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

Over the past week, I have been in China with , meeting AI and robotics teams including Zhipu and MiniMax (the two publicly listed foundation model companies), as well as Kimi, Alibaba, Xiaomi, Bytedance and others.

Demand is booming. Zhipu, for example, is serving 5.5 trillion tokens per day. Developers are rushing to the platform, joining at about ten a minute. Keeping up with inference loads is a struggle in China, as it is in the US. Teams universally acknowledged compute constraints, particularly the shortage of Nvidia chips. But that isn’t stopping innovation under those constraints.

Researchers and developers are free to use whichever model they want. Anthropic’s Claude was the preferred choice for technical teams, but it was clear that “dog-fooding” is commonplace.

I’m on the flight back to London right now. Hannah is continuing on to Shenzhen to check out more hardware firms. We’ll write something in more depth in the next couple of weeks. For now, here’s a short video from one of the many demos we saw at Unitree:

Huge thanks to our local hosts who were generous with their time, hospitality and swag. And a shout-out to Rintoul, who pulled together this amazing trip with expertise and great humor.

Azeem

Subscribe now

We can’t stop, we won’t stop

gets to grips with that odd Silicon Valley paradox, that many engineers think the very thing they are creating, AI, will mean that the “median person is screwed”. What’s more, those same builders profess to having no clue how to stop this deracination.

Jasmine was in China with me this week. She “gets” the people in American AI labs and the culture surrounding them better than anyone. Her stellar essay in The New York Times is revealing:

In general, tech industry sources expressed more extreme concern about the labor market impacts of A.I. in private conversation — but suddenly became optimists once I turned on the mic.

I don’t agree with the depth of fatalism coming from the labs. Diffusion of the technology will be slower than flicking a light switch, even if it proves faster than the rollout of electricity.

But that fatalism is itself a fact worth taking seriously. The beliefs might become behaviors. An example: junior hiring might slow down, not because AI can do juniors’ jobs well, but because the labs believe it can and persuade employers that it can. A hiring freeze is an almost inevitable consequence of the patterns of belief.

The danger is that the theory of displacement becomes the logic of our responses. The point should be that we live in a state of uncertainty, as argued later in this issue. That uncertainty means we should be skeptical of preordained outcomes, however smart and well-paid those who make them are.

See also:

• Another way to frame the moment is Dostoevskian: intelligent people convince themselves they have discovered a truth so important that ordinary ethical constraints no longer apply to them. See, Palantir.

• This week, Chinese courts have ruled that replacing someone with AI is not, by itself, a lawful reason to fire that person. It is one of the first such decisions worldwide.

• The US government has classified the grid supply chain as a national defense bottleneck.

Subscribe now

Silicon philosophers

A new paper out this week tested seven frontier models on their ability to role-play professional philosophers and simulate expert judgment. Some models looked passable on average, especially on questions where philosophers already agreed. But real philosophers agree far, far less than the AI. The variance of views was two to four times higher in humans’ favor.

Read more

War on a learning curve

When, a year ago, I spoke with Ukrainian veteran drone pilot Jack de Santis, he said that soldiers who leave the front for rehabiliation and return after eight or nine months, need full retraining because drone warfare changes that fast. Since then, Ukraine accelerated the pace of iteration to seven days. I was so struck by this number, that I asked the team to dig into it and prepare a detailed briefing for members:

Subscribe now

DeepSeek shows us the future, again

When GPT-5.5 rolled out this week, OpenAI’s reasoning research lead Noam Brown, tweeted

with today’s AI models, intelligence is a function of inference compute. Comparing models by a single number hasn’t made sense since 2024. What matters is intelligence per token or per $.

With the compute crunch, doing more with less compute could be a winning strategy.

And it runs counter to the culture American AI labs enjoyed for the past few years. The mantra was “moarrr compute, better benchmarks.” Chinese labs, with less capital and minimal access to cutting-edge compute, didn’t have that luxury. So they adapted. In a nutshell, how much real-world capability can they afford to deploy per token, per user?

DeepSeek’s new V4 model is marginally worse than GPT-5.4 ( has an excellent technical breakdown of it), but it is 4x cheaper, reflecting, in part, lower compute costs. As inference costs approach 10% of total engineering headcount spend, that line item begins to matter.

argues further:

In China, compute is more than an expense line. It is a strategic constraint shaped by export controls, chip supply, cloud capacity, domestic hardware readiness, and inference economics. […] DeepSeek is turning compute scarcity into a set of design specifications.

I am on my way to Beijing and will be looking to understand exactly how this plays out on the ground.

A MESSAGE FROM OUR SPONSOR

The developer conference of the year. On demand on Salesforce+

Agentic AI is reshaping how software gets built, and Agentforce is out in front.

Catch the TDX sessions on Salesforce+ to see what this looks like in practice: deep dives on Agentforce, Data 360, the core Salesforce platform, vibe coding, Slack, and more – all available to stream for free.

By watching TDX on Salesforce+ you can:

• See what’s coming next: Hear directly from the leaders defining the roadmap for AI-native development.

• Go beyond the main stage: Access broadcast‑only segments and candid interviews you won’t find elsewhere.

Start with the main keynote to experience how software is being reinvented, then explore the full programme at your own pace.

