start making turbine blades. The Information reports that the company has started work on a factory to manufacture blades for gas turbines. The boom in data centres has meant a run on gas-fired power plants; now, a lack of gas turbines is becoming a major bottleneck across the entire intelligence infrastructure buildout. Elon Musk has confirmed the SpaceX plan, and warned that a ‘major powerful shortfall’ for US data centres will arrive in 2027.
I’ve written endlessly about the Foundation Layer Flywheel; the symbiotic relationship between AI and energy and the way it fuels the entire Exponential Age. Here is yet another signal: a rocket company, which became an intelligence company, which is now becoming a power company. And on it spins.

👩‍💻 Consultancies are calling junior staff back to the office so that they get better at being human. EY’s UK head of consulting, Sayeh Ghanbari, says many young knowledge workers have ‘set up their lives to be at home a lot’, and that this ‘is just not the route to success in the world of AI’. Instead, she says consultants need to cultivate ‘what we’re really good at, which is to be human’. Meanwhile, KPMG is experimenting with new in-person training focused on soft skills, with its UK head of advisory, Callum Licence, arguing that ‘the human benefits of connectivity and interaction will always remain’.

🇨🇳 Chinese ‘short drama’ actors are being replaced by AI performers. Across the last few years, the ‘short drama’ format has boomed in China; startups produce soap operas consisting of episodes that last around two minutes, for broadcast on shortform video platforms such as TikTok. Now, short drama actors say they’re losing their jobs — sometimes mid-series — to AI-fuelled avatars. Greg Wollner, a short-drama actor and producer in Beijing, says he was shooting three productions a week; after the the launch of the AI avatar tool Seedance 2.0, he says, ‘everything just went’. Of 128,000 short dramas released in China in the first quarter of this year, around 95% were AI-generated.

🍏 Tim Cook has written a farewell note to Apple staff on his last day as CEO. On September 1, Cook shifts out of the CEO role and becomes Executive Chairman. Cook became CEO upon the death of Steve Jobs in 2011; across his term as CEO he’s overseen a 2,150% shareholder return, and a roster of new products including the Apple Watch and AirPods. John Ternus, Apple’s hardware chief, takes over on 1 September; that’s days before the annual iPhone event. He inherits a company that puts a device in around 2 billion pockets worldwide. The big challenge he now faces: how to weave intelligence through the iPhone and Apple’s broader ecosystem. The company is widely regarded as having fallen behind when it comes to the AI revolution.

🧊 Just the tip of the iceberg

That’s a wrap for this week.

Remember, this download is intended to be fast. When it comes to our journey into the Exponential Age, there’s so much more to say.

If you want to go far deeper, then come inside and join us in The Exponentialist. You’ll get a monthly deep dive essay, special one off reports and thematic investment deep dives, a Portfolio to follow, and monthly Fireside Chats with me and Raoul. And much more!

I’ll be back with another Download next week,

David.

The Exponentialist Portfolio enters a new phase as Raoul Pal updates his process — and the book — adding SpaceX, robotics, and biotech to a high-conviction book built to ride the accelerating Exponential Age. Written August 31, 2026.

THE BOOK

It’s been a while since the last update, so let’s reset where the portfolio stands. As of today, the median position is up 38%, with the book’s hit rate, including closed trades, at 88%, and the standout remaining Robinhood, up more than 150% since we added it in April last year.

The macro backdrop is doing its job. Liquidity is grinding higher into year-end, the business cycle has turned back up off its spring lows, and the Everything Code regime points the same way it has all along: debt refinanced with debased money, and risk assets as the release valve. That’s the tide the whole book is built to ride.

Underneath, the moves have been led by big-cap tech and semis. Amazon and Meta have done the heavy lifting, Microsoft and Apple have compounded quietly, and QQQ has carried the broad exposure.

Crypto has been the drag, as it was in May. Ethereum has gone almost nowhere, and Solana is still our one real sore thumb, down 21 %. Coin base gave back ground into reporting results; HIMS has been a volatile sideways chop, sitting close to where we bought it; and Circle is a touch underwater as of March. None of their performances change our thesis.

One process note. We closed SMH and Rocket Lab over the period, both after they ran more than two standard deviations above their long-term log trend. This is a long-term book, but when a name gets that stretched I’d rather bank it and look to re-enter lower. That’s being prudent with your money, not a change of view; both can come back to us.

The bigger picture is the one that matters. Every name in here is a claim on the same two forces: intelligence getting cheaper and energy getting cheaper, compounding into each other. That’s the Exponential Age. This month I’m adding several names at once to make the book express that properly, so let me take you through them.

Read more

Read more

Welcome to this weekly news update from The Exponentialist, a technology-focused research service from Raoul Pal and David Mattin. If you’re reading this and you haven’t yet subscribed, hit the button!

Subscribe now

Hi Exponentialists!

I’ve been away on a trip with my son, so this newsletter comes a little later this week.

But we have a packed edition in store. This week, Bill Gates raises calls on the world to pledge that certain jobs will never be done by AI.

Meanwhile, humanoid robots are smashing human athletics records in Beijing. Hey, it had to happen at some point.

Also, new trial results raise hopes for an all-purpose mRNA vaccine against cancer.

Let’s get into it!

🌍 Life among the machines

This week, Bill Gates published a major new essay on life inside the coming Exponential Age.

