Demand for Computation Skyrockets Due to DeepSeek’s AI Capabilities
At a recent investor conference, Nvidia founder and CEO Jensen Huang announced that the open-source rollout of DeepSeek R1 has significantly increased the demand for compute resources, largely due to the broader implementation of reasoning AI techniques. The technology is catalyzing growth beyond traditional AI by enhancing analytical capabilities.
As pointed out by Huang, the computation demands associated with post-training customization and inference scaling have now eclipsed those required for initial training. Reasoning AI, which offers more detailed computational power compared to conventional one-shot inference models, relies heavily on reinforcement learning and model fine-tuning, processes that require exponentially higher computational resources.
“AI is transitioning from basic perception capabilities to more advanced reasoning functions,” Huang articulated. “With the deployment of reasoning AI, we are witnessing a new scaling law that emphasizes the need for greater computation during model execution. Improved reasoning leads directly to smarter outputs.”
Understanding the Compute Demand for Reasoning AI
Nvidia reported an impressive increase in sales for the quarter ending January 26, 2025, reaching $39.33 billion—an increase of nearly 78% year-over-year. The net income rose to $22.09 billion, translating to $0.89 per share, a substantial jump from the previous year’s figures. However, Nvidia’s stock experienced a slight drop of 1.49% in after-hours trading, falling to $129.32 per share, and has decreased over 9% since the DeepSeek R1 launch.
Models including OpenAI’s o3, DeepSeek R1, and xAI’s Grok 3 showcase the shift from passive perception to active reasoning, necessitating a hundredfold increase in compute capacity per task compared to traditional inference-driven AI. Huang remarked that “the scale of post-training and model customization is enormous, requiring vastly more computational resources than what was needed for pre-training.”
The evolution in computational requirements is not just a temporary trend; it’s emerging as a core focus for data centers, which, traditionally centered around CPU capabilities, are now adapting to prioritize GPU-accelerated environments. The future of data centers will lean towards being AI hubs, optimized for machine learning and AI deployments, rather than merely serving as storage facilities.
“Looking ahead, most capital expenditures in data centers will pivot to support accelerated computing and AI capabilities,” Huang stated. “We foresee that a significant portion of software development will center around machine learning.”
The Horizon for Upcoming AI Applications
According to Huang, Nvidia’s Blackwell architecture is strategically built to facilitate seamless transitions between pre-training, post-training, and inference scaling, enabling more efficient processing of AI workloads. This architecture is designed to link GPUs in a manner that maximizes throughput, allowing reasoning AI models to operate 25 times faster compared to Nvidia’s previous setups.
While consumer and search-based generative AI applications have surged in popularity, Huang emphasized that the next phase of AI development is just on the cusp of emergence. This forthcoming era will be characterized by AI-driven autonomous agents capable of making decisions and executing complex tasks independently of human oversight. As this technology unfolds, governments and organizations are expected to create localized AI ecosystems to address specific privacy concerns.
“The next generation of AI applications is on the horizon, featuring agentic AI for business, robotic physical AI, and region-specific sovereign AI,” Huang explained. “While these developments are still in their early stages, we are closely observing substantial activity within these spaces.”
