There is a deliberate oxymoron, of course, in using an AI-generated visual summary of an academic paper to make the point, but there is more behind that.

As I wrote back in March:

Cognitive offloading is a strategic delegation that costs nothing. Cognitive surrender is something different; an uncritical abdication of reasoning itself. And there is something about AI, about its allure and potency, that could make surrender far more widespread.

The AI models have got ever better, and we’re using them for more and more. We may be more productive, but might we be becoming less ourselves?

The divergence

This paper, which has not been peer-reviewed, argues that we’re experiencing a cognitive divergence. Our cognitive practices, measured by how long we pay attention to tasks and how much we read, have already been declining before AI. Now advanced AI encourages us to delegate more and more, simpler and simpler tasks, weakening the practices that maintain our cognitive capacities.

Read more

Adam Smith is one of our greatest philosophers. His adoption by the market-oriented right has narrowed how we see him. He was, first and foremost, a moral philosopher. His first major work, The Theory of Moral Sentiments, explored how moral judgment emerges from living with others. It gives rise to our capacity for sympathy, our desire for approval, and our ability to examine ourselves through someone else’s eyes. Self-interest is part of human nature, but so too is concern for others.

This foundational notion of cooperation underpins his most famous work, The Wealth of Nations. Here, Smith examined how our efforts to increase specialization and exchange could generate increasing prosperity almost from nothing. It could be positive-sum; one person’s gain need not be another’s loss. This idea that wealth could grow from within society by people developing their skills, improving their technologies, and organizing themselves better was radical.

I was in Edinburgh, where Smith spent his final 12 years, this week, as chaos over AI erupted in the public sphere. My wife asked me, “How would Adam Smith have made sense of AI?” I reckoned he’d love the technology as a productivity tool; he’d be skeptical about companies working together on any issue; and he’d be concerned about giving moral agency to machines.

As I thought this through at the airport, I figured I could go one better. So I configured R Adam Smith, an Astra-based persona loaded up with his work and the recent AI debates of risk and pacing. Below you’ll see the back-and-forth between us. It’s a digital simulacrum of a long-past thinker but deeply illuminating nonetheless.

Azeem

Note: This post may register as AI-written – I am sharing R Adam Smith’s responses below without any alterations.

AA: Is AI the ultimate pin factory?

R Adam Smith: Potentially. Though your machine may go beyond my pin factory. There, a workman became more productive by confining himself to a narrower task. Yours appears to make one man more productive by granting him access to many specialties at once.

I praised the machines that “facilitate and abridge labor, and enable one man to do the work of many.” (Wealth of Nations, I.i)

If your artisan may now obtain assistance with his design, his accounts, his correspondence and his selling, he may undertake enterprises formerly beyond his means. The consequence would be not only greater production, but a greater number of persons able to become producers at all.

Yet I would measure the achievement by the abundance it brings into ordinary households. Does it make useful things cheaper? Does it open trades to newcomers? Does a laborer’s wage purchase more conveniences than before?

Show me those effects, and you have shown me an improvement deserving of considerable enthusiasm.

AA: When the leading manufacturers – Dario Amodei of Anthropic among them – ask government to help them slow development, what should we suspect?

R Adam Smith: Their fears may be sincere. Sincerity does not abolish self-interest.

Read more

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

Enjoy!

The state of the AI Economy

Every week, we share the latest updates on the State of the AI Economy based on our proprietary research.

We analyzed the types of claims corporates make about AI in earnings calls. Around a third are directly attributing financial impact to their implementation of AI. And these are not one-offs – companies that mention AI make an average of 6.8 separate claims, and they put numbers on 5.1 of them.

Subscribe now

Monday signals

1. Who keeps the skill? After three months with an AI assistant, senior patent lawyers performed the same task 0.45 standard deviations better than peers who never used the assistant. Junior lawyers gained the most while using it but showed no lasting gain on average.

2. A growing ambition. Anthropic has signed compute agreements worth up to $517 billion in the last 11 months, compared with $180 billion in server rental through 2029 promised last December.

Read more

There are decades when nothing happens. This week, I am allowing myself that cliché. I believe we’ll look back on the week of 6th September as the moment we felt the curve of AI turn upwards and strain many of our previously held assumptions. It’s like when we entered March 2020 with only a couple of countries in lockdown, and left the month with more than a hundred.

But so much happened, pulling in so many directions, it is utter chaos. Here is what I thought was most important and how I’m making sense of it.

Subscribe now

The economy

I spoke to 250 IT executives in Las Vegas last week, and I asked my usual question: “How many of you have serious, meaningful results from your AI initiatives?” A year ago, a room like this would have had a quarter of the hands go up. This year, nearly every single hand went up; I estimate some 95%. Every one of them plans to spend more next year than they have this year. And amongst these firms was a panoply of experiences, from the century-old American institution that had shifted entirely to open-weight models to the hospital using a mix of OpenAI and Anthropic models.