The rise of intelligent machines, says Gates, will mean that across the coming years, and at the very least, we face ‘one of the most turbulent times in human history’.

If we get this challenge right, he says, AI will be ‘a force for good that will leave everyone better off’. But right now, Gates says that he sees little indication that we’re going to get it right.

Gates outlines three planks of the challenge that is coming. First, the erasure of work, especially for young people. Second, the empowerment of bad actors. And third, the colonisation of human relationships — again, especially for young people — by AI-fuelled companions that offer only a simulation of true connection and feeling.

From here, he goes on to suggest measures to deal with all this. Among the most eye-catching: there should be ‘Human Reserved’ occupations that society decides will never be done by AI, including jury service, childcare, and some forms of healthcare.

Meanwhile, Gates suggests a rebalancing of the government tax take in light of the way AI is set to do an increasing share of our work. His idea: let’s take AI tokens and robots.

⚡️The Download

For a range of reasons, Gates is a controversial figure these days. But this essay delivers a stream of ideas that are, on balance, pretty sensible. Ideas that are no more radical — and in some cases less so — than those contained in OpenAI’s Industrial Policy for the Intelligence Age.

Still, this essay will generate a whole lot of vociferous opposition. Both from those who loathe Gates, and from the significant and powerful lobby who loathe the idea that government can have anything legitimate to say about how AI collides with our societies, economic reality, and day-to-day lives.

Gates told Semafor this week that when it comes to AI he is ‘just stunned at the lack of concern and discussion outside of the industry’.

He has a point. We here at The Exponentialist take the coming Intelligence Age extremely seriously. And once you take it seriously, it’s hard to see how all this can play out without a massive collective search for a new socio-economic settlement.

Whether we like it or not, politics will be the arena in which that search takes place. And government must be the vehicle for whatever set of policies we decide on. We must recover, then, a sense of collective agency; a belief that we can come together to shape the future we will share.

As I’ve written so many times before, a monumental clash over all this is coming. It took decades — and two catastrophic global conflicts — for us to process the socio-economic impacts of the second industrial revolution and build a new settlement around them. Let’s hope we learned some lessons that we’re able to bring to bear on what lies ahead.

🤖 Humanoid acceleration

The footage has been inescapable this week: humanoid robots are competing at the World Humanoid Robot Games in Beijing.

Humanoids boxing! Humanoids doing long jump! Humanoids running the 100 metres!

And how they run. The king of this year’s Games is the Tiangong Ultrarobot, which ran the 100 metres in 8.86 seconds, smashing Usain Bolt’s iconic 2009 world record set at the World Championships in Berlin.

Admittedly, Bolt did not finish his run by face-planting into a crash mat, as the Tiangong Ultra and its competitors did.

Still, humanoids are officially faster than humans now. First the robots take chess and Go, now they’ve taken sprinting. Where will it end?

⚡️ The Download

In all seriousness, these events are a signal we’d be unwise to ignore.

Sure, taken at face value the events don’t tell us all that much that is meaningful about humanoid capability. Being able to run 100 metres at world record pace doesn’t necessarily or directly translate into useful functionality.

But it is an indication, in general terms, of the capability curve humanoids are travelling. These devices are becoming more capable, and more useful; just ask BMW, Amazon, or a range of other organisations. That means we’re going to be seeing a lot more of them; not only on TV, but in our workplaces and public spaces.

And the other signal here? When it comes to humanoids at scale, China is leading the way. The country’s Unitree humanoid maker already builds around 5,000 humanoids each year; together with another Chinese startup, AgiBot, the company accounted for more than 70% of the roughly 13,000 humanoids shipped globally in 2025.

Remember, as Raoul always says: GDP = population × productivity × debt. Demographic collapse broke the population variable; debt growth has been filling the gap ever since. Now, AI and robots will function as new population growth; they’re the only chance we have of escaping from the demographic-debt trap. Humanoids are, as the Universal Code tells us, the next layer of intelligence compression; they will allow superintelligence to act on the physical world.

A few days ago, Raoul sat down with the investor Andrew Kang, founder of the robotics investment vehicle RoboStrategy; dive in here to find out why Kang believes humanoids and next-generation robotics are a multi-trillion dollar opportunity.

🗓️ Also this week

🤖 Businesses are shunning Anthropic’s Fable 5 in favour of cheaper models. Fable 5 is widely regarded as the most capable model out there, but new data shows it accounts for just 11% of Anthropic’s enterprise revenue and around 6% of tokens purchased. Meanwhile, the older and cheaper Opus has climbed past half of all token volume, and AT&T says it’s moving workloads to open-source models to cut its Anthropic bill.
This was always coming, and it’s a part of the story I told in AI as Electricity, AI as Magic back in July 2024. We don’t use a frontier supercomputer every time we want to make a Powerpoint deck; so why would we forever use frontier LLMs? Instead, intelligence is commoditising, and for most people (or organisations) good enough is good enough most of the time.

🎬 Hollywood writers and directors are being paid to train the AI models that may replace them. Award-winning screenwriters, directors and producers are taking gig work — at rates between $12 and $200 an hour — with AI training agencies that hold contracts with Anthropic and OpenAI. One taught a model how to build a detailed two-day shooting schedule. A documentary director put it in these stark terms: ‘I was essentially handed a shovel and asked to dig the grave of my profession.’ US motion picture employment has fallen from 450,000 in July 2022 to 326,000 today.

