Here’s our Monday roundup of data signals across AI, energy and markets.

Enjoy!

1. Job boundaries are blurring. Nearly half of job-specific ChatGPT tasks fall outside users’ primary occupation.1

   

2. Productivity follows use. Employees who use AI across several different use cases are twice as likely to report a positive impact on productivity than employees who use it for one or two types of tasks.

   

3. Agentic patents. Globally, patents for agentic AI use have grown 59% in the last year – now making up 9% of AI application patents.

4. Value chain growth. While the S&P 500 companies are beating expectations by 27% this Q2, Bloomberg’s AI Value Chain companies come out at 71%. Companies along the supply chain are outperforming incumbents.

Read more

“It is a lot to cope with the rollercoaster of the last decade and deep uncertainty of what’s coming next. Over many years, the quality and depth of the newsletter has ensured I am better informed and inspired.” — Hugh K., a paying member

Get 30% off for 1 year

AI adopter’s decision trap

We modeled three types of companies adopting AI. They have the same starting economics, the same 5% hit rate, but different learning practices. Two years into their investment, all three are losing similar amounts of money. In year five, the eventual loser looks best. It takes eight years to see which approach leads to outsized ROI.

Most CEOs are facing the decision trap right now – you likely don’t know if your firm’s spending is learning that will compound to ROI, or waste. The FT calls Zuck “the king of the side quest” as he works away on a portfolio of bets:

Mark doesn’t need each project to succeed, as long as the experiments deepen Meta’s infrastructure and inform the next move. In our model, system builders that consistently compound their learnings over time have the winning formula. See our framework and the interactive model:

Too brittle to go exponential

A $45 billion fund at its peak, started by an AI researcher with no hedge fund experience, betting on the AGI capex build-out at roughly four times leverage, was forced to liquidate this week. Leopold Aschenbrenner started Situational Awareness LP in late 2024 on a thesis that the path to superintelligence would put trillions into compute, chips and power.

He did great. And then, the trade turned. The Philadelphia Semiconductor Index fell 28.6% from its June peak, and software gained.

Situational Awareness’ unraveling is not proof that Leopold’s thesis is wrong.

Read more

One of the surprising effects of working with AI is that I spend much more time reading long-form, books and journal articles in particular. More than before, I appreciate a book as a complete thing, an idea held together by an author, chewed over, possibly for years, and presented in a new way.

There are over 350 titles on the bookshelf in my study. The oldest one is a 231-year-old first edition. The newest is likely an advance reader copy of something I have been sent. As I’ve been working on my new book this year, I’ve really been focused on reading things relevant to that, and I haven’t read as broadly as I normally would.

I do want to share some suggestions of things you might want to read over the final month of summer while trying to be a little bit non-obvious, focusing on older titles or the left field.

As a side note, I’m wildly impressed by how well-read Exponential View readers are from what you share in the community Slack (join if you’re a member on the annual plan) and when we meet in person.

Here you go:

FICTION

Perfection by Vincenzo Latronico

Vincenzo Latronico’s Perfection looks at the meticulously curated life of a pair of digital nomads currently holed up in Berlin. The reality isn’t the Instagram feed. Perfection is my top choice.

I Do Know Some Things by Richard Siken

I don’t read much poetry. But I turned to Richard Siken’s latest collection in an attempt to learn. He wrote I Do Know Some Things after suffering a stroke that wrecked his language skills. It takes you through his journey of relearning who he is. It is prose as poetry, flat and unrelenting.

Notes from Underground by Fyodor Dostoevsky

I’ve also really fallen back into Russian literature for the first time in decades.

Dostoevsky’s Notes from Underground is just a standing rebuke to anyone who thinks humans make great decisions. It’s the protagonist standing tall against every version of utopia.

Fathers and Sons by Ivan Turgenev

Scientific nihilism clashes with romantic liberalism in Turgenev’s novel. This is a tale of generational disruption and of what even disruptors can’t remove. I dropped my copy in the swimming pool.

Read more

Reuters ended 2025 with “Companies still waiting.” Seven months into 2026, the waiting continues. Barclays says that broad adoption of AI has not yet lifted productivity.

Executives are under pressure to show they can deliver – half of all CEOs BCG surveyed worldwide say their jobs depend on getting their AI strategy right. Public disclosures of net AI returns are patchy. JPMorgan’s estimate of $1-1.5 billion in value from its AI use is a rare case of a company naming a number.

Adopters are spending a lot, so where are the returns? That is the question most are asking right now. And yet, in a successful technology rollout, the first visible economic signal may not be the returns. Winners and losers might look the same. We have created a model to show why this is the case and what signals to follow to understand if your AI adoption is going well.

Members of Exponential View get access to the full interactive model to test the assumptions behind today’s essay.

Subscribe now

Learning costs

All investments follow a path. You might begin by buying an office building, starting in the red. You earn a return by leasing it to tenants, which can bring you into neutral and, if all goes well, you’ll climb to profitability.

Investing in new technology can follow a similar path. Much of the upfront cost is learning how to use the technology – new processes, skills training, making changes to the organization. Mistakes are almost guaranteed, and learning is expensive. The learning bill will almost certainly arrive before your returns.

Learning is a continuous practice, not a one-time exercise. It happens through a series of projects, each with its own investment J-curve. A company might have dozens of projects at different stages of maturity running at once.

If we adopt the premise that the AI economy is going through a J-curve – firms are investing upfront, learning through deployment, and scaling what works – at the aggregate level, this can make a successful rollout look expensive, even irrational, before it looks productive.1

Our model has three archetypes of companies experimenting with a general-purpose technology, in this case AI:

Archetype 1: The bounded adopter

Bounded adopters find something that works, put it to work, and then stop experimenting.

In 1976, the NYSE’s Designated Order Turnaround system allowed member firms to send small orders to the floor electronically, bypassing the human broker who would normally carry them. Even as the system caught on – by 1999, more than 90% of orders arrived this way – it automated only the delivery of orders; human traders still executed the trade. In 2000, NYSE’s market structure committee rejected a fully electronic order book and chose to keep the floor and its specialists.

But competitors didn’t wait. By 2005, Nasdaq, which already had automated execution, was handling about 15% of trading in NYSE-listed stocks. Eventually, NYSE switched. It merged with the all-electronic Archipelago in 2006 (a combination then valued at $9 billion), and in 2008 the SEC approved a plan that phased out specialists.

