Dear Aventine Readers, 

It’s long been taken as a given that the US and China are in a winner-take-all race to develop artificial general intelligence. But is that true? This week we hear from Substack writers questioning that narrative and offering alternative versions of what's actually driving each country's pursuit of advanced AI. 

Also in this issue: 

Thanks as always for reading! 

Danielle Mattoon
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

When is a race not a race? When each side is running a different event. When there's no finish line. When one side doesn’t think it’s a race at all. 

Those are not serious answers. But they are versions of the arguments raised on Substack in the last month contesting the popular narrative that America and China are locked in a desperate, winner-take-all race to build advanced AI. This framing, its critics argue, is an inaccurate but convenient construct peddled by tech companies and perpetuated by lawmakers in Washington that is now actively damaging our ability to build AI that benefits wider society.

Here is how a group of Substackers — think-tank wonks, analysts, AI researchers and tech writers, several of whom recently spent a week touring Chinese AI companies — say we should think differently about the geopolitics of building artificial intelligence.

Can only one country win? 

In a long post for Transformer, a Substack dedicated to AI safety and policy, the tech reporter Yi-Ling Liu traced the history of the AI race narrative, arguing that it has been a self-serving story sold by Silicon Valley to Washington.

Most accounts agree that the framing began in 2017, when China's State Council set a goal of becoming a world leader in AI by 2030. In response, Booz Allen, a consulting firm specializing in cybersecurity and defense, published a piece of sponsored content in The Atlantic declaring this contest more consequential than the space race. The tech industry seized on the storyline and hasn’t let go. Mark Zuckerberg invoked the Chinese threat as a means to head off regulation for facial-recognition technology. Palantir and Scale AI used it to win billions in Pentagon contracts. Sam Altman pitched US intelligence officials on an "AGI Manhattan Project" to compete with China, as early as 2017.

The arrival of ChatGPT in late 2022 lent the narrative greater urgency. Liu wrote that after its release, "key members of the administration became increasingly concerned with the prospect of AGI. A distinct set of ideas became popular among US policy circles: 'short timelines' (AGI was imminent), 'fast takeoffs' (it would arrive suddenly) and 'decisive strategic advantage' (whoever got there first would seize lasting dominance)." US policymakers, she wrote, are now "obsessed with the idea that China is engaged with them in a race to AGI," and have shaped their policy accordingly, including aggressive chip export controls and virtually no domestic AI regulation.

Challenging this US narrative, Graham Webster, a research scholar at the Stanford University Program on Geopolitics, Technology, and Governance, argued that the race formulation doesn't make sense because AI isn’t a neat contest but a complex sprawl of technology, and it’s unclear where within it a race is being run. He also argued that there is no well-defined finish line. And if artificial general intelligence is considered to be the finish line? Well, he wrote, “the concept of AGI has always been speculative and diffuse. Some ideas of the thing have clearly been achieved. Others are so far outside of what we witness today that it’s hard to see any ‘race track’ ahead. A race to AGI simply is not a specific thing unless a specific definition is agreed [upon].”

Alvin Graylin, an honorary senior fellow at the Asia Society's Center for China Analysis, dug a little deeper, pointing out that the US vs. China framing assumes a winner-takes-all dynamic, but that might not be the case at all. It might turn out that reaching some arbitrary AGI threshold matters less than deploying "good enough" AI in ways that boost the economy.

Practical vs. superintelligent AI

The idea that AGI might not be AI’s most desirable endpoint bubbled up across other recent posts that focused on the fact that the US and China think very differently about the purpose of artificial intelligence. 

Afra Wang, a tech journalist, pointed to a key difference between the US and China in the way AI is built and deployed. In the US, Silicon Valley builds new technologies and decides how they should be used with little oversight, she wrote in her Substack, Concurrent. In China, on the other hand, companies build the technology but it is the state that decides how to use it. “This social contract suggests the hardest philosophical and economic questions — AI and job replacement, AI and inequality, AI and human meaning — do not belong to those tech companies,” she wrote. “They belong to the state and to the academic institutions the state funds and endorses.” That helps explain why Western researchers, such as AI researcher Nathan Lambert who writes on the AI Substack Interconnects, found Chinese AI scientists more focused on building models and less interested in their long-term social implications than their US counterparts.

