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

LLM councils and groupthink

By Rohit Krishnan, originally published on Strange Loop Cannon

One way to get the best out of LLMs is to use model diversity. The models are not all the same so if you use their unique natures, you can get better responses. We saw it with the work on MarketBench. And we also saw this when Karpathy came up with LLM Council as a way to get multiple models to work with each other and get us a better answer.

But I started wondering, with people, when you put a bunch of them together in a committee, some things get better but some things do get worse! And relying on an LLM to audit is also error-prone. “Design by committee” is a four letter word for a reason. LLMs are better than us probably, but surely this process is also somewhat lossy. So what do we lose?

To test it, I set up an experiment, where I set up a few committees of models:

• First, I took each answer, then gave those to a fourth model and asked it to write the final version.

• Then, the llm-council – essentially peer review and then a chairperson summarises

• And a “best answer” picker – just a direct pick.

With people, the problem with committees is that they “smooth out” all idiosyncrasies. They take out any “spiky” points of view, and make things much more normie. Same thing here. So to test how we do I had to find some way to grade how the various final responses were. I broke each answer into small “cards” using Sonnet. A card could be a mechanism, observation, metric, failure mode, image, or some other important detail.

Then I clustered cards that appeared to mean the same thing. If a cluster appeared in one solo answer, we called it a single-model idea. If it appeared in more than one, it’s shared. And two judges scored the solo-derived clusters without knowing which model produced them or whether a council kept them.

Now, it’s not perfect, but it’s the cleanest way to test the problem of “how to rate which answer is better” that I could find without doing human ratings.

First, the result: the council does not simply keep the best bits from everyone. It keeps a minority of the good ideas, while peer review seems to give consensus ideas an extra push.

Obviously, the final summarized versions usually read better. It is calmer, more complete, less jagged, all things you’d expect. But we had misses. Examples.

• A field report noticing that salvaged retail scent cartridges had become status symbols in a squatted mall, used to mask the smell of communal living.

• An incident report arguing that logged-but-deprioritized risks are more dangerous than unknown ones, because they manufacture a false sense of control.

• A data-recovery plan that asks users to re-confirm suspect fields at their next login (”please re-confirm your shipping address”), quietly crowdsourcing recovery from the one authoritative source.

In the final runs, the blended council kept only about a quarter of the good ideas that appeared in just one model’s answer. Remember, these were ideas that two blind judges rated as useful, non-obvious, and worth keeping, and still roughly three quarters did not make it into the final answer.

The peer-review version did not solve this either. The rare ideas survived at about the same rate as in plain blending: 24% versus 22%. But if several models had raised the same idea, the peer-review council kept it about a third of the time, but if only one model raised it, a quarter.

To test this, I ran sixteen open-ended prompts: eight strategy problems and eight writing tasks.

Figure 1. The experiment path from solo answers to idea coverage.

I plotted what happened with the ideas. The red dot below is good idea that only one model came up with. Blue is good ideas that multiple models came up with. And the X-axis shows how many of each actually showed up in the final answer. So the selector for instance showed about 37% of all good single-model ideas, and 24% of the multiple-models ideas, which makes sense because it picks one full answer and discards the others.

Figure 2. Coverage of blind-rated high-value ideas.

The consensus tilt is smaller here, but interesting. In the peer-review council, shared high-value ideas that survived had an 11% uplift over single-model high-value ideas. Or put another way, a 50% relative lift!

The denominator for shared ideas is small, though. What’s interesting is that this shows us how the specific topology of the “council” changes what you’re likely to get, like a peer-review round ends up becoming a consensus detector even above a single model blending the answers from all other models.

This is a problem with all cognitive beings. In group decision-making research, back in the 1980s, Stasser and Titus called it biased sampling of shared information - groups are more likely to discuss information that several members already know than information that only one has. That line of work led to the “hidden profile” problem, where a group can miss the best answer because the crucial evidence is scattered across individuals rather than shared up front. We’re seeing the same thing here.

The work on LLMs, meanwhile, so far has mostly come from the other direction. Multi-agent debate papers ask whether multiple models can improve the final answer, and yes, they often can! But depending on the topic and the question, a council can absolutely improve the average answer and still drop some of the best ideas.

As users, we want to get better answers, cheaply. That’s the whole goal. Councils are great ways to make some answers better depending on how you structure it. But they’re not cheaper. So it is important to make sure they are actually better! If they’re not, or at least not universally, then how the council should be structured is an incredibly important problem!

What we still see here is that there is no free token lunch. If you use councils to get the benefits of model diversity, don’t assume it will preserve the best ideas. To do that we have to work harder, and understand how to work with these models.

For instance, one thing we know is that the best way with LLMs usually is to be explicit, since otherwise, even if they’re aligned, they cause emergent problems. So the best protocol might be to explicitly gather and store the best ideas from each solution separately and ensure they’re stored, ranked and assessed, before a final answer is written and revised.

It does much better, though it’s slower and heavier. I don’t know if this is the best we can do. The structure might change depending on the question asked, the domain, or the types of answers expected.

Humans have gone through thousands of types of “councils” until we reached interim solutions which give us decent results nowadays. And even then, we have to change the shape of the councils constantly, as we evolve, and society evolves.

To figure out how to get the best results from our work requires a lot more effort into designing the councils. If you’re working with them, you will need to experiment and eval against your individual problem sets, which is the only way to know if this specific council setup will help with your specific problem. Copying someone else’s homework won’t work!

Homo Agenticus are odd enough creatures that using them well requires much, much more experimentation than one might assume. Especially when the problems of using them suboptimally is that we lose actual functionality, often without knowing it!

Leave a comment

This essay was originally published in Rohit’s publication Strange Loop Cannon.

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

In the Global Justice Report, Piketty and his co-authors recommend the world converge on a national income of €60,000 (around $68,000) per capita in 2025, in PPP terms, by 2100.

They, too, propose we:

• Keep annual growth in Europe and the US to roughly 0-0.5% while poorer countries catch up,

• Cut working hours to about 1,000 per worker per year,

• Consume fewer material goods and more education, health, care and culture.

• Fund the transition with steep taxes, including a wealth tax on billionaires and very high top income-tax rates.

Reports like this arrive with a splash: they profess to be blueprints for a just global economy. But this isn’t that. It’s really a roadmap to Animal Farm.

Subscribe now

Nine problems

1. The report misreads the history. It embraces growth for poor countries while freezing it in rich countries, as if the two were separate. The great escapes from poverty happened through trade, technology, supply chains and frontier knowledge. Much of it originated in wealthier nations. If the economic frontier slows down, the authors need to explain how the channels that allowed poor countries to converge would keep working as before. For a deep dive on growth and extreme poverty, read this.