Borders, an American book retailer, is another example of bounded adoption. In 2001, it entered into an agreement with Amazon to run its e-commerce site. At this time, Borders was one of the top operators of bookstores in the world, and the Amazon deal helped it maintain an e-commerce site. But that’s where Borders stopped developing its in-house online capability, and its growth remained anchored in physical stores. Only in 2008 did Borders bring its own e-commerce site back in-house, ending the Amazon agreements after nearly seven years. By then it was too late and Borders filed for bankruptcy in 2011.

Archetype 2: The project accumulator

The project accumulator keeps exploring, but rarely or never learns. It launches new projects without figuring out what separates the winners from the losers. Nothing carries forward, so each project starts with the same odds as the last.

In the 1980s, GM made multiple automation bets at once. It bet on factory robots, modernized plants, a $2.5 billion acquisition of a data-processing firm; it created Saturn, a new car brand subsidiary with a new factory and labor arrangements; and it bet on NUMMI, a joint venture with Toyota. By 1986, GM’s capital spending was going to hit $10 billion.

Of all the projects, NUMMI seemed the least likely to succeed. Toyota got GM’s worst-performing factory and rehired the same workforce that was let go when the factory closed down in the past. Under new management, NUMMI outperformed every other GM factory. GM saw this happen, knew what was working well, but for various reasons, the learning traveled too slowly to be transformative.

Read more

Hi,

Here’s our Monday roundup of data signals across AI, energy and markets.

Enjoy!

Subscribe now

1. Rents for the incumbents. OpenAI, Anthropic and Google take 90% of spend on Vercel’s AI Gateway, but only make up 52% of the tokens. For the average token, the Big 3 generate more than 8x as much revenue as the rest of the field1.

   

2. Selective use. AI reaches 68% of occupations in the US; within professions, it covers ⅕ of a typical job’s tasks.

3. Cheap fables. Users seem to find output from Cursor’s Router on Auto Intelligence mode just as good as Fable at ~60% lower cost.

4. Utilization, not transformation. US labor productivity is up, largely because firms are running the existing capital harder, with little change in total factor productivity (TFP)2.

   

5. Drugs good for the economy. GLP-1s cut long-term sickness leave by 17% over four years in Denmark. It didn’t move employment or pay; the benefits accrued to employers and public finances.

A MESSAGE FROM OUR SPONSOR OKTA

Secure your AI agents

Governance and oversight are the biggest concerns among teams adopting agentic AI – for a good reason. Your agents can read sensitive information, take actions for you and work for a long time without oversight.

The companies least at risk will be those with governance frameworks that help them scale AI with confidence.

Okta’s 5-minute assessment will give you a score of how secure your AI agents are today and show exactly where the risks are.

Start your assessment

6. Unicorn central. The Bay Area accounts for 91% of the generative AI unicorn market cap and 39% of the market cap, including all unicorns.

7. In the hands of the few. In the US, 5% of VCs generate 90% of investment profits.

8. Cities without kids. The number of kids under 5 in large US cities has dropped 15% in the last decade, even in cities where the population is growing.

9. Policy impact. Despite a late start, electric vehicle sales in Latin America are catching up to the US thanks to tax breaks and other incentives.

   

Thanks for reading!

Hi,

Welcome to the latest Sunday briefing! I am off on holiday for a couple of weeks. The team will continue to tend to Exponential View while I’m gone, so you won’t miss a beat.

Azeem

Distillation grey zone

On 1st September 1789, Samuel Slater set sail from England for New York. He’d learned that the Pennsylvania legislature had recently passed an act awarding £100 to a British textile worker who smuggled high-end machinery into the state. America was going to great lengths to acquire industrial know-how, by any means possible.

Britain had made it illegal to export textile machinery and technical drawings. The ban extended to prevent textile workers from emigrating. Slater had worked in mechanized textile mills and saw his chance.

He committed the entirety of Richard Arkwright’s system – the first factory method for spinning cotton – to memory. He disguised himself as a farm laborer and broke the laws of his nation as he carried himself across the Atlantic, valuable know-how secretly distilled into his brain.

By 1790, Slater was a partner in a cotton mill in Pawtucket, Rhode Island, built from the plans he memorized. President Jackson called him the “Father of American Manufactures.” He died in 1835 worth around a billion dollars in today’s terms.1

Today’s question is to what extent are Chinese AI labs distilling the outputs of American AI models – and is it really a problem? The easy answer is the hawkish one: hugely and yes, it is. But it’s not the only answer.

First of all, distillation is a decades-old machine learning technique in which a larger model can train a smaller, more efficient model. A lab running distillation internally is not an issue. But it is possible to distill a model from the outside (even without the provider’s permission). This is the accusation against Kimi and other Chinese labs.

Almeida, a four-year veteran of OpenAI, explains that effective distillation is much harder today than a few years ago, when AI models helpfully provided their reasoning traces. What Almeida calls “behavior parroting” is to learn from final answers, the least powerful approach but might still work well to help bootstrap another model.

There is substantial evidence that both Chinese and American researchers have trained models on the outputs of frontier systems. Stanford’s Alpaca project has admitted as much. Anthropic has alleged that DeepSeek, Moonshot and MiniMax have used more than 16 million Claude chats via 24,000 fake accounts. Michael Kratsios, Trump’s science chief, says he now has evidence of how Moonshot ran distillation attacks.

Anastasios Angelopoulos, the CEO of Arena, a benchmarking company, makes the case that Kimi K3 is exceeding the performance of some of the top US models, something distillation alone doesn’t allow, and he predicts that “American labs will start distilling Chinese intelligence.”

It isn’t clear that distillation is illegal yet. points out that “[t]here’s no legal precedent that model outputs are IP.” The US Copyright Office’s 2023 statement on AI confirms as much:

When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship. As a result, that material is not protected by copyright.

Unlike the case of Samuel Slater, who knew he was breaking British law, the problem here is a case of the exponential gap: the technology has stepped ahead of the law.2 If labs have IP in outputs of their models, they will essentially have IP in every future economic activity of all their users.

It also weakens the labs’ own position – why should AI models be prevented from consuming Anthropic’s IP when Anthropic is allowed to consume yours and mine?