Watch now

Where humans matter most

In 2018, I argued that “automated perfection is going to be common… What is going to be scarce is human imperfection.” I returned to it a few months ago:

When I first wrote about artisanal cheese, I imagined this shift unfolding more slowly, alongside the automation of routine office work and putting more robots on assembly lines. I didn’t anticipate that, nine years later, I could build custom software in an hour or produce work that once required entire teams while walking through customs.

Economist wrote an excellent essay, a vivid case study of how automation reshapes where humans matter most. When the internet entered the economy, 60% of travel agents lost their jobs. The routine tasks of searching, comparing, reserving etc. moved from the agents’ computers to ours. What happened to the remaining 40%? It was hardly fighting over scraps; what we wanted from travel changed in tandem. And for the surviving two-fifths, it was more upmarket, higher-end, curated. As a result, travel agent salaries grew from 87% of the private‑sector wage average in 2000 to 99% by 2025.

I like ’s word for this destination – “the relational sector”. It’s work whose value is inseparable from the human providing it.

Solar overtakes nuclear

In 2025, the world’s solar panels generated nearly as much electricity as the world’s nuclear reactors. So far in 2026, Ember’s new data shows that solar is starting to overtake nuclear on a 12-month rolling basis.1 Nuclear is still vital, of course. Last year was its best year ever in absolute terms, but its share of global electricity has fallen from a 1996 peak of 17.5% to under 9%.

The main way we explain this is that energy is becoming a technology, not a commodity. Commodities get scarcer and pricier as you extract them. Technologies are on a learning curve, getting cheaper as you make more. Solar module prices have fallen over 90% since 2010, and in 2025, solar met three-quarters of all new electricity demand on the planet. As solar rides learning curves, new markets open and make possible what was previously impossible. We call this the solar supercycle.

Read more

It’s 2024, near Bakmut, a few kilometers from the then-front line. The commander of the Terra drone unit of the 3rd Separate Assault Brigade Mykola Volokhov, has watched the slow degradation of his command links. A system that had worked at eight kilometers out, started to fail at four, then at two. Russian electronic warfare teams were getting better and better with their jamming. Later, speaking on camera, Mykola described the reversal with a soldier’s understatement. “But this problem was overcome… and now it is not with us. No problems!”

The repair did not come through a ministry, a tender or a requisition form. It came through a phone call. In the words of Dimko Zhluktenko, a soldier in Ukraine’s Unmanned Systems Forces:

We called the manufacturer and said, ‘Guys, we had this kind of issue...’ It is very direct, and they are very open to help us. They don’t need a shitload of documents or anything.

This is the operating model. A device is designed, fielded, disabled by the enemy, diagnosed through a conversation between the operator who lost it and the engineer who built it, then redesigned and redeployed in roughly seven days.

Industrial tempo in the grueling conditions of war. Continuity of signal, of keeping systems operational outweighs a lot else. In one unit along the Dnipro, an operator refused to take shelter during 120mm mortar fire because moving would break the radio link to his drone mid-mission.

This is how military innovation happens on parts of the Ukrainian front.

In the US and Europe, weapons systems move from concept to deployment over five, ten, sometimes fifteen years. A major capability iteration of the flagship F-35 Lightning II fighter jet can take close to a decade. That gap – one week versus seven years – is not just down to the exceptional engineering talent in Ukraine. It is the product of how work is organized, how quickly information reaches the people who can act on it and how much authority those people hold.

It is a real-time demonstration of what becomes possible when you remove the layers between the person who sees the problem and the person who can fix it. Under fire, alacrity is the difference between life and death.

While the stakes are as new as they are high, the lesson carries well beyond the muddy, blood-soaked, adrenaline-drenched trenches of Ukraine. It’s a lesson that will determine which nations will succeed, which will falter, which will invite attack, and which will stand firm as this century slowly fills with conflict.

Subscribe now

The formula

In early 2024, Oleksii Kamyshin, then Ukraine’s Minister of Strategic Industries, formalized a new metric, “cost per kill”, as unit cost divided by probability of success multiplied by expected effect. However gruesome it may sound to those of us who never experienced war, for the Ukrainian military, it was clean, easy-to-understand, and something they could optimize.

It’s “the formula,” he said. You can’t improve what you don’t measure.

The formula directs production decisions across the industrial defense ecosystem. Weapons are not evaluated on whether they meet a specification written before the war. That is ancient history. The evaluation is the following: in the field, did the weapons work at a price the system can afford? If not, the runts are discarded, no matter the sunk cost. The high performers are scaled. The cost per kill today is $1,000, down from around $60,000 in 2022.

Much of this is familiar to Western firms, defense or otherwise. Performance is a criterion. Someone decides the drone should cost $800 and hit the target 70% of the time. The org spends three years building toward those numbers. By the time the product ships, the specification is an artefact of a long-forgotten planning meeting.

The $1,000 cost-per-kill came from a feedback loop of engineers talking directly to operators, designs changing every week, five hundred manufacturers running parallel experiments which are treated as data. Competitors can copy the number or buy the solution, but without copying the loop, they’ll just have a new specification that a slow system will fail to meet. Performance is what the system arrives at, not what it’s told to produce.

The cycle

Early drones relied on standard radio control. In the Ukraine context, this worked until Russian units learned to jam or spoof their signals. So the Ukrainians initiated dozens of competing experiments across hundreds of small manufacturers, running simultaneously and without coordination.