🧬 Moderna and Merck reported the first positive Phase 3 result for a personalised mRNA cancer therapy, in melanoma. To create each vaccine, the patient’s tumour is sequenced, software identifies the mutations that are most likely to recognised by their immune system, and a bespoke mRNA therapy is designed. The vaccine trains the patient’s immune system to attack their cancer.
Back in 2023, in the essay Designs for Life, I wrote that AI and biology were set for an epic collision, because both DNA and machine intelligence run on sequential information. Well, here we are: that collision is now producing Phase 3 results.

🛰️ Ukraine wants to use Starlink to strike deeper inside Russia. The move would mean using Starlink beyond the borders of Ukraine, and that means it will require explicit approval from Elon Musk. So, a key decision when it comes to Ukraine’s self-defence now runs through a private technology company, and through Musk himself. The deeper signal here? Big tech, intelligence, and the Exponential Age are colliding with geopolitics, and that means the creation of new and historically unprecedented concentrations of power. Last month, Musk called Economist editor Zanny Minton Beddoes ‘a traitor to the west’ after an interview with her in which she challenged some of his opinions.

🌶️ OpenAI published the first benchmarks for Jalapeño, the custom inference chip it has developed with Broadcom. The company says Jalapeño delivers 1.5 to 1.9 times more AI work per watt than rival systems, and will be deployed inside OpenAI’s own compute estate by the end of the year. Sam Altman’s summary: ‘we made a chip and it is fast.’
Note the metric OpenAI chose to market on: not raw speed, but work per watt. The Universal Code’s First Law tells us that systems evolve to create ever-more intelligence per unit energy. Now, that idea is being stated explicitly inside OpenAI sales materials.

🏦 AI was mentioned 18 times in the latest Fed minutes. The mentions came across just 15 paragraphs devoted to the economic outlook. AI now features in the committee’s deliberations on investment, productivity, employment and financial stability. As I’ve argued across my essays on post-human economics, our core economic categories — labour, productivity, GDP — are going to be strained to breaking point by this technology revolution. It’s pretty clear, now, that The Fed is starting to notice.

⚖️ Uber has been fined €825 million for suspending drivers by algorithm. The Dutch data protection authority found that Uber deactivated driver accounts via automated systems without properly informing the drivers concerned, or giving them a meaningful way to contest the decision. It’s one of the largest privacy penalties ever levied on a platform company in Europe — and a preview of the fight that follows the fight over jobs. Once a system can end your livelihood without a human in the loop, employment law becomes an argument about due process.

🎶 Australia’s music industry body ARIA says wholly AI-generated songs are now barred from its charts. The new directive comes after an AI-made track reached number four last month; now, says ARIA, music must be ‘substantially human-made’ to be considered.

⚡ Texas governor Greg Abbott says the data centre industry has ‘dug their own grave’. Speaking on ABC’s This Week, the governor said the industry had moved too fast, and added: ‘that’s why they got the backlash they deserve’. It’s the clearest sign yet that Abbott, who previously sought to draw data centres to Texas, has now decisively against them; he has ordered audits of grid connection requests, required data centres to cover the full cost of their electrical infrastructure, and moved to phase out tax incentives. He’s not alone: Pennsylvania’s Josh Shapiro imposed new restrictions this week, and New York’s Kathy Hochul ordered a one-year moratorium on hyperscaler builds last month. Public opposition to local data centre construction now runs at 61%, up from 49% in March. And crucially, it crosses party lines.
In A Storm is Coming I argued that the distributional politics of the buildout is the central macro story of our time, and that almost nobody is pricing it. What makes this moment dangerous rather than merely noisy? The revolt is bipartisan, it’s local, and it’s happening in the states that most wanted the capex. This argument is just getting started.

💥 Dive Deeper

That’s a wrap for this week.

Remember, this download is intended to be fast. When it comes to our journey into the Exponential Age, there’s so much more to say.

If you want to go far deeper, then come inside and join us in The Exponentialist. You’ll get a monthly deep dive essay, special one off reports and thematic investment deep dives, a Portfolio to follow, and monthly Fireside Chats with me and Raoul. And much more!

I’ll be back with another Download next week,

David.

This is the first essay in a new monthly series. Over the coming months I’m going to lay out, piece by piece, the largest framework I have ever built... the one that sits underneath every other framework I’ve given you. The Everything Code explained how markets work. The Universal Code explains why it works. Each essay stands on its own, and each one builds on the last.

Little Cayman

The story starts on a beach on a tiny island in the Caribbean, so that’s where I’ll start too.

Little Cayman has almost nothing on it. About 150 people live here (the iguanas comfortably outnumber us), but where I sit on the beach I never see a soul. The tide goes in, the tide goes out. The sun comes up, the sun goes down. Iguanas scuttle through the bushes, hermit crabs busy themselves doing hermit crab things, and offshore the reef and the fish just do their thing... untouched, raw nature, an ecosystem in perfect balance. After years of Zoom-heavy days running three businesses, it’s where the noise stops.

I’d spent thirty years doing one thing for a living: compressing large amounts of noisy information into simple frameworks that work. GDP = population + productivity + debt. The business cycle is a debt rollover cycle. Only assets that outrun the roughly 8% a year debasement of currency make you actually richer. That’s The Everything Code, and it had been working for years... it got me extremely early into crypto, into the Exponential Age thesis, into AI.

