This article by Max Bessler was published at Palladium Magazine on October 30, 2026.
Become a PALLADIUM member today! You will receive our beautiful quarterly print edition and help support our non-profit publication.
Since World War II, Washington has leveraged the global presence of U.S. companies to advance its national security interests. AI models are the latest arena where U.S. and Chinese technology firms compete for market share. The U.S. government controls where Silicon Valley companies can export the best chips; intelligence agencies create backdoors on digital platforms to spy on adversaries worldwide. The world relies on these same technology firms, from Nvidia to Meta, which also grant Washington significant power. Many American technologists and policymakers hope AI technology could strengthen this tech advantage. However, changes in the AI sector might make this playbook stall before it starts, regardless of whether China truly “catches up” in its AI model capabilities.
The two AI superpowers fight for market share with distinctly different approaches. OpenAI and Anthropic keep their best models “closed,” not revealing the neural net weights, the numerical parameters that generate answers and form the core of modern large language models (LLMs), or the proprietary training techniques used. Silicon Valley labs then charge users for access to these proprietary models. Yet, this business model is not the only one. Leading Chinese firms Moonshot, Alibaba, Zhipu, and DeepSeek make the weights of many of their LLMs publicly available, allowing users to download the entire model and run it on an owned or rented graphics processing unit (GPU) cluster. Beyond selling AI usage, these Chinese AI companies earn revenue indirectly, through consumer-facing applications, hosting their models in the cloud, offering consulting, and granting licensing agreements.
Today, any user with access to enough computing power can download an LLM from the open internet, watch tutorials on how to “tune” its weights, and run the model themself, bypassing the guardrails that closed models impose. If current trends continue, global demand for AI products will likely fragment into a “divided market” split between a larger commoditized pool and a smaller premium echelon of offerings. That weakens the power of the inventing state. Currently, common tasks are still sent to expensive models, like simple advice, web research, and writing, but they will become table stakes for any future off-the-shelf model. Pricier models will dominate a smaller sliver of complex, innovative AI applications unknown today. Washington has in the past used critical Silicon Valley technology to coerce other states, but a divided AI market filled with free, open-weight models undermines future U.S. ability to use AI models as an economic weapon.
Such divided markets exist elsewhere today. Take medicine: 90% of all prescriptions are generic drugs but make up only 12% of spending; brands capture only 10% of prescriptions but take home 88% of all spending. The commoditized tier holds most of the volume while the smaller premium tier captures profit. Buyers will pay a premium when the gap in capabilities is worth it, but below that, they decide based on cost. Like in the drug market, we are headed towards a familiar future where the products split between commoditized and premium AI offerings.
A Dividing AI Market on the Horizon
Are these open models as capable as closed models? No, but they’re good enough. Benchmarks like Epoch and METR suggest even the best Chinese open models continue to trail the American frontier, but they don’t have to surpass it when cost is the concern. A divided market just needs the commoditized pool to satisfy basic demands, not rival the tier above it. Companies that use cheap open-weight models for routine tasks can still selectively pay for Claude when needed for tougher jobs.
Motivated businesses and savvy consumers are already waking up to the advantages in splitting their use between open and closed AI models, and open models are flooding the market from Chinese firms. Looking at “gateways”, companies that route LLM usage as a service, analysts can compare spending between open Chinese models and closed domestic AIs. Vercel’s September 2026 report showed the market dividing the same way drugs do: “Open-weight models processed 56% of August’s gateway tokens and accounted for 14% of spend.” In July, a CNBC report found that the share of tokens used by U.S. companies on Chinese models through OpenRouter, which lets developers access different models, has sat above 30% since February and risen to as high as 46%. Downloads tell a similar story. Hugging Face, a popular website to download model weights, found the majority of its 2025 trending models were “either developed in China or derivative of a model developed in China.” Sophisticated buyers reserve expensive models for the demanding work that requires them and route everything else toward the cheapest capable models, themselves predominantly Chinese.