And, as other Substackers have noted, the Chinese state has a very specific view of what it wants AI for. Kyle Chan, a fellow at the Brookings Institution, summarized his recent testimony to Congress in which he described China as caring more about building good models and diffusing AI through all of society than the US. "They want to use AI to turbocharge a wide range of sectors, from manufacturing and research to education and healthcare as well as of course their military,” he said. That is in contrast to the US, he pointed out, where AI companies seem more focused on the pursuit of building superintelligence, with practical applications more a means of offsetting the cost of building all-powerful AI.

China’s practical approach flows back into how AI research itself is done, wrote Zilan Qian, an AI policy researcher, on the Substack China Talk. In Silicon Valley, she wrote, AI companies are heavily invested in the belief that so-called recursive self-improvement, by which AI improves itself by writing better and better code, will transform the technology into what some in the field only half-jokingly call the Machine God. In China, the path to improving AI is focused on making it more accessible and useful. There is a view that models will only become more practically effective if they interact with the real world, leading to a focus on training on multimodal data and making them physically present in the physical world through robotics.

Graylin contrasted the approaches in the US and China to hunting for different animals. “Two hunters enter a forest and must decide: cooperate to catch a stag that feeds both for a month, or hunt hares separately that provide a week’s sustenance each,” he wrote. “The stag requires both hunters working together. If one goes for the stag while the other chases hares, the stag hunter goes home empty-handed.” The US, he said, is hunting for a stag in the shape of AGI; China, meanwhile, is hunting for hares in the form of economic wins.

America’s obsession with bagging a stag has led it to embrace policies designed to limit China’s capabilities, Graylin wrote, such as strong limits on the export of cutting-edge chips to the nation. But Hannah Petrovic and Azeem Azhar, writing on the Substack Exponential View, argued that the restrictions imposed by those policies has in some respects made China stronger, by pushing its AI research to be scrappier, more collaborative and more creative. The result is that “Chinese labs are extracting 4-7x as much intelligence per unit of compute as naive scaling predictions would suggest,” they wrote, adding that “the constraints have inadvertently created the conditions for the most formidable competitors to develop exactly the capabilities that will matter most in the coming years.” 

Matt Stoller, on his antitrust Substack BIG, added an economic wrinkle. “The US goal is to keep the stock market up, and to do that you create monopolies with large market capitalizations. The Chinese goal is to create useful technology, with market capitalization mostly irrelevant,” he wrote. And China, he argued, doesn’t care all that much: It is creating a thrumming AI ecosystem that is invigorating its economy. Over time, he went on, China may end up exporting its AI capabilities — cheaper than America's, by necessity — to other nations, including the US, much as it does with solar panels and batteries.

It’s time to change the terms of the race

Graylin proposed a change of approach, which would be particularly dramatic for the US. He argued that both countries should be hunting hares first — developing their economies through practical AI deployment — before later collaborating on the stag. This reframes the relationship "from a Cold War arms-race mental model to a Space Race innovation model," he explained. While the space race was a competition, it was also mutually advantageous — spinning off new technologies — rather than being potentially mutually destructive. Such an approach needn't preclude export controls, he pointed out, but it would limit them to genuine military, cyber and biorisk threats rather than blanket bans on advanced technology.

Liu, too, made the case for cooperation: The “AI race narrative is undermining prospects for international co-operation, at precisely the moment when co-operation is most needed,” she wrote, adding that it’s getting in the way of cooperation on safety and obscures problems like domestic AI-driven labor disruption and energy bottlenecks. 

Perhaps people in the corridors of power are paying attention, at least a little: During President Trump’s recent trip to China, Treasury Secretary Scott Bessent said that the US and China will start discussing AI safety, though no timelines were mentioned, and there are currently no plans for either side to change the way they’re developing the technology.

So when is a race not a race? At least according to these Substackers, it’s when it shouldn’t really be a race at all.

OpenAI, Anthropic, and the War of the JVs, from The Change Constant

OpenAI and Anthropic sell AI as a way to automate work inside companies, but they’re also building service businesses that look an awful lot like consultancies to make that happen. As Saanya Ojha explains in this post, both firms are assembling private equity-backed ventures staffed by engineers and consultants whose job is to help companies integrate AI systems into their operations. These are not small side projects. Anthropic’s is reportedly a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs. OpenAI is said to be seeking roughly $4 billion from private equity investors for a similar entity valued at $10 billion. Partnering with PE firms, she points out, provides direct access to networks of portfolio companies, all of which are under pressure to improve productivity ahead of exits or IPOs, while PE firms get a mechanism for injecting AI into their holdings to boost valuations. Everyone wins! Well, maybe not everyone. Traditional consultancies like Accenture, Deloitte and PwC used to advise companies on how to deploy new technology, but why bother with those firms when you go straight to the source?