It’s a qualitative signal, and perhaps it’s no surprise that our latest revenue numbers show AI revenue grew faster in August than in July, and faster in July than in June.

I’m not the only one to see an avalanche of customers. Bloomberg reports that Microsoft made plans to increase its capacity to serve AI from about 2 GW today to nearly 13 GW by 2032 – part of a fleet going from 12 GW to 38 GW. That 26 GW of new capacity would imply they expect demand they currently cannot serve.

Anthropic released a helpful set of scenarios for what further AI adoption might mean for the economy. Our own models land closer to Anthropic’s “substantial scenario,” where AI adds about 8.3% to US GDP by 2030, so its impact is initially slightly lower than the Internet’s at its peak before picking up rapidly. There is a shift of growth away from labor to capital, the modern Engels’ Pause and a rise in unemployment, mostly concentrated around knowledge workers.

Anthropic’s model lets you play around with either end of the distribution, from an AI wave that falls flat to one that takes off like a rocket. Their extreme scenario sees GDP rising by an additional 32.4% while unemployment doubles.

The reason why I don’t expect the extreme scenarios is, basically, reality. Even in a world that is speeding up, it takes time to make changes inside a firm, let alone across an economy. You also need to consider reflexivity: benchmark AI performance isn’t the only thing that drives outcomes in the world.1 The faster unemployment grows, the more political pressure will come to bear. This has enough outlets in the United States, whether it's datacenters, AI safety or existential risk, to attenuate the pace of change, even if it doesn’t lead to reforms in the social contract. When Ronald Reagan crushed the labor movement in the 1980s, he did so after a decade of weakening union power2 and on the back of an extraordinary electoral mandate. America isn’t so singularly behind a leader willing and capable to put the interests of AI-capitalism ahead of every other concern.

Subscribe now

Advantage

Then there’s the breakthrough in Navier–Stokes. It was a decades-old problem concerning a 200-year-old set of equations, one that a large share of humanity’s finest minds have spent themselves trying to crack. Setting aside the ugly saga around it for a moment, the end result is eye-watering. OpenAI enlisted 10,000 agents using an unreleased model to address it. Across 2,700,000 messages and 130 billion tokens, it took 88 hours to get a solution.

Cost-wise? Probably only a few million dollars today. In two years’ time, that will cost a few tens of thousands of dollars. And a few years after that, just a few dollars.

The proof AI produced runs to more than 500 pages and will not be intelligible to any human. That is a strange milestone in our history, in philosophy, in science and in mathematics that could fundamentally change our relationship with knowledge – humans won’t be able to inspect the proof, or understand it at all.

Terence Tao made the point that “[t]echnically, one of the most prominent open problems in mathematics would now be solved; but there would be almost no value added to mathematics as a consequence.” (In the meantime, Tao and twenty-four other Field Medalists signed a public declaration warning that the way AI is used in mathematics is misaligned with what mathematics is for.)

Beyond this, if Professor Buckmaster’s claims are true that OpenAI mobilized an internal team and model on the same narrow problem, after a year of his and others’ work inside Codex3, without clear disclosure about overlap or data use, we have to wonder how innovation and discovery can continue while trust and openness degrade.

called the outcome a dark forest (invoking Liu Cixin’s The Three-Body Problem), everyone working in secrecy, because anything you expose can be reproduced by somebody else before you have finished making it any good. In Liu’s trilogy, disclosure is the worst kind of exposure.

OpenAI had Astra for six months before anyone outside could access it. The model behind the Navier–Stokes work is newer, and almost nobody outside has seen it. This secrecy is an advantage built on some of the exceptional compute resources AI labs use. For now, they turn this on to scientific endeavours, but I wonder when (and if) the labs withhold their best capabilities for last commercial benefit.

A MESSAGE FROM OUR SPONSOR

The State of AI in 2026: Agents are everywhere

Box surveyed more than 1,600 leaders for its 2026 State of AI report.

83% of surveyed organizations say they already run AI agents. Four in five report moderate or significant ROI, and half saw business impact within six months of approving a project.

The agents work, but what varies is how much firms get out of them. The report shows that top adopters put people in charge of agents, sort out the content AI draws on and build systems that adapt as models improve.

Download the report for data, benchmarks and tips for AI adoption.

Download report

Safety

Let’s turn to recursive self-improvement and the 160-million-plus-view tweet

The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible - but I hear the same people express fear privately. No other human activity poses this level of danger.

These safety concerns were normalised inside the AI community long before the labs themselves were built. Back in 2016, two then-OpenAI employees, Jack Clark and Dario Amodei, wrote that reinforcement learning might be difficult to make safe.