Measuring AI usage is inherently difficult: tech reporter Neil Woodford pointed out that “OpenAI’s API alone moves roughly six and a half times as many tokens each month as every model on OpenRouter combined”. Finding accurate industry-wide numbers is near impossible. Adding to the difficulty, when companies and employees do choose to disclose how they use AI, it can come back to bite them. As open models compete for a higher share of tasks, firms want to stay quiet about it: revealing AI usage can disclose sensitive aspects of corporate strategy. And politically, with an administration obsessed with technological competition, employees holding the AI checkbook in the U.S. gain little goodwill if they reveal they are using models controlled by Beijing.
The stakes are high for a market transformation arriving unannounced. Divided markets such as U.S. pharmaceuticals didn’t imply the risk of losing a war. AI does. In his Palladium article “American AI May Not Survive Chinese Open-Source,” Daniel Francis frames the flood of Chinese open models as an “economic attack on U.S. military development.” These military concerns should be front of mind when considering national security, but there are economic weapons at stake too. Francis argues that as AI capabilities improve exponentially, AI will become “a permanent strategic advantage technology that must be pursued at nearly any cost.” If the commoditized layer in the private sector continues to grow, this starves the frontier labs of crucial profits they rely upon to further develop the best models for the military.
Francis is right to focus on the U.S. military. When policymakers consider how to secure America, “national security” concerns often lead them to prioritize military might. But since the Second World War, policymakers have also realized that thanks to the United States’ lead in the global economy, they wield tools capable of coercing other states. After a long run of U.S. technology dominance, a dividing AI market weakens the next U.S. economic weapon on the tech frontier.
Losing Coercive Leverage
Throughout the Cold War and after, the United States has played a delicate balancing act of leveraging new waves of technology to serve the national interest. Foreign policy makers in Washington benefit each time Silicon Valley invents anew for the world—each wave of innovation centers Bay Area businesses at nodes of webs that span global manufacturing chains. Through desktops in the 1970s, operating systems in the 1980s, the internet and e-commerce over the 1990s, search engines, social networks, and smartphones in the 2000s, and streaming and cloud computing in the 2010s, U.S. firms have gone abroad with Washington behind them. Technology companies seek to internationalize their businesses to improve profit margins and reach new customers, but the U.S. mandate to protect national security follows behind them.
When technology firms go abroad, they do not set out to create ways to coerce geopolitical adversaries, but they bring home economic weapons for Washington all the same. When U.S. policymakers became concerned about the Chinese telecommunications giant Huawei’s international expansion, Washington cut Huawei off from Google’s Android in 2019 and TSMC’s chips in 2020. Subsequently, Huawei’s global smartphone market share collapsed from 20% in 2020 to virtually nothing outside mainland China by 2024. The webs that U.S.-aligned technology companies wove to maximize profit end up making even foreign champions reliant on them. Henry Farrell and Abraham Newman describe this underground empire as a world of “weaponized interdependence.” Firms chasing profit “inadvertently provide states, which are concerned with political as well as economic considerations, with the necessary levers to extend their influence across borders.” Washington gains policy levers with its ability to force technology companies to coerce others on its behalf.
In the late 20th century, Cisco found itself in this web of weaponized interdependence. Over the course of the 1990s, Cisco commercialized an array of profitable products that built the backbone of the internet and provided a node that Washington could control. Its routers trafficked most of the world’s online data. In fiscal year 1998, Cisco sold $8.5 billion of equipment that cost it $2.9 billion to make. Cisco sold network infrastructure to make money, not to spy. But so much of the internet ran through Cisco hardware that a means of U.S. control emerged. Glenn Greenwald’s No Place to Hide revealed photos of National Security Agency (NSA) employees planting “beacons” into Cisco equipment.
For this strategy to be successful, it requires control of an economic “chokepoint.” Former Treasury official and international sanctions expert Edward Fishman defines these chokepoints by three conditions: “a single country or coalition of close allies possesses a dominant, concentrated market share… substitutes are unavailable in the short term… and the country or coalition can weaponize its position in ways that impose asymmetric pressure, inflicting substantial pain on the target while suffering minimal self-harm.” Washington remains hopeful: U.S. AI frontier labs are leading innovation and chasing profits that pose to create a new chokepoint for coercion. Fishman himself hypothesizes that “if the United States can sustain its lead in artificial intelligence, the models that Silicon Valley firms produce could become the backbone of global business, handing Washington another chokepoint of extraordinary reach.”