2. Bad science. For their climate assessment, the policy baseline is set at 4.8-4.9°C by 2100, which roughly aligns with the now discredited and “implausible” RCP 8.5 scenario. RCP 8.5 ignored that learning curves for solar and other renewables exist entirely. Robust academic research dating back to 2016 had already discredited it. (Read on this here.)
   Piketty has argued he used his own climate modelling projections, which happen to align with RCP 8.5. It makes at least one titanic assumption that we fail to “accelerate the energy transition.” That latter assumption is absurd given this is already happening. Solar, batteries, electric vehicles and many other technologies are on steep learning curves and spreading like topsy.

3. Implausible politics. In the proposal, global incomes would converge at €5,000 per month (about $5,700) by 2100, with lower inequality within countries. For Western Europe and America, this would essentially mean a freeze on economic growth. In the case of the US, a growth in incomes of 0.12% per annum for 75 years. It is impossible to imagine 18 US electorates, from 2028 to 2096, agreeing to this. There would have to be some kind of external force majeure, an extraordinary top-down imposed policy against the will of citizens in nations across the globe. This is anti-democratic in its own right.

Read more

American companies spending the most on AI have seen their revenue grow five times faster than the economy as a whole. Non-spenders are tracking the economy closely. This is according to Ramp, a fintech with a side hustle in excellent firm-level data across the US.

Ramp’s finding echoes this 2020 AEA paper by James Bessen and colleagues. Bessen examined broadly defined automation investment at the firm level across the Dutch non-financial economy between 2000 and 2016. Firms that automated grew sales 2% faster than those that didn’t.

But which firms automated? Other studies by Acemoglu et al. give some kind of answer: automotators were already largely more productive and had higher output than non-automators. In my book, I took this further, arguing that any type of automation is a complex undertaking and requires a better management team than not.

AI also seems to be fuelling a surge in new business formation, according to Torsten Slok. It’s much easier and cheaper to launch a company today with the help of LLM-based AI. A mediocre ChatGPT lawyer, finance director or marketing manager is, after all, better than none.

In the hysteria of the AI narrative, all of this is boringly plausible. It’s a technology which reduces the costs of certain classes of cognitive activities; so we’ll do more of them.

Governing self-building AI

Anthropic released some fascinating data on how Claude has changed the way Claude itself is built. The amount of code contributed per developer is eight times higher in the current quarter than the average to the end of 2024. There is a meaningful uptick starting in 2025 as Claude 4 and Claude Code were rolled out. We discuss whether lines of code are a meaningful metric below.

I found this chart more interesting—showing how well Claude Code can complete different classes of tasks. It appears that the Mythos, Anthropic’s latest release, which is only available internally and to select organisations such as the NSA, has demonstrated a step change in capabilities.

Anthropic took the opportunity to warn against the risk of recursive self-improvement, going as far as suggesting “it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up.” To be fair to Anthropic, they’ve been reasonably consistent with that idea for a few years—as have others.

A multi-lateral process can’t hurt things much, especially in today’s fractious environment. However, I’m sceptical about this threshold for “recursive self-improvement” as a notion of a runaway technology. For two reasons. First, we are dealing with increasingly abstract layers of automation that accelerate the product roadmaps of human-directed companies (not out-of-control machines). So there may be a risk, but it isn’t recursive. It’s a strategy. The second is that commercial realities act as a natural attenuator: securing the capital, the chips, and the power will prove less tractable than pouring out reams of Python code.

When we all do it, it’s harder to stand out

Read more

AI companies have, until recently, been a lot like gym owners, offering generous bundles. And most AI users have been like gym-goers, occasional, sometimes frequent users. But others have been voracious and incessant, churning through their Claude and ChatGPT subscriptions.

And now, the AI labs have changed their tune, introducing more usage caps and metered pricing, particularly for coding tools.

So does moving from bundled pricing to metered pricing expand or shrink markets?

It depends.

Bundles expand markets when marginal costs are low and when customers get a diverse set of benefits from the bundle. A gym membership is a good example: it doesn’t matter whether the gym-goer visits twice a month or 10 times a month; it costs the same for the fitness company to serve. At the margin, if everyone went all the time, the model might break due to congestion and increased cleaning and maintenance costs.

However, if a product has variable costs, a pricing bundle is really a decision about who bears the risk of over- or under-use. If the bundle favours the buyer, the buyer has no incentive not to max it out. Few of us can spend more than 20 or 30 hours a week in a gym, but many of us can run riot with AI coding models. And AI models do run up substantial variable costs for those peddling them.

Chatbot usage is all over the map. Sarah Friar, OpenAI’s finance chief, noted that ChatGPT Pro users hit the app 11 times more frequently than active free users. That 11x is wide enough. But consider what happens with agents. If I am hunting-and-pecking in a chatbot, I’ll struggle to consume 100,000 tokens a day. R Mini Arnold, my agent, won’t get out of its SSD for less than 100,000,000 tokens a day. I spotted one user racking up 130 billion tokens in a month.

Moving from bundles to usage-based pricing brings with it sticker shock. But it is manageable. Consider Uber. The taxi company has put an AI cap on its 5,000 developers. Today, they are limited to $1,500 per month or $18,000 dollars per year per agentic coding tool. Like Oliver Twist, developers can ask for more. If all 5,000 developers spent the full $18,000 a year, the bill would be $90m. Against Uber’s 2025 free cash flow of $9.8bn, that is less than 1%. It’s essentially immaterial; a good CFO can deal with it in a heartbeat. And Uber is likely to be a heavier spender than most of Main Street. The pricing debate isn’t really about affordability.

It’s really about whether the customer can connect spend to value.

Back to the future

Fortunately, we have a good precedent that can help us think this through. In the early days of Internet advertising, going back to AT&T’s banner advert on Hotwired in 1994, advertising was sold on a cost-per-mille basis. The advertiser paid for a bundle of impressions. Perhaps users clicked, perhaps they didn’t.

Today, Internet advertising has moved to a metered model. Customers don’t buy a bundle of impressions; they buy a metered outcome.

The question is: what happened to ad pricing as we moved to the metered models? And what happened to the size and profitability of that market?

Of course, you know the answer, but it is worth seeing the pattern. When pay-per-view was introduced, it helped grow that market and is now vastly preferred.

Read more

Since we first ran this analysis in September last year, a lot has changed. Some 170 AI models have been released. The best now handle tasks four times as long as last year’s top performer. Usage has responded; quarterly token consumption has tripled.

The money has followed. The NASDAQ is up 20%; quarterly capex commitments have jumped 43% to $158 billion; and AI sector revenues have nearly doubled to $25 billion in Q1 of this year.

