NVIDIA Invests $26 Billion to Dominate Open AI

NVIDIA Challenges Its Clients with a Colossal Investment in Open AI

NVIDIA, the chip giant for artificial intelligence, has launched a bold offensive that could disrupt the industry. With an investment of $26 billion over five years, the company aims to develop open-weight AI models. This strategy, while risky, could transform NVIDIA into an essential player not only as a hardware provider but also as a creator of AI models.

NVIDIA has traditionally been recognized for its hardware contributions to AI, but documents filed with the SEC reveal a much broader ambition. The company plans to directly compete with heavyweights like OpenAI, Anthropic, and DeepSeek by developing its own AI models. This information, initially reported by Wired, has been confirmed by NVIDIA executives, underscoring the validity and scope of this project.

$26 Billion for Open Models: What is NVIDIA Planning?

The figure of $26 billion is impressive, especially when compared to the training costs of existing models. For instance, GPT-4 from OpenAI reportedly required around $3 billion for its development. NVIDIA thus has considerable leeway to create multiple cutting-edge AI models. The Nemotron 3 Super model, already in place, has 128 billion parameters, and another model reaching 550 billion parameters is currently in pre-training.

The concept of "open weights" is crucial to this strategy. Unlike total open access, this approach allows developers to download and modify the final model parameters while keeping the training data proprietary. This represents a compromise between OpenAI's closed models and the total transparency advocated by other labs.

This strategy is not entirely new. Companies like DeepSeek, Alibaba, and Baidu have already adopted open models to capture the ecosystem. However, NVIDIA brings to this approach financial resources that few can match, as well as a unique advantage: control over the hardware on which these models operate.

Why NVIDIA is Taking Big Risks by Targeting Its Own Clients

NVIDIA's choice to venture into the development of open models is not without risk. Among its largest clients are Microsoft, Amazon, and Google, all of which are also investing in their own AI models and seeking to reduce their reliance on NVIDIA's CUDA technology. By developing open models, NVIDIA is directly competing with these giants, which could complicate its business relationships.

NVIDIA's CEO, Jensen Huang, is betting on a well-known strategy for the company. For years, NVIDIA has built a free software ecosystem around its processors, such as with NemoClaw and the NIM microservices. Open models optimized for NVIDIA GPUs create a new type of dependency: the models are free, but they work optimally on NVIDIA hardware, much like a free razor with expensive blades.

This situation echoes the impact of CUDA, the proprietary platform that has tied developers to NVIDIA chips for two decades. Open models could become the CUDA of software, an ecosystem so convenient that it becomes hard to do without. Companies adopting Nemotron are then incentivized to use NIM for inference and NeMo Guardrails for security, further strengthening NVIDIA's grip.

The timing of this initiative is also significant. NVIDIA has just announced a potential $100 billion partnership with OpenAI for inference hardware. By playing both sides, as a supplier and a competitor, NVIDIA is engaging in a balancing act that may not be sustainable in the long term.