Azeem Azhar’s Exponential View #571 argues that DeepSeek’s newest model illustrates a structural difference between Chinese and American AI labs — one shaped by export controls and hardware scarcity rather than strategic preference.
DeepSeek V4 and the compute constraint thesis
OpenAI’s reasoning research lead Noam Brown wrote, according to Azhar, that “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 $.”
Azhar frames DeepSeek’s new V4 model in that context. He describes V4 as marginally worse than GPT-5.4 but approximately four times cheaper, a gap he attributes in part to lower compute costs. He cites analyst Poe Zhao on the structural logic: “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.”
Azhar notes that as inference costs approach 10% of total engineering headcount spend, that cost differential becomes a material line item for enterprises making procurement decisions. He was, at the time of writing, travelling to Beijing to report further on how this dynamic plays out in practice.
The newsletter’s framing contrasts American and Chinese lab culture. American labs, Azhar writes, operated for several years under a “moarrr compute, better benchmarks” assumption. Chinese labs, without access to that capital or hardware, adapted toward a different question: how much real-world capability can be deployed per token, per user?
Drone warfare on a seven-day learning curve
Azhar opens the newsletter with a data point from drone warfare. He recounts a conversation from roughly a year prior with Ukrainian veteran drone pilot Jack de Santis, who said that soldiers returning from rehabilitation after eight or nine months needed full retraining because drone warfare had changed that fast. The newsletter reports that Ukraine has since accelerated its pace of iteration to seven days. Azhar says he was struck enough by that figure to commission a detailed briefing for members.
The drone section is connected to a broader theme in the newsletter about organisations on steep learning curves and the pressure those curves place on people and institutions. Azhar frames it alongside his longer-running argument about automation and where humans end up once routine tasks shift to machines.
Where human work goes
The newsletter returns to a 2018 Azhar argument that “automated perfection is going to be common… What is going to be scarce is human imperfection.” He revisits it with reference to economist Ernie Tedeschi’s essay on travel agents as a case study.
According to Azhar’s citation of Tedeschi’s work: when the internet entered the economy, 60% of travel agents lost their jobs as routine search, comparison, and reservation tasks moved to consumers. The remaining 40% shifted into higher-end, curated travel. As a result, Azhar reports, travel agent salaries grew from 87% of the private-sector wage average in 2000 to 99% by 2025.
Azhar credits Alex Imas with the term “the relational sector” to describe work whose value is inseparable from the human providing it — the destination he argues automation tends to push surviving jobs toward.
Solar overtakes nuclear
Exponential View #571 cites Ember data showing that in 2025, the world’s solar panels generated nearly as much electricity as the world’s nuclear reactors. Azhar reports that so far in 2026, Ember’s data shows solar beginning to overtake nuclear on a 12-month rolling basis.
He notes nuclear is not declining in absolute output — 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 newsletter attributes the divergence to learning curves. Solar module prices have fallen over 90% since 2010, Azhar writes, and in 2025, solar met three-quarters of all new electricity demand globally. The framing Azhar uses: energy is becoming a technology, not a commodity. Commodities become scarcer and pricier; technologies get cheaper as production scales. Azhar calls this the “solar supercycle.”
The issue does not make predictions about when or whether solar’s share will stabilise, nor does it quantify the pace of that overtake beyond the rolling 12-month figure from Ember.