Bahrad Sokhansanj has some other sensible suggestions, summed up as follows. Address distillation but scope it narrowly, only focusing on the real harm done, with the right instruments. Is it about national security? Or something else? Regulate more broadly, and this will play into the labs’ desire to skew the regulatory field in their favor.

Again, there is precedent: in 1824, once he was settled as a prominent American industrialist, Samuel Slater lobbied for protectionist tariffs to stifle foreign competition. By then, he was also known as “Slater the Traitor” back in his hometown.

See also:

• Will Kimi K3 change the economics of AI? Our analysis.

• Satya Nadella was one of many tech leaders to force the case for open-weight models to quieten rumors that the American administration was considering limiting them. Jensen agrees:

A MESSAGE FROM OUR SPONSOR, OKTA

To get AI security right, take care of identity

When you and your team deploy AI agents, identity becomes your strategic infrastructure. An agent is an actor that reads your data, calls APIs, and does things on your behalf. You need to give it clear permissions and to audit its trails.

Okta’s AI Identity Readiness assessment will score the security of your AI agents and show exactly what you need to fix to go to production with peace of mind.

Run Okta’s 5-minute survey to evaluate your agents.

START HERE

Want to sponsor Exponential View? Get in touch.

China’s chip chase

In 2015, Made in China set a target of 70% self-sufficiency in semiconductors within a decade. It was often mocked as fanciful and missed wildly. Data from Morgan Stanley now shows that domestic suppliers will have met about 41% of China’s AI chip demand in 2026, up from 20% in 2023. Beijing may hit its 70% threshold five years late.

Compute-weighted, the 41% figure delivers less compute compared to Nvidia, but for the purposes of strategic autonomy, the quality gap is increasingly a non-issue.

The spark was Washington’s export restrictions. “If the U.S. hadn’t forced our country, our company and our industry into a corner, we would never have done something like this”, says Huawei’s deputy chairman. It has become an “all-out push” according to this excellent reporting.

A leaked conversation between DeepSeek’s boss, Liang WenFeng, and several investors supports this. Liang says:

What’s the gap with the U.S.? Only one thing: resources. We don’t have enough GPUs – our count is still small. […] Domestic chips now have a historic opportunity. Previously, adaptation was hindered by poor ecosystem… But that’s changing. NVIDIA CUDA’s moat is eroding rapidly.

Full transcript and context at ’s blog.

Subscribe now

We the people

Peter McCrory, Anthropic’s Head of Economics, points out that the US labor market has shrugged at AI. Unemployment is at 4.2%, and Anthropic’s data finds no worsening unemployment even in the most exposed occupations.

AI augments rather than replaces, for now. Not a single profession has been 100% handed over to machines yet. Every job still needs human effort. It is changing how work gets done, and if workers get more productive, value shifts inside existing roles, and those who use the technology best stand to benefit.

The final hard-to-automate tasks, the “weak links”, as Professor Chad Jones calls them, are the things only a human can do. Companies will need people to get them done, and this protects employment, keeping a decent share of income in human paychecks

Here is another take. People still matter and will continue to matter. Europe creates far fewer successful innovative companies than the US. One reason I’ve often argued is the simple cost of changing the workforce. I think of startups as exercises in making mistakes and learning from them. Every additional cost to making a mistake means an opportunity to learn not taken. Yoram Wijngaarde finds a simple relationship (correlation is not causation) that shows that the more expensive it is to let go of staff, the lower the rate of unicorns per capita.

Prophet motive

A new biography of Jean-Paul Marat, one of the leaders of the French Revolution, reviewed in the current LRB, is worth reading for anyone trying to make sense of today’s AI debate.

Stanford historian Keith Baker3 has a new biography of the journalist and politician. He argues that Marat hates mediation of any type, from Newtonian formulas to parliamentary assemblies and calm discussion- anything that stands between the people and the truth. He tried to write a daily pamphlet, shouted rather than argued and manufactured intimacy. “By making his journal ‘more interactive, more dynamic, more personal’, he fashioned an intimacy that allowed him to speak for the people.”

In amongst this, his paranoia did help him identify real corruption and institutional betrayal. Baker calls him the first modern populist.

High-frequency publishing, paranoia as analysis, a parasocial closeness and the constant insistency that any complexity is just conspiracy in disguise… Well, AI discourse is now selecting for exactly this Marat-like temperament.

While today’s keyboard warriors carry none of the physical violence of Marat’s Terror, there is a similar underlying logic against nuance and complexity. Doomers and accelerationists share patterns – purge rhetoric, aggressive polemics, and the framing of every whiff of nuance as corrupt. What is left are the extremes. Call to mind imminent economic disaster; catastrophic fraud; utopian abundance… or, simply, the transformation of the human condition.

What can get lost in these extremes is the reasoned position that admits and examines evidence. A position that balances probabilities and accepts answers might be complex, incomplete and – contingent.

See also:

• 💪🏼 Really stoked that AMD’s CEO Dr. Lisa Su opened her keynote with Exponential View data. You can get the same data here.

Short morsels to appear smart at dinner parties

Why AI-assisted bioweapons won’t kill us. via EV member

🏋🏼‍♀️ The share of UK businesses using AI has nearly tripled since 2023, but most firms are still dabbling.

Arsenal FC is building AI models for football (soccer).

Young people are more likely to gamble in financial markets when important life goals (like buying a house) feel out of reach.

Global air and sea surface temperatures are headed for a new record.

Good news from the Amazon: wildfires are at a record low this year and deforestation is at a 10-year low. h/t EV member

Vintage LLM 😎Training AI models only on pre-1931 texts help researchers study what AI can do without contamination from the modern web.

👀 Stripe is in talks to buy OpenRouter.

Intel is shipping the first chips with layers printed on ASML’s $380 million High-NA EUV machines.

Know thy maths. breaks down why the EU’s 46% electrification target by 2040 is mathematically unachievable. Relevant for anyone working in policy, really.

🐳 Orcas preparing food for their young? Amazing.

Thanks for reading!

Kimi K3 has caused quite an uproar since its release last week. It’s the first time a Chinese model has taken the lead on the frontend Code Arena benchmark. And that’s three months since Moonshot AI’s previous impressive flagship model, Kimi K2.6, was released.

Following in Moonshot’s steps, Alibaba announced over the weekend that Qwen3.8 – a 2.4 trillion-parameter model – is coming soon, and unlike its last release, this one will be an open-weight model. No benchmarks or further details have been released as of yet.