We’ve reviewed the data against our five empirical indicators, derived from analyzing 300 years of investment booms and busts. In our model, two indicators turning red indicate bubble conditions. Only one of these indicators is in the red; the rest range from green to amber.

Here is what these biomarkers show—and what it would take to change our minds.

Economic strain

For economic strain, we look at the total capex into AI, which is predominantly in building out the data centers and power connectors, the chips and cooling, and other things that go in them. Here, we consider the US portion of that divided by US GDP.

Capex has risen dramatically since September 2025, from $110 billion to $157.7 billion per quarter. Neoclouds are now contributing meaningfully to this number, with their share of total capex rising from 12% to 18%.

This rise in capex has led to economic strain exceeding 1% of US GDP, marking the first time it has entered the amber zone. AI capex now matches the scale of the late-1990s telecom build-out at its peak. Goldman Sachs estimates that aggregate AI capex will approach $1 trillion dollars in 2027. If we assume about 70% of that is in the US, this would be the indicator into the red zone towards the end of 2027.

An important note on our methodology: we used trailing twelve-month numbers for both Capex and GDP to measure economic strain. When capex rises so quickly, this may attenuate the indicator. On a quarterly basis, it actually doesn’t change the prognosis much. Economic strain remains in the amber, closer to green for this quarter.

Industry Strain

Industry strain assesses whether the industry can afford the investments it’s making. Investments always come ahead of revenue, but speculative build-outs that predate real revenues, as seen in the telecoms and dotcom bubbles, are deeply problematic.

Quarterly revenues have roughly quintupled year on year: OpenAI from $1.7 billion to $6 billion, and Anthropic from $400 million to $4.8 billion. Based on our own bottom-up tracking of deduplicated sector revenue, we now put the figure at $25 billion a quarter, up from $13 billion in September.

One note on that number. Since the first piece, we’ve refined how we calculate sector revenue, moving to a more nuanced value-added approach that reduces double-counting across the cloud, model-lab, and application layers. Our original September figure for Q3 2025 was $25 billion; under the new method, that same quarter comes out at $13 billion.

This keeps industry strain in the red for now. But if revenue growth holds on trend – and capex forecasts don’t revise upward – it should exit the red by the end of the year.

Revenue Momentum

Growing revenues address industry strain. Ultimately, if real customer revenues accelerate fast enough, they will pay for the investments being made. We measure this through doubling time: how long it takes revenue to double.

As a market matures, you’d expect revenue growth to slow, and that was our assumption in both the original September work and the early-2026 update. We forecast a deceleration. That isn’t what happened.

Revenue growth has actually accelerated, particularly by Anthropic at the turn of the year, surprising us, analysts, and even Dario Amodei.

“We tried to plan very well for a world of 10x growth per year. And yet we saw 80x. And so that is the reason we have had difficulties with compute.”

Rather than worsening as we expected, the indicator improved and is now well into the green, with revenue doubling every 0.73 years. Even if revenue growth for the rest of the year drops by 75 percentage points, this indicator would still sit in the green going into 2027.

Read more

This week, we published a framework to explain why individuals are getting more productive with AI, but firms often are not. Make this your weekend reading.

If you want me to speak to your board about this, now is the time to book me for the fall.

Subscribe now

Linear knives in exponential gunfights

A column in the Financial Times made a fascinating observation about “the impossible maths of the AI boom.” It pointed out that in the five years to 2030…

AI capital investments are expected to rise by 20% a year, a growth rate never seen before in this industry. Meanwhile, revenues are expected to grow 15 per cent annually…if the hyperscalers continue on the current trajectory, the AI boom will become a story of one of the largest destructions of shareholder value in history.”

Punchy stuff. To get these conclusions about the hyperscalers' ignominious future, the author used “the consensus estimates of analysts for the capital expenditures and revenues between 2025 and 2030.”

This all sounds reasonable. Equity analysts are, of course, paid very well to understand the companies in their sector and work long hours to do so. A consensus estimate is the average of what these well-heeled professionals think.

The problem is that they are, on average, late to regime shifts.

In particular, when a technology hits a learning curve and demand explodes, producing that familiar hockey stick, analysts often stay stuck with their old assumptions.

Yes, yes, exponential curves don’t, in the real world, go on forever. But when they are going on, they are, in fact, non-linear. And the straight line and the exponential curve don’t fit well.

You can see this by a simple evaluation of just how well analysts have made sense of booming AI demand. When you or I spend money with OpenAI or Anthropic, the majority of it ends up with the hyperscalers, who, in turn, are pouring hundreds of billions into buying chips, memory, and other trinkets. While Wall St analysts don’t formally track Anthropic because it is a private company, they do track those firms in the supply chain.

Micron for on. Micron is an American company that makes advanced memory popular with hyperscalers. LLMs chew through memory, and Micron has done well as a result. But how well did the equity analysts covering Micron do?

The chart below shows the consensus estimates for Micron’s earnings per share. Pay attention to the purple line, which is FY2026. As late as December 2025, the consensus view was that Micron would deliver about $18.25 EPS. Five months later, with some early data to help them, the median view is now about $58. So what was the point of the $18.25 forecast?

Of course, Micron isn’t a hyperscaler. But it is unusually exposed to AI. Only three companies make fancy high-bandwidth memory; Micron was the minnow and the only American player. Maybe this pattern of revisions was an exception, not the rule. Maybe Micron was just … weird.

So consider a larger company like Google, with a 20-year track record. AI, understandably, is only a small part of the firm’s $300 billion revenue business. If you and I are buying heaps of tokens from Google, it’s just a drop in its overall business. But even then, between May 2025 and May 2026, the median analyst view on Google grew by 40%.

What to expect with your expectations

We all know what a share price is. It’s the price right now that's tradable on a stock exchange. But it represents a view of the future: a compressed probability distribution over future outcomes for the business.

If a stock price reflects the market’s view now, an analyst's target represents something specific. It is their expectations for the stock’s value.

Michael Mauboussin, a bit of a guru in the world of finance, has one of the best definitions of expected value: “the weighted-average value for a distribution of possible outcomes” of a company.

So when you, I, or an analyst determines a share price target, we aren’t just choosing a number. We are expressing the outcome of the probability calculation, a distribution across possible futures for that asset. The keyword is distribution. And the shape of that distribution depends on your assumptions. If you think this is a linear, business-as-usual technology, you will make one set of assumptions. You’ll be ignoring the adoption rate, cost declines, performance improvements, or the possibility of an extreme outcome. You haven’t just chosen the wrong point on the curve. You’ve chosen the wrong curve.

Sorry, sir, I misjudged the speed of that iceberg.