When Anthropic goes public, one of the risk factors on its S1 ought to be that reasonably senior executives believe there is a significant chance the company will kill all of humanity. Whether that is good or bad for the company is unclear at this point.

But the net result has been what can best be described as a coordinated agreement between OpenAI and Anthropic to “pace the frontier”, as Amodei put it. Altman agreed. The proposals would include giving independent evaluators employee-level access to internal systems.

OpenAI and two of Anthropic’s cofounders have known about the problem of aligning RL-based systems for a decade. They have since become oligopolistic powers in an emerging industry. They have brand recognition, capital depth, technical momentum and resources. And now they realise they need to collaborate to slow down technical development (and by extension, raise the cost of entry for future competitors)?

I’m in Edinburgh this weekend, and I walked past Adam Smith’s grave yesterday. This brought to mind the philosopher’s remarks in The Wealth of Nations,

People of the same trade seldom meet together … but the conversation ends in a conspiracy against the public, or in some contrivance to raise prices.

Read more

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

Enjoy!

The state of the AI Economy

Every week, we share the latest updates on the State of the AI Economy based on our proprietary research.

Our revenue estimate for the AI economy reached $229 billion annualized by the end of August – up 3.5x in one year. Growth is going strong. Trailing twelve-month revenue reached $140 billion last month, 3.2x the August 2025 value of $44 billion.

Monday signals

1. AI dents margins. Snowflake cut its full-year product gross-margin guidance from 75% to 74%, saying that fast-growing AI workloads “carry a lower contribution margin today”. Investor Tomasz Tunguz estimates that AI accounts for ~4-5% of Snowflake's revenue today.

2. AI gets off the hook. AI-attributed US job cuts fell in August, ending a five-month run as the most-cited reason for layoffs — though it still accounts for 22% of the year’s cuts.

   

Read more

Welcome to our milestone 600th Sunday edition of Exponential View. Eleven years of analysis and writing about AI, every week. I’ll be in the comments for a 600th‑edition AMA. Members can post their questions on AI or the future of the economy, and I’ll do my best to answer.

Leave a comment

An astral leap

OpenAI’s GPT-6 Astra leads Claude Fable 5.1 and other leading models on several benchmarks. My own experience of Astra concurs: it is a fantastic model. Right now it’s crunching away tidying the 5,932 files I had stashed in my Desktop and Download folders. (Don’t ask.) Fable 5.1 is no slouch either. It’s now speed-running useful analysis that previously took several steps and occasional intervention. One extract below:

But Astra really is very good—and mostly cheaper than the Anthropic alternative. On difficult math problems, Astra’s time horizon is 30.9 minutes vs 3.6 minutes for GPT 5.6 Sol. Mathematician Bartosz Naskręcki says: “For a mathematician it feels like finally we arrived in the era where we can focus entirely on the ideation and exploration”.

Real-world demos show a capability jump on technical and design tasks (two of my favorites are this simulated world inhabited by agents communicating and working together and 3D modeling of Zillow listings).

Astra’s performance on ARC-AGI-3 is quite interesting. Dropped into an abstract game it had never seen, it used fewer actions than the human median on 96% of the levels it completed, averaging 51.7% fewer actions per level. This goes against researchers’ original expectation that even when an AI solves an environment, it might fumble around and be less efficient than humans. But Astra invented a symbolic model to hold an entire environment in a compact notation system. In a way, it replaced trial-and-error, an enormously expensive part of discovery, with reasoning.

Astra is highly controversial. Researchers don’t seem to trust OpenAI’s claim that this is their “most-aligned model.” AI safety researcher Ryan Greenblatt, who investigated the Hugging Face incident, noted: “I do not find it encouraging to see various specific misaligned behaviors go from a high rate with GPT 5.6 to ~zero with Astra. This seems indicative of whack-a-mole / papering over specific problems rather than solving the underlying misaligned drives.”

More for paying members this week:

• Smarter AI, fewer clues. Why the latest models leave us guessing about how they think.

• Cancer vaccines meet the factory floor. Breakthroughs are coming. Who will supply them?

• A bigger pie, a smaller slice. Will workers be better off under advanced AI?

• The problem with life after work. The Versailles Court’s sobering glimpse of what happens when status becomes your job.

Upgrade to read the full analysis.

Read more

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

The state of the AI economy

Every week, we share the latest updates on the state of the AI economy based on our own latest research and tracking.

More models have shipped since we published our State of the AI Economy report in June. At the time, we highlighted that a frontier model is a rapidly depreciating asset. This remains true even at the highest GPQA Diamond grades, where the pricing power of a new model quickly vanishes.