Yet, a divided AI market denies Washington the use of AI as a chokepoint, crippling its viability as an economic weapon for U.S. policymakers. Although labs in Silicon Valley may lead the frontier, a plethora of open-weight Chinese good-enough alternatives spreads the world’s usage across models Washington cannot control. Substitutes exist, are free to download, and will continue to improve in the future. Because of these factors, there’s little chance for asymmetric pain to be inflicted by U.S. control of AI models. Even worse, trying to use AI as a chokepoint could backfire: sanctioning Chinese models could spur more open-source development within China, Europe, or even the United States itself. Fishman warns of this feedback loop, in which “every time Washington weaponizes a chokepoint, other countries take steps to insulate themselves from it”.
Today, both Washington and Silicon Valley are facing the economic pressure of China’s AI industry. Washington cannot restrict access to open-source models without upsetting thrifty players at home and abroad. If companies utilizing Chinese models were put on some sort of entity list, it would, as Fishman warns, needlessly disrupt U.S. firms and weaken their influence. A ban would also spook U.S. allies into pushing for their own sovereign models, probably themselves chasing open-weight Chinese models that could be run locally.
In Silicon Valley, labs are locked into a drive for profits that disincentivize them from diverting resources to open, unprofitable models. Unlike their more public-spirited Chinese counterparts, U.S. labs are embedded in an economy where profit capture creates power, and there is little incentive to spend money building top-rate open-weight models. Consequently, the frontier labs have done little to support an open-model ecosystem that rivals China’s, and it seems likely that for the foreseeable future China will dominate the open-weight sector.
Other American firms—that aren’t frontier labs—do benefit from the arrival of a divided AI market. The trending equilibrium where Chinese labs consistently release inexpensive open-weight models capable of tackling simple tasks is a boon for Nvidia, which benefits directly from the demand for domestic data centers running Chinese models. It also benefits U.S. corporations, especially startups, which can use cheap open-weight models to automate tasks that would be too expensive to run on Claude or GPT. In July, after Anthropic accused Chinese AI labs of distilling Claude models and releasing them as open-weight, Nvidia led more than 270 tech companies in signing an open letter against restrictions on Chinese open-weight models. OpenAI and Google begrudgingly signed late after it was clear that opposition to open-weight models was a losing strategy; Anthropic still hasn’t signed.
To overcome this erosion of the frontier labs’ economic position, Francis advocates that American AI capabilities must be defended with government support. He lands on a federal subsidy for the labs to reserve their most advanced models for the government—which also prevents model distillation—while leaving “heavily neutered” models for the public and keeping open-weight models out. That does not make the frontier a source of leverage. A model that never diffuses is no chokepoint, and leaves no lever for Washington to pull. The U.S. would be preserving AI for military and cyberwarfare use, while forfeiting the arguably more useful means of peaceful economic coercion.
Defending Against a Dividing Market, Lessons of Japan
This story has played out before. The U.S. had lost the commodity half of a strategic technology in the 1980s, when Japan took over the DRAM semiconductor market. Japan, like China today, didn’t out-innovate Silicon Valley, but rather dominated the commodity pool that emerged out of a frontier technology. DRAM semiconductors were invented by IBM, but after specifications were standardized and chips became interchangeable, the DRAM market was commoditized: little differentiation, low prices and margins, high reliability, and easy replication. Over the course of the 1970s and 1980s, Japan would take advantage of its manufacturing expertise to flood the open market. By 1987, chipmakers NEC, Toshiba, and Hitachi had pushed Japan’s share of the global DRAM market to 80 percent.
In response, Washington engaged in a series of protectionist policies after intense lobbying from Silicon Valley semiconductor firms, frustrated by being beaten by Japanese competitors. Invoking national security concerns, Bay Area chipmakers Fairchild, National Semiconductor, and Intel went to Washington and asked lobbyists to broker domestic and foreign policy responses. The Semiconductor Industry Association produced both SEMATECH—a joint industry-government research consortium—and a Ministry of International Trade and Industry-negotiated agreement whose side letter set an expectation of “slightly above 20 percent” of the Japanese DRAM market for foreign suppliers within five years. Buoyed in part by the industrial policy of SEMATECH, American firms had regained half of the world semiconductor market by the mid-1990s.