And yet you constantly revise your expectations upwards, chasing the puck rather than getting ahead of it. Below, you can see the median analyst consensus for three companies with varying exposure to the AI wave. I’ve plotted that median expected value from 12 months ago to a month ago. They keep revising higher and are still well below what the market thinks right now.

These revisions are significant. Analysts aren’t just missing the number once. They are missing it repeatedly. That isn’t bad luck; this is systematic, wilfully blind to what happens with an exponential technology. We’ve seen this before with the IEA and their two decades of comedically ignoring solar’s falling costs.

Analysts covering stocks affected by the growth of AI use are doing that. Their model of the world doesn’t accommodate a range of “possible outcomes”, so the expected value (their price target) is wrong. One problem is that no-one in the major banks really covers AI in its entirety. You have chip, cloud, and software analysts. But the action, the demand action, is in private companies.

A second problem is that time and again, the market is good at digesting financial information and poor at stomaching technological change. There is a sociological dimension at play, not exactly a kind of groupthink, but an intellectual tic at work. One example is reversion to the mean. Mean reversion, the idea that you can’t stay off a long-term trend for long, is a powerful heuristic. And it works most of the time, but there are some processes, especially compounding ones, where it is less helpful.

Reversion is mean

JPMorgan published a chart showing consensus forecasts for hyperscaler capex and R&D as a proportion of revenues. This is a data set similar to the one above, although it tells a different story, as it shows not raw dollars but share of revenue.
But notice what the consensus says about 2027, a massive crash; a 25-40% reduction in those investments as a share of revenues.

Assuming the hyperscalers are acting rationally, only three things could drive that slump.

1. A substantial slowdown, or perhaps retrenchment, in AI revenues during this year (which would then affect next year’s planning)

2. Some kind of scientific or technical breakthrough that allows hyperscalers to meet demand without the same levels of investment, and to roll this out to AI systems serving quadrillions of tokens a week within a single planning cycle across the hyperscalers' complex, heterogeneous infrastructure.

3. Some massive exogenous shock that seizes the global economy and pushes every firm into defensive posturing.

Which of those three reasons can it be? On the second, it is not the job of an infrastructure analyst to forecast scientific breakthroughs, which even Demis Hassabis or Yann LeCun can’t. So it can’t be the second. Nor can it be the third — forecasting a massive exogenous tail-risk shock, a black swan that suddenly arrests the global economy, is also not their job. They are not paid to be Nostradamus.

So the only possible reason is reason number one. But we’re halfway through the year; everyone is caught by growth rates, including the AI labs themselves, so what are these analysts seeing that no one else is? Nothing. The reasoning is inchoate.

Truth is, exponential shifts involve regime changes, not merely rerating1. That is uncomfortable. It is easier to assume reversion to the mean than to update your models, which worked so well in the past, with one which is dynamic, uncertain, full of feedback loops and, crucially, depends on the future being radically different to the past. That seems like hard work at best, crackpot at worst. Easier to tweak and tweak old faithful.

Let’s go back to the column in the FT which questioned the rationale of AI capex investments based on consensus analyst forecasts. There is an important caveat this particular argument needs to acknowledge: it relied on consensus forecasts from a class of models that have been wrong, sometimes by an order of magnitude, on the very thing their forecasts are used for.

The lesson isn’t that analysts are too pessimistic and should just revise higher. It is that consensus forecasts that are built on the intuitions of mean reversion are the wrong thing to use during regime change. It’s not about whether the numbers will surprise to the upside. It’s whether the framework can even see what’s coming.

Don’t take a linear knife to an exponential gun fight.

Subscribe now

AI is making us dumber smarter … who knows? 🤷‍♂️

Research across American college students, looking at 370,000 personal statements, showed that “after ChatGPT became available, their essays suddenly used diverse and colorful language, but lacked truly creative ideas.” Rebecca Winthrop’s essay is a must-read. “For the first time in human history, we have a technology that can generate words separately from the thoughts they represent.” It’s a shortcut. Students love it. And with that, we’re losing our grip on thinking.

Or are we?

shows that being roundly trounced by AlphaGo has, in fact, improved the quality of human Go play.

Yes, maybe AI is making us dumber.

Here is The Economist with an interesting survey on whether AI is making us stupid. I love the irony that the site offers me an AI summary on an article that, in essence, discusses whether AI tools are making us intellectual wimps.

You probably recognize, feel, many of the conclusions. As we hand certain skills over to AI, those skills might atrophy. This happens in other areas, like motor skills, so it could apply to more complex cognitive activities. Some of the best things humans seem to be able to think involve integrating lots of ideas from our experiences, listening, formal instruction, doing and joining the dots.

The word intelligence is derived from the Latin intellegere. That Latinate root itself comprises inter, meaning between or among, and legere, meaning to choose, select or to read. Intelligence means “to choose between” — to connect, to join the dots.

Perhaps if we hand the actual act of intelligence (the connection-making) to AI, we’ll weaken that muscle.

If you don’t use it, you’ll lose it.

It sounds like common sense. The studies cited in The Economist provide some scientific backing for that idea.

Caution, though: these are small-scale studies. Lots of things sound like common sense until we can study them properly. That video games drive real-world violence, for example. Common sense, of course. But it doesn’t stand up to the research.

Now I don’t know if AI will destroy human thinking. My gut sense is that some things will get worse and some things might get better.

I am acutely aware of these risks. And we should take seriously what educators and students are telling us now about how their attention and quality of thinking seem to be eroding.

Thinking about thinking

Readers may imagine that with our billion tokens plus a day, Exponential View Towers is like something out of Futurama. Half of us in VR headsets, another half plugged in to our Neuralinks, and the last half deep in the Matrix.

But that’s not the case. Thinking is very, very important at Exponential View; we care about it a great deal. Many of our early experiments with ChatGPT focused on how it could improve the quality of those cognitive efforts, not on how we could take a shortcut through the valuable part of the process.

AI has allowed us to take away lots of “computer stuff”. Typing my essays and responding to tedious emails are now handled by RMA or Claude.2 Back in December, I used to get up early to fire off new workloads. Since the models have become more reliable, I’ve been able to leave them unattended for hours at a time, whether they are running coding tasks, simulations, literature reviews or adversarial attacks on our arguments.

Paradoxically, AI has made me much more aware of my thinking processes. I have become deliberate in reflecting on my reflecting. LLMs are absolutely phenomenal and show the banality of the average. I’ve never read as critically as I do today.

Because thinking matters to us so much, we spent a lot of time figuring out how to keep mentally fit.

Here is a picture of one of our whiteboards.

You can see the scoping work on the thesis on why companies struggle with their AI transformations in green marker. And in the yellow box is a little exploration of Durkheim and suicide for something else we are working on. It’s more whiteboardmaxxing than tokenmaxxing.

