The core of Nvidia’s strategy centers around promoting open-source model recipes, mirroring the success of projects like Linux. This approach aims to distribute the capability of creating AI models, rather than concentrating it within a few large organizations. Nvidia’s investment, totaling $26 billion, is predicated on the belief that widespread adoption of these open recipes will generate substantial demand for Nvidia’s inference hardware. The company’s goal is to establish a landscape where countless entities can build ‘token machines,’ effectively reducing the dominance of proprietary AI models.
Currently, models like Llama 3 and Olmo, developed by Ai2 and EleutherAI respectively, demonstrate the viability of this open-source approach. These models, along with their associated training recipes, allow companies to adapt and deploy them for specific applications. The open-weight models themselves, offering just the model weights and inference code, represent a more accessible entry point for developers and researchers. This contrasts with the closed-source models offered by companies like Anthropic and OpenAI, which typically include the entire training process.
However, the path to widespread adoption of open-source AI faces significant challenges. Training large language models is capital-intensive, and the open-source recipe development process is resource-demanding. Many companies are hesitant to invest in the necessary infrastructure and expertise. The success of Nvidia’s investment depends on the emergence of a viable economic model – one where the profits generated by open-source model training and inference ultimately return to Nvidia.
Looking ahead, two potential futures exist. The first, if successful, would see Nvidia’s investment fueling massive demand for its hardware and generating substantial profits. The second scenario involves a fork in the open-source ecosystem, with models prioritizing efficiency, modifiability, and specialization, particularly in areas like enterprise-specific agent deployments. This shift would likely see open models filling a ‘long-tail’ ecosystem, complementing the more dominant closed-source models in areas like knowledge work and drug discovery.
Recent trends indicate a move towards post-training customization, with users fine-tuning base models like DeepSeek V4 Flash or GLM 5.X for agentic tasks. This shift is accompanied by a decreasing number of open model builders releasing base models, alongside experiments with revenue-share licenses to sustain the development of near-frontier open-weight models. The success of these experiments will be crucial in determining the long-term viability of Nvidia’s investment.