Open models are now estimated to be 4-7 months behind the frontier in cyber capabilities, down from 6-10 months in 2025. And despite compute constraints, efficiency improvements mean these labs are doing more with less. Comparing the compute availability and model performance between US labs and Chinese labs, we estimated Chinese labs to be getting 4-7x more out of their compute.

and I spent some time with the Moonshot AI and Alibaba teams in China back in April and May, and we’ve had time to think about the economics of open-source models and how they affect the entire ecosystem.

Does Kimi K3 break the economic case for AI?

Some have claimed that Kimi K3’s performance breaks the economic case for AI as it lowers the cost to complete various tasks at frontier standards. For instance, Microsoft engineers are reportedly testing whether Kimi K3 can be used within Copilot.

We don’t think this is the case, and in today’s post we’ll work through what might happen next.

In The State of the AI Economy report, we found that token usage is elastic across providers. This means that every drop in token price leads to a larger increase in token volume, more than offsetting the difference.

For every 10% price cut, token consumption rises 12-18%. A paper by Demirer et al, found a similar effect: a 10% price cut resulted in an 11% or so increase in volumes, which economists call an elasticity of -1.11.

The net effect is a rise in total token spend. But note that the effect is a weak one, not the cantering Jevons’ paradox sometimes presented. Reality might tilt the scales further in favor of more, not less, demand. Workflows are becoming more token-intensive as we rely on reasoning models and verification and approval loops. And the early evidence suggests that firms that adopt AI early tend to increase their relative spend alongside growing headcount. These effects might be short-term elasticities rather than ones that can be sustained for decades, but for now they indicate that falling prices increase volumes and, with that, revenue.

Subscribe now

Flowing down the stack

The model weights may be free, but the inference is not. Kimi K3 has 2.8 trillion parameters. The weights alone occupy 1.4 TB. It needs to be served on something like a 72-GPU NVIDIA GB200 NVL72 rack or equivalent. That’ll cost $3-4 million to buy and install. Operating it consumes about 120 kW continuously, over a million kWh per year, before you consider networking, storage, cooling, and humans. If you rented these in the open market, it would cost about $7 million a year.

But for infrastructure providers, the economics of hosting open-source models can be very attractive compared to serving closed-source models. A simple way to understand this is to think of the hyperscaler as needing to pay a license fee for a closed-source model but not for an open-source one1.

Read more

Here’s our Monday roundup of data signals across AI, energy & markets.

Enjoy!

1. Kimi can code. Kimi-K3, released last week, overtakes Fable 5 and GPT-5.6 Sol to take first place on the frontend code arena benchmark.

2. Demand unlocked. DeepSeek nears $500 million in annualized revenue with gross margins as high as 70-80% on its V4 model.

3. Ask and receive. Tests of leading AI models on prompts based on real terrorist cases show that a third of responses would’ve provided useful help to attackers, and labeling the same prompts as “research” increased compliance from 17% to 42%.

Read more

“You inspire me to think more exponentially. — Robin D., a paying subscriber

Subscribe now

AI AND GEOPOLITICS

Deep Kimpact

Moonshot AI’s new open model, Kimi K3, might be more of a shock to the US than the original DeepSeek model. As a model, it is very good.

I’m wary about benchmarks because benchmarks aren’t the real world. But the emerging consensus appears to be better than Claude Opus 4.8 and, in some cases, on par with Claude’s Fable and OpenAI’s GPT 5.6. The real question, of course, is on which dimensions does K3 beat the frontier labs, and for which workloads are those dimensions important?

Price-wise, it is expensive for an open-weight model. According to Artificial Analysis, it’s about the same price as GPT 5.6 Sol but about 24x more expensive than DeepSeek V4 Pro. Indeed, on a per-token basis, it is only half the price of OpenAI’s GPT 5.6 Sol, far from the usual price advantages of Chinese models.

For the AI economy as a whole, for companies around the world, for governments that aren’t rich, this is probably a net positive. The inference margins that OpenAI and Anthropic enjoy are significant, and they can maintain them because they have the very best models. But it’s pressure, not displacement. Enterprises don’t buy on price alone. They value security, support and possibly the fancy professional services on offer. And the harnesses OpenAI and Anthropic have built remain a differentiator.

As we argued in the State of the AI Economy, token demand is elastic. Falling prices drive demand, and that demand drives infrastructure usage. Hyperscalers and neoclouds will serve these higher-end open models, further fueling demand for compute and everything around it. This pushes more of the revenue pool towards the compute layer and away from the model layer margin. This strengthens rather than weakens the infrastructure payback case—and, of course, the chip and memory suppliers that sit below them. A tempering note: cheaper intelligence still waits for monthly management meetings and a slow-moving approval process.

A MESSAGE FROM OUR SPONSOR, OKTA

Get your AI agent security right

Your AI pilot worked, but now comes the hard part – deploying AI agents across a company that wasn’t designed for autonomous software.

To deploy your agents safely, you need to know what AI can access, what it can act on, and where accountability sits.

Okta’s AI Identity Readiness assessment will score the security of your AI agents and show exactly what you need to fix to go to production with peace of mind.

Run the 5-minute assessment to benchmark your organization.

Start here

ENERGY

Solar will be fine, or will it?

Lazard’s latest energy cost report shows that the levelized cost of solar photovoltaic electricity has risen in the US on a year-on-year basis. Back in 2021, this was $38 per MWh; in 2026, it’s $69. The price of gas generation has also risen from $60 to $90.

You’d be right to point out that we’ve long argued that because solar panels – one of the key cost elements for solar power generation – are on such strong learning curves, the price will keep trending down.

To make sense of this, I looked at the evolution of solar electricity costs in thirteen markets between 2020 and 2025 using data from IRENA, the International Renewable Energy Agency. The headline story is that, yes, PV modules remain on an aggressive learning curve, with unit costs dropping as production increases. Overall systems costs continue to trend downward, but the levelized cost for delivering electricity has risen slightly since 2023.

See also:

• In June, solar supplied a record 25% of the EU’s electricity, more than nuclear, gas, wind and hydro.

• Households in three Australian states are getting three hours of free power every day during peak solar hours.

