When DeepSeek’s artificial intelligence (AI) model became widely known early this year, it prompted an enormous debate in the tech industry: Would the Chinese AI model, with its frugal use of compute resources, lower the need for NVIDIA’s high-performance GPUs?
Certainly the stock market thought so, with NVIDIA’s stock plummeting by nearly 17% in one day. And now as lower cost Chinese AI models continue to proliferate—Chinese tech firms released at least 10 new models or upgrades in the last two weeks–the question takes on even greater urgency.
After all, the fortunes of far more than NVIDIA are at stake: Much of the US-based tech industry is built on the primacy of the NVIDIA GPU. A cohort of leading AI vendors have announced plans to spend more than $300 billion in 2025 CAPEX on infrastructure that rests largely on NVIDIA GPUs.
Yet even as the debate continues, it appears a potential answer lies in NVIDIA’s recently unveiled open-source software framework Dynamo, which will drive what NVIDIA calls AI Factories. These Factories are aggregations of specialized data center infrastructure assembled to support AI training and inference.
Also core is NVIDIA’s Cosmos, which is a platform that enables developers to create AI models for physical systems like self-driving vehicles and robots; Cosmos includes so-called ‘world foundation models’ that will accelerate this physical AI development.
To understand how this conglomeration of technologies will enable NVIDIA to thrive in a world filled with DeepSeek models, it’s important to understand how they work in concert with one another.
Measured by the number of tokens per second, NVIDIA’s GPUs, supported by Dynamo, can process far more throughput than DeepSeek’s R1 model, according NVIDIA. “Dynamo can capture that benefit and deliver 30 times more performance in the same number of GPUs in the same architecture for reasoning models like DeepSeek,” said Ian Buck, VP and general manager of Hyperscale and HPC at NVIDIA, at the company’s recent GTC conference.
Central to this exponential improvement is a key advance in AI computing: Dynamo distributes the heavy compute load of inference across an infrastructure that can include up to 1,000 NVIDIA GPUs. This massively parallel processing enables a far greater amount of computing to be accomplished in the same time frame. Fueled by Dynamo, AI factories will offer a higher premium service at a premium dollar per million tokens, according to the company.
The bottom line, from a revenue perspective, is that when performing inference work, Dynamo allows more tokens to run each second, which creates more revenue per second for vendors deploying NVIDIA GPUs. Hence, NVIDIA remains competitive in a world of low-cost DeepSeek models.
“This is a great example of how DeepSeek will really impact the AI market,” said Stephen Foskett, president of the Tech Field Day business unit of Futurum Group. “Their approach to training AI models is being widely adopted, reducing the cost and improving the speed of large language models. An AI factory based on NVIDIA Cosmos is more practical and profitable for service providers. These optimizations will also allow the industry to deploy larger and more complex models than previously thought possible, enabling NVIDIA’s ‘world foundation model’ vision.”
NVIDIA, to support its claims, has posted its own modified version of DeepSeek R1 on HuggingFace, a platform for AI and machine learning. The company claims its altered version improves the model’s performance significantly when comparing early vintage Hopper GPUs to the recent Blackwell, with no loss to the model’s accuracy.
