Washington’s new AI fight is not about chatbots. It is about who becomes the default stack.

Washington’s new AI fight is not about chatbots. It is about who becomes the default stack.

By Amara Diallo

July 8, 2026

The newest front in the U.S.-China technology fight is not a chip factory, an app store, or a single viral chatbot. It is the quiet layer underneath a fast-growing number of products: the artificial intelligence model a company chooses when it needs code written, customer support summarized, listings moderated, fraud patterns detected, or a new internal tool built quickly.

That choice is now drawing scrutiny in Washington. U.S. lawmakers are weighing how to slow the adoption of Chinese-developed AI models inside American companies, CNBC reported Wednesday, citing an ongoing joint investigation by the House Committee on Homeland Security and the House Select Committee on China. The committees have been examining the use of models built in China by U.S. firms, including letters sent earlier this year to Cursor and Airbnb about their exposure to those systems.

The political concern is straightforward: if Chinese models become cheap, capable defaults for American startups and corporate engineering teams, Beijing’s AI ecosystem could gain influence not by winning a consumer popularity contest, but by becoming infrastructure. The technical reality is harder. Open-weight and open-source models can be downloaded, modified, hosted through intermediaries, or embedded deep inside software workflows. That makes this a different kind of national security fight than blocking a phone app from federal devices or restricting exports of advanced semiconductors.

The immediate trigger is performance and price. Chinese models have narrowed the quality gap with U.S. rivals while often costing less to use, according to CNBC. For companies watching AI bills become a real line item, the pull is obvious: if a model is good enough and cheaper, engineers will test it. If it improves developer speed or support quality, they will try to ship with it. In software, defaults harden fast.

That is why the story matters beyond Capitol Hill. The AI race is often framed as a contest between national champions — OpenAI, Anthropic, Google, Meta, DeepSeek, Alibaba, Baidu, Moonshot AI. But the adoption fight is happening in messier places: procurement forms, cloud dashboards, code editors, API wrappers, and the judgment calls of CTOs under pressure to deploy useful tools without exploding budgets.

CNBC’s report says some U.S. government departments have banned Chinese AI models including DeepSeek, but there is no broad prohibition on American companies using them. That gap is now the policy battleground. The question is not only whether Washington distrusts Chinese AI. It is whether the U.S. can offer an alternative that is open enough, affordable enough, and capable enough that companies do not feel pushed toward Chinese models by economics.

Andrew Garbarino, chairman of the House Homeland Security Committee, told CNBC that the Chinese Communist Party is “racing to close the gap” in capabilities that could shape the future of cybersecurity. He pointed in particular to reports that a Chinese open-weight model could match leading U.S. models in certain vulnerability discovery and cybersecurity tasks. That is a sharper worry than generic chatbot competition: a model strong at finding software weaknesses can be useful for defense, but also for offense.

The State Department also framed the issue in ideological and security terms. A spokesperson told CNBC that the growing use of Chinese AI models by U.S. companies raises “serious concerns,” arguing that those models are designed to advance Beijing’s narratives, censor dissent, and reflect Chinese Communist Party ideology and values. A spokesperson for the Chinese embassy in the United Kingdom rejected that framing, telling CNBC that China “opposes baseless allegations and malicious smears against its AI development” and that the country’s AI sector is built on self-reliance and scientific strength.

That exchange captures the political split. Washington sees AI models as strategic systems that can carry hidden risks: data exposure, censorship patterns, security vulnerabilities, dependency, and the possible transfer of U.S. user behavior into ecosystems shaped by a rival state. Beijing sees the criticism as an attempt to contain Chinese technological progress while U.S. firms keep their own advantages.

For readers, the key is not to treat either side’s broadest claim as settled fact. There is a real security debate here, and there is also a real market force. The models are being considered because they work well enough for some tasks and cost less. If policymakers ignore that practical reason, they will write rules that sound tough and then get routed around by companies looking for cheaper compute.

Airbnb’s response to CNBC showed how nuanced corporate use can be. The company said its AI activity runs “overwhelmingly on U.S.-origin models,” while acknowledging use of a limited number of China-origin models. Airbnb said those models are open-source, run only through approved U.S.-based service providers, and keep data and operations separate and protected.

That is the kind of arrangement many companies will argue is materially different from sending sensitive data straight to a foreign platform. A model’s country of origin is one thing. Where it runs, who hosts it, what data it sees, whether its weights are inspectable, and how it is monitored are separate questions. A serious policy response has to distinguish among those layers.

Cursor, according to CNBC, built its Composer 2 model using Kimi, an AI model developed by Moonshot AI in China. CNBC reported that Cursor declined to comment on the probe. That example lands because developer tools sit close to valuable code. If a company’s programmers rely on an AI assistant for software generation, debugging, or vulnerability analysis, the model may touch intellectual property, security assumptions, and internal architecture patterns. Even when raw code is not exfiltrated, model behavior can shape what gets built.

