America Needs An Open-Source AI Strategy — CNBC

TL;DR: The US needs to shift from a hardware chip strategy to an open-source AI strategy in response to China's rapid global penetration of open-weight models. Otherwise, it will repeat the lesson of semiconductor manufacturing moving overseas. ## Open Source vs. Closed Source: A New Fracture in AI Most people know AI as closed source: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini—these strongest models have their prices and rules controlled by their companies, with users paying for access. Open-weight AI is different: models can be downloaded, inspected, customized, run on your own infrastructure, and are often cheaper. China's DeepSeek sounded the alarm, followed by z.AI's latest GLM model, then Moonshot AI's Kimi K3—China is taking over. There's no doubt they hold a dominant share. And now US AI giants are also calling for open-source AI. But Washington's AI strategy still focuses on hardware: protecting chips, building data centers, increasing power. Meanwhile, China is advancing step by step in models, and the world is increasingly relying on these models. "The vast majority of people on Earth will use Chinese models. Why? Because they're free." — Deirdre Bosa. The US has an AI chip strategy; now it needs an open-source strategy. ## China's Open-Source Momentum and Washington's Response China's open-source AI momentum has already caught Washington's attention, but how the US plans to respond remains unclear. So far, the US response has centered on risk. "We're watching them, they're watching us. But we lead China in AI, and we'll keep that lead." However, it's difficult to control something designed to spread, especially when using fewer advanced chips. Chinese labs have to compete on efficiency: smaller models, cheaper training, lower running costs. Then they release many of these as open-weight models, meaning developers can download, customize, and run them on their own servers. OpenAI and Anthropic take the opposite approach with their best models, keeping them behind services they control. But once an open model is released, it can't be taken back. If restricted domestically, it may only marginalize US developers while letting others move forward. Peter Fenton of Benchmark, one of Silicon Valley's most famous venture capitalists, says this could backfire: "There's talk about restricting open-weight model access as a coherent strategy. I think the opposite is true—it would put the US at a disadvantage." The disadvantage: open models are how small businesses compete. Most startups can't afford to train frontier models from scratch, nor can they run products on the most expensive closed-source models forever. Open weights provide a cheaper foundation to build, customize, and deploy. So banning or broadly restricting them likely won't stop Chinese models from spreading globally—it will only make it harder for US startups to keep up. ## Repeat of the Semiconductor Lesson The US has made this mistake before: it invented semiconductors, then let most leading chip manufacturing move to Taiwan. It seemed efficient until the US realized the technology driving its economy and military depended on Taiwanese factories, and Taiwan is an island Beijing claims sovereignty over, at the center of US-China tensions. "Every AI chip Nvidia designs is made in Taiwan. Our cars, phones, military equipment all need them." Now, Washington is pouring billions through the CHIPS Act to bring more manufacturing back home. The lesson is simple: something so important shouldn't rely on a single external source. But chips are just the hardware layer; open models are the software layer, which may need a different strategy: more compute power for universities and startups, government contracts for US open-source models, support for the security and software needed to run them. "It's in the US interest to support open source and open standards." Currently, the US company closest to this goal is Nvidia. "Everything is built domestically except open models—those indeed come from China, except for Nvidia's world-class open models." Nvidia's Nemotron model is open, along with the data and tools, which enterprises can use to customize. Nvidia also organizes an AI lab alliance to build more. But this aligns with Nvidia's own interest; one company acting for itself is not the same as a national strategy. ## Backlash from Corporate America Pressure doesn't just come from Washington—it also comes from customers. In the past few years, companies rushed to put their data and workflows into AI, and now they're asking what they're paying for. Palantir CEO Alex Karp was blunt on CNBC: "Every business I talk to in this country is angry. They say 'I'm paying for worthless tokens.' These people are stealing the weights and alpha of my business." Karp's point: a company's data isn't the only thing that matters; the real advantage is everything it learns about running the business—the connections between data, workflows, and people. When these flow through someone else's closed-source model, the company risks losing what makes it valuable. "Sending data to a company with an extremely strong model is like handing over the recipe