Nvidia's open-source strategy upgrade: From "supporter" to "builder", plans to develop the world's top open-source AI models itself

Nvidia is quietly undergoing a strategic identity shift, upgrading from a "supporter" of the open-source AI ecosystem to a direct "builder."

According to Wallstreet News, on August 11, The Information reported that Nvidia hopes the next-generation Nemotron 4 will rank among the world's strongest open models.

Employees involved in the project estimate that the largest version will have at least 1 trillion total parameters, about twice that of the Nemotron 3 Ultra released in June.

The pre-training data and architecture for the model have been preliminarily determined, but the final specifications and release date are yet to be confirmed, and the most crucial round of training has not yet begun. The so-called "trillion-parameter Nemotron 4" is currently still an R&D plan and should not be considered a released product.

This effort can easily be understood as Nvidia wanting to become like OpenAI or Anthropic, but in reality Nvidia does not need to recover costs by charging for Nemotron's API.

As long as the free model encourages more enterprises, governments, and startups to train, fine-tune, and deploy AI, Nvidia can layer its charges through GPU, NVLink networks, complete systems, CUDA, and enterprise software.

Nemotron 4 acts more like a demand generator for compute: the model is free, but the computations needed to use it are not.

Nvidia Needs to Expand Buyers from a Few Giants to the Entire Market

The current GPU boom is highly concentrated.

Nvidia's 10-Q as of April 26 shows that three direct customers contributed 21%, 17%, and 16% of the quarter's revenue, a total of 54%; the company also disclosed that an AI research and deployment firm indirectly contributed "significant revenue" via cloud services.

These direct customers may be system integrators, distributors, or cloud providers, and cannot be simply equated to three end users, but the customer concentration risk is already on the books.

More troubling is that major customers are building Nvidia alternatives at the same time:

• Amazon AWS uses Trainium to carry out Anthropic's training and inference, claiming nearly 1 million Trainium 2 units now serve Claude;
• Google's Ironwood TPU targets large-scale training and inference;
• AMD is also competing for open model workloads with the MI350 and ROCm.

Closed-source models focus demand on a few labs and cloud platforms, which are massive in scale, have strong bargaining power, and are motivated to move mature and stable workloads to their own chips.

For Nvidia, the safest market structure is not one model company dominating AI, but thousands of different models and applications growing simultaneously. The more fragmented the demand, the harder it is for clients to cover all architectures with one custom ASIC; the faster the workload changes, the more advantageous mature general-purpose GPUs and the CUDA software ecosystem become.

This also explains why Nvidia cannot just support the open ecosystem with words alone. An open model that badly lags behind closed-source counterparts only attracts research experiments, making it hard for banks, manufacturers, and government agencies to budget for production clusters.

The model must be strong enough—so strong that enterprises are willing to use their own data to continue training it, and customers are willing to hand over high-frequency workflows to it. Only then will open models shift from developer download counts to GPU utilization rates in server rooms.

The role of trillion parameters lies here. It is not a trophy to show intelligence, but a tool for Nvidia to cross the usability threshold for enterprises.

The existing Nemotron 3 Ultra has 550 billion total parameters, with 55 billion parameters activated per token, supporting up to 1 million token contexts.

Nvidia's test results show it approaches other leading open models on some tasks and has clearly higher inference throughput. However, The Information cites third-party rankings indicating it still lags behind the strongest Chinese open models and is far from the top tier globally.

Therefore, parameter size only signals ambition; it cannot substitute for actual performance.

In sparse MoE architectures, total parameters and those actually called during inference can differ greatly. Data quality, training token count, post-training, tool-calling ability, and inference systems collectively determine the final result.

If Nemotron 4 only has a flashy "1 trillion" figure but does not enter the top tier of real-world tasks, it will hardly change any procurement decisions.

A Free Model Can Sell an Entire AI Factory

Nvidia has already paved the conversion path:

1. The open Nemotron provides model weights, data, and training recipes;
2. Companies use NeMo for continued pretraining, fine-tuning, reinforcement learning, and evaluation;
3. For deployment, they connect to NIM inference microservices, TensorRT-LLM, and CUDA;
4. Once scaling up, the compute lands in DGX Cloud, public cloud GPUs, or enterprise-built HGX/DGX clusters.
5. Beyond GPUs, NVLink, InfiniBand or Spectrum-X networks, CPUs, DPUs, and NVIDIA AI Enterprise subscriptions also get into the bill.

Early prototypes of this pathway have already appeared. Palantir brought Nemotron into isolated U.S. government networks, allowing clients to train on local infrastructure, retain model weights, and iterate continuously, with production deployment supported by NVIDIA AI Enterprise.

The Japanese market is even closer to a "sovereign AI" template: research institutes and companies use Nemotron's data and NeMo to train Japanese models, HGX B300 serves private infrastructure, and Jetson is responsible for edge deployment.

Closed-source APIs cannot enter isolated networks, nor can they easily satisfy requirements for data residency, model ownership, and local auditing. Open models bring loads that would never occur on public cloud APIs into government server rooms, enterprise data centers, regional AI clouds, and edge devices.

For model companies, this is a relatively fragmented, hard-to-monetize market; but for Nvidia, which sells infrastructure, every new deployment location means an additional entry point for hardware and software revenue.

Model routing is making the business even larger. The Information reports that along with Nemotron 3.5 Lightning, Nvidia released a free model routing software. The router assigns simple tasks to small, cheaper models and sends harder ones to large models.

On the surface, it helps clients use less expensive compute; in reality, the bet is that as per-task costs fall, enterprises will activate more resident AI agents, turning a single question-and-answer exchange into a continuous workflow involving planning, search, tool calls, verification, and rework.

Gartner predicts that by 2030, the inference cost for trillion-parameter models will drop by more than 90% from 2025, but the tokens consumed per AI agent task could be 5–30 times that of an ordinary chatbot. If token usage grows faster than price falls, total inference expenditure will still rise.

Nvidia's bet is not "the more expensive the large model, the more GPUs will be sold," but rather "the cheaper intelligence gets, the faster usage scales." Covering lightweight models, flagship models, and routing tools simultaneously, Nemotron is proactively creating this demand elasticity after price reductions.

The Purpose of Making Models Is to Get the Next-Gen Chips Right in Advance

Nemotron has another value not directly reflected in model revenue: it is Nvidia's own full-stack stress test.

If a chip company waits for client models to be finalized before adapting to new attention mechanisms, MoE routing, long context, and lower-precision formats, hardware iteration will always lag behind.

By training cutting-edge models in-house, engineering teams can spot bottlenecks in compute, memory, interconnects, data throughput, and inference scheduling earlier, feeding these insights into next-gen GPUs, NVLink, and software libraries.

Nemotron 3 Ultra already shows clear signs of "hardware co-design": the model uses a hybrid Mamba-Attention with MoE architecture, pretraining adopts the NVFP4 format designed for Blackwell, and multi-token prediction is used to speed up generation.

Nvidia says the quantized version can be deployed on four B200 GPUs. Such models are both products and demonstration workloads for Blackwell's low-precision capabilities, memory systems, and inference software.

Once open-sourced, external developers will further optimize this workload for Nvidia.

The new model architectures will first run on CUDA, PyTorch, vLLM, SGLang, and TensorRT-LLM. The optimization flows back to the Nvidia platform, leaving newcomers faced not just with a performance gap in hardware, but with a production system refined by thousands of projects.

In cases Nvidia showcased this year, Blackwell’s per-token cost when running DeepSeek V4 fell by up to 80% within a month; with multiple software optimizations stacked, throughput could rise by up to 20 times.