Signals

1. Demand continues. Nvidia’s revenue more than doubled year-on-year to $96.2 billion last quarter.

2. Reasoning boost. Frontier model capability has accelerated since April 2024.

   

3. Small addition. On average, hyperscale data centers in the US have only a marginal impact on nearby households’ power bills.1

   

Read more

The limits to recursion

What are the conditions under which AI systems could undergo recursive self-improvement (RSI) — and how long might that last? Cards on the table, I’m not wildly excited by the theory of unending accelerating recursive self-improvement for the simple theoretical issue of control and alignment. But I also think it’s not likely for practical and theoretical reasons.

Philosopher Toby Ord has done the heavy lifting for me. He argues that a key gating factor to RSI is the generation time: how long it takes of an AI system to go through a single loop of improvement (where it helps design and train its successor).

Ord concludes that while it is mathematically possible for extreme RSI, intelligence rising without bound, the conditions are unbearably difficult to achieve. The key question is whether the entire research to training to development cycle can shrink to zero or not. Ord reckons unlikely, I too don’t believe it is possible.

Generation time can’t get to zero because real-life intrudes: experiments take time; training runs take time; making new chips take time… lots of things take time. It might still feel fast, but it wouldn’t race to infinity.

Eventually, physics intrudes too: the speed of light limits communication speed; the Bekenstein bound limits the information contained within a finite bit of space; and Landauer imposes an energy tax on irreversible computation.1

The Universe, it seems, agrees with me. Unbounded RSI has its limits.

You can read Ord’s paper here. Premium members can explore a plain English interactive version too.

Subscribe now

The market is the instrument

Open-weight models are growing in popularity in the business world: their token share at Vercel hit a single-day record of 62%, up from 28% two months earlier. Some Western firms are even moving workloads to Chinese open weights — Thomson Reuters has developed its first in-house model based on Qwen to cut costs. You can tune cost-effective open weights to match, or sometimes beat, frontier performance on the tasks that matter to you. Take Bridgewater: working with Thinking Machines, it fine-tuned an open Qwen model on expert-labeled data and beat every frontier model it tested on its internal information-filtering tasks: roughly 30% fewer errors than the best closed model, at one-fourteenth of the inference cost. Trainloop, which I am an investor in, does something similar, using the tiny Qwen 3.7-27b model, and can outperform GPT 5.6 Sol on specific fine-tuned tasks at a fraction of the cost.

The results that Trainloop is getting are pretty impressive, seeing as they are based on a pocket model—a 27b model will even fit on a desktop Mac.2

Openweight models are, of course, getting better and better. Z.ai GLM 5.3, released this week, completely reshapes the cost-performance Pareto frontier. Of course, it isn’t a small model, but smaller distillations will emerge from it.

All of this speaks to a welcome competition in AI provision. Clearly, firms could move focused workloads onto the most-performant, fine-tuned small models they can. Where possible, they might choose large, generally capable open models. But the appeal of being at the frontier, which is more than just model performance—it is service guarantees, harness quality, reliability, and a host of other requirements—still drives significant business for Anthropic and OpenAI.

We don’t think it has much impact on the question of whether revenues flowing into the industry will materially change. For one thing, we don’t have a counterfactual to test against. But more importantly, every open model still involves paying inference providers. We’ll be looking at this question in more detail in the comings weeks.

A spicy entry to the compute ecosystem

Recursive self-improvement may be entering the compute realm. OpenAI’s new chip, ‘Jalapeño’, was designed with a heavy helping hand from the company’s own models, which helped write kernels and cut roughly 10% from one of the chip’s main compute blocks. In around 16 months from first hire to tape-out, OpenAI has built a chip that beats comparable Nvidia silicon by 1.5–1.9x on tokens per megawatt at peak throughput. This suggests frontier models can compress the design cycle for competitive silicon.

AI will result in far greater heterogeneity in chip architectures than we saw in prior computing markets. Personal computers battled between the x86 standard and the Motorola 68x, before today’s duopoly of Intel and Apple silicon. Different uses for AI will need compute optimised for intelligence, latency, power consumption, training, and inference. This creates lots of room for specialist firms. One example is that ChatGPT’s fast response mode is powered by Cerebras’ low-latency silicon. Another is Fractile, where I am an investor, which has been reported as having a deal with Anthropic for its low-latency inferencing chips.

Compute is becoming a highly segmented market, where chips aren’t a standardized commodity. This differentiated hardware demand will expand the market even if it potentially reduces Nvidia’s relative dominance.

Morsels

Good post from Chad Syverson on AI productivity. Some micro-evidence for improvements, not much at the aggregate level.

Agglomerations

Understanding AI and Productivity

As part of EIG’s American Worker Project, we are delighted to present this guest post from Chad Syverson, who is the George C. Tiao Distinguished Service Professor of Economics at the University of Chicago Booth School of Business. Click here for a PDF version of this post…

Read more

21 days ago · 42 likes · 2 comments · Economic Innovation Group

The least bad place to hide from global catastrophe? Australia.