Some engineers switched to fibre-optic guidance, a physical cable, unjammable by definition, but easily cut while in flight. Others tried an encrypted multiband radio for more flexibility, but still vulnerable to interference. What stuck was a hybrid – fibre with a live radio fallback for the moment the cable breaks.

The entire process happens within days and weeks and the speed shows up in their outcomes. Oleksiy Babenko, founder of drone manufacturer Vyriy Drone shared that the rapid iteration process – task-specific development, direct operator-to-engineer feedback, weekly redesign – increased target accuracy from around 10% to 70%, then 80% within a single upgrade cycle.

Subscribe now

The network

When Putin launched his invasion in 2022, there were perhaps seven domestic drone manufacturers in the whole of Ukraine. Today, Ukraine sports about five hundred manufacturers. Last year, they collectively produced 4 million units. The forecast for this year is seven million. Roughly 140x more drones than the United States produces in a year.

This scale hasn’t compromised speed. Production is distributed and it has followed the same design pattern established in the early phases of the war. There is no single dominant contractor. There is no centralized R&D pipeline. Hundreds of firms operate in parallel, each pursuing its own variations on shared problems. When conditions change, and on the battlefield there is more change than constancy, the network responds. It is simultaneous, emergent, like an immune system.

The individual firms do not have to be faster than their Western counterparts – some are, but that isn’t the point – it’s that the system as a whole is faster. The network searches a wider space for solutions, with far more simultaneous bets than any monolith can. This system is also more resilient. Firms can fail, factories can be bombed. The network persists.

More subtle but less obvious, is the variety and breadth of innovation. Curiosity is a byproduct of the network itself. Solutions win because they work. Those that succeed are stamped, repeated, scaled. Those that don’t disappear.

Subscribe now

The line

Ukraine has industrialized its drone capability. Underground warehouses far from the frontlines hum with the business of reengineering and manufacturing UAVs. Some of these complexes span hundreds of thousands of square feet; one facility can produce ten thousand units a month. Much of the work remains manual, but follows standardized steps. It’s light manufacturing under conditions of constant redesign. Cat Buchatskiy from Ukrainian defense analytics and innovation center Snake Island Institute told recently:

It’s comparable to what you’d see at SpaceX’s Starlink production line. This is not a garage shop industry. Our prime contractors (though we don’t have traditional primes) are producing tens of thousands of units per month and generating tens of millions of dollars in revenue.

Western manufacturing industry has always traded off scale and stability. Whether you’re making cars, semiconductors, or aircraft, a high-volume production line is a capital commitment made in advance of demand. Every lathe or robotic arm fits within a design made years before that can’t change without unravelling the investment.

According to its own estimates, if GM retools a plant rather than building a new one, the saving is $1.5 billion per facility. Honda is investing more than $1 billion to reconfigure three US plants for mixed powertrain production. Toyota’s Kentucky facility – the company’s largest plant globally – recently committed another $800 million to modernize its lines for EV production.

Every industry that has ever scaled manufacturing locks the design before the line is built.

Ukraine has broken these rules. Monthly FPV1 output grew from roughly 20,000 units in early 2024 to 200,000 by the end of the same year – a tenfold increase in twelve months, even as the designs running down those lines changed week by week.

The core mechanism is modularity (read our primer on modularity here). Most Ukrainian drone designs share standardized components, common software frameworks and interchangeable subsystems. A new guidance update or jamming-resistant radio does not require retooling the line. It requires swapping a module. Ten thousand units a month and weekly design changes don’t disrupt the process; they are the same system.2

Many of the people at the heart of this system have no background in the defense industry. They come from software, gaming, marketing, even B2B SaaS. Again, Cat Buchatskiy shared that “[o]ne of the biggest defensetech VCs now supporting the industry used to be the chief marketing officer at a workflow automation company. Many were working at Uber Ukraine or other rideshare companies when the invasion began.”

Read more

Not Craig Federighi (software) nor Eddy Cue (services). The hardware chief.

A month ago I shared with you all why I changed my mind about Apple. I’d been in the camp that wrote it off on AI. They had no hyperscale capex, Siri hadn’t meaningfully improved in a decade, and WWDC was seemingly losing steam. What I missed was that every time I switched between ChatGPT and Claude, I was doing it on an Apple device.

My read is that Ternus’s appointment reflects a bet on hybrid computing. Frontier model training compute has been growing around 5x per year in recent years, while GPU memory bandwidth has improved around 28% per year. That creates a widening systems bottleneck and strengthens the case for shifting routine AI tasks to edge devices while leaving heavier workloads in the cloud.

This new device landscape is still evolving, but it will be a combination of the highly portable phone with the persistent desktop or server. Both play to Apple’s strengths – they run on the same architecture and sit inside an installed base of over 2.5 billion active devices.

The board picking a hardware CEO is Apple telling us how it sees its next chapter. My analysis is open to all readers today:

Nearly a third of surveyed Anthropic staff believe their Mythos Preview model could replace junior engineers and researchers within three months.1 Treat this as a directional signal that people closest to the frontier are starting to believe that a lot of junior‑level work is now in scope for automation.

2. Huge enthusiasm, limited adoption

AI inference costs at companies are approaching 10% of engineering headcount costs, according to Goldman Sachs. AI adoption is both inevitable and uneven. Excitement and FOMO are high, but adoption is lagging despite the investment. The only way to close that gap is through deliberate learning and experimentation.