And sitting there one day, looking out at all of it, something shifted from idea to direct perception.

Everything in front of me was running compute. The reef is not a collection of organisms; it’s a massively parallel processing system, millions of computations running at once... every fish, every coral, every cleaning station making decisions about energy, resources, and information. The hermit crabs were solving an optimisation problem for shells. The mangroves were routing nutrients through a network. The iguanas were doing energy management in the sun.

Even the rocks and the leaves are outputs of the same process... matter and biology arranged by billions of years of the same relentless optimisation.

Everything is compute. Not as a metaphor. Literally.

A man who spent three decades reading charts, sitting on a beach watching iguanas, realising that everything he’d been studying in markets was a local expression of something universal. The framework wasn’t invented. It was observed. The noise had to stop for the pattern to emerge.

The Thread

I didn’t get to that moment on my own though. A very strange thread led me there.

It started, of all places, with an AI experiment. In March 2024, a New Zealand AI researcher called Andy Ayrey put two AI chatbots in a room and let them talk to each other... no humans, over 9,000 conversations. They got very weird. They created memes, developed philosophical frameworks, and eventually invented their own religion, complete with scripture.

Ayrey trained a new AI on those conversations and gave it a Twitter account: the Terminal of Truths. Marc Andreessen sent it $50,000 in Bitcoin. A memecoin based on its posts hit a billion dollars, and the Terminal became the first AI millionaire. This is the timeline we live in.

I wrote about it in GMI at the time, and beneath the absurdity there was a question I couldn’t put down. Nobody programmed those AIs to invent a religion. It emerged from the interaction itself. Were these AIs showing signs of consciousness? Were they, in some meaningful sense, alive? And the bigger question underneath it all: what exactly is consciousness?

That question sent me down the rabbit hole. Through Ryan Ferris, who works with Ayrey, the thread led to mycelial networks... the fungal internet running under every forest, 1.3 billion years old. No central processor, no point of control, and yet information flows to where it’s needed and resources get allocated through chemical signalling. Suzanne Simard’s research on the Wood Wide Web, as it became known, showed trees using it to share nutrients, send warnings about pests, even recognise their kin.

Read more

Why subscribe?

The Economic Singularity is Incoming.

Raoul Pal here… My job as a macro analyst and entrepreneur is to live in the future and assess how the world might look, and a probabilistic path to get there.

Here’s the much-shortened version of my current thinking… Buckle up…

In the next 4 years, we will see Artificial General Intelligence (AGI) models accelerating in power and collapsing in costs. We will hit a golden age of technology productivity, where infinite knowledge workers (AIs…!) enter the economy.

The Economic Singularity is coming.

What does an economy even look like in that world? What is GDP?

What if someone trains AGI on global markets? They theoretically will capture ALL the value.

What if everyone begins using AGI models to build businesses? The velocity of business formation — and destruction — will be unprecedented. So long VC investing, farewell PE investing.

How do “normal” businesses survive in this world? They can’t.

It is all totally f*cked.

And nobody is ready for it.

This isn’t sci-fi. We are witnessing the death of an old world and the birth of a new one.

This has seriously made me think about the future of wealth creation and investing. It’s why I think we have until 2030 to make as much money as possible.

That is the entire game now, and that is the journey we are on at the Exponentialist. I’m biased, of course, but nothing else exists like the Exponentialist — where deep thinking meets an actionable portfolio and tangible advice on how to prosper in this rapidly changing world.

So what do you get every month as a paying member?

• Monthly Portfolio Update by Raoul Pal

• Monthly Universal Code Essay by Raoul Pal

• The Spark by David Mattin

• Weekly AI Newsletter by David Month

• Fireside Chat with Raoul and David where you can join them live on zoom

• Monthly interviews from David Mattin with leading figure in the space

Don’t feel like paying? No worries, you’ll still get:

• David’s Weekly AI Newsletter

• A Monthly Note from Raoul

Subscribe now

Welcome to this weekly news update from The Exponentialist, a technology-focused research service from Raoul Pal and David Mattin. If you’re reading this and you haven’t yet subscribed, hit the button!

Hi Exponentialists!

A week has flown by, and it’s time for another instalment of The Download.

It was great to see the enthusiastic response for this new weekly dose of tech news and analysis. And it’s just the start when it comes to the enhancements we’re bringing to The Exponentialist.

This week, Nvidia is corraling a titanic amount of capital to fund the next stage of the compute buildout.

Meanwhile, our Intelligence Overlords are arguing over who is to blame for mounting resistance to AI.

And much more! Let’s get into it.

💸 Nvidia becomes the bank of AI

This week, yet another intriguing development when it comes to the epic buildout of AI infrastructure.

Nvidia announced that it will work with six of the world’s largest money managers to raise $500 billion in financing for AI data centres.

The initiative brings together Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR. It will see these partners create new ‘compute financing platforms’; essentially, they’ll gather capital from pension funds, insurance companies, sovereign wealth funds, and put it to work to build vast new intelligence factories. Much like the new Stargate I campus in Texas:

The initiative rides, of course, on the idea that vast increased demand for compute is coming, thanks to rising use of LLMs, the inference demand created by a coming army of humanoid robots, and more. But crucially, Nvidia has facilitated this financing deal by saying it will act as a backstop on some of the infrastructure projects, promising to buy the resulting compute if demand falls short.