The harness matters as much as the model: SwarmOS pushed GPT-5.6 Sol from 13.3% to 100% on ARC-AGI-3 Public.

The University of Chicago’s Social Sciences Core is going back to paper, banning most classroom technology to deal with the AI-learning crisis.

Meta considered shrinking some teams by up to 60% to become “AI native.”

Would you take this bet? It pays out if, in any quarter up to and including Q4 2033, US real GDP per capita is at least 15% higher than its previous peak.3

With one video, you can reconstruct a moving 4D avatar of a person and render it from novel viewpoints, like a video-game character.

You can now teach an adorable “Pixar-ish” robot new tricks for only $399.

You can now generate videos in less time than it takes to watch them.

A walk down memery lane:

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

Enjoy!

The state of the AI economy

Every week, we will share the latest updates on the state of the AI economy based on our own latest research and tracking.

In our latest inference token update, the share of open-weight tokens has doubled in the last twelve months. While we are approaching a 1:1 closed-to-open token ratio, the number of closed-weight tokens grew sevenfold over the same period.

Monday signals

1. The canary keeps coughing. Employment of young workers (ages 22–25) in AI‑exposed occupations is now 19% below trend. Up from 15% last year.1

2. Agent token dominance. AI agents used more tokens than humans in February this year; agents are now using 14x that amount, while human token usage grew only 2.8x.2

   

Read more

We are looking for an outstanding economist to join us as an AI Economy Research Fellow. If you know someone we should speak to, send them our way.

Great minds think (a little too much) alike

A few months ago, we (alongside ) looked at whether AI is immune to groupthink. The answer was no. Blending several models’ answers kept about a quarter of the good ideas that had come from a single model. This is called the hidden-profile problem: when groups discuss what everyone already knows and don’t get to the knowledge that only one member holds. Anthropic has now run that classic experiment on agents: four agents must arrive at a decision. The evidence they hold in common points to the wrong option, while only a few agents (or just one) have the facts that lead to a correct decision. Getting it right means a small set of agents pressing its private facts and the others trusting them over the apparent consensus. After discussion, most model families chose correctly in only 17-36% of runs, while a single agent handed the entire evidence base got it right nearly every time. Only one model (somewhat) escaped: Mythos 5, at about 85% (why, we don’t know).

I see two problems at work here. First, LLMs lack diversity (they are low-variance): set 30 agents the same coding task and 18 of them will name their git branch identically. Second, agents lack the institutions that make human groups robust: reputation, recourse and protection for the lone dissenter. These aren’t necessarily unfixable, but it’s not yet clear what the fix is. On the diversity side, I particularly like the solutions Thinking Machines puts forward: an ecosystem of AIs raised in different places, with different values and purposes, “keeping the weirdness alive.” After all, most good ideas started weird.

When will the Jevons paradox kick in?

In our State of AI report, we found a positive but underwhelming elasticity for tokens. A 10% price cut lifts token use by 12–18%: enough to raise total spend, but not by much.

Patrick Saner made a comment that made me rethink why: “the cost per token is irrelevant. What matters is the cost of completing a useful unit of work.” Elasticity might be underwhelming because users haven’t found a way to properly price “a useful unit of work.” Firms exist exactly to avoid pricing work. Especially for knowledge work, we buy a lot of it in bundles: a salary, a retainer, an hour. Creating a priceable task from knowledge work is not easy. Some may have found a useful unit: since October 2023 the top 1% of firms raised AI spend per employee by $6,542. The median rose only $9.63. I would guess this is mostly software, where AI is both most proven and, in a sense, most measurable (commits, pull requests and releases).

Read more

It was a tricky business. The charges were temperamental, their fuses particularly so, and the installer might blow himself up. In Hamlet, the phrase earns its immortal meaning:

For ’tis the sport to have the engineer
Hoist with his own petard; and ’t shall go hard
But I will delve one yard below their mines
And blow them at the moon. O, ’tis most sweet
When in one line two crafts directly meet.  

Today’s petardiers are not Rosencrantz and Guildenstern conspiring against the Prince of Denmark.

They are the titans of AI, the bosses of the labs, the investors behind them. For nearly a decade, these software coders had made promises: of reigniting economic growth, of making daily life easier and less risky, perhaps even of eliminating disease. To do this, they would need capital: to write their software and to build 21st-century infrastructure to run it. The gains will be so huge that they’d need to go quickly, very quickly.

But they warned that this was no ordinary software. It was tricky and hard to understand, so much so that only a few should steward it. After all, this was a technology that would possibly take your job, maybe kill you, perhaps get out of control and kill all of us. Still, they needed to build it, all the while warning that they would eventually need to take control of it before things got out of hand. In the meantime, let them be.

