오픈웨이트 AI, 쿠버네티스의 시대를 열다

Open-weight models are becoming the foundation for the next AI ecosystem. The US should compete in it, not wall itself off. I have seen a version of this story before. In 2013 I co-founded Mesosphere, an open source cloud-native software company. We built on Apache Mesos, which my co-founder Ben Hindman had helped create at UC Berkeley. We later built DC/OS (Data Center Operating System) around Mesos, released it as open source, and commercialized it through an enterprise distribution with support and proprietary features. After several years of massive growth, Kubernetes disrupted us. It was newer, fully open source, and it quickly galvanized the cloud-native community. Many of the world’s best distributed-systems and infrastructure engineers bet their careers on it, and even some of our most loyal community members changed horses. Once that happened, innovation moved to Kubernetes. Whatever the platform was missing, someone started building: networking, storage, observability, deployment tools, policy engines. A large number of startups were created, and legacy vendors joined in as well. Almost every component required to run Kubernetes in production became available as open source. Cloud providers and companies including Mesosphere/D2iQ, Rancher, Red Hat and Nutanix then built businesses around integration, enterprise features, support and operations. Kubernetes did not win simply because its repository was public. It became a neutral substrate that engineers, cloud providers and enterprise vendors could all extend to fit their customers’ needs. Common interfaces and vendor-neutral governance gave everyone confidence that they could build on it. The lesson I took away wasn’t that open source always wins. It was that once an open platform that people can customize becomes the industry’s center of gravity, no single vendor can match the combined rate of innovation around it. I think AI is approaching the same point. Open weights turn a model into a platform First, a terminology note. Most models we casually call “open source” are more accurately described as open-weight . You can download and modify the trained parameters, but the training data and complete training process usually aren’t available. That falls short of the Open Source Initiative’s definition of open source AI . The distinction matters. It doesn’t, however, prevent an ecosystem from forming around the artifact people can run and modify. The Kubernetes analogy isn’t perfect. Kubernetes contributors could inspect and change the actual source, and improvements could flow back into a shared upstream project. Model fine-tunes usually don’t work that way. Frontier weights may be downloadable but still require expensive hardware, and there is no AI equivalent of the CNCF providing neutral governance and common interfaces. Those are real differences. The common mechanism is that a sufficiently capable, portable substrate can attract complementary innovation far beyond what its original creator could build alone. The first reason to use open-weight models was self-hosting. Companies wanted control over their data. They wanted to run models in their own cloud or data center. And as usage and inference costs grow, they increasingly want control over cost as well. That demand produced a healthy open source serving stack: vLLM, SGLang, llama.cpp, Ollama, MLX and others. But self-hosting is only the beginning. Open weights turn the model itself into something developers can adapt and redistribute. Hugging Face now hosts more than two million public models . Around popular families such as Qwen and Gemma, developers produce: quantized and converted weights for different silicon architectures and scale; fine-tunes and LoRA adapters for coding, medicine, law, math and agentic workflows; model merges that combine different fine-tunes; adaptations for runtimes such as TensorRT-LLM, vLLM, MLX and others. Until recently, it was easy to dismiss this activity at the frontier. Open models were useful, but the base models weren’t good enough for the hardest coding and agentic tasks. That gap is narrowing quickly. Z.ai has released GLM-5.2 with public weights under an MIT license. Its own evaluation reports 62.1% on SWE-bench Pro versus 58.6% for GPT-5.5, although results vary across benchmarks and agent harnesses. Moonshot says Kimi K3 approaches the closed frontier on long-horizon coding and has promised to publish its weights on July 27. Artificial Analysis supports the performance claim, scoring it alongside Opus 4.8 and GPT-5.5 in its independent evaluation. Once the base model is good enough, the ecosystem can compound. I expect new projects around agent runtimes, coding harnesses, sandboxes, evaluations, observability and specialized fine-tunes. Together they can become a production-grade stack: an open-weight model running on open source software, customizable for a team’s workload, hardware and economics. Will that stack beat every closed model on every benchmark? Probably not. But I would not bet on any single vendor out-innovating the combined open ecosystem over time. Frontier technology is a talent war, and open ecosystems give talented people everywhere a reason to build on the same foundation. Banning Chinese models would be an own goal This brings us to the current debate in Washington. After the release of Kimi K3 and other capable Chinese models, the Trump administration is reportedly considering restrictions on Chinese open-weight models. The exact form of a potential ban remains unclear. A broad ban on American researchers and companies using Chinese open-weight models would do something else entirely. It would cut the US off from an ecosystem that is already attracting many of the world’s best AI researchers and engineers, including a large number of Chinese researchers. The rest of the world would keep building. American developers would be the ones locked out. We’ve seen this dynamic with Qwen already. Hugging Face reports that Chinese models accounted for 41% of model downloads over the past year. If the best open-weight foundation models increasingly come from China, innovation will accumulate around them in the same way it accumulated around Kubernetes. How the US should compete The US should compete in that ecosystem, not retreat from it. Release frontier-grade American models American labs need to release frontier-grade open-weight models under licenses that startups can actually build on. There has been progress. NVIDIA’s Nemotron models are commercially usable under NVIDIA’s own permissive license. Thinking Machines released Inkling under Apache 2.0, as did OpenAI with gpt-oss and Google with Gemma 4 . But OpenAI’s and Google’s strongest models remain closed, as do those from most American frontier labs. Use procurement to create an open market The government should use procurement to create demand for portable, interoperable systems rather than permanent dependence on one API vendor. The Department of Defense has done this before. Platform One provides open source tools and enterprise products that different military programs can build on. The same playbook can accelerate innovation around open-weight models. Build the rest of the stack American companies need to build the rest of the stack. Startups can customize and extend the models, embed them into products, and provide the serving, tooling, support and operational layers. Our leading silicon companies will keep improving the hardware. Hyperscalers and neoclouds can serve the models and their ecosystems. Set standards instead of banning models Safety is the strongest argument for restrictions, but a blanket ban is too blunt and would sacrifice access to the entire ecosystem. A better approach is independent testing and standards for frontier models. The analogy isn’t exact: Kubernetes conformance tests co