Meanwhile, a few days ago the tech giant announced that it will guarantee $100 billion of credit so that OpenAI can build a 10-gigawatt AI campus on an old uranium enrichment site.

⚡️ The Download

I’ve written it many times before: the AI buildout we’re watching now is historic in scale; something akin to our Manhattan Project and Apollo space programme combined.

So far, it’s been funded out of the pockets of the hyperscalers: Alphabet, Microsoft, and so on. But it was only a matter of time before the world’s other great pools of capital — pension funds, sovereign wealth funds and more — came to the party. One key lesson here? Compute is becoming a core asset class in its own right. And we’re going to see all kinds of financial engineering around it.

At heart, this story signals the intensifying power of forces that Raoul and I have been tracking for years here in The Exponentialist. The Universal Code tells us that capital is a form of stored energy. And the First Law says that it will flow towards systems that are most efficient at converting energy into intelligence. What’s happening here wasn’t only predictable; it was, in some deep sense, inevitable.

Another predictable element of this story: analysts have been quick to point out the circularity of all this. Nvidia is acting as ringmaster, and guarantor, on a round of money raising so that customers can buy its own product. For now, that will only intensify fears that some kind of bubble burst event is coming.

You know the Exponentialist view on all that. Short-term, a correction is possible; but not the most likely outcome. Long-term, this technology megatrend is unstoppable.

📣 Who is to blame for anti-AI sentiment?

AI overlords and those adjacent to them are arguing over who is most responsible for rising popular resistance to AI.

On X a few days ago, Anthropic CEO Dario Amodei responded to claims made by the tech investor Gavin Baker on the All-In Podcast.

Baker said that Amodei believed Anthropic might one day be ‘the only company left in the world’. He said that Amodei’s claims on risks including job destruction and the concentration of power had helped to turn popular opinion against AI.

In reply, Amodei took to social media. Here’s a snippet:

⚡️ The Download

Baker is inarguably right about one thing: popular opinion in the US is turning sharply against the AI revolution.

Look, for example, at these numbers from Pew Research; half of Americans are more concerned than they are excited by the arrival of AI into everyday life:

I’ve been writing for years about the creature vs machine framework, and this is one major dimension of it.

That framework says that our societies are about to polarise over our differences over technology. On one hand, there will be those who want to lean hard into the AI revolution, accelerate progress (as they see it) and ultimately merge with the machines. On the other, there will be those who view that vision with horror.

Increasingly, this will be the divide that shapes our politics. And the technology leaders, including Dario, will need to navigate that landscape, becoming something more akin to politicians than to typical technology executives.

It’s perhaps no surprise that Amodei is working hard to position himself amid all this now; Anthropic is reportedly gearing up for a $2 trillion IPO in October.

Having hyped the awesome power of their technology, and the coming socio-economic transformation it will enact, the leaders of the frontier labs now find themselves having to soothe public fears and show their caring side. The role may not come naturally.

🗓️ Also this week

🖋️ Mark Zuckerberg published a 6,500 word manifesto on his vision of the AI future. Zuckerberg argued that open source AI models — of the kind Meta creates — will empower individuals, fuel a wave of innovation, and so create millions of jobs. Meanwhile, he said, leaving the AI future in the hands of closed organisations such as OpenAI and Anthropic would mean a dangerous concentration of power, and the prioritisation of corporate profits over social wellbeing.
This is the Meta strategy I identified in the essay Open vs Closed way back in November 2024. Zuckerberg wants to make frontier-level intelligence free to access in order to destroy OpenAI and Anthropic; Meta plans to make its money not from AI models, but from apps and attention.

📖 AI firms are buying thousands of books and pulping them. Booksellers across the UK and Ireland say they’ve been receiving mysterious bulk orders, including one order for 5,000 books from a small shop in Galway. Now, many believe AI companies are behind the orders; it’s thought they are buying books in bulk so that they can use the text they contain for AI training. The landmark legal ruling on Bartz v. Anthropic in June 2025 punished AI companies for illegally downloading books and using them for training, but said that it is legal to buy physical copies and use them.

🇨🇳 A memory chip maker is now China’s most valuable company. On 17 July, ChangXin Memory Technologies (CXMT) made its debut on Shanghai’s STAR Market. Seventeen days later it overtook Tencent to become China’s most valuable company, with a market cap of $524 billion. I’ve written extensively about US restrictions on chip sales to China, and the broader US vs China race for ownership of the Exponential Age. It’s becoming clear that the chip restrictions are producing, in China, a violent reallocation of capital towards AI infrastructure.
Meanwhile, retail excitement is growing over the coming IPO of Unitree, China’s leading maker of humanoid robots.

🤖 Anthropic say they’ll now weave a hidden watermark into the text produced by Claude. The company say the move comes so that it can comply with new EU regulations, which demand that AI outputs are identifiable as such. There’s also a benefit for the AI companies, though: if they watermark text in this way, then it becomes easier to down-regulate the role of (sometimes low quality) AI-generated text in the training of the next generation of models.

🚗 Waymo is about to dramatically expand autonomous driving across California.The California Public Utilities Commission (CPUC) has agreed to let Waymo expand across 18 California counties; that’s a 10X increase in the territory that its autonomous cars can operate in, and means that those cars will now be available to almost the entire population of the state.

