It's Time to Push for Domestic, Open Models in AI

Note: I use "Open" to refer to "open-weight". It's my hope that "Open-source" models become more prevalent with time!

There is something in the air right now in AI land. A perfect concoction of awe in the power of state-of-the-art models, cynicism at the amount of financial leverage frontier research labs hold, and a sense of dread of how much volatility exists; on the future of jobs, the markets, and on human creativity itself.

AI is, by its nature, a technological force that thrives on centralization. In turn, our financial markets have consolidated an equal force into the labs and companies building AI. Since the advent of LLMs, dogma has been that the model providers will always hold a moat by being the utility providers; it doesn't matter how much of a commodity they become. With this leverage, they can easily absorb more capital, build more complimentary products and absorb SaaS competitors, and become true hegemons of technology.

But open-weight models, both domestic and abroad, are attempting to push the pendulum back the other way. The releases of Kimi K3 from China's Moonshot AI and Inkling from America's Thinking Machines are potent reminders that the gap in model performance is shrinking.

"History does not repeat itself, but it often rhymes"

This quote is often attributed to Mark Twain, and while origins are disputable, its meaning has always been relevant for the ongoing evolution of computers.

Every major wave in modern-day computing strongly benefited by finding low-cost and decentralized versions of its flagship technology.

• Supercomputers and mainframes slowly phased out as personal computers became more practical, reliable, and powerful. This transformation continued with smart phones, where they went from luxury to commodity and Android now being the dominant operating system for mobile devices.
• Cloud computing ate the on-premise computing marketplace alive. While still relevant, it was clear that the cloud solutions provided a lower barrier of entry to products and innovation via the Internet.
• Open-source software in general has been transformative. Linux powers the world, but in the 90s everyone thought operating systems were done; you had Windows for PCs and Unix-derivatives for the enterprise (see supercomputers above). Today, we all benefit from open-source software that powers the infrastructure of the Internet itself.

The democratization of compute is what inevitably pulls our technology forward and brings true benefits to society at large. Why should AI be any different if we really care about its impact on the world?

Swinging the pendulum back

There are clear ways where a democratized approach to AI models pulls us away from the high-risk, centralized ownership of AI today:

1. Local inference would break the API rent model. The more we push for smarter, higher-performing models on smaller devices (aiming for a similar transition of mainframes to android devices), there is no need to be metered as much for AI workloads that happen external to you.
2. Fine-tuning and customization would let competitors build differentiated products on top of the base model rather than being absorbed as "SaaS features" for a frontier lab.
3. Price competition! This is one we already feel the impact of in the present moment. With downloadable models and weights, the marginal cost of inference will keep collapsing towards compute cost. Vendor pricing power will diminish rapidly.

Why domestic open models?

Open models break the utility oligopoly. But crucially, where those models are trained and made, who controls the release pipeline, and how easy it is for US entities to run them locally without sending data overseas, determines whether democratization strengthens American innovation or just keeps shifting our dependency to Beijing.

Data sovereignty & operational security

If only high-performing open models come from Chinese labs, US companies, hospitals, and government agencies face a binary decision: either pay the toll to a closed US API, or download and run a model whose provenance, training-data governance, and update pipeline sits outside US jurisdiction. Domestic open releases lets US entities run frontier-capable inference within the US without these foreign risk factors.

Breaking API dependency within the US market

Open models within the American ecosystem pushes AI from a service subscription to an infrastructure component domestic entities can fully own, fine-tune, and deploy; Just like how Linux broke dependencies on proprietary UNIX vendors for enterprise computing.

Protecting our innovation ecosystem

When models are open and domestic, US universities and startups can experiment, publish, and build without NDAs, usage restrictions, or fear that a foreign state-linked entity controls the foundational layer of their work. We get to keep downstream value creation rather than forcing researchers and engineers to either pay the OpenAI/Anthropic tax or rely on Chinese open releases.

National Competitiveness

If open AI becomes synonymous with "Chinese AI", the global open ecosystem standardizes on Chinese architectural choices, evaluation standards, and eventually even hardware optimization targets. A credible presence from US open models would ensure that the nation has a voice in the open-standard stack and not just the closed-API market.

What about security?

If there's one point the frontier labs keep making as a peril of powerful AI models, its the security implications of unfettered access—bioweapons, distribution of false information, and autonomous hacking agents (to name a few).

It is true that open models can be fine-tuned to become more dangerous, and this is harder to do with a closed API. But the response is not to hide US AI technology while Chinese ones circulate freely, because we know that dealing with security risks will be inevitable. Our open models need to be built with strong underlying alignment, defensive capabilities, and be paired with significant legal deterrents, like DNA-screening laws the AI CEOs themselves are lobbying for.

Inking is a good example of this, which ships with explicit safeguards for weapon-related, cyber, and CBRN queries. It has a refusal architecture that was evaluated by external testers. It's not a perfect, airtight solution, but it is auditable.

The stage is set

The frontier labs and their models have been immensely useful and important. Without them and the financial engineering backing them, the development of AI models would have been slower, and there would have been less incentives for people to work on open-weight models.

But its time to swing the pendulum back the other way, and we need to do so at home, of our own volition. Every piece of historical and present evidence shows that it remains the only way for our long-term success as a technology and innovation hub.

The alternative is that we lose it all to China. Is that objectively a bad thing? For America, at least, I believe it is.