Last year, I bought the team fountain pens. Everyone is expected to spend at least two hours a week working with pen and paper and without a computer. I don’t want to overclaim that this makes a difference; the science suggests it might.

I’m writing longhand more than I have since those halcyon days when Boris Yeltsin was a communist, Margaret Thatcher was Prime Minister and Donald Trump was married to Marla Maples. It has been a long time since I picked up a fountain pen in anger.

I’ll work on paper, while RMA will go off and do its own research. It goes through papers, my notes, and transcripts, finds what I might have missed, and counters. When I return to my screen, I am several steps further than I would’ve been without it. But I’m sceptical about its arguments. Sometimes they clear a mental block. Sometimes they are helpful because I can treat them as directions not worth pursuing.

While I don’t have scientific evidence that definitively tells me whether AI is adding to our mental faculties, you can sense which way I have buttered my bread.

Asking the right questions

The question “Is AI making us dumb?” is not the right question. The real challenge that technology has changed the basis on which we expect thinking to occur, and we haven’t made our new expectations clear. It’s an exponential gap.

Few marathon runners would complete a race using a car. The point is the race.

Getting ChatGPT to do your thinking is a bit like finishing the marathon in a car. So let’s investigate the motivations for doing that. Perhaps people don’t value the thing they are doing and just want to get it done? Perhaps they can’t judge what excellent is, so they will be satisfied with good? Perhaps the activity is simply instrumental, a means to an end, not a constitutive activity with value in itself?

We’re asking people to think about college applications. Ostensibly, the application determines their suitability for the next four years. To get an AI to do it—rather than, perhaps, act as a Socratic critic—speaks to the perceived stakes of that application. The point is not that modern life will be full of AI; it is that to use AI well, you need to think well. Unless, of course, you’ve concluded that AI will always outthink you in all the ways that matter.

It’s also worth examining in the workplace. Of course, there are times when its appropraite to get a machine to do a task. I don’t grow my own cacao beans. But perhaps it reflects workplace environments geared to throughput rather than quality, agency and ownership?

To reverse Charlie Munger, show me the incentive, and I’ll show you the outcome. The outcome here is offloading important thinking to a chatbot. So what incentives produced it? We’ve developed tests, credentials and career ladders and cemented them across centuries of practice. They were built when cognition was scarce and scarcity paid well. Thinking proved you had that scarce asset, and we never needed to ask if you were actually the one doing the thinking. That isn’t true now, but the payoff for perfecting the scarcity theatre of a perfect GPA, Ivy education, and the rest remains.

So what system of incentives do we need that values (and motivates us) to use our actual brains?

And that, rather than whether AI is making us dumb, is the right question to ask.

And it is a much thornier one to answer.

Subscribe now

Short morsels to appear smart at dinner parties

AI software engineering roles are growing and concentrating in the top-paying US companies according to the latest state of the software engineering job market by and .

Implementation expertise is in high demand.

And Anthropic is seen as the most desired company to work for, followed by starting one’s own business. via

digs into China’s AI optimism: “perhaps every respondent genuinely believes that AI is good for both society and for themselves. Or perhaps they see AI as another surgery necessary to survival, knowing full well that flesh will be cut away and discarded.”

ByteDance is designing its own Groq-style AI accelerator chips to lessen its reliance on Nvidia.

Non-fiction book publishing is absolutely not ready for AI.

China ETFs are sinking.

Lack of sleep changes our gut microbiome.

How to weigh a cell. A nice piece by

Anti-drone nets in Ukraine:

Thanks for reading!

She is not alone. Uber’s COO Andrew Macdonald went on record this week saying that the relationship between AI investment and results is not there yet:

I think maybe implicitly there is more that is getting shipped, but it’s very hard to draw a line between one of those stats and, ‘Okay, now we’re actually producing 25% more useful consumer features.’

AI has delivered something. I have felt it; my team has felt it; most users have felt it, which is why we keep returning and using more of it. Two years ago, only a dozen Anthropic customers were spending over $1 million a year on Claude1; today, more than 1,000 do. More impressively still, Anthropic’s average corporate customer increased their spend by a factor of five in the past year.

But in more than three years since ChatGPT’s release, only 27% of executives say AI has met their ROI expectations. What do we make of the other 73%? Could their expectations be too high? Or too low? Do they even have the right class of expectations?

In a way, we can’t tell, but we can feel the vibes. And the vibes are that individual workers are getting faster and more productive. But for now, those individual gains from AI do not compound into firm-level ROI.

That is the puzzle we are going to solve in today’s essay.

Let’s go!

Subscribe now

The productivity puzzle, restated

Back in 1987, Robert Solow famously pointed out that you could see the computer age everywhere but in the productivity statistics. He was right for the next few years, then wrong. Paul David realized this was a common problem with general-purpose technologies. They systematically depressed measured productivity in their early stages because companies needed to invest in all sorts of hard and soft know-how before the gains appeared. calls this the productivity J-curve: general-purpose technologies are a drag in their early years because firms have to make complementary intangible investments before the gains materialize.

Paul David’s 1990 paper on electrification is the canonical account of why a general-purpose technology can sit inside firms for ages before we see the results. The story he tells – building on Warren Devine’s earlier survey of the shift from shafts to wires – runs through three phases. And those phases map onto where AI is now.

Stage 1: The lightbulb

One of electricity’s first factory roles was the simplest – lighting. A brighter floor was safer than one lit by gas, and cleaner than one lit by oil. But work still flowed through the same sequence of people, machines, shafts and belts. Electricity had improved the workers’ immediate environment, but it did not change the factory’s operating logic. When ChatGPT was first released, it did something similar. It increased how quickly we could write emails; individuals sped up on some tasks, but the firm did not. This is Stage 1 of AI transformation: the lightbulb.

Most of the AI products we see today are all about individual productivity. Yes, there are enterprise plans for ChatGPT and Claude and whatever else. But the unit of work is still the task that the individual has to hand. The enterprise plan just lets them quickly access the corporate skills repository.

Stage 2: The group drive

The next stage of electricity adoption in factories focused on cost savings rather than productivity. Louis Bell wrote in 1891 that large central steam plants could be five to seven times more coal-efficient than small engines. Factories bought power from central stations and installed electric motors to drive their existing shafts and belts.

Then, a professor of electrical engineering, F. B Crocker, and his colleagues found another application. Electric power frees the shop floor from the shafting. Mechanical power belts, tools and oil, the mess of the factory floor, could all be moved; machines no longer had to be arranged in parallel lines beneath shafts.

The open question was how many motors a factory needed: one per tool, or one per group of tools? The latter became known as group drive. It was a single motor that powered a cluster of machines via a shared shaft.