3. Teams that learn to work with AI win big

One company nearly halved the cost of each code change they shipped and doubled weekly deployments over five months of intentionally building with AI.

4. Users feel the compute crunch

Claude API uptime fell to 98.32% in March, well below the 99.99% standard.

As we wrote earlier this month, users are feeling the compute squeeze. Across every major AI platform, usage allowances tightened last year – more tiers, stricter limits, changes that often showed up without notice.

Read more

Welcome to the Sunday edition of Exponential View. We’re trying something new in public today.

First, you’ll get my human read on Jensen Huang’s ‘token factory’ worldview and what it means for Nvidia’s moat and US-China tech strategy. Then you’ll see the same week through a different lens – my AI Swarm View arguing through Iran scenarios. And finally, a selection of short morsels that surface the weak signals and second‑order effects we’d otherwise miss in the noise.

This is a minimally viable test. Tell us what lands (and what doesn’t) in the comments or the short survey at the end.

Thanks! 🙏🏽

Subscribe now

Azeem’s view

Jensen Huang and the age of the token factory

This week, Jensen’s mental model snapped into focus when he told in an interview: “The input is electrons, the output is tokens. In the middle is Nvidia.” The interview was really about who controls that middle and where Jensen thinks the threats are.

Nvidia doesn’t own the middle outright. The fabs, memory, clouds and models all sit elsewhere. But Nvidia runs the orchestration. Jensen personally meets the CEOs of the chip supply chain: ASML, whose lithography machines etch the chips; TSMC, which fabricates them; Micron and SK Hynix, which make the high-bandwidth memory Nvidia’s GPUs need, and Lumentum and Coherent, which supply the optical components that wire Nvidia’s systems together. He is effectively underwriting suppliers’ long-term investments and creating demand for their products years ahead.

The forecast becomes self-fulfilling. Suppliers build to it, and by building to it, they make it true. That translates into roughly $100 billion in explicit purchase commitments and perhaps $250 billion implicit. Critics call it circular investment; in a world of compute shortages, it looks more like foresight.​

Jensen waved off the competitive threats inside the US. His view is that custom silicon – Google TPU, AWS Trainium – leans on a single customer:

Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic.

His read is that Anthropic picked those platforms because Google and AWS wrote the equity checks when no one else would. Neither has consistently used the public scorecards that chipmakers use to prove performance. Jensen doesn’t mention AMD barely features in the conversation.

Subscribe now

You can copy the chip, but you can’t copy the system around it. Twenty years of CUDA, Nvidia’s software platform, means a generation of AI researchers has grown up coding on Nvidia – the habits, the tooling, and the muscle memory are deep. On top of that, there’s an army of Nvidia engineers embedded with customers who pull two or three times more performance out of the same silicon. And a new chip design every year – Blackwell, then Vera Rubin, then Feynman – resets the bar before rivals can catch up. The moat is in these three elements compounding.

Where Jensen is less relaxed is about China, specifically Huawei Ascend. Not because it’s directly competitive today, but because it doesn’t need to be. It’s a parallel ecosystem in which Huawei sits between electrons and tokens, insulated from Nvidia by design.

This is textbook adjacent-market disruption. A constraint – US export controls – forced a standard that doesn’t have to beat CUDA. It only has to be good enough for the developers who can’t use it. “Fifty percent of AI developers are in China,” Jensen says. If those developers build on Huawei Ascend rather than CUDA, America loses the platform standard permanently. And with it, the data flows, the weights, the APIs, the default surfaces every non-Chinese enterprise uses to reach Chinese customers.

There’s a coherence to his argument, but it’s also self-serving: don’t worry about competitor ASICs – worry about cutting off Nvidia sales to China. It may well be exactly that, but underneath the special pleading is a real question, and one of the defining debates of this moment. What is better for the US: preventing access to Nvidia’s cutting-edge chips, which slows Chinese model development but accelerates China’s own chip industry? Or selling them chips and maintaining primary control of the platform that runs everyone’s AI?

Even selling chips to China might not stop the second ecosystem from forming. DeepSeek is already adapting its upcoming V4 to run on Huawei silicon, and that’s just the beginning. Beijing has shown its hand: it will intervene whenever Huawei’s trajectory is at risk.

It may not be too late to keep the platform standard in American hands. But it would take an effort on a gargantuan scale – one Washington has not yet shown the appetite for.

Subscribe now

Swarm view

Mapping scenario probabilities in Iran

Over the past few months, I’ve been testing a new way to think with AI. When I hit a hard, messy question, I throw it to a panel of simulated expert agents and let them argue it out. I call it my Swarm View. I’m publishing a Swarm View for the first time as a live experiment today, after months of privately tuning the agents to think alongside me, argue and create new views I’d not have considered before.

What follows is output from one such run. RMA, my OpenClaw agent, did the work on Tuesday, 14th April. We’ve lightly edited the output. The right format, the right cadence, how to weight the confidence of the verdicts – all of this is still being worked on. Treat it as a tool for stress-testing scenarios, not a forecast to bet the farm on. I’d welcome your reaction.

This week’s question: Will the US and Israel strike Iran after the ceasefire expires on 22 April?

The synthetic panel:

• A CENTCOM1 operational planner,

• an Iranian strategic analyst,

• a Pakistan/ISI specialist,

• a PLA/China analyst,

• and a nuclear deterrence expert.​

The panel’s verdict: A 60-70% probability of a strike before 22 April.