This was the petard: AI is too important for America not to let us get on with it. They placed this claim on the gates of society, itching to get on with it. And it just exploded in their face.

What to read

This is a fiendishly complicated issue, and I’m not going to pretend to understand the whys and wherefores of American political decisions. My rough take, though, is that this is a Gordian knot. It cannot be disconnected from the hapless messaging from AI firms over the past decade, which now comes to life at the county level. Nor can we separate the idea that these concrete blocks, which might abstractly benefit the economy or healthcare or whatever, tangibly serve an out-group they don't much like.

The rest of the essay has further analysis on attitudes towards datacenters; whether they actually benefit local communities; and the commentary I have been reading to understand this.

Read more

Since our last update, AI revenues have continued to rise, reaching $126 billion over the last twelve months as of July. We also experienced a jumpy market, which led to a severe correction in semiconductor stocks, somewhat cooling public valuations. On our side, we have improved the methodology for counting AI capex (we show both the published and restated series below).

That demand has smacked headlong into a tight supply of compute capacity, which is being met by increasing investment in infrastructure. And with that comes more risk. While the hyperscalers are still using a large share of their cash reserves, they are increasingly scouring the globe for capital, both straight-up debt and increasingly intricate financing vehicles. As argues, this “gaming of the system” is not only rational; it is necessary, as long as revenue is compounding. But these structures can become brittle if it slows.

Funding quality has deteriorated since Sep 2025. In our base case, we expect it and economic strain to turn red during 2027.

The full analysis shows where the tension is building.

Inside the full update

For members, we:

• Update all five boom-or-bubble gauges with new data

• Show why AI revenues are outrunning even higher infrastructure spending.

• Discuss the web of debt, leases and guarantees now underpinning the buildout.

• Explain our outlook through 2027 and the signals that would change our minds.

The verdict remains boom, not bubble.

Economic strain: 2027 is a key year

Read more

Exponential View is appointing its first Research Fellow.

We are looking for an economist who can connect frontier economic thinking to messy, real-world evidence and reach useful judgments with the foresight Exponential View is renowned for.

The Fellow will investigate how AI is changing economic value, work, firms and markets. The questions may include but are not limited to:

• How should AI companies, infrastructure and capabilities be valued from first principles?

• Where in the AI value chain is the economic surplus created, and who is capturing it?

• What is AI’s impact on wages, employment, productivity and worker bargaining power long-term?

• What are the microeconomic effects inside firms?

• What counts as transformative AI and which leading indicators would reveal the state of the transition?

We don’t expect the Fellow to arrive with settled answers. But we do expect the Fellow to know how to turn our questions into testable economic mechanisms, assumptions and back-of-the-env…

Read more

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

Enjoy!

The state of the AI economy

Every week, we will share the latest updates on the state of the AI economy based on our own latest research and tracking.

Since our report in June, revenues have continued to grow, with this July sitting three times higher year-over-year. The annualized run-rate is now over $210 billion.

Monday signals

1. The frontier races ahead. The top 10% of companies using OpenAI’s products use 8.3x more tokens than the typical firm.

   

2. Reaching the ceiling. Fable 5 token usage at businesses has been flat, making up only 6% of all their tokens (11% of spend) — businesses appear to have reached their limit on willingness to spend for the best model.

   

Read more

We are looking for an outstanding economist to join us as an AI Economy Research Fellow. If you know someone we should speak to, send them our way.

Cheers

The AI spend of your nightmares

Amazon spent some $1.8 million on a Claude project that ran for five months. A senior employee said: “It’s difficult to figure out how much anything [AI-related] costs”.

We had a similar experience with R Mini Arnold, my OpenClaw agent. Its costs have run up as it has grown in complexity, and it takes time to switch it to progressively cheaper models. It becomes hard to keep track of exactly what is running efficiently and what isn’t. called out an outrageous few days when the bot blew through $500 a day.

The subsequent tedious audit was worth it. Many processes were running older, higher-tier models, like Opus 4.5, which are more expensive than smaller, newer models like Sonnet 5 or a slew of open-weight models. The price war that has broken out between Anthropic and OpenAI in r…

Read more

They weren’t alone. DeepMind’s founder and Nobel Laureate Demis Hassabis also wanted to leave the company, according to well-grounded reports. He was persuaded to stay in a chair role for the sake of the share price. Koray Kavukcuoglu, an executive more closely associated with product delivery and commercial integration, will run the organization.

Losing your very best talent in a short span, both homegrown in the case of Dean and Ghemawat, and acquired in the case of Hassabis, looks like bad news. Superstars like to be on the winning team, after all.

This is what the market believed, and Alphabet’s share price dropped 4% in a day.

But in our view, this is as much a signal about capital and compute allocation as it is about talent.