AI AND SOCIETY

An Anne-style bargain for AI

Britain’s first experiments with copyright began in the decades after William Caxton introduced printing in the 1470s. The Crown licensed a single company to police what went to print, and in return, its booksellers had an exclusive right to copy the texts with no end date. This license manufactured scarcity for over a century.

In 1695, Parliament refused to renew the Licensing Act, which enabled the booksellers’ monopoly. That 15-year interregnum brought an explosion of ideas: London gained 70 political periodicals (from one) and print culture spread to the provinces and American colonies. Then, in 1710, Parliament enacted the Statute of Anne: the world’s first copyright law. It gave exclusive rights for a fixed term, just 14 years, renewable once, and then the work went into the public domain. The law’s title was “An Act for the Encouragement of Learning.” Enough scarcity to incentivize creation and freedom after that so that knowledge would compound.

Twenty-first-century economics agrees. Joel Mokyr took the 2025 Nobel Prize for showing how useful knowledge becomes self-generating. Scientific understanding enables new technologies. The problems encountered in applying those technologies stimulate further science. And a society open to new ideas allows each advance to become the foundation for the next. Knowledge does not simply compound; it helps produce more knowledge

In a world of AI, this compounding will be doubly true. And will set up greater tension between content industries, who, like Britain’s booksellers in the 17th century, will want to protect their old business models, and the potential to drive open-source AI models and a raft of complementary startups. Brian Williamson argues that the EU’s Anne-style settlement- its text-and-data mining exception is key to the EU staying at the AI frontier.1 It allows AI to learn from lawfully accessible material:

Machines as well as humans should be free to learn; what matters in terms of protecting creators is whether outputs, not inputs, duplicate existing work.

For Europe, woefully behind in semiconductors, compute infrastructure and foundation models, yielding to copyright lobbies would further weaken its relative position in AI.

Subscribe now

MISC

Short morsels to appear smart at dinner parties

AI impact on jobs: Junior roles are being “seniorized,” and employers are more open to humanities graduates.

Russian soldiers’ average survival time after reaching the front is 20-30 minutes.

💸 Prediction markets are starting to bet on AI compute costs.

estimates that China will have a Mythos-like model in February 2027. His full analysis is worth reading.

Claude’s personality changes depending on the model and the language you use with it.

💪🏼 There’s now a promising candidate for the first vaccine to prevent pancreatic cancer.

Apple is testing PrismML’s tech to run big AI models directly on iPhones.

Good short essay by : “We tend to conflate power-seeking AI and superintelligent AI.”

😷 How Palantir embedded itself in the UK state, an investigation: “Despite having no real history of working with health data, Palantir began positioning itself as the go-to expert and Global Counsel started hiring Westminster insiders who had contacts in healthcare.”

Thanks for reading!

Here’s our short Monday roundup of signals to kick off your week.

Subscribe now

1. Another scaling law? ByteDance researchers found that newer AI models learn on the job1 about twice as fast as models from just three months earlier.

2. Execs talk down job cuts. The share of CEOs expecting significant headcount cuts from AI fell from 46% in January 2025 to just 20% in May 2026.2

3. AI’s audience splits. ChatGPT’s user share fell below 50% for the first time in March.3

Read more

“With so much hype around the tech, your no-nonsense unbiased assesment is essential.” — RB, a paying member

Subscribe now

Reading is dead (again)

Roberto Serrano, a professor at Brown, suspected his economics class was relying on ChatGPT to do their exams. He made the final paper a closed-book exam, and scores for 56 of his 59 students collapsed (by as much as 100%). Kudos to the two students who seem to work unaided.

This is a problem of incentives. The students appear to value the high score rather than mastering the subject. Perhaps because the value of a degree is increasingly less about intellectual excellence and more about job-market signalling.

The percentage of Americans who read for pleasure on any given day fell to 16% by 2023, down from 28% in 2004. Literacy, not a natural human state, is a learned skill that needs practice. Americans aren’t practising – postliteracy loves short-form video, after all. These data are from before the rise of generative AI.

Reading in America is a different problem. The data, even pre-AI, is terrible and doesn’t support the notion that post-literacy arrived after ChatGPT. Rather, the trend has been ongoing for at least two decades.

When access to intelligence is uncapped, AI could divide people based on our willingness to think and engage with what is difficult. David Brooks makes the case :

What really matters, therefore, is not brainpower but the willingness to run the mental marathons that produce high-quality results. […] The crucial task before us is to cultivate people’s desire to seek out cognitive complexity. How do we train people to see their life as a hero’s journey in which they take on difficult missions that they may fail at and that will certainly involve pain and suffering?

Reading long-form, constructed arguments that have been closely fought through by an author forces the reader to engage with the material far more deeply than a stream of summaries does. (These may give the illusion of thinking, but that isn’t the case.) I’ve been working on my second book over the past few months, and yes, R Mini Arnold, my agent, has been an extraordinarily helpful research associate. The latest AI models, Fable and GPT 5.6, can be prompted to produce outstanding (almost) research, but you really need to know what to ask and how to ask for it. In my case, that’s meant building up my mental map the old-fashioned way. Which means I need to sit quietly, read original material, consider it critically and handwrite my notes.

Elsewhere:

• LinkedIn is awash with AI-generated posts; Substack is less so. I have built a Chrome extension that hides AI-generated content. It makes X more manageable to browse.

• The University of Chicago Law School is piloting device-free first-year core classes and requires students to learn to use AI effectively.

• “Fable is better than me at my job, but Fable alone would be a mediocre investor,” says one VC.

This isn’t the demand softening you are looking for

No, GPU demand is not softening. Silicon Data’s one-year H100 contract index bottomed near $1.70/hour last October and has since rebounded ~38% to $2.35. Spot prices are up 10% this year.

The SpaceX S1 offers some clues about how future demand might shape up as it discloses the infra deals the firm has signed: a three-year tenor with unusually permissive 90-day cancellation terms.

Read more

Here’s our short Monday roundup of data signals across AI, energy and markets:

1. A GPU wave is ahead of us. More than 95% of the Grace-Blackwell GPUs have not yet been deployed, even though the chip has been shipping since December 2024.1 (h/t )

   

2. Freelance hunting. Fable 5 now completes 16% of real freelance projects at a quality just as good as that of human professionals, according to the Remote Labor Index. This is about double the previous best2.

   

3. David vs Goliath. A 35-billion-parameter model, trained differently3, now matches 1-trillion-parameter models on some long-horizon benchmarks.