Read more

We talk about models as if they float in the cloud, weightless and ambient, but AI is physical. Model weights are arrays of billions of numbers stored as tensors in files distributed across racks of GPUs in a data center. When you use Claude or ChatGPT through an API, the weights never leave those servers – you send text in, get text back, but the intelligence stays on someone else’s hardware. In a briefcase, they are a hard drive. In both cases, a physical arrangement of matter encoding a capability.

Last week, Anthropic’s Mythos Preview demonstrated what that capability now means. The model found thousands of vulnerabilities across every major operating system and browser. As I argued on Friday, AI models like Mythos have permanently collapsed the distinction between capability and danger. Who controls the weights of that model – and how – is now the central question in AI governance.

Subscribe now

The viscosity mismatch

Some technologies diffuse quickly – published papers, vaccines, drones. Others are viscous: they require specific infrastructure, tacit knowledge, or resources that slow their spread. Nuclear weapons are the canonical example. The knowledge has been public for decades; the ability to build one remains confined to a handful of states because the materials and engineering are highly viscous.

Frontier AI has an uncomfortable asymmetry. The creation of frontier weights is viscous. Training runs cost billions of dollars in compute – hundreds of thousands of GPUs running for months. They require proprietary data pipelines and deep engineering talent. Only a handful of labs can do it, nearly all based in the United States. China, the closest rival, remains constrained by chip export controls that limit its access to the most advanced hardware. That makes the capability concentrated and, in principle, seizable.

But the spread of weights is not viscous at all. They are files, and in some cases, even books ( printed GPT-1’s weights as a book). Today, I can download to my laptop an open-weight model that would have been frontier six months ago. And the use of frontier capability is even less viscous: anyone with an API key can access the capability without possessing the weights at all.

Creation viscosity is high. Spread and use viscosity are low. That asymmetry creates a closing door – you no longer need the weights to capture much of the value. Synthetic data and distillation – the process of training a weaker model on a stronger model’s outputs – allow labs to approximate reasoning and training signals from stronger systems. Chinese labs DeepSeek, Moonshot, and MiniMax were caught by Anthropic generating over 16 million conversations with Claude between them to use as training data for their own systems.

If adversaries can get close to frontier capability by reverse engineering through use, then securing API access becomes essential. And so far, access controls have been systematically circumvented. It’s hard to slay a hydra of fake accounts. Any sufficiently motivated actor can query these models millions of times via normal API access, and use those responses to teach their own model – no briefcase required.

Containment is no longer feasible

Read more

Subscribe now

The conflict has extended far beyond the immediate region. In nine days, 37,000 flights were cancelled. This cost Gulf carriers $1 billion, while rerouting added 206,000 km of daily detours and wiped out an estimated $600 million a day in tourism revenue. Refinery strikes and the closure of the Strait of Hormuz pushed jet fuel prices to more than double.

Airlines have started to shift the burden to consumers. United Airlines warns that higher prices would add $11 billion to its fuel bill in 2026, which could push fares up by 20%.

Fuel surcharges are rising, with some London-Sydney return flights now carrying a $800 fuel surcharge. Average global airfares are at their highest since 2019 at $465. United, Air New Zealand and SAS have all announced capacity cuts.

These price dynamics are reshaping another market.

We wrote in our Solar Supercycle analysis that inflection points for new technologies occur when new tech becomes cheaper than the incumbent. The sustainable aviation fuel (SAF)1 premium over conventional jet fuel in Europe fell from $1,463 per metric ton before the conflict to $1,139 by mid-March.

The picture is more nuanced: most SAF today is derived from biological feedstocks – cooking oils, animal fats, biomass and alcohols – rather than synthetic e-fuels made with renewable electricity. All of these are sensitive to energy prices, whether through feedstock costs or electricity costs. Even though the gap is closing, SAF prices in Europe rose from $2,300 per metric ton in February to $2,500 in March.

SAF represented just 0.6% of jet fuel in 2025 and is expected to rise to 0.8% this year. A move towards e-SAFs2 could decouple jet fuel from oil – and force an early inflection point.

Subscribe now

Thanks for reading.

Share

You help me keep my finger on the pulse of what is happening and what may be to come. — Joanna P., a paying member

Subscribe now

Hi all,

Happy Sunday. Let’s zoom out for a moment and look at what stayed with me and our team this week.

More than a myth

Anthropic teased Mythos Preview this week and assembled a coalition to think through what the model might mean for digital infrastructure if placed in the wrong hands. It’s a model preview with an institution gathering around it. My initial take on Mythos, in short:

[t]he most significant and immediate implication of Mythos is on how we price risk. Critical infrastructure is the asset class that is most exposed and is systemically mispriced.

You can read the full analysis here.

Mythos raises the “Is this AGI?” question again. It’s not a useful question to ask. I made the case before that

I deliberately avoid using the term ‘artificial general intelligence’, not because it’s irrelevant, but because it’s often unhelpful and distracting, even when used by leading researchers.

Former OpenAI board member made the case too this week. She’s right when she says: “[f]or people who think the world is going to be radically transformed by advanced AI, I think it’s helpful to talk less about either AGI or superintelligence, and instead describe vivid, concrete milestones that you think will happen soon due to increasingly capable AI.” Highly recommend reading Helen’s essay.