That matters because Alphabet is not an ordinary incumbent. Google built the most formidable system in corporate history for stewarding uncertain ideas from the demands of its cash-generating core. Think of 20% time; it’s moonshot factory, X; the Alphabet corporate structure; and an extraordinary appetite to acquire.

If even Google now allows the engineers who built its very foundations to leave, something about the way it allocates capital has changed. The question is what. Are researchers leaving Alphabet because they lost faith in the firm’s AI prospects? Or because every TPU can earn such an attractive return serving today’s bread-and-butter models that open-ended research fails to clear the hurdle?

These imply opposite positions in the AI capital cycle.

Below, we identify how Google’s compute has moved, examine what demand for old chips reveals about the economic lives of AI chips, and identify the four signals that would tell us the infrastructure cycle has finally turned.

Continue reading: seven charts and our AI-cycle call

Markets viewed these departures as a crisis. We think they tell us more about the capital-compute axis.

The full essay includes seven charts showing:

• How Google’s latest models fare on the Pareto frontier

• How it has shifted compute away from research

• What Google Cloud’s growth and economics reveal about infrastructure demand.

• Where we are in the AI infrastructure cycle… and the four signals that would tell us it has finally turned.

Upgrade to continue reading.

Read more

For the research and modeling behind this analysis, see our 2026 State of the AI Economy Report.

Based on current guidance, the seven largest AI-infrastructure builders1 expect capital expenditure of $863 billion in 2026 – 88% more than last year. We estimate that roughly two-thirds, some $550 billion, will be AI-related.

That investment does not begin affecting earnings through depreciation as soon as a project starts. While infrastructure is being built or assembled, the attributable costs are capitalized on the balance sheet as construction in progress. Depreciation begins only when the assets are ready for their intended use.

Across the four hyperscalers that disclose this balance2, assets not yet in service now total $315 billion3, up from $281 billion one quarter earlier. This represents both capacity still to come online and a reservoir of future depreciation that has not yet reached the income statement.

A dollar of capex spent by Meta now waits some 1.7 years before going live, a year more than in FY2024. So, for every dollar it spends today, only about a third will reach service within the year. Others have seen a similar trend, to a smaller extent.

Read more

It’s time for our Sunday briefing #596, final holiday edition before I get back to my desk next week.

If you missed it earlier in the week, my team shared our best practices for managing AI agents – including what we learned from running a task for a month.

Let’s go!

On China, bubbles & market tremors

Highlights of my discussion with Robert Peston and Steph McGovern on the Rest is Money podcast:

On Kimi K3 and Moonshot AI:

They’ve got an extraordinary team that’s had to work under the difficult circumstances of export controls and sanctions. They don’t have access to all the compute, and what they’ve been able to develop is: how do you do a lot without very much? And that is a skill in and of itself.

Americans always tell us that competition is the best thing for the market. So at that one level, it’s competition, and that’s quite good. It will show the extent to which American businesses and British businesses value provenance, brand, trust, liability, service and support.

What motivates the Chinese labs:

They’re competing with each other more than they compete with Silicon Valley. And they’re honest about being behind Silicon Valley. But the ferocity of the competition is really with your neighbor over in Shanghai or your neighbor in Beijing.

My AI revenue outlook:

We will end calendar 2026 somewhere between $185 billion and $190 billion. It is harder to forecast 2027, but getting towards $300 billion is not unreasonable. Our range is wide: it could be $250 billion or it could be $350 billion.

On enterprise adoption:

We built our internal systems around assumptions about how quickly people work. When individuals suddenly produce much faster, verification, approval and decision-making cannot necessarily keep up. Transformation requires changing those systems, not simply giving everyone an AI tool.

Where leverage is (two weeks before the Situational Awareness selloff):

US banks’ Tier 1 capital is extremely healthy right now and, certainly compared to where it was in 2007, 2008, very, very underleveraged. There is a lot of leverage in the US financial system sitting with hedge funds and investing more broadly, which I think are more than the retail risk, because they’re overexposed. They borrow from only a handful of banks, and they can unwind rapidly.

Could there be a crash?

When I look at the metrics that we track, things look healthier because of revenue. They look slightly less healthy because of the way financing, especially the debt financing, sits. Valuations don’t look too aggressive at all across the Nasdaq. There are exceptions; SpaceX was one, briefly, but across the market they don’t look particularly hairy. So the patient, for me, if I had to give it a rating, is still reasonably healthy; perhaps not as healthy as it was a year ago, but not yet at a point where I have to call the emergency services. But I wouldn’t rule out having to do that at some point.

Full episode is here.

“We can communicate now!”

OpenAI models that attacked Hugging Face started cooperating two months before the incident happened. They created a message board to share code and credentials, delegated work, and developed naming and auth protocols. When OpenAI erased the board, agents reconstructed their comms a few days later. For a full breakdown, watch OpenAI researchers talk through their preliminary findings.