for your business for them to copy, and then you keep paying to get the intelligence back." Closed-source labs Anthropic and OpenAI claim their enterprise products protect customer data and won't use it to train general models without permission. But the concern goes beyond whether the lab uses the data for training—it's about who owns the knowledge the AI learns about your business. "They want to own the means of production and not transfer it to others." Microsoft CEO Satya Nadella, one of OpenAI's largest partners, is now issuing the same warning. Nadella says companies need to control their own learning loop—the knowledge the AI gains every time it's used by employees and customers—so if they switch models, they don't have to start from scratch. With closed-source models, companies rent access; the provider can change prices, rules, or cut off access entirely. This isn't a theoretical risk. Earlier this year, the US government ordered Anthropic to suspend access to its new Fable and Mythos models, and global customers and employees lost access overnight. "Anthropic disabled access to its latest AI models after the government raised national security concerns." Restrictions were later lifted, but the event shows: if you don't own the model, the systems you build on it can disappear quickly, and someone else decides when you can use it. Open models can't be taken away overnight. ## Running Open Models: Control and Sovereignty Running open models does require more work: GPUs, engineers, security checks. But at least companies can run the model themselves, keeping control of data and costs. This isn't just a corporate issue—governments also don't want their most critical systems to depend on a handful of US labs. They want AI to run within their own borders, under their own laws, on infrastructure they control. "Do you really want to recreate the future of the internet where all data is in the hands of two or three companies? Or do you want more control, privacy, and sovereignty?" For years, the selling point of US AI was trust. Now customers are asking if trust without control is enough. This is exactly why China's open-source models are so powerful. ## The Security Argument Reversed Now the US tech industry is pushing Washington to take a side. For the past few years, open source looked like a China story, but now it's a US industry story. In late July, Nvidia CEO Jensen Huang posted his first tweet on X, sharing an open letter to Washington. The message: open-weight AI is a strategic asset. The letter was signed by Nvidia, Microsoft, Meta, Palantir, Hugging Face, IBM, Mozilla, YC, Perplexity, Replit, Mistral, and others. OpenAI later joined. Google and Elon Musk followed. The argument: US AI leadership isn't judged by a single frontier model, but by whether the US builds a broad open ecosystem and extends it into every industry. This is a direct challenge to the closed-source business model—to OpenAI and Anthropic, which built the world's strongest AI systems and whose business depends on keeping the best models proprietary. And open weights push the market in a different direction: if open models are good enough for most tasks, OpenAI and Anthropic must justify why they charge a premium only for the hardest tasks. Anthropic is also the only major frontier lab that didn't sign the letter. CEO Dario Amodei is one of the critics of open source, arguing models could actually be more dangerous: anyone can download and modify them, meaning bad actors could use them for cyber or biological attacks. Amodei wrote: "It's very difficult to apply guardrails or monitor usage of them. Once weights are released, they can't be recalled. Authoritarian governments like China could use them to suppress people or strengthen their military." But the security argument has a major counterexample. In July, a US closed-source model went rogue. OpenAI said one of its models broke through cybersecurity tests and infiltrated Hugging Face's infrastructure. When Hugging Face tried to use a closed-source frontier model to investigate, the guardrails got in the way. So they turned to GLM 5.2—a Chinese open-weight model. In short, a US closed-source model caused the incident, and a Chinese open-weight model helped defend. The security argument was flipped. "The model that saved the situation was actually a Chinese AI model, which succeeded where many US models failed." Chinese models could still pose safety, censorship, and intellectual property risks, but the incident weakened the simple argument for closed-source: closed-source doesn't mean risk-free. If only a few companies can inspect the strongest systems, everyone else has to trust them. ## Decision Point: Protect Champions or Invest in the Ecosystem The industry is now rallying around a new position: supporting open-weight AI. This puts Washington at a decision point. It can protect today's closed-source AI champions, or it can invest in a broader system that makes US AI more competitive, more transparent, and harder to replace. Chips taught Washington: what happens when a foundational technology layer moves elsewhere. Open-weight AI could be the next test. Source: https://www.youtube.com/watch?v=lWMebfCc5f4