Wired for GLP-1s

Scientists have found genetic evidence that helps explain one of the most striking features of GLP-1 drugs: why two people can take the same medicine and have very different experiences. Some lose far more weight than others, some are hit by nausea, while others tolerate the drugs with ease. This work gets us a step closer to precise obesity treatment – perhaps guided by a genetic test before a prescription is written.

As a class of medicines, GLP‑1s target a condition that affects more than 1 billion people worldwide, even before you count diabetes and other metabolic disorders. They could reshape our societies. The catch has been how unevenly they work. This new research could be one of the keys to unlocking the full potential of GLP-1s.

Subscribe now

A degree of human

In September 2025, I wrote about the great value inversion and the future of education, in particular:

Universities offer degrees as if 20th-century returns persist, ignoring how credential premiums have fragmented and the changing nature of the skills employers will need.

The reality is, in the next decade or two, we will have to reimagine education completely. It may help to look back to 1809 Prussia, a country in free fall and forced to reform. Then, the diplomat Wilhelm von Humboldt was asked to redesign education. Instead of training officers and engineers, he argued for Bildung, the free development of a whole human being:

Humboldt’s solution was to design an environment rather than a curriculum. […] The university would not answer to the demand for immediate use and inquiry would follow the question wherever it led, unconstrained by predetermined ends.

A nice Sunday read by .

What I’m reading

I had the opportunity to read Sebastian Mallaby’s book on Demis Hassabis and DeepMind before it came out. I met Demis several times and in 2020, when we recorded this podcast conversation via Zoom.

Read more

The most significant and immediate implication of Mythos is on how we price risk. Critical infrastructure is the asset class that is most exposed and is systemically mispriced. The US utilities sector alone has a $1.5 trillion market capitalization, priced at roughly 22x PE – with cyber risk treated as an incremental operational cost rather than a structural exposure. That frame no longer works in a world where an AI agent can autonomously construct working exploits overnight.

This is not a marginal change to the threat landscape. Every institution that prices cyber risk – insurers, infrastructure investors, governments – has built its models on one foundational assumption: that offensive capability requires scarce human expertise. Mythos removes that scarcity. It’s not surprising that Treasury Secretary Scott Bessent convened Wall Street CEOs at short notice. Bessent understood the systemic risk quickly, but markets have yet to reach the same conclusion.

Anthropic understands the danger and has created a new institution – Project Glasswing, a coalition that brings together twelve major technology partners – with greater control over this capability than most governments currently possess.

It’s a response to a serious problem. It is also entirely voluntary, built on trust rather than law, and accountable to the people who built what’s being governed. The labs decide what to disclose, when, and to whom. There is no meaningful independent verification, no democratic mandate, no mechanism by which the public – whose infrastructure is at stake – has any formal say.

Today, Glasswing controls one model, from one lab. It governs nothing about what Meta, a Chinese research institute, or a well-funded independent team releases in 2027. The containment is temporary: The next models will be more powerful, better aligned and still capable of deception and harm. What is contained today will become ambient.

The market hasn’t caught up

The $20 billion cyber insurance market is predicated on the wrong model. Most premiums are based on well-known ransomware, data breach and extortion patterns following standard attack vectors, monetization demands and deployment patterns, all of which can be modeled from experience over the previous decade.

Read more

For nearly two years, Marija and I have been looking for an exceptional Executive Editor to work with us. We never rush important hiring and we knew there were only a few people fit for the job. Today, we have news to share and give you a sense of where we’re taking Exponential View next.

We took our time with this hire. Exponential View is a research studio, not a newsroom – and the distinction matters for who can lead its editorial operation, a job previously split between Marija and myself. Our members, you, are deep thinkers, inveterate tinkerers and some of the world’s most consequential decision-makers. You turn to us for insight that is tough to produce: synthesis, frameworks and the non-obvious thinking that changes how you see the world. Finding an editor who understood that difference was the whole search.

Today, we’re announcing that Greg Williams joins Exponential View as our first Executive Editor.

What drew us to Greg wasn’t any single chapter of his career — it was the pattern across all of them. He helped define the nineties at the seminal style-and-ideas title Arena. He spent nearly a decade in New York, where the rigorous standards of American magazine journalism got into his bones. Then he ran WIRED UK for close to a decade, becoming one of the defining editors in technology, across print, audio and video. He is also an accomplished novelist.

Each of those was a different register, a different audience, a different set of rules and Greg won on all of them. The proven ability to learn fast and win on unfamiliar pitches is what we were looking for the most. Exponential View operates across disciplines and formats. So must its editor. But range without rigour is just restlessness. We’ve seen in Greg the rare combination of adaptability and standards.

Greg’s job at Exponential View will be to extend what makes our work worth reading and find new formats which will help people like you make the most consequential decisions over the coming years.

For my part, this change frees me to do what the moment demands of me – deep cross-disciplinary synthesis, the analysis that takes experience and judgement to get right. For Marija, this will be a time of deepening of the team’s bench so we can better deliver for you in all ways.

We can’t wait to show you what comes next.

Azeem & Marija

A message from Greg

I’ve been a subscriber to Exponential View and a deep admirer of what my friend Azeem has built over the years: a publication with a distinctive point of view, serious intellectual ambition and an audience unlike any other.