In Defense of AI Slop, from Theory of the Game

Reid Hoffman joined Substack in April and is already getting traction. This essay offers a surprisingly compelling defense of AI slop, which, he argues, marks a predictable stage in the adoption of a new general-purpose technology. His analogy is electrification: Before electricity reshaped industry and everyday life, it was often used for spectacle, such as the gigantic illuminated Heinz pickle once mounted on New York City’s Cumberland Hotel. At the time, critics dismissed the pickle and lots of other electric displays as garish and wasteful; they became emblematic of what many people thought was a dangerous and unnecessary technology. Hoffman argues something similar is happening with AI. Public frustration over giant data centers and soaring energy use is colliding with the perception that, for many people, AI produces little more than stupid memes and fake images. But, he writes, this “mistakes the ‘pickle’ for the ‘grid.’” Transformative technologies, he writes, often arrive first as novelties before the infrastructure underpinning them becomes economically indispensable. His take is predictably techno-utopian, but it’s hard to outright dismiss the historic comparison. 

Seductive Salience, from Threading the Needle

One common theory of political change is that if you want policymakers to act, you need to make a lot of noise. Anton Leicht, a researcher focused on the political economy of advanced AI, argues that this strategy isn't guaranteed to yield success for artificial intelligence regulation. His core idea is that high public engagement rarely produces careful policy. Instead, it tends to create volatile, emotionally charged politics in which the issue at hand — in this case, AI — is used as a vessel for broader preexisting grievances, like anti-capitalist or anti-tech views. Once that happens, nuanced policymaking gets more difficult. Leicht concedes that some issues are inevitably going to become politically salient, with AI-driven job loss being an obvious example. In those cases, policymakers and advocates have little choice but to engage with the public debate and try to shape it as best they can. But he argues that other areas — such as frontier model safety and technical governance — aren’t suited to widespread and simplistic public debate, because they demand careful negotiation, institutional coordination and highly specific rules. In recent years, the AI industry has often drawn attention to politically volatile issues around AI like job losses, existential risk and the impact of a future AGI. Leicht’s argument is that it might be time to shut up in order to build the kind of useful AI governance that could benefit society.

Will we know when AI is taking our jobs? from The Argument

One of the difficulties in understanding the impact of AI on employment is that technological change rarely swaps humans directly for machines. An accountant probably won’t arrive at the office one morning to find Claude sitting at their desk; instead, their work may gradually evolve — month by month, task by task — until the role becomes something fundamentally different. The problem, this post argues, is that our tools for measuring labor markets aren’t equipped to capture that sort of change. Standard employment statistics can tell us how many jobs were created or lost, but not why. By the time economists can confidently isolate AI as the cause of a shift, years may have passed. So researchers are experimenting with new ways to track what’s happening: analyzing usage data to understand how workers are adopting AI tools; conducting interviews to explore how employees perceive changes to their responsibilities; scraping job postings to identify which tasks companies are increasingly hiring for; studying pricing and sales data to see how falling production costs driven by AI reshape demand. None of these approaches is perfect, but they may help inform our understanding of what’s happening inside the labor market before significant job losses. 

Do renewables make electricity cheaper or more expensive? from Bright Spots

They do both. That is an unsatisfying answer, but it’s also correct according to this analysis. Jan Rosenow, a professor of energy and climate policy at the University of Oxford, studied energy markets in Europe and the US and found that higher shares of renewables generally push down wholesale electricity prices. But that relationship breaks down when you look at household bills. Retail electricity prices are shaped far less by generation costs than by network charges, taxes, levies and infrastructure spending. In other words, even if renewable energy makes electricity cheaper to produce, consumers may not feel much of the benefit. That points to what Rosenow sees as the real policy challenge. “If we want clean electricity to translate into clean heat, clean industry and clean transport,” he writes, “the wholesale market is doing most of its job. The retail market mostly isn’t.” He suggests redesigning retail pricing so consumers benefit more directly from cheap renewable power, for instance by shifting taxes and levies away from electricity prices and onto any fossil fuels used to generate electricity, or by exposing households to more real-time pricing so they can save money by using power when renewable supply is abundant. The complication is that decarbonizing the grid requires enormous investment in transmission, balancing and storage infrastructure. Those costs are typically passed through to consumers as network charges. And if electrification stalls, those fixed infrastructure costs get spread across stagnant or declining electricity demand, potentially creating a vicious cycle in which electricity becomes even more expensive.