4. Chinese pharma boost. Chinese biotech companies’ licensing deals are up 87% in the first five months of the year compared to the same period last year.

5. Wet-lab assistance. In a near-autonomous loop, GPT-5.4 helped Molecule.one’s lab run 10,080 reactions and increased the average yield in the Chan-Lam process4 by around 50%.

6. Holding swap. Central banks worldwide now hold more gold than US Treasuries – this is the first time since 1996.

   

7. Wealth concentration. Nearly half of all income earned in the US belongs to the top 10% of earners, the highest since WWII.

8. Fewer babies, richer adults? Across countries, a one-percentage-point lower birth rate in 1950 was associated with almost 27% higher GDP per working-age adult over 1970-2020, with no change to aggregate GDP.

9. Extreme heat. French utility EDF took almost 10% of its nuclear fleet offline (some 6.2 gigawatts) as river temperatures breached the threshold for cooling.

Thanks for reading!

“Always an excellent perspective on emerging systems and their impact across the human landscape.” — Neill K., a paying subscriber

Subscribe now

The latest on AI and jobs

Our friends at Ramp and Revelio Labs released fresh data on AI jobs impact, based on more than 21,000 US firms. They find that heavy adopters grow headcount faster, not slower. These firms increased employment by about 10% over two years after adopting AI. Entry-level roles grew even faster, at 12%.

[H]igh-intensity AI firms are selecting different kinds of candidates. In this case, we believe they are selecting for a new set of skills, specifically, people who know how to use AI and use it well. Entry-level workers, especially recent graduates and college students, are a natural place to look.

We’ve written before that the widely accepted narrative that AI replaces jobs is too simple – this still holds. The opposite claim, that there’s nothing to worry about, is simplistic as well. Labor markets are complex but we can make some assumptions about what’s going on:

First, complementarity. If AI makes workers more productive, firms may want more workers because the return on each additional hire rises.

Second, supervision. As AI-generated work increases, firms may need more people to manage, review and quality-control that output.

Third, demand expansion. If the cost per task falls, more tasks become viable. Latent demand becomes actual demand. (Exactly what has happened with computing since the 1970s.)

Some firms are now learning that they fired people too quickly, mistaking task automation for human obsolescence. As we argued, the initial gains from AI show up in individual productivity, but the harder prize comes when firms redesign entire workflows and decision-making loops around it. There’s no evidence so far that humans aren’t needed in this redesign.

See also:

• Medicine is trying to protect against “never skilling,” the risk that trainees rely on AI so much that they never develop clinical judgment.

• Goldman Sachs economist Joseph Briggs expects AI adoption to temporarily displace about 9% of the US workforce over a 10-year transition.

Subscribe now

Open by necessity

US chip controls have driven China to treat open-source as resilience infrastructure, a new paper argues. Following each major US export control event since 2022, forking of LLM repos on GitHub jumped among China-linked developers but barely moved among US developers – 0.143 additional forks per repository-week for China vs 0.012 for the US, an 11x gap:

When uncertainty around upstream inputs rises, developers appear to increase engagement with open, locally runnable model infrastructure… [This is] a broader shift toward distributed innovation ecosystems that can expand participation, accelerate diffusion, and increase resilience under geopolitical and technological constraints.

Qwen and DeepSeek spread into research and commercial work globally almost as quickly as the best US models. But when authors examined US patents, the use of Chinese-origin models was rarely disclosed.

Another new paper suggests that Chinese innovation is becoming more self-reliant. The share of science produced in China that underlies domestic patents has grown from 1% in 2000 to 26% in 2025. China still builds on research done elsewhere, but domestic research is growing.

See also:

• Notes from our 10-day visit to major AI labs in China.

• One company you likely never heard of controls a large share of the optical transceivers that make every major AI data center work.

• Wars are depleting global tungsten stocks. China mines 80% of it, and the West is scrambling to reopen its mines.

Progress needs maintenance

Read more

Here’s our Monday roundup of data signals across AI, energy and markets.

Enjoy!

1. First, a chart from our inaugural state of the AI economy report. AI quarterly revenues are now exceeding the quarterly AI capex depreciation expense, but have not yet covered cumulative historic depreciation — let alone the additional headroom needed for a healthy margin.

   

2. China’s young guns. On average, Chinese AI labs hire talent with just 1.6 years of experience vs. 5.5 years for comparable roles in the US. (h/t )

   

3. The weight-loss dividend. Previously not-employed women who start GLP-1 treatment are about 27 percentage points more likely to be working after 18 months.

4. A new giant in town. SK Hynix’s market value passed Samsung Electronics’ for the first time.

Read more

Om Malik died on Wednesday. He was one of tech’s truest voices, as a journalist, a founder, an investor, a questioner and a photographer. He understood, before most, that technology is a human endeavor, not just an engineering one. Over the past 15 years, I’d seek him out regularly on my trips to the Bay Area.

He will be greatly missed.

Azeem

The state of the AI economy

We published The State of the AI Economy report this week. It’s the first research report we know of to dissect the demand side of the AI economy, which is critical for understanding where AI is heading.

I’ve been in this research for months, and after all that time, one chart still stands out to me. It’s the chart that shows the break in the 50-year compute growth trend.

So, what’s so special about it… I’ve been tracking the global stock of compute for the past five years. This has involved building a model of the total number of computers of all types (mainframes, minicomputers, PCs, laptops, servers, phones, IoT devices) in the world and making reasonable, bounded estimates of compute horsepower. Up until 2023, the trend line was pretty clear: roughly 66% compounded growth in global compute stock, even across major platform shifts, from minicomputers to PCs and from PCs to phones.

Trends with a five-decade history have momentum, so to break this trend requires something special. The last time this happened was back in the mid-1990s, as businesses around the world inched past Solow’s paradox1, then accelerated by a growing consumerism of computing via Windows 95 and the Internet. Technology helped; the Intel Pentium had substantially better floating-point capabilities than earlier x86 chips. It arrived on the scene in 1993 – out of reach for my student budget.

The trend line reverts back to the mean around 2006. Dennard scaling2 broke: chip makers had to move to multi-core architectures, which don’t give a smooth scaling of FLOP capacity.