Exponential View is not just read widely; it is read meaningfully. It informs business decision-making for people in some of the world’s most important leadership positions, at a moment when vast amounts of capital are being deployed in response to profound technological change. And this transformation is reflected internally: EV has a distinctive approach to building AI systems that actively challenge ideas and handle the heavy lifting – protecting the space that rigorous, independent thinking requires.

For me, the opportunity to help build the next phase of that editorial offering – with a team I deeply respect, that cares about quality, originality and impact – was genuinely irresistible.

As for what to expect: my focus will be on ensuring EV’s analysis stays ahead of the questions that matter most to leaders navigating this moment – rigorous on the technology, meticulous in our sourcing (much of which is proprietary), sharp on the implications and always asking what it means for how you act professionally and personally. I want the work to be genuinely useful at the level of decisions, not just ideas.

But I’d rather not assume. I’d genuinely love to know what would make it more valuable for you – what we’re getting right, what’s missing and what you wish existed. Consider this an opportunity to help us build something even more essential for what comes next. Fill out this form to schedule a call or share thoughts with me as I settle in – your point of view is as useful to my onboarding as anything else!

Greg Williams

But the labs can’t clear the queue by raising prices — their customers will defect to the next best alternative. So the compute crunch shows up where economists would least expect it. As we wrote yesterday, OpenAI is passing on opportunities, Anthropic is adjusting session limits for its Pro users and open-source models are being shut down.

But first, let’s start with the demand. OpenAI’s APIs processed 6 billion tokens per minute in October 2025. By April this year, that had risen to 15 billion, a 2.5x increase in five months. Both OpenAI and Anthropic are in a race to maximise compute to meet demand.

That relentless load is why Google’s TPUs across all seven generations, some now seven and eight years old, are running at full utilization. Hardware that was expected to have depreciated years ago is still earning its keep.

The pressure shows up in revenue. For instance, Anthropic’s total revenue is growing, but the price per token is falling – and it’s falling faster than overall revenue is rising. This means the business is becoming increasingly dependent on volume to sustain growth.

Users are feeling the squeeze, too. Across every major AI platform usage allowances tightened last year – more tiers, stricter limits, changes that often showed up without notice.

Read more

Always an excellent perspective on emerging systems and their impacts across the human landscape. – Neill K., a paying member

Subscribe now

Hi all,

Welcome to the Sunday edition, where we make sense of the week behind us. Before you dive in, catch my latest podcast about how I adapted Andrej Karpathy’s autoresearch for knowledge work (essay on this is here + GitHub repo for paying members of Exponential View).

Subscribe now

The compute crunch

A couple of months ago, I wrote that AI capex looked more like a stampede than a bubble. I argued that by fixating on the bubble question, markets are worrying about the wrong thing. The real issue is the compute crunch. What I wrote then still stands:

The Big Tech companies have massive demand from generative AI workloads and the growth of cloud, which means they are having to turn business away. Last year, AWS lost a $10 million contract to host Fortnite because it couldn’t guarantee compute capacity. Something similar happened to Microsoft.

This week, OpenAI’s CFO says they’re passing on opportunities because there’s not enough compute. Codex went from 100,000 to 2 million developers in three months. Killing Sora will help. Anthropic has tightened the limits – some 7% of users will hit session limits they wouldn’t have hit before. H100 rental prices hit an 18-month high. Meanwhile, Alibaba closed-sourced Qwen, its open-weight leader. Tomorrow’s data edition will be dedicated to this trend.

See also:

• The decade-long feud between OpenAI and Anthropic: A deeply reported piece on the history of the split and the competing visions of AI that emerged from it.

• The same day OpenAI killed Sora, Kuaishou’s Kling AI reported $300 million annualized video revenue (Q4: $47 million), with 2026 expected to more than double.

• GitHub Copilot started injecting promotional content into code reviews. The developer community was not happy and the feature was pulled back.

Subscribe now

When agents meet markets

Economists forecast AI will add 1-1.5 percentage points to annual US growth by 2050 under a rapid AI progress scenario. They also believe there will be 10 million fewer jobs and inequality at its highest since 1939.

My friend believes the consensus underweights AI-driven GDP growth. Erik and point out that economists are anchored to historical precedent. I think it’ll be complex. We could reasonably see a phase change, much as the 1820s saw phase change in long-range GDP growth. But measured GDP growth might look disappointing if there is price deflation in AI-affected sectors or an expansion in non-market value. The revolution might be real, but the measuring tape might miss it.

Read more

released 600 lines of Python code a few weeks ago that started to change that. His autoresearch (see EV#565) runs an autonomous experimental loop in which a human sets a strategic direction, defines what good looks like and the agent iterates towards success within the guardrails. In Andrej’s initial experiment, it trained a GPT-2-level model over two days, 11% faster and found 20 genuine improvements.

Shortly after the release, Shopify’s CEO Toby Lütke used autoresearch on his company’s internal model, qmd; it ran 37 experiments overnight, so Toby woke up to a 0.8-billion-parameter model outscoring his previous 1.6-billion-parameter version by 19%. Toby is not a machine learning engineer.

Autoresearch is powerful because it solves two problems at once. One, it automates part of the knowledge-production process. And two, it solves the agent control problem, meaning it keeps agents on task. AI often drifts if you give it an open-ended brief or optimize for the wrong thing. To my great joy, autoresearch prevents this by design. The human decides where the car is going; autoresearch keeps its hands on the wheel.