Group drive preserved existing layouts, reused sunk capital, needed fewer motors, and gave many of electricity’s benefits without the cost of rebuilding the plant. It was cheaper and easier than the alternatives, so it won and dominated factories through the 1890s, 1900s and into the First World War.

AI agents are better than chatbots. They can handle whole workflows rather than single tasks. But, like group drive, they are attached to the existing organizational geometry. In the case of electricity, this was the shop floor layout. For AI, it’s the web of processes designed by the companies well before anyone knew what an LLM was.

An AI recruiting agent speeds up a process that was previously done by humans and an applicant tracking system. Their recruiting pipeline may shrink from weeks to hours. A customer service agent takes on more tickets than the support team could before. The logic behind these examples is essentially cost-saving – fewer support tickets need humans, recruiters screen more candidates with the same headcount and marketers create more variants without additional staff. The firm isn’t making consequential decisions any faster. This is Stage 2, the group drive, machines turning faster on the same shafts.

Stage 3: The unit drive

It was only later, once the organizing logic of the factory moved to throughput from cost-saving, that the deeper value of unit drive, one motor per machine, became clear. In 1913, Ford’s Highland Park plant decided to orient machines and workers around the workflow rather than the geometry of shafts and belts. Over the next decade (1919–1929), as more factories adopted unit drive, US manufacturing labor productivity grew by 5.4% a year.

The pattern for AI will mirror the pattern David spotted for electrification. Stage 1 speeds the individual, Stage 2 a workflow and Stage 3 the firm.

Subscribe now

The ladder

You’ll be familiar with other maturity models, like Carnegie Mellon’s Capability Maturity Model. These are useful, but they treat each stage as a monolithic capability tier. A capability tier tells you how well a firm does a “fixed thing.”

Our ladder is different. The stages are organizing logics. The logic is the goal the firm is pursuing with the technology. A workshop that installed electric lighting was no worse than Ford’s factory, but it was pursuing a different goal: a safe way to illuminate the workshop floor. You didn’t get to an assembly line by adding more lightbulbs. Factories were pursuing cost savings and everything was organized to fit that goal, from layout to staffing to supplier relationships.

A maturity model will often tell you to do more of what you are doing to move forward. We reckon that what matters is what you are trying to do. And our stages hint at why companies deploying AI might get stuck. A firm does not graduate stage by stage on every axis. An individual developer who is 50% more productive with AI tools but must submit to traditional review cycles will find themselves in a queue. The product team that can prototype faster than ever but needs to wait for sign-offs, will build up a backlog of features. A sales team, supercharged by AI-assisted proposals, may close deals faster than legal can review them.

As firms move towards Stage 2 on execution, those managerial layers, how decisions get made, stay where they were. We call that mismatch congestion. It is the buildup of individual and then team outputs waiting for somewhere to go. To become a Stage 3 firm, you need to rebuild around decision speed between workflows, not the speed of individual workflows. If you are already suffering from congestion, adding more workflows and more output to a blocked decision pipeline will only make things worse.

We discovered this at Exponential View. We could prototype, even build, new features faster than our usual processes for releasing them. We’re figuring out how to fix that – and we’re only a team of eight, so we are sympathetic to the challenges of larger orgs.

To speed up those decisions, AI needs to be able to take them. The role that managerial oversight has may be the thing preventing the cycle time from shortening. AI will need a new cognitive layer that allows the firm to interpret signals without a worker as an intermediary. If a customer sends a feature request to customer support, it typically passes through a support agent and then a product manager, who decides whether it belongs on the roadmap, before a developer codes it weeks or months later. With AI, the signal is observed by an agent. It orients against the roadmap and the codebase. It decides whether the feature is worth drafting. It builds it. Hours, not weeks. This is the Stage 3 firm, the unit drive.

Subscribe now

How to move to Stage 3

Read more

Last week, we explored what tokenmaxxing means for CFOs and how firms can buffer unexpected AI costs.

We go further today to show you why AI bills are hard to forecast today and what will happen as we crack that problem.

Subscribe now

The token explosion paradox

We estimate that the number of tokens processed per quarter has grown by around 17,000x over four years.1

Token prices have collapsed during this time. Demand for machine intelligence is highly elastic, meaning that as prices fall, consumption increases by more than the decline in price.

One reason is that cheaper tokens have made agents economically viable. At the same time, agents use tokens at rates that are orders of magnitude higher than those of chatbots for single-turn queries. That shows up as the total tokens processed per output token—advanced models do a lot of processing below the surface that a user doesn’t see.

A lot of this growth is driven by China’s domestic demand and its model providers, especially ByteDance and Alibaba.

Subscribe now

The cost of the ghost token

When you use an AI agent, the final result you get is really just a summary of all the work the agent has undertaken. There may be dozens of tool calls to browse the web or load up a file to check and validate the work it has done. All of these are steps consume tokens: they become hidden multipliers.

The first of these is token amplification.2 A coding agent that operates over 10 turns might need to re-read its full context every turn. That repetitive reading of context could use as many as 55x more tokens than a single-turn query for the same task.

Actual active inference is probably only 15-20% of the total token consumption. The rest is invisible work that you, as a user, and possibly the company paying for it all, haven’t modelled.

The long tail of tool calls

Agents make anywhere between five and twenty-five tool calls per task. And each call adds more context, tokens and API costs. It also increases the likelihood that the model will need to retry the task to get it right.

The price of safety

Governance and safety costs3 are between 20% and 40% of total spend. Amazon Bedrock charges $0.15 per 1,000 text units for content filters, and Azure charges $0.38 per 1,000 text units. When you add those charges on top of model inference, you could be looking at 15-30% on top of the existing costs for every call.

For firms that use a frontier model as a judge to evaluate another LLM’s output, the evaluation step will cost almost as much as running the main model again. One study found this approach to be 18x more expensive than using smaller, purpose-built guardrail models.

Subscribe now

Forecasting is off

What makes this particularly hard to manage is that the tools themselves cannot yet forecast their own costs. A new paper found that frontier models fail to accurately predict their token use.4

Runs on the same task can vary by as much as 30x in total tokens. So even those of us who have developed a sense for token consumption run into surprises. In the study, 6.7% of tasks lasting less than 15 minutes cost more than the average 1-hour task.

Subscribe now

The cost management ladder

So, what does this mean for companies? We will see firms climb the maturity ladder as they figure out how to track token spend and likely evolve into more sophisticated observation and monitoring of model calls.

Firms will benchmark their internal processes against their previous experience and, when the data is available, against those of other companies. The ambition will be to optimize price-to-outcome ratios. This will all help clarify the unit economics for specific commercial outcomes. The value of observability and monitoring will increase, as evidenced by the share price rise of Datadog, a leading observability platform.