The debate is unfolding as Chinese AI firms push further down the stack. Ars Technica reported Tuesday that DeepSeek is planning a move into data-center inference chips, citing Reuters reporting. The goal, according to that report, is likely to reduce reliance on both Huawei and Nvidia. Ars noted that the focus is inference — running AI models after they are trained — not training itself.

That matters because inference is where AI becomes an everyday utility. Training a frontier model is capital-intensive and spectacular. Inference is the bill that arrives every time a user asks a model to summarize a contract, generate code, search a database, translate a support ticket, or screen a transaction. If Chinese AI companies can pair competitive models with cheaper inference hardware and aggressive distribution, the market pressure on U.S. companies grows.

The chip context is broader than DeepSeek. Apple has begun testing DRAM chips from China’s ChangXin Memory Technologies for devices sold in China and has been lobbying the U.S. government to allow broader use of CXMT products, CNBC separately reported Wednesday, citing the Financial Times. CXMT is expected to be central to Beijing’s push for a more self-sufficient AI supply chain, according to that report. CNBC said CXMT is currently the world’s fourth-largest DRAM producer and cited SemiAnalysis data projecting its market share could rise to 15 percent by 2028 from roughly 11 percent last year.

Taken together, these are not isolated headlines. They point to a Chinese tech strategy that spans models, chips, memory, cloud infrastructure, and enterprise adoption. The U.S. response has often focused on export controls — trying to limit China’s access to the most advanced chips used to train AI systems. But if Chinese firms keep improving open models and building more of their own AI hardware, the policy question shifts. Washington cannot only ask what China is prevented from buying. It has to ask what the rest of the world is choosing to use.

One option under discussion is procurement pressure. Kyle Chan, a fellow at the Brookings Institution’s John L. Thornton China Center, told CNBC the administration could consider federal procurement bans that restrict government agencies and private companies serving the U.S. government from using Chinese AI models. Daniel Remler, a senior fellow in the technology and national security program at the Center for a New American Security, told CNBC that procurement requirements could discourage government contractors from relying on those models, while officials might also disseminate findings about risks and vulnerabilities.

That path would be familiar. Government purchasing rules often become de facto market signals. If federal agencies and contractors cannot use certain tools, enterprise vendors may redesign their stacks around compliant alternatives. The effect can spill beyond the government market because companies dislike maintaining separate versions of the same product.

But the limits are equally clear. Chan warned CNBC that it is “ultimately impossible” to ban China’s open-source AI models because model weights are freely available online, and that such efforts could raise First Amendment issues. Remler also noted that action against Chinese models could hurt startups that use them or chill support for open models more generally.

That is the part Washington has not fully solved. Open AI models are not normal imports sitting in a container at a port. They are files, code, weights, derivatives, fine-tunes, and hosted services. A ban can address federal use. It can regulate contractors. It can require disclosure. It can shape risk management. But it cannot easily erase the underlying software from the internet.

The most durable U.S. answer may be less dramatic than a blanket ban: build and support better alternatives. CNBC reported that committee aides are examining whether the United States has a sufficient open-weight AI strategy so American companies and cyber defenders are not forced to choose between expensive or restricted U.S. models and cheap, capable models developed in China. That is the heart of the story.

If U.S. policymakers want companies to avoid Chinese models, they need to understand why those companies are reaching for them. Some are chasing cost. Some want open weights. Some want fewer usage restrictions. Some need models that perform well on coding or cybersecurity tasks. Some are experimenting because engineers experiment. A successful strategy would pair risk disclosure and procurement rules with investment in trustworthy, auditable, competitive open models that domestic firms can actually afford to run.

There is also a transparency problem for users. Most people interacting with an AI feature do not know which model sits behind it. A travel listing description, support chatbot, fraud alert, coding assistant, or workplace search tool may be powered by a U.S. model, a Chinese model, a mixture of models, or a company’s fine-tuned derivative. If origin and hosting matter for security, privacy, or censorship, companies will face pressure to disclose more about their AI supply chains.

That disclosure should be useful, not performative. “Made with AI” is too vague. “Uses only U.S.-hosted models for customer data” is more meaningful. “Open-weight model derived from a China-origin base model, hosted by a U.S. cloud provider, with no customer data retained for training” is clunky, but closer to the truth users and enterprise buyers need. The next phase of AI governance may look a lot like supply-chain labeling.

For now, the immediate stakes are political and practical. Lawmakers can make Chinese AI adoption more costly for federal contractors. Agencies can issue guidance. Companies can tighten internal review. But if the technology keeps improving and remains cheaper, informal adoption will continue unless credible alternatives exist.

That is why today’s story is bigger than a probe. It is a warning that the AI race is moving from splashy model launches into the plumbing of the digital economy. The winner may not be the model with the loudest demo. It may be the model that becomes the easiest, cheapest, good-enough default inside thousands of products.

Washington is trying to stop Chinese AI from becoming that default. The harder task is proving that American and allied alternatives can win on more than fear.