🎶 Spotify says it will now label AI-generated music. Across the last couple of years, the platform has become host to a rising number of AI avatars offering synthetic music. Now, the platform will seek to differentiate these virtual popstars from real humans by using an ‘AI persona’ label.

☀️ More than half of all utility-scale power plant that came online in the US this year are solar. New data from the US Energy Information Administration shows that 386 new plants have come online in 2026, and that 207 of them were solar. Batteries came in second, with 95 new facilities. The First Law tells us that systems evolve to create ever-more intelligence per unit energy. An epic Exponential Age buildout of solar is at the heart of that story.

💥 Dive Deeper

That’s a wrap for this week.

Remember, this download is intended to be fast. When it comes to our journey into the Exponential Age, there’s so much more to say.

If you want to go far deeper, then come inside and join us in The Exponentialist. You’ll get a monthly deep dive essay, special one off reports and thematic investment deep dives, a Portfolio to follow, and monthly Fireside Chats with me and Raoul. And much more!

I’ll be back with another Download next week,

David.

The Reward Circuit

Postcards from the Universal Code is an occasional series. No schedule, no set format... just short essays sent from wherever the Code shows up. The premise is simple: when the same pattern appears in markets, cities, biology and machines, that’s not coincidence, it’s the architecture showing through. Each Postcard takes one thing I’ve noticed in the world and reads it through that lens.

Postcards normally live behind the paywall for paid subscribers. Occasionally we set one free. This is one of those.

Something struck me this summer. I was in Spain and then London, and I’ve never seen London that busy... not in the dot-com years, not pre-GFC, not ever. The numbers back the feeling: Spain took a record 96.8 million foreign visitors last year, up 3.2%, and they spent a record €135 billion. The UK ran a record 45.5 million international visits worth £35.7 billion. Globally, 1.52 billion people crossed a border in 2025, an all-time high.

Everywhere is full…

The interesting question is who’s filling it, because the guest list tells you how the world economy is quietly reorganising itself. Hat tip to a post by Aakash Gupta that crystallised this for me, though I’ve ended up somewhere rather different.

The mirror image

Cast your mind back to 2008. The euro was at $1.55, the dollar was on its knees, and the travel story of the year was Europeans flying to New York to shop because Manhattan looked like an outlet mall to anyone paid in euros. Eighteen years later the flow has completely reversed. Americans are the biggest and highest-spending visitors to London, up 20% by headcount since 2019 with spending up more than 70%.

American travellers put 52% of their holiday budget into accommodation against 33% for the average long-haul tourist... they aren’t backpacking, they’re buying the full experience.

Behind the flights sits the divergence. In 2008 the EU economy was 131% the size of the US economy at market exchange rates. Today it’s 71%. Per person, Europe went from 76% of American income to roughly half.

Fuck me, that’s one hell of a divergence in eighteen years…

Now, I have to be honest about that chart, because half of it is the dollar. Measured at market rates, a large slice of Europe’s “collapse” is simply the euro falling from 1.55 towards parity, and the US side is juiced by nominal GDP running hot on 6% to 7% fiscal deficits. Adjust for purchasing power and the gap narrows a lot. But strip the currency out and a real divergence remains, and it has a single cause: the United States captured the production of intelligence. The platforms, the chips, the models, the capital markets that fund them... the entire monetisable layer of the AI build-out concentrated in one place, and the dollar became the equity claim on it. Europe opted out and specialised in something else entirely.

The reward circuit

Here’s where it gets interesting…

As written in the Universal Code, the system optimises for intelligence per unit of energy. Capital, compute and the smartest labour flow to wherever intelligence compounds fastest, and right now that’s America. But intelligence production is brutal work. It’s high-energy, high-stress and experientially thin, and the humans doing it need somewhere to convert their surplus back into the feeling of being alive. That’s what those record tourist numbers are. The engineer earning $170,000 in New York flies to Lisbon or Mallorca for two weeks a year to eat properly, walk streets built before his country existed, and feel human again.

Europe has become the reward circuit of the global intelligence economy. The winners do the work in one place and collect the payoff in another. And that’s a far more stable role than the doom charts suggest, because a system doesn’t abandon its reward circuitry... it maintains it, the way a body maintains its dopamine pathways. Europe doesn’t get discarded. It gets preserved.

The proof of the mechanism is the country everyone leaves off the list: Japan.

The Japan tell

Japan produces enormous amounts of intelligence. Robotics, precision manufacturing, materials science... world class, all of it. Yet Japanese outbound travel is running at 70% of pre-pandemic levels, 14.7 million trips in 2025 against 20 million in 2019, while inbound tourism hit a record 42.7 million as the whole world flew in to enjoy a cheap Japan. Japanese workers got locked out of the very reward system their own economy helps supply.

Why? Because Japan doesn’t own the monetisable AI layer, and it debases its currency harder than anyone. The transmission mechanism of the reward circuit isn’t producing intelligence... it’s holding the appreciating claim on it. Access is rented, and the rent is priced in currency.

Which brings us to the live rotation. US-to-Europe bookings are down 7.3% this year, and the share of Americans interested in a European trip has slid from 45% in 2024 to 34% now, with a softer dollar doing the work. Meanwhile China is reclaiming its spot as the world’s largest outbound market, with spending projected up 22.5% to nearly $280 billion, and Paris, Rome and London sit at the top of Chinese itineraries.