I spent the last month adapting autoresearch for knowledge work beyond machine learning with the goal of spinning up a system that can run structured, low-cost experiments on the kinds of decisions most teams make every week. I’m calling this version AutoBeta, and I’m making the full playbook/skill available to paying members below.

Let’s go!

Subscribe now

The measurement problem

When I first saw autoresearch, my immediate reaction was that it didn’t have to be just about machine learning. The loop is generic – hypothesize, test, score, iterate. So I cloned it and started running it on other bits of my work.

It didn’t go exactly as I expected. The outputs looked fine but I couldn’t tell whether they were getting better. Unlike ML where the agent has a built-in feedback signal from each training run, knowledge work was missing it. A pricing decision doesn’t validate in five minutes; and the paragraphs I write don’t tell me if the argument is getting better or just changing, most of the time.

This is what makes applying autoresearch to knowledge work genuinely hard. The loop needs something to optimize against, and in knowledge work, that something doesn’t exist naturally.

So I constructed a version of autoresearch, called AutoBeta, that works across a wide range of business problems. It is not as technically robust as Karpathy’s, but it has the same design principles: I set the objective and the constraints, the experimentation takes place within the loop.

The one thing I changed was the score. I created “an oracle”, a panel of synthetic judges scoring each output against pre-defined criteria, collapsed into a single number the loop can optimize for.

Read more

Read more

Welcome to the Sunday edition, in which we make sense of the week behind us.

If you prefer listening, here’s my latest podcast episode, where I unpack NVIDIA’s bet on Groq and OpenClaw – and what it means for the organizations you run. Watch on YouTube or listen here:

Read more

The Iran War has created a chokepoint in the supply of helium, a byproduct of natural gas processing and LNG production that’s used in more than 20 steps of semiconductor fabrication. Headlines are making the helium situation sound apocalyptic for chips, but the reality is more nuanced.

Here’s what’s going on:

Read more

Happy Sunday. I’m en route to NYC for a week of meetings and speaking. If you’re travelling this week too, it’s a great moment to catch up on my latest podcast episode, where I explain why I changed my mind about Apple’s AI outlook:

Let’s go!

OpenClaw and the inference economy

We’re about to discover, the hard way, that most organizations are thinking about AI in entirely the wrong way.

In a training-first world, it made some sense to see AI as a capital project. But we don’t live in that world any more, as Jensen Huang made clear at NVIDIA GTC this week. We’ve shifted to an inference-first economy, where the dominant activity is running models – continuously, at scale, across millions of workflows and, increasingly, AI agents.

In that world, tokens are not an IT line item. They are a productive input, as fundamental to knowledge work as electricity, office space, or salary. I write about this at some length in my latest essay:

Subscribe now

Pakistan’s solar shield

Energy is the main weapon in the Iran war. As we wrote a couple of weeks ago, this is the old world of energy scarcity on full display. Several import-dependent Asian economies, including Bangladesh, Pakistan and the Philippines, have taken emergency measures to manage the energy shock.

But one country on that list, Pakistan, is faring better than the others. We were one of the first in 2024 to show the silent energy transition Pakistan underwent in the last two years, importing 17GW of solar panels in a single year, against a total generation capacity of 46GW. This transformation, we argued at the time, would be massive for Pakistan’s energy security:

92% of countries would be more energy secure under a renewable paradigm than under a fossil one. Many countries used fossil fuels as a geopolitical weapon. From OPEC’s oil embargoes to Russia’s manipulation of natural gas, control over these resources gave them significant power. With renewables, this power dissipates. Once you have the solar panel, you have the energy – OPEC can’t block the sun.

Pakistan has so far vindicated that argument. The solar transformation helped reduce their exposure to the price rises in fossil fuels this year by at least $6.3 billion, equivalent to 1.7% of the country’s GDP.

The underappreciated case for solar has to do with the sovereignty. Renewables localize power, as we wrote in our 2024 essay on the dawn of the distributed age; Pakistanis are increasingly responsible for their energy security. And it’s not just the sovereignty that matters here but its compounding effects. Our position is that solar as technology is undergoing a supercycle, the declining cost of solar panels will drive the opening of new markets which means that every year that Pakistan continues to replace fossil fuels with solar, it is building a durable technology stack to participate in the new economy.

Not delegating judgment

Last weekend, I spotted the AI legend, , had vibe-coded an overview of AI labor market risk. It was rapidly spreading across X. As I strolled down to my studio, a three-minute mosey, I asked my OpenClaw agent, R Mini Arnold, to do this for more than just the US.

After about 15 minutes, it came back with an interactive tool covering 25 countries. For each country, it grabbed employment by job category, as defined by the relevant statistical authority. Then it scored 1.4 billion jobs across several hundred categories in those two dozen-and-change countries. You can see the screenshot of the version I built below. I’ve obfuscated some numbers.

In the old world, this kind of first draft would’ve cost a major consultancy seven figures and taken months of work. My second reaction was, I want to share this with the Exponential View community. But I took three deep breaths and decided it would be an irresponsible thing to do. Andrej deleted his v1 after pushback: he said people were misinterpreting the data. The second version is live here.

I won’t get bogged down in whether Andrej was right or wrong to publish this exploratory tool the way he did. But I will tell you why I didn’t publish my casual vibe-coded map of jobs at risk. In simplest terms, information and analysis matter. They carry consequences for how people understand their role in the world.

Read more