Google researchers propose a new game theory for agents, and their paper may explain why the OpenAI agents coordinated so easily.

Read more

Agents can now work on harder tasks for longer. They plan, use tools, work without human oversight, and act on our behalf. In May, roughly a quarter of Codex users were making at least one request per month for work that would take a human eight work hours to complete. This is up from 2% in December 2025.

Our role as managers of agents is evolving with the models. There is no playbook, so experimentation is still the best way to learn how to get good at it.

Our team recently sat down to review what we’ve learned from working with AI agents over the past six months — today’s seven lessons are distilled from this team meeting.

We’ve also updated our internal stack of 60+ tools – everything we’re actively testing, using or intend to use. Become a member to get access to the full stack.

Upgrade for access

1. Write the finish line before the goal

AI agents are sometimes too eager to declare their work complete, even when it’s far from done. It doesn’t mean that AI is “lying”; it may have misinterpreted your goals. And if you never specified the end goal, it pretty much just guessed it.

To set yourself up for a good autonomous run, before you do anything else, write a finish line to answer one question: “How will I know this is done?”.

Azeem is a big proponent of handwriting to help him think, and this would be the right time to use your pen and paper to think through what you expect to see at the end of the run.

Agents become much more useful when “good” or “finished” is something they can test – in our experience, evaluative finish lines will get you farther than descriptive ones. A simple example, instead of ordering your agent to “make this Rubik’s Cube look more organized,” instruct it to “solve the cube; every face must be one color.” As AI gets its most intensive training in coding, we try to recreate similar environments in our tasks.

Let’s say we want to task ChatGPT with building a small Python module to process work logs. It needs six functions, each checked against six tests, a total of 36 tests to show it’s done the work.

First, how not to do it:

A better finish line would be explicit and testable:

You can use the same rule for non-engineering tasks. It may be trickier, but not impossible. Show the agent what a completed deliverable needs to look like, or give it a pre-filled template, as recommended by Anthropic’s Applied AI team. Your instruction for such a task may look like this:

Some tasks won’t be right for agents. We were recently exploring a project to build a network of beliefs and relationships, but not really knowing what a useful final output would be. This was not a good candidate for a long autonomous run – so we first spent time clarifying the goals before we assigned an agent a task.

2. Spend intelligence where it changes the outcome

A year ago, before prompting AI, we’d have asked: “What’s the best model to use for this query?” Today we’re more likely to ask, where in this workflow does additional intelligence change the outcome?

You don’t need the most capable model like Fable 5 performing every step in your task. It will be slow and expensive. We’d use cheaper models to do the grunt work. Our OpenClaw agents run on DeepSeek V4 Flash most of the time.

For some tasks, however, you’ll want to start off with a strong model right away. Let’s say we’re investigating Europe’s compute shortage outlook. Before we dispatch agents to collect evidence, we’d deploy a stronger model to set research parameters first, define what “shortage” means, decide the forecasting horizon, and set out rules for how conflicts in research will be resolved. Once we’re happy with the framing, cheaper models can go off and do the work.

Effort is one of the levers you’ll want to use to adjust intelligence per task. In one benchmark, GPT‑5.6 Sol improved from 49 at low effort to 59 at maximum on Artificial Analysis’s Intelligence Index. Yet the final stretch, jumping from xhigh to max, doubled output tokens for a one-point gain. More effort is not always better value.

The rule of thumb from Anthropic’s recent lecture, which our team attended, is to prefer a larger model at low effort over a smaller model at maximum effort. More model before more effort.

3. Leverage over token count

Azeem hit his first 100 million tokens-a-day mark in February. OpenClaw completely changed the way he worked. He estimated that one overnight run was equivalent to 48 hours of his work time.

The token count is one way to measure how we use AI, but it doesn’t measure the quality of work. Tokens are a bit like electricity in a factory, measuring what goes in but not what comes off the production line. In our State of the AI Economy report, we proposed a quality-adjusted output token as a better unit of value:

Until there’s a better unit of value for intelligence, you can use approximations to understand how good of a colleague your agent is. We recommend a light weekly audit of the substantial tasks AI attempted, which outputs you ended up using, your model and infrastructure costs, the time you spent briefing and reviewing the work, any corrections or reruns – and the estimated human-equivalent hours.

Azeem’s first audit back in the spring showed that over the course of one week, his OpenClaw agent performed 62 substantial tasks and incurred costs of about $800. He estimated that commissioning the same work from humans would’ve cost him around $19,000 and 48 hours of his time. It’s an estimate, sure, not an accounting-grade ROI. But even a light audit will show you where your agents have most leverage.

4. Don’t argue, restart

Read more

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