2026-04-26 10:40:51
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:
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.
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.
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.
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.
2026-04-25 01:11:12
Along Ukraine’s eastern front, a drone stutters mid-flight. On the operator’s screen a few kilometers away, the grainy view of splintered treelines and trenches, fades into static. The control link drops. He removes his headset, he knows what’s happened. Russia has found a way to jam the frequency. Again.
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.
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.
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.
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.
The Exponential View House View: Hardware iteration speed is an organizational phenomenon, not an engineering one.
The one-week cycle is not a product of exceptional engineering talent. Ukraine’s engineers are skilled, but they are not uniquely so. The cycle is a product of feedback loop architecture: how quickly failure data reaches the people who can act on it, and how many parallel experiments are running simultaneously. Any organization that treats hardware R&D as a capability problem rather than a structural problem will misdiagnose it – and find that importing the metric without importing the structure produces neither the speed nor the cost.
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.”
2026-04-21 23:40:54
Apple’s board picked John Ternus, senior vice president of Hardware Engineering, to succeed Tim Cook on September 1.
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:
2026-04-20 22:02:51
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.
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.
One company nearly halved the cost of each code change they shipped and doubled weekly deployments over five months of intentionally building with AI.
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.
2026-04-19 10:25:43
Hi,
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! 🙏🏽
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.
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.
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.
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.
2026-04-15 17:45:58
In May 2025, an OpenAI representative arrived at Los Alamos National Laboratory bearing locked metal briefcases. They were accompanied by armed security officers. Inside the cases were the model weights for ChatGPT o3 – at the time, OpenAI’s leading reasoning model. The weights were physically transported because the destination, the Venado supercomputer, operates on a classified, air-gapped network. You cannot download a frontier AI model onto a classified government system – you have to walk it in.
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.
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.