At the same time, the PC market had matured, and the development of mobile computing exploded in volume, but mobile processors are optimized for their physical constraints, rather than processing power. Cloud computing has centralized a lot of operations; it also emphasizes efficiency over raw power, and the shift from owned-and-operated servers meant higher compute utilization over a (relatively) lower base than otherwise.

Which brings us to the shift starting in 2020 as AI accelerators begin to make their mark.

Today’s AI runs on compute – floating point operations (FLOP) – as its central input. And we’re bringing more FLOP-factories online. Will this trend also revert to the long-term mean, just as the PC wave did? My guess is that this current pace would be sustained for a few years, perhaps even a decade or more, before reverting back to the long-term trend line.

See also:

• My Bloomberg Technology interview about our report.

• Coverage across Bloomberg, WSJ.

We are all managers now

Last week, we wrote that AI-native firms are flatter and have fewer managers than non-AI-natives. While this is true, for the sake of precision, we should say fewer managers of humans. With AI, every frontier employee becomes a manager of agentic teams.

Read more

These numbers took us several months to construct, and as far as we know, it’s the first bottom-up, deduplicated measure of consumer and enterprise AI spending across the full stack. We are releasing this research today in our first The State of the AI Economy report.

The supply side of the AI market is well-understood. The picks and shovel suppliers, the computer chips, the memory, the power transformers, the cooling, all of the components of AI data centers are largely public companies. We get a sense of what is being spent on the buildout through their disclosures, sales and forward order books.

But understanding the demand side is much harder. And this is what we’ve spent the last few months tackling. We built a proprietary AI economy model that looks at total AI spend, whether enterprise or consumer, to answer the hardest questions of the AI wave:

• How big is the market really?

• Are the revenues growing?

• How far do the revenues go to cover the investment expense?

• What happens to the economics in the future as token prices fall and the quality of those tokens improves?

The State of the AI Economy report

Before we get into the report, we think it’s important to break down how we did this work.

THE METHODOLOGY

How we count the demand side

One of our central design choices was not to count the same thing twice. We report the dollar spent by an end customer. So if you spend one dollar with Anthropic for Claude and Anthropic spends 50 cents with Amazon to serve it, we track both figures internally, but we will report it as our de-duplicated number: one dollar. This avoids double-counting the value that flows through the supply chain.

This isn’t straightforward. While it’s easy to count the supply side, the demand side is trickier to entangle. Much of the revenue flowing into AI comes from privately held companies such as OpenAI, Anthropic, Cursor, ElevenLabs, and hundreds of others. They don’t legally need to disclose anything.

The remainder flows to the big hyperscalers that serve these models: Amazon, Google, and Microsoft. While they are public, they don’t consistently disclose their AI segment revenues.

To shed light on this, we examine public statements from hyperscalers and neoclouds, their suppliers, and their customers, using only high-confidence, detailed facts to inform our modeling. We also look at well-reported leaks and self-reports, to which we assign a confidence score.

The result is an item-by-item financial model for the largest contributing companies and business units. Each model is effectively a deconstructed financial plan, a P&L, balance sheet, and cash flow, and these are triangulated against other external sources and internal consistency checks. This makes our numbers auditable. We can identify which data point, with which confidence weighting, contributed to any given estimate.

What we don’t count

We don’t include internal AI uplift, which is how much recommendation systems have improved, increasing ad revenue at Meta or Google. We do have models for those, but we’re not reporting them here.

Nor do we consider efficiency savings that the bigger tech companies might realize with their internal tools. We’re not tracking that yet.

We don’t include professional services and systems integration. When a Fortune 500 company spends or invests in AI, only a portion of that spend will go to an AI company. It won’t represent the full extent of their commitment, because a large part of it will be paying professional services to support the implementation.

We have got models for revenues in China, but this v1 of our report doesn’t include Chinese data yet.

THE TOP LINE

Are the revenues real?

Over the past 12 months, the AI ecosystem generated $110 billion in revenue when you remove double-counting. The growth rate is healthy. Annualizing the most recent month’s revenues indicates a $175 billion revenue run rate.

These revenues are growing faster than previous IT-oriented waves, roughly three times more rapidly than the mobile or Internet waves.

While many companies have moved beyond occasional pilots, they are still in the early stages of scaling and deepening. In conversations Azeem has had with senior execs across a range of industries in Europe and the US (from industrials to insurance, from finance to pharma), the consistent message is that they intend to invest more heavily in AI in the coming years. Companies are also becoming more vocal about the impact of AI on earnings calls, with the caveat that half of the surveyed CEOs believe their jobs depend on getting AI right.

Can AI revenues pay the GPU bill?

The next question we wanted to track is whether AI revenues can cover the capital investment that’s required to build the infrastructure. Our model separates AI-oriented CapEx from ordinary CapEx across the major hyperscalers and neoclouds, the specialist AI cloud providers. This adjustment is important because hyperscalers were already spending around $120 billion annually1 on CapEx before ChatGPT.

We capture the additional investment in AI infrastructure, then depreciate compute assets over 6 years and other infrastructure over 14 years. Our modeling shows that revenues attributable to hyperscalers just about clear the depreciation expense.

Six years is defensible. That longer useful life reflects two things. One, demand still exceeds available AI compute; and two, operators are getting better at managing GPU fleets. Both help. The second alone is enough to justify a longer economic life.

Subscribe now

What is the future of the token?

We also examine how market size changes as token prices fall. The elasticity of demand shows that lower prices are met with increased spending. We estimate that across providers, every 10% price cut leads to 12-18% more tokens in use, so the total spend still rises.

We suggest that although a token is a useful billing metric, it is still not the unit of value we need to measure the economic value of intelligence circulating through the sector. Quality‑adjusted output tokens give us a better “intelligence quotient” for the AI economy by combining how many tokens are produced, how many of them are actually user‑visible outputs, and how capable the underlying models are.

What else is in the report

The report also covers:

• What AI demand has done to the US power industry and how power efficiency is changing,

• What is happening to token costs, and how consumption-based billing may expand the market,

• Four scenarios for how fast AI demand could grow under different price and capability trajectories.

The State of the AI Economy report

Happy Sunday.

GenAI drives almost 2% of traffic to Walmart and Target; home and electronics categories lead. A quick show of hands, please:

Here’s Sunday edition #579.

Subscribe now

Product is the team

A working paper by Harvard Business School’s Rembrand Koning and INSEAD’s Hyunjin Kim studies how AI-native startups1 are different to non-AI startups. At similar funding and growth, AI natives are 25% smaller; they have more engineers, denser expertise and fewer managers.