Long-term, attributing costs to specific teams, workflows, and customers will help finance teams budget for AI spending and decide how to route and tune model use.

Taken together, the firms with the most mature approach to their AI deployments should start to understand the true unit economics of their outputs in the coming quarters.

Consistently insightful, with no agenda beyond understanding how technology shapes our world. — Vincent N., a paying member

Subscribe now

The backlash is worsening

American college graduates are angry. In my Saturday column, I write:

The narrative around AI has been about promises for tomorrow, but sacrifices today. […] Engaging with this resistance won’t be effective if AI leaders throw distant fantasies at people dealing with muddy water coming from their kitchen tap today.

AI’s creative expression

An OpenAI reasoning model solved an 80-year-old open problem in discrete geometry. The interesting point here is that the reasoning model found an unexpected bridge between two different fields of math. Discrete geometry and algebraic number theory are meaningfully separate cultures. An expert in one field might know something about the other, but not at the depth required for breakthrough research.

The human mathematicians who validated the proof reckon that the connection the AI found was both unexpected and non-obvious. In a sense, it’s reminiscent of AlphaGo’s Move 37.

Because AI systems can span different domains of knowledge, my hunch is that one of the biggest dividends we will see from them in these fields will be connecting domains that live in isolation because of the structure and specialization of modern science.

But of course, another way AI will impact research is by speeding up the scientific method: the full hypothesis-experiment-analysis loop.

Robin, a multi-agent system, completed the full cycle of hypothesis generation, experimental selection, data analysis and refinement. Human scientists ran specified experiments in a wet lab. The outcome was the identification of an existing drug that could be repurposed to treat macular degeneration.

Read more

But the backlash against AI is growing ever stronger.

Earlier this week, Google’s former boss, Eric Schmidt, was soldily booed at a college commencement address

At another college commencement, boos shocked the commencement speaker who proselytized AI. Joanna Stern interviewed one of the college students, Houda Eletr:

Joanna: Do you use AI?

Houda: Throw it away….if you create this big thing that is supposedly going to save a human’s life, then you should give it to me in a human way.

A reader from New York sent me this photo:

There will be more booing.

Consider data centers and their local impacts.

At a hearing, American politician Alexandria Ocasio-Cortez showed a jar of mud-brown water collected from a tap in Georgia, shortly after Meta started building a data center.

Poor ambassadors

The narrative around AI has been about promises for tomorrow, but sacrifices today. AI leaders have warned that the world could be destroyed by this technology.

They warned that large numbers of people, particularly graduates, would lose their jobs. But everyone knows the founders are making more money than Croesus. And now, there is construction, a physical manifestation, a real inconvenience, something you can point at.

And for what? So we can explore the stars using intelligent spaceships and ‘colonize the galaxy’ according to Demis Hassabis. (Demis is by far the most sympathetic of all the AI leaders, but even his messages feel out of touch.)

Engaging with this resistance won’t be effective if AI leaders throw distant fantasies at people dealing with muddy water coming from their kitchen tap today.

Neither will facts. The US energy system is getting a much-needed boost in investment to upgrade energy generation and the grid to more modern technologies. I am certain that in a decade we’ll look back on this as the time when the US grid prepared for the 21st century, becoming more resilient, efficient and greener. But that well-informed argument is completely irrelevant in the face of today’s resistance.

Fantasies or facts, these are all fodder for the technorat. Globalization or NAFTA made a graph on some economists’ slide deck go up. Grid renewal and ‘productivity’ excited the same neurons, but don’t speak to the heart.

And the AI industry has a real problem because the resistance rests on tangible issues today. Just a few years ago, the sense of unease many people had about AI was fissiparous and intangible (remember the asinine PauseAI movement?).

Wage devaulation for freelancers and gig workers is here and now. Data center construction, the threat of rising power bills and dirty water, here and now.

Pressure on graduate employment, whether driven by AI or not, will increase. I’ve spoken to several senior execs at US firms in finance and consumer services in the past few weeks. Their AI projects emphasise headcount reductions or, euphemistically, productivity. While these don’t show up in aggregate statistics yet, the perception that they do is what counts.

The AI companies, in their San Francisco bubble, pursue increasingly abstract technical milestones. Each new release beats a benchmark but comes with another warning of the lasting harm the model will do to what it means to be human. No wonder American voters find AI less appealing than ICE, the frequently masked paramilitary agency that enforces immigration law.

Data centers have become the first issue in decades to unite both sides of American politics. Resistance has become a truly bipartisan issue. Washington can’t be relied upon to settle things. The federal government has chosen to delay, often theatrically. The White House wants a light-touch national framework and has pushed agencies to challenge state rules, but even its own recent AI executive-order has been derailed.

The states are moving instead. From Colorado to New York, California to Texas, and a thousand more state-level bills around AI have been introduced.

Local politicians will discover that an anti-datacentre vote is a cheap way to electoral success. The midterms this October will be one to watch. But nothing meaningful will happen in Congress until at least after the 2028 election.

This will not stop AI. It will not stop the construction of data centres, but it will be boon time for lawyers as they first challenge the legislation and then work out how to comply with it.

Meanwhile, companies and consumers will continue to spend on AI. The technology will improve. American firms will report successful large-scale AI deployments. Revenues will rise, records will be broken.

And the crescendo of boos will rise and rise and rise.

This week’s Monday edition looks at one of the fastest-growing, least predictable costs in the AI stack: tokens.

Let’s go!

The bill no one budgeted for

Uber CTO Praveen Neppalli Naga shared last month that his 5,000 engineers had depleted their entire 2026 token budget in just four months. So has ServiceNow.

Agentic adoption was bound to drive this kind of demand and the finance has to respond. CTOs are increasing their tech budgets this year – nearly 50% say their budgets are up by 10%. (As a side note, we believe that 10% is marginal given the token explosion we are experiencing.)

A token explosion is great news for engineering teams, but a real headache for CFOs who signed off on modest pilots months ago. 71% of companies exceeded their AI budgets in 2025 and over half of the surveyed finance bosses say cost management is their greatest concern.

Partially because AI costs can be highly variable and tricky to manage. The cost is not fixed as it was with good, old fixed-seat SaaS.

Is this the new normal?

Exponential token use is diffusion in action. In the US, the average monthly spend on AI by large enterprises grew 36% to $85,000 between 2024 and 2025.1

Labs in China told us that coding is where most labs are throwing their resources right now, their P0 in engineering terms.

Read more

This is the most relevant newsletter I receive in my inbox every week. – Martin T., a paying member

Subscribe now

Hi all,

Suffocated by export controls, Chinese AI labs have developed an efficiency moat that may define the AI market’s development over the coming years. My report from a trip to China.