Did America win again? In one sense, no: it lost the technological competition for DRAM. Japan kept the commodity, and Korea took it from Japan next. What ultimately saved Bay Area firms in the 1990s was leaving the DRAM market and moving up the value chain into more profitable offerings. Japanese firms captured a leading sector that was dividing. Their failure to capitalize on the brief tailwind and build upon the next wave of innovation is a lesson for the geopolitically concerned technologist today. Japan won the battle without winning the longer war. As the hardware layer of semiconductors became commoditized, Bay Area startups began innovating a new wave that used them: personal computers, microprocessors, workstations, and networking all rested on commoditized DRAM. Even as they lost the commodity layer, by innovating on top of DRAM, these firms came out ahead.
DRAM, while an economic success, fails as a national-power chokepoint: too many manufacturers, too easy to substitute, and nothing to withhold for creating power asymmetries. Despite holding 80 percent of the market, Japan could not coerce anyone with it, much less afford to impose its own trade terms on the global market. Were the AI market to divide and selectively commoditize along a similar line, then the story would play out the same. Washington would not be able to win it as a lever, but neither would anyone else. So where, then, do the real chokepoints lie?
Even in a divided AI market, the U.S. government does not lose every lever in the AI industry. Outside of models, the export controls on semiconductors produced on Western supply chains are keeping China behind the frontier: 39.7 million petaflops of AI compute live in the United States, compared to just 400,000 in China. Compute passes Fishman’s chokepoint criteria with flying colors: the American firm Nvidia accounts for more than 85 percent of the market for AI semiconductor chips. Open models still need chips to run. Their lack of access to cutting-edge Nvidia chips slows the ability of Chinese labs to improve their AI models, but as the ongoing success of Chinese open-weight models shows, this isn’t enough to knock China out of the race.
Export controls work on chips because they are narrow. At home, the constraint on developing the best models is not how many chips Nvidia can sell, but how fast the country can power and permit the data centers to run them. Abroad, selling too little can spur indigenization and undermine Nvidia’s profits. Selling too much is the greater risk, accelerating China’s pace to match U.S. labs at the frontier. The second chokepoint today lies in the downstream industry of AI model hosting and inference. U.S. clouds hold over 70 percent of the market; Moonshot is negotiating to put Kimi K3 on all three of the largest; Thomson Reuters and Harvey have already built on Qwen and Kimi K3. By controlling both the chip layer, where Chinese open-weight models run, and the application layer, where the money is made, the U.S. can maintain a dominant position despite the divided market in the AI model layer.
Japan demonstrates what happens when future innovations rely upon a commodity layer of technology. Intel’s microprocessors, Apple’s computers and Cisco’s routers all ran on Japanese memory. Profits remained in California, and Japan gained little coercive capability from it. If the AI market substantially divides, the same can happen one layer up: customers of Harvey and Thomson Reuters pay the companies directly, while Moonshot and Alibaba struggle to profit even as their models do all the work.
The future of the world AI industry will center on its two largest economies: the United States of America and China. Divided markets weaken the inventing state, and the United States might have to accept that this market for AI will divide no matter its policy. How much this will hurt the profitability of frontier firms is unknown, but companies elsewhere in the AI stack remain, and with well-designed policy they can remain a chokepoint for Washington’s economic weapons. To ensure that the United States holds this chokepoint, Washington should keep chip controls narrow and keep the clouds American, the place that hosts the world’s cheap models. Maybe the U.S. can no longer tell you what model to use, but it can own where that model runs.
Max Bessler researches how technology companies play geopolitics. He is an MPhil candidate in International Relations at the University of Oxford and a Visiting Fellow at the Bay Area Council Economic Institute. You can follow him at @MaxBessler.
PALLADIUM is the first magazine of the third millennium. We are non-partisan, non-profit, and on a mission to discover what the future of governance and society will look like.
Become a PALLADIUM member today! You will receive our beautiful quarterly print edition and help support our mission.
We thank Pardon, for underwriting our work in 2026 with their generous gift.