Now, what’s interesting in this paper is that AI makes a real difference in a business when it’s folded into the product. Some of the knowledge work that used to be done at the edge of the product by human teams is now done directly within the product. When the work is done directly in the product layer, the feedback loops close. Kim and Koning write:

Capturing real value from AI often requires re-engineering processes around it—and, as our results suggest, re-engineering the product itself so that AI does the work directly.

This complements my perspective that firms must redesign around AI‑closed loops. Giving workers a Copilot or coding tools is most useful for understanding what the technology can do. But real gains will come from the much harder task of rebuilding how you work centered on AI. Read our analysis here.

• See also, Sam Altman speaking to the Stanford CS153 class this week said:

  everything about starting a startup has changed so much. … with an affordable amount of spend on tokens, you can do what a 100 person incredibly great engineering team would do as a startup and that was just totally impossible. I think what you can take on, the level of ambition you can have, the speed of which you can move, the amount of stuff you can do at once, is just totally different.

• Joanne Chen, general partner at Foundation Capital (which backed Netflix and Uber) and I discuss what it takes to build AI-native companies here.

  Subscribe now

The bitter lesson always prevails

Generalist frontier models are outperforming best-in-class specialist medical tools in head-to-head tests. The specialist models that were tested are gold standard — almost two-thirds of US physicians use OpenEvidence. My go-to expert on AI and medicine, , says: “This was not anticipated.”

But this is precisely Richard Sutton’s Bitter Lesson from 2019

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. […] We have to learn the bitter lesson that building in how we think we think does not work in the long run.

Read more

I am a fan of AI councils. As many readers will know (from previous posts), I use AI expert panels every day to deliberate on hard questions, challenge my thinking, and do research.

These AI councils, or panels, start from an assumption that a group can produce better work than a single agent. My friend and one of the sharpest AI practitioners out there, , tested this hypothesis.

I thought Rohit’s experiment was genuinely useful for anyone thinking of designing or already working with AI councils, so I asked him to share the findings with us today.

Over to Rohit.

– Azeem

Subscribe now

Read more

Here’s our Monday roundup of data signals across AI, energy and markets.

Enjoy!

1. AI & jobs. AI was cited as the top reason for nearly 40% of US job cuts in May1.

2. Mind the gap. The top 1% of US firms spend $7,450 per employee on AI each month, roughly 650x the typical firm’s $11.38.

   

3. AI productivity guarantee? Cognition says they will fund up to $10 million in credits if their AI agents fail to deliver the engineering value enterprise customers paid for.

Read more

I sent a note on Thomas Piketty’s blueprint for global justice to members of Exponential View yesterday. I call its recommendations a “blueprint for managed decline”. The commentary touches on the future of modernity, scientific dynamism and democratic legitimacy. Read it here.

In today’s Sunday briefing, we look at:

• The iPhone reduced fertility; what will AI do?

• What’s driving the consensus behind pausing AI development?

• The largest solar factory, gene therapy for rejuvenation & Chinese Gen Z finds inspiration in Western memes++

Subscribe now

But, first…

How will we distribute AI wealth?

, Sam Altman, Donald Trump and Vinod Khosla all agree that the public should own a slice of AI. Bernie proposes transferring 50% of ownership of leading AI companies into a sovereign wealth fund; Trump is supportive; Sam agrees in principle but pushed back on the 50%; and Vinod wrote an op-ed in the FT advocating for a new tax code on wealth AI creates and eventually pooling it into a sovereign wealth fund:

A sovereign fund with ownership of AI companies makes every American a capital owner, not a bystander to the AI economy.

The same day the Senate held a hearing on AI, Anthropic published a framework that has sovereign wealth funds as one of the mechanisms to deploy in case of the worst-case scenario, when “[t]he search for work stretches past a year, then past two, and for some, eventually stops”.

The core assumption here is that AI will tilt the economy’s income from labor to capital. We don’t have evidence that this is happening yet and we may not know for a long time.

Vinod believes that AI will be capable of doing 80% of jobs by 2030 and that $15 trillion of US GDP, which is labor compensation, will mostly disappear. Economist Chad Jones, in contrast, thinks the transition will take closer to thirty years because of weak links – even if we automate most tasks, the ones that don’t get automated will slow everything down (this aligns with what we see happening with AI diffusion into organizations right now). Slow doesn’t mean painless, though. And for a long time, we may not know which scenario we’re in.

My view is that there will be human jobs around for much longer than the prevailing narratives might suggest. Virtually all of these jobpocalypse scenarios of recent years have been mooted theoretically but not in the messy world of life. We’ll create more roles and that last mile (what Chad calls “the weak link”) will command decent wages. That doesn’t mean we won’t see increasing returns to capital over labor. So, societies might decide to change how they tax and redistribute. The different experiences between the UK and Norway over North Sea oil suggest that endowments seem to work better than transfers.

See also: I was on ’s podcast this week. We covered the AI bubble question, the productivity paradox and why Chinese AI engineers are Claude-pilled. Watch or listen here.

Subscribe now

Time to pause

Washington has issued an export control order that prevents Anthropic from offering its Fable and Mythos models to non-US citizens, including Anthropic’s own employees. Apparently, researchers at Amazon discovered a jailbreak and it seems like Amazon’s boss, Andy Jassy, passed concerns on to the Commerce Department. (Amazon recently invested $5 billion into Anthropic and indicated it would invest a further $20 billion, having already put $8 billion into the firm and acting as its major distributor.)

The reaction seems a little over the top, given the minor trigger. I’d half expect it to be diluted in the next few days. Or it might just be the next stage in the feud between the Defense apparatchiks and Anthropic.

But it comes at a time when Anthropic, OpenAI and DeepMind are signaling they’re in favor of slowing down frontier AI development. rightly asks, what did they see? Is it the fear of the growing backlash or a genuine capability explosion?

My brief thoughts on what’s behind this.

The most charitable reading is that the labs feel they are reaching a point at which their AI models challenge the security assumptions that underpin modern life. This isn’t unreasonable: they have consistently worried that this point might arise. That actual point may still be years away, but Demis and Dario do think in long arcs – and we might need years to prepare.

There is a less charitable reading that is also plausible.

Read more