In this Sunday’s edition:

• Anthropic powers into Main Street as Microsoft and OpenAI consciously uncouple

• Who should really control your AI?

• Voltaire, the Enlightenment entrepreneur

The demand

Anthropic’s CFO Krishna Rao gave a fascinating interview on Patrick O’Shaughnessy’s podcast. It’s a cornucopia of insights: for one, the firm’s enterprise customers increased their spending by a factor of five over the past year. That 90% of Anthropic’s code is written by Claude Code is well known, but Krishna says that 90% of finance reporting is now AI-driven as well. Cowork is growing faster than Claude Code (which grew from zero to $1 billion in just six months). Krishna’s been on quite the ride himself – he joined Anthropic two years ago when revenues hit $250 million and now presides over annualized revenues approaching $50 billion – a fivefold increase in as many months. Recommend a listen.

Elsewhere: Anthropic leads OpenAI in business adoption, according to Ramp.

A profitable, unhappy marriage

OpenAI and Microsoft are no longer exclusive. OpenAI now gets to play with other resellers and Microsoft gets better access to the startup’s AI, as well as drops complicated revenue-sharing arrangements.

It’s been a profitable marriage, if not an entirely happy one.

Microsoft’s $13 billion OpenAI investment has yielded over $30 billion in revenue in the two years since ChatGPT’s launch. OpenAI was the biggest customer of Microsoft Azure’s AI business, ploughing $23 billion into the hyperscaler. (Our estimates suggest up to 60% of Azure’s AI revenue came from OpenAI). Microsoft’s operating margin expanded by 3.9% over the last three years.

Read more

On the first day, IPO pops bring out the bears and cynics for good reason. The dotcom heyday was full of them. Calico Commerce, VA Linux, TheGlobe.com: each up 300%, 605% and 697.5% on day one, respectively. Remarkable performances, long-term disasters. VA Linux, the “best” of them, lost about 98% of its value. TheGlobe.com and Calico were effectively wiped out.

What did they have in common? Low revenues, fast growth rates and a booming, booming, booming market.

So when Cerebras opened 75% above the expected IPO price, it was tempting to file it away as another overhyped stock riding the frenzy around a new technology, itself bundled in layers of hype.

What I saw instead was a sign that the market is finally starting to grasp the demand for AI inference.

I’ve been following Cerebras since 2018, when I first spoke with Andrew Feldman, the founder. Last year, and I visited their headquarters in Sunnyvale.

Read more

We participated in dozens of hours of discussions with researchers, founders, product leaders, and business owners across the infrastructure, hardware, models and application layers. Every lab is obsessed with ByteDance’s Doubao, and respectful of DeepSeek’s scientific process. Claude is the model of choice for coding, universally rated as the best thing out there. The researchers we met were humble, welcoming, focused purely on technical priorities and building the next big model. Researchers were also very young: at one lab in particular, the average age wa…

Read more

Moore’s Law was always two things: a physical observation about transistor density, and an economic bargain about cost. The physics has been slowing for years.

Is TSMC’s hesitation the first sign that the economics are reversing, too?

Subscribe now

I. The bargain breaks

Semiconductor manufacturing experienced one of the steepest learning curves ever recorded in any industry. For five decades and 10 generations of technology, each more expensive tool delivered cheaper chips – reliably, every 18 to 24 months.

The consequences are everywhere around us. The smartphone in your pocket has more raw compute than a 1990s supercomputer and costs less. The cloud infrastructure that runs modern AI exists because each successive generation of chips was cheaper per operation than the last. Software could eat the world because the hardware kept getting cheaper.

The cost per transistor – the broadest measure of chip economics – stopped falling in 2011. The narrower measure that tracks what lithography itself delivers, transistors per wafer-dollar, kept improving for another decade, helped by the industry’s shift to extreme ultraviolet light (EUV) around 2019, a shorter wavelength that let chipmakers print finer features and restored the cost-down curve. Now it has reversed too. High-NA should be the next step in the bargain, but if the best customer won’t take it, the bargain doesn’t hold.

II. What changed

Read more

One of the best to understand AI capability curves. — Mark C., a paying member

Subscribe now

Hi all,

Back home after three weeks on the road and easing back in.

In this week’s issue:

First, AI self-improvement. thinks there is a real chance a frontier model trains its successor by 2028. If that is true, what should we already be seeing in how the labs hire, spend and build?

Then, jobs and AI. In some of the occupations most exposed to AI, postings are rising.

Finally, what will more capable AI agents mean for token budgets? Exponential View members get access to my interactive model.

Let’s jump in!

The signs of self-improvement

Anthropic’s has argued that there is a 60% chance that a frontier model will train its successor by 2028. It is an exciting claim, perhaps revolutionary, perhaps frightening; the prospect of a recursive intelligence explosion.

There are plenty of reasons to read Jack’s essay and conclude that the picture might not be so clean.

One objection is that frontier training now looks less like a pure research challenge and more like an industrial scaling problem. The bottlenecks are not only about optimizing CUDA kernels. They are about negotiating land leases in Wyoming, securing power infrastructure, obtaining chips, and hiring the electricians to wire it all together. Over a three-year horizon, those physical constraints may matter more than algorithmic advances.

So what can we infer from the revealed preferences of frontier labs? If automated R&D were truly likely by 2028, what would we expect their behavior to look like now?

First, hiring would change. Labs would still want elite researchers, but the profile would shift towards people who can make research agents useful. Fewer pure researchers, more research multipliers, people who can build an automated research factory.

Second, labs would overinvest in compute before the automation arrives. Because of those physical constraints, they would want more GPUs, more memory, more power, more data centers, more inference capacity, and better internal tooling.

If a lab believes the R&D loop is about to accelerate, then waiting becomes expensive. You would expect it to tolerate ugly near-term cash burn in order to secure the pre-commitments it needs.

Read more

As we wrote three months ago, we are in the midst of a compute crunch – demand is running ahead of supply and our position has stayed the same:

The real risk isn’t that we’ve invested too much in AI. It’s that we haven’t invested nearly enough.

Today’s quick look at the data shows that much of the supply remains latent, waiting on enterprise funding. When firms start spending serious money, the compute crunch will get crunchier.

Subscribe now

On the supply side:

Nvidia B200 GPU rental prices grew 114% in six weeks. Frontier model releases pull demand toward the newest chips; the premium customers pay per hour to get a B200 instead of an H200 has grown by more than 6x.

Infrastructure provider Lightning AI says that some forty of its customers are seeking 400,000 GPUs, ten times the current fleet of 40,000. Providers are already rationing GPUs by customer size. Microsoft now requires Blackwell customers to lock in at least 1,000 chips for a year and is cutting off smaller customers whose servers sit idle.

Read more