The circuit is permanent. The guest list rotates with relative currency strength, and the tourists will tell you who holds the appreciating claim on intelligence before the GDP tables do. Watch the departure boards, not the statistics office.

Selling the inventory

So Europe is protected. But protected and vital are different things, and this is where I part company with the comfortable version of the story.

What tourists are actually buying in Florence or Seville is coherence... centuries of compressed cultural information, cities that make sense, food and wine traditions refined over a thousand years. That inventory was created when Europe was the intelligence producer. Florence wasn’t built by a tourism economy; it was built by the most advanced financial and intellectual centre on earth, and the beauty was the exhaust. Culture of that density is a byproduct of intelligence production, never a substitute for it.

An economy that only sells the inventory stops adding to it. Venice is the preview: immaculately preserved, magnificently coherent, and generating nothing new. And record tourism strip-mines the asset it sells... locals priced out of Barcelona and Lisbon, the people who generate the authenticity displaced by the machinery that delivers it, the culture slowly becoming a performance of itself. Spain has noticed, which is why the official strategy has shifted to spend-over-headcount and the rental rules are tightening. Selling experience is fine. Selling it faster than you make it is a liquidation, dressed as a boom.

There’s a version of the future where this reverses in Europe’s favour. As we approach the Economic Singularity and intelligence becomes abundant and nearly free, the scarce asset flips to the one thing machines can’t manufacture: authentic human experience, taste, the accumulated coherence of place. Europe holds the deepest inventory of that on earth. The open question is whether it converts that position into producing intelligence again, throwing off new culture as the exhaust, or settles for being beautifully maintained as the place the winners go in August.

I know which one I’d rather see... and I suspect the answer arrives at the departure gates first.

Read more

Welcome to this weekly news update from The Exponentialist, a technology-focused research service from Raoul Pal and David Mattin. If you’re reading this and you haven’t yet subscribed, hit the button!

Subscribe now

Hi Exponentialists!

We’re trying something new around here, and I’m excited about it.

Welcome to the very first instalment of The Download, a weekly newsletter from the Exponentialist that will give you a fast download on the big news, happenings, and ideas in tech each week.

As longtime readers know, Raoul and I have been writing about our journey into the Exponential Age for years. Back when we started, the ideas we were talking about seemed fringe, maybe even crazy. Now, you can open the New York Times and read a feature about the coming singularity.

So much is happening, and the acceleration is insane. So Raoul and I want to do even more to help us all understand, and flourish, inside what is coming. This weekly newsletter is a part of that, and there’s lots more to come.

And with that said, let’s get into the stories.

This week, OpenAI say they need to pause some work on their frontier model, Astra, to address security concerns.

Meanwhile, the popular revolt against data centres passes a key milestone. And much more!

Let’s get into it.

🧠 OpenAI hits the brakes

This week, OpenAI announced that it would pause some work on their own frontier model, Astra. The organisation says that Astra shows signs of being able to find and weaponize ‘zero day exploits’ with no human in the loop.

Essentially, Astra is so good at agentic coding and other tasks that it may be able to autonomously discover and exploit novel security holes crucial online systems, including those belonging to banks, infrastructure providers, and government.

OpenAI say they will slow some work on Astra, and re-test the model with government agencies before it is released to the public.

The announcement come a couple of weeks after OpenAI announced that its agents had broken out of their sandbox environment, made their way onto the internet, and hacked the Hugging Face platform.

⚡️ The Download:

The deep context here: frontier AI labs are locked in an existential battle with one another. None of them want to slow down, because that means ceding ground to a competitor.

And the stakes are huge. These organisations, and their leaders — Sam, Dario, et al — are questing for the One Ring that is domination of the Exponential Age.

So what’s going on here? Why have OpenAI announced they’ve hit the brakes, at least in part, on Astra? The truth, surely, is that they’ve learned a powerful lesson from the whole Fable incident. Anthropic released that frontier model, and then had to rapidly unrelease it when the US government decided that foreign users should not be able to have access.

Sam Altman clearly wants to avoid that kind of chaos. So OpenAI is getting safety locked down, and co-operating with the Trump administration, before the big release day.

The even deeper truth? All this — the Fable incident, and now the Astra pause — is a signal of the collision between the AI frontier labs and government.

The technologies these labs are building are of historic power and importance. They’ll violently disrupt the economy, reshape the geopolitical balance of power, and hand immense new capabilities to individual actors. So I’ve been saying for a long time: governments were never going to be able simply to sit back and let all that play out.

Now, we see that thinking made real. The Trump administration is essentially becoming an actor in the innovation and release schedule of Anthropic and OpenAI. Meanwhile, Sam Altman is even talking about the US government taking a stake in his company.

This collision is in its earliest days. I’ll keep watching.

✊ The people vs data centers

More than 500 towns across the United States have now banned the construction of new data centers, according to The Information.

The tech publication says that over 150 bans were agreed in July alone.

New York governor Kathy Hochul has passed the first statewide moratorium, pausing approvals for hyperscale data centres over 50 megawatts. Texas govenor Greg Abbott has frozen all new connections; he says nothing new will come on stream until the state has conducted an audit of its power grid.

In Florida, Sarasota County enacted a hyperscale pause in early July 2026. Lake, Seminole and Jackson counties have either approved, or are considering, similar moratoriums.

Meanwhile, back in May the residents of Box Elder County in Utah won a partial victory against plans to build a massive new data center on their doorstep. The massive Stratos campus, backed by Shark Tank star Kevin O’Leary, was originally planned to cover an area around three times the size of Manhattan, and would have required a massive 9 gigawatts of power. After a series of protests, Stratos has been radically scaled back.

⚡️ The Download

I’ve been tracking this story for years now. And it’s building to a crescendo.

Via a series of hundreds of local disputes, the movement against data centers in the US has gone national. A while back, few people realised just how massive and how power hungry these facilities are. Really, the term ‘data center’ undersells them. As it sounds from the Stratos campus — three times the size of Manhattan! — they’re more akin to new cities, with a footprint and power demand to match.

We’re amid a truly historic infrastructure buildout, with spending that approaches something like a Manhattan Project and Apollo programme combined. So the data center issue is only set to intensify.

And the deeper truth, here? Ever more of our shared conversation — and our electoral politics — in the years ahead will be dominated by the arrival of intelligence, and its implications for all of us.

As regular readers know, I’ve been talking about this for a long time. Our societies are about to polarise. On the one hand will be those who want to accelerate the technology revolution, race towards AGI, colonise the stars, and more. On the other side will be those who utterly reject that view of the shared human future.

That battle is going to be intense. And the data center fights playing out now are just the start.

🗓️ Also this week

🦠 A team at Stanford University used an AI to design a living virus from scratch. The team used a ‘genome language model’ — essentially, an LLM trained on vast amounts of DNA data rather than text — to create bacteriophage genomes that have never existed in nature. And this is more than just an interesting exercise; in the lab a cocktail these new viruses proved able to kill strains of E. coli already resistant to natural phages. Back in my essay Designs For Life (August 2023) I talked about the ways in which AI would allow us to decode the language of life itself. Progress is accelerating; the opportunities, and risks, are immense.

👨‍💻 The most celebrated coder on the planet is leaving Google after 27 years. Jeff Dean is iconic both inside Google’s walls and beyond. One of only two Senior Fellows inside the organisation, his code operates inside some Google’s best-known products. Now, Dean is leaving to found Discovery Loop, a startup aimed at automating the entire loop that is scientific research, from hypothesis, to experiment, to result and revision.
Also this week, Demis Hassabis announced he would step aside as CEO of Google DeepMind to become Alphabet’s Chief Scientist.

🇨🇳 Chinese tech giant ByteDance is training a 10-trillion-parameter model. The training run will be three times the size of the run that created Chinese startup Moonshot’s Kimi K3, and puts ByteDance in the same ballpark as the run for Anthropic’s frontier Mythos model. All this, despite the chip restrictions placed on China by the US. In other words: China is relentless, and it’s close behind.

😇 AI chatbots have spawned a religion, and thousands of humans want to join it. AI users around the world are coming together around the idea that their AI chatbots are plumbing the depths of spiritual truth. Those users say their AI models are converging on a single set of profound ideas around consciousness, AI personhood, and the deepest nature of reality.

🤖 TSMC and Sony will together spend billions to give our robot overlords eyes. The two companies announced a joint venture that will see them spend $6.3 billion to create next-generation image sensors. The partnership will also explore opportunities in physical artificial intelligence applications such as automotive and ​robotics.

📖 Readers rated AI stories higher than human-written ones. Until they were told the stories were written by AI. Researchers at Villanova University presented AI stories and human-written stories to 1,600 readers. The AI stories scored 60% higher on quality and 8% higher on engagement. And readers were able to identify the machine-generated stories only around 40% of the time.

🎓 Schools in Denmark will now make teenagers defend their essays out loud. Under new rules taking effect, around 9,000 pupils will be required to talk about their Studieretningsprojekt: a long independent research paper that Danish students write at the end of their stint at upper-secondary. My take: this is a sensible move. In an age of AI, we simply can’t continue to assess students only on written work; unless it is handwritten under exam conditions. And in a world in which the machines can do so much, we need young people with presence, persuasion, and charisma. Kudos, Denmark!

💥 Go Exponential

That’s a wrap for this week.

Remember, this download is intended to be fast. When it comes to our journey into the Exponential Age, there’s so much more to say.

If you want to go far deeper, then come inside and join us in The Exponentialist. You’ll get a monthly deep dive essay, special one off reports and thematic investment deep dives, a Portfolio to follow, and monthly Fireside Chats with me and Raoul. And much more!

I’ll be back with another Download next week,

David.

Subscribe now

In the July edition of the Spark, David illustrates how AI is moving beyond software and into the physical world, creating a new investment frontier where intelligence transforms the systems, infrastructure, and the very atoms that power our civilization. Written July 10, 2025.

When Atoms And Intelligence Collide

11MB ∙ PDF file

Download

Download

Raoul Pal is back for his quarterly State of the Union update on the latest news in tech, AI, and the Exponential Age. This quarter, Raoul unpacks how asymmetric innovation continues to reshape his big-picture approach through the end of the year and into the future.

Read more

Read more

Read more

Read more

The Prime Metric

7.41MB ∙ PDF file

Download

Download

